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
|
|---|---|---|---|---|---|---|---|---|---|
1602.07781 | 1 | 1602 | 2016-02-25T03:09:47 | A Markov chain model for the search time for max degree nodes in a graph using a biased random walk | [
"cs.SI",
"cs.DM",
"physics.soc-ph"
] | We consider the problem of estimating the expected time to find a maximum degree node on a graph using a (parameterized) biased random walk. For assortative graphs the positive degree correlation serves as a local gradient for which a bias towards selecting higher degree neighbors will on average reduce the search time. Unfortunately, although the expected absorption time on the graph can be written down using the theory of absorbing Markov chains, computing this time is infeasible for large graphs. With this motivation, we construct an absorbing Markov chain with a state for each degree of the graph, and observe computing the expected absorption time is now computationally feasible. Our paper finds preliminary results along the following lines: i) there are graphs for which the proposed Markov model does and graphs for which the model does not capture the absorbtion time, ii) there are graphs where random sampling outperforms biased random walks, and graphs where biased random walks are superior, and iii) the optimal bias parameter for the random walk is graph dependent, and we study the dependence on the graph assortativity. | cs.SI | cs | A Markov chain model for the search time for max
degree nodes in a graph using a biased random walk
Jonathan Stokes and Steven Weber
Department of Electrical and Computer Engineering
Drexel University, Philadelphia, PA 19104
6
1
0
2
b
e
F
5
2
]
I
S
.
s
c
[
1
v
1
8
7
7
0
.
2
0
6
1
:
v
i
X
r
a
Abstract-We consider the problem of estimating the expected
time to find a maximum degree node on a graph using a
(parameterized) biased random walk. For assortative graphs
the positive degree correlation serves as a local gradient for
which a bias towards selecting higher degree neighbors will
on average reduce the search time. Unfortunately, although the
expected absorption time on the graph can be written down using
the theory of absorbing Markov chains, computing this time is
infeasible for large graphs. With this motivation, we construct
an absorbing Markov chain with a state for each degree of
the graph, and observe computing the expected absorption time
is now computationally feasible. Our paper finds preliminary
results along the following lines: i) there are graphs for which
the proposed Markov model does and graphs for which the model
does not capture the absorbtion time, ii) there are graphs where
random sampling outperforms biased random walks, and graphs
where biased random walks are superior, and iii) the optimal bias
parameter for the random walk is graph dependent, and we study
the dependence on the graph assortativity.
Index Terms-graph search; Markov chain; biased random
walks; greedy search, assortativity.
I. INTRODUCTION
A graph representing Facebook's network of 1.4 billion
users would require 1.4 billion nodes and hundreds of billions
of edges, stretching the capacity of current hardware to hold
the graph in memory. This inability to represent large graphs
makes them difficult to study. One way of sidestepping this
issue is to study a representative subsample of the entire graph.
How one takes this sample often depends on the properties of
the graph being studied. The simplest method of sampling a
graph is to select nodes uniformly at random, however, many
graphs, specifically those representing social networks, exhibit
a power law degree distributions the probability of selecting
a high degree node using this method is small. Intuitively a
better sampling method would use the information gained by
a sample to increase the probability of selecting a max degree
node on subsequent samples. In the context of social networks,
sampling methods for a max degree node that exploit local
information exploit the friendship paradox; on average your
friends have more friends than you do. One goal is to study
the impact that biasing the random choice of the next neighbor
in a random walk towards selecting higher degree neighbors
has on the time to reach a maximum degree node.
This research has been supported by the National Science Foundation under
award #IIS-1250786.
Previous work has developed analytical bounds for the
hitting time of a biased random walk, the time it takes to get
from one node in a graph to another, and the cover time of the
walk, the time it takes a walk to visit every node in a graph.
Ikeda shows that the hitting and cover time of an undirected
graph of n nodes is upper bounded by O(n2) and O(n2log(n))
respectively [1]. Cooper shows all nodes of degree na for
a < 1 in a n node power law graphs with exponent c can be
found in O(n1−a(c−2)+δ) steps [2]. Maiya evaluates several
biased sampling strategies on real world graphs numerically
showing that a walk which always transitions to the max
degree node attached to its current node is a good method
of exploring these graphs [3].
Our approach is different. We first observe that the basic
theory of absorbing finite-state Markov chains (§II) yields
expressions for the mean and variance of the random time to
reach an absorbing state, which in our case is a maximum de-
gree node on the n-node graph. However, computing this mean
and variance requires inverting an n×n matrix, which for large
graphs is prohibitive (§III). Consequently, we consider the case
of assortative graphs where the degrees of the endpoints of
an edge are positively correlated, and recognize that for such
graphs an intelligent strategy for minimizing the search time
to find a maximum degree node is to exploit the local gradient
provided by the assortativity, and to select the next node in
the walk with a bias towards higher degree neighbors. We
construct a random walk on a significantly reduced state space,
with one state for each degree in the graph, and the transition
probability matrix derived from the joint degree distribution
of the graph and the random walk bias parameter (§IV). The
advantage of such a model is that we can analytically compute
the absorption time of the random walk on this reduced state
space quite easily. A second goal is to study how the statistics
of the random absorption time to find maximum degree nodes
on large graphs may be captured using this model.
As shown in §V, we investigate three natural questions,
for which this paper offers only preliminary and numerical
results: i) for which graphs does the above Markov chain state
reduction accurately capture the mean absorbtion time?, ii)
for which graphs does a biased random walk find a maximum
degree node more quickly than does random sampling?, and
iii) how does the optimal bias parameter (β) depend upon the
graph? For our preliminary numerical investigation of these
three questions, we employ Erdos-R´enyi (ER) graphs, rewired
(using a standard rewiring algorithm) to ensure the graph is
connected and has a target assortativity (α ∈ [−1, +1]); the
assortativity plays the role of an independent control parameter
in our investigations. Our preliminary results suggest
the
following answers to the above questions. First, there are
certain (α, β) pairs for which the above model does and does
not work well, in particular α < 0 and β large gives a poor
match. Second, for certain values of α the optimal β∗(α) are
such that the resulting optimal absorbtion time of a biased
random walk is superior to random sampling, but otherwise
not. Third, the optimal β∗(α) appears to have an increasing
trend in α, so that "following the local gradient (choosing a
maximum degree neighbor)" is optimal for highly assortative
graphs, while choosing a neighbor uniformly is superior for
highly disassortative graphs.
II. ABSORPTION TIME FOR MARKOV CHAINS
We will have cause to use the theorem and corollary below
on the mean and variance of the random absorption time TX of
an absorbing finite-state discrete-time Markov chain (DTMC)
X = (X(t), t ∈ N) taking values in a finite state space X
(with X = n). Partition X into absorbing states X (with
X = n) and transient states X (with X = n = n − n), and
so too partition the n×n transition matrix PX into submatrices
(cid:21)
(cid:20) X
X
A O
R Q
X
X
,
PX =
(1)
where A is the n × n submatrix of transition probabilities
between absorbing states, O is the n× n zero matrix, R is the
n × n submatrix of transition probabilities from transient to
absorbing states, and Q is the n × n submatrix of transition
probabilities between transient states.
Definition 1. The fundamental matrix for the DTMC X is the
n × n matrix NX = (In − Q)−1, for In the n × n identity
matrix. The absorption times TX = (TX (x), x ∈ X ) are the
collection of n random absorption times starting from each
possible initial transient state, with TX (x) = min{t ∈ N :
X(t) ∈ XX(0) = x).
Theorem 1. ([4]) The absorption times TX have means µX
and variances σ2
X:
µX = NX 1n, σ2
X = (2NX − In)µX − (µT
X µX )1n,
(2)
where 1n is the n-vector of ones.
Let pX = (pX (x), x ∈ X ) be the initial distribution of X
on the transient states X , i.e., X(0) ∼ pX, and define TX as
the corresponding random absorption time.
Corollary 1. Given the initial distribution pX for X(0) on
X , the resulting absorption time TX has
E[TX ] =
pX (x)µX (x), Var(TX ) =
(cid:88)
(cid:88)
pX (x)σ2
X (x).
x∈ X
x∈ X
(3)
III. BIASED RANDOM WALK ON THE GRAPH
(cid:83)
Denote the undirected graph on which the search takes place
as G = (V,E), with V the set of vertices and E the set of
undirected edges. The order and size of G are V = n and
E = m, respectively. The neighborhood of any node v is
denoted Γ(v) = {u ∈ V : {u, v} ∈ E}, and the degree is
d(v) = Γ(v). The set of degrees found in the graph is D =
v∈V d(v), the number of distinct degrees is δ = D. The
nodes V are partitioned into subsets (Vk, k ∈ D) with Vk =
{v ∈ V : d(v) = k} the set of nodes of degree k, and the
resulting degree distribution is pD = (pD(k), k ∈ D), with
Vk
n . In particular, the max degree found in
entries pD(k) =
the graph is dmax = maxD, and the set of nodes with max
degree is Vmax = Vdmax. An absorbing biased random walk
on G is defined as follows, where the notation is adopted from
the general case presented in §II.
Definition 2. The absorbing discrete-time Markov chain
(DTMC) V = (V (t), t ∈ N) on G with bias parameter β ≥ 0
has states V, partitioned into absorbing states V = Vmax and
transient states V = V\Vmax. The n×n transition probability
matrix PV has entries PV (u, v) = P(V (t+1) = vV (t) = u).
v(cid:48)∈Γ(u) d(v(cid:48))β for v ∈ Γ(u)
For u ∈ V we set PV (u, v) =
and PV (u, v) = 0 else. For u ∈ V: PV (u, u) = 1 and
PV (u, v) = 0 for v (cid:54)= u. The absorption times from each
transient state are TV = (TV (v), v ∈ V), with
initial
TV (v) = min{t ∈ N : V (t) ∈ VV (0) = v}. Note the event
V (t) ∈ V is equivalent to d(V (t)) = dmax.
(cid:80)
d(v)β
This random search for Vmax uses a transition probability
that is biased towards selecting higher degree neighbors of
the current node, as introduced by Cooper [2], with the bias
monotonically increasing in β. For β = 0 the next step is
uniform among all neighbors of u, while as β → ∞ the
next step is uniform among the maximum degree neighbors
of u, i.e., Γmax(u) = argmaxv∈Γ(u) d(v). A key goal of this
paper is to characterize E[TV ] and Var(TV ) as a function of
the initial distribution pV on V and on the bias parameter
β. Although numerical estimates can (and will) be obtained
by simulating the random walk, for large graphs (large n) it
is computationally infeasible to compute E[TV ] and Var(TV )
analytically using Cor. 1, due to the need to compute NV as
the inverse of the (large) n × n matrix In − Q. This difficulty
motivates us to define an approximation of the biased random
walk using a significantly smaller state space, discussed next.
IV. APPROXIMATE BIASED RANDOM WALK
In this section we develop our key approximation giving
a more computationally feasible means to estimate the mean
and standard deviation of TV . Our derivation consists of the
following steps: i) define the joint degree matrix J and the
conditional degree distribution matrix J for G, ii) define the
biased walk degree transition probability distribution p(n) and
average biased walk degree transition matrix ¯P , and iii) define
the approximate biased walk degree transition matrix P .
A. Joint degree and conditional degree distribution matrices
We define the joint degree matrix J and the conditional
degree distribution matrix J for G.
Definition 3. The joint degree matrix J is the δ× δ symmetric
matrix with entries J(k, l) = #{{u, v} ∈ E : {d(u), d(v)} =
{k, l}}, the number of edges in G with endpoints of degrees
k and l, and entries J(k, k) twice the number of edges with
endpoints both of degree k, for (k, l) ∈ D2.
Example 1. The joint degree matrix J for the graph G shown
in Fig. 1 (left) is found by grouping the edges E by the degrees
of the endpoints, as shown in Fig. 1 (right):
J =
1
2
3
4
1
2
3
1
2
1
1 3
4
1
3
2
2
.
endpoint, and (cid:80)
(cid:83)
l(cid:48)∈D J(k, l(cid:48)) = kVk since the sum is the
number of edges with a degree k endpoint.
Row k of J is denoted J(k,·) = ( J(k, l), l ∈ D). The
locations of the non-zero entries of J(k,·) are denoted Dk =
D(v), i.e., l ∈ Dk means there exists some {u, v} ∈ E
v∈Vk
with d(u) = k and d(v) = l.
B. Biased walk degree transition probability distribution
We define the biased walk degree transition probability
distribution p(n) and average biased walk degree transition
matrix ¯P .
Definition 5. The biased random walk degree transition prob-
ability distribution p(n) = (pl(n), l ∈ D) from a node with
l∈D nl ∈ D) is
degree neighborhood n = (nl, l ∈ D) (with(cid:80)
(4)
pl(n) =
(6)
nllβ(cid:80)
l(cid:48)∈D nl(cid:48)l(cid:48)β .
The following is an immediate result of Def. 2.
Proposition 2. The biased random walk V in Def. 2 obeys
the degree transition probability
(7)
P(d(V (t + 1)) = lV (t) = v) = pl(n(v)), l ∈ D(v).
The above proposition states that the biased random walk
V degree transition probability distribution has the property
that the probability of the degree of the next node of the
biased random walk depends upon the current node v only
through n(v). The next definition gives an average transition
probability from nodes of degree k to nodes of degree l under
the biased random walk.
Definition 6. The average biased walk degree transition
matrix ¯PV is the δ × δ matrix with entries
¯PV (k, l) =
1
Vk
pl(n(v)), (k, l) ∈ D2
(8)
(cid:88)
v∈Vk
giving the average probability of transitioning to a node of
degree l over all starting nodes of degree k.
A key point, developed below, is that ¯PV is of size δ × δ
whereas the transition matrix PV is of (potentially) signifi-
cantly larger size n × n.
C. Approximate biased walk degree transition matrix
, n ∈ Nk,
(cid:1) =(cid:0)
P(N = n) =
k(cid:81)
A random δ-vector N = (Nl, l ∈ D) has a multinomial
distribution with parameters (k, p), denoted N ∼ mult(k, p),
if
(cid:19)(cid:89)
where (cid:0)k
N has support Nk = {n ∈ Nδ : (cid:80)
(cid:18)k
(cid:1) is the multinomial coefficient. Here,
the set of all possible δ-vectors from Nδ that sum to k.
l∈D nl = k}, defined as
The approximate biased walk degree transition matrix PW
has entries defined using the expectation of the biased random
walk degree transition probability distribution pl when the
l∈D nl!
pnl
l
n
l∈D
(9)
n
Observe(cid:80)
Fig. 1: Graph G (left) with edges E grouped by the endpoint degrees
(right).
k,l J(k, l) = 2m, and that if an edge from E is
selected uniformly at random then the probability that it should
have endpoints {k, l} is given by J(k, l)/m for k (cid:54)= l, and
J(k, k)/(2m) else. The conditional degree distribution matrix
is obtained by normalizing each row of J into a probability
distribution.
Definition 4. The conditional degree distribution matrix J is
l(cid:48)∈D J(k, l(cid:48)), for
(k, l) ∈ D2.
the δ × δ matrix with J(k, l) = J(k, l)/(cid:80)
of each neighbor of v, i.e., D(v) =(cid:83)
It is important to motivate J for what follows. Towards this
end we define the following. First, the set of degrees D(v) ⊆
D of some v ∈ V is formed by taking the union of the degrees
u∈Γ(v) d(u). The number
of neighbors of some v ∈ V of each different degree D(v),
which we term the degree neighborhood of v, is given by
n(v) = (nl(v), l ∈ D(v)), with entries nl(v) = #{u ∈ Γ(v) :
d(u) = l}.
Proposition 1. The entries J(k, l) of the conditional degree
distribution matrix in Def. 4 are the averages of the fraction
of neighbors of degree l over all degree k nodes:
(cid:88)
Proof: Observe J(k, l) = (cid:80)
J(k, l) =
1
Vk
v∈Vk
nl(v) by partitioning
all edges with a degree k endpoint by the degree of the other
v∈Vk
nl(v)
k
.
(5)
abcdefghhcabgdefABCDEFFEEEEAABBBBCCDDD2143241423243444 degree neighborhood n is taken as a random multinomial
vector N with parameters (k, J(k,·)).
Definition 7. The approximate biased walk degree transition
matrix PW is the δ × δ matrix with entries
PW (k, l) = E [pl(N)] , N ∼ mult(k, J(k,·)).
That is:
PW (k, l) =
(cid:88)
(cid:18)k
(cid:19) (cid:89)
n∈Nk
n
l∈Dk
J(k, l)nl
(cid:80)
nllβ
l(cid:48)∈Dk
nl(cid:48)l(cid:48)β .
(10)
(11)
Our model is suitable for those graphs for which ¯PV (k, l) ≈
PW (k, l) for each (k, l) ∈ D2; as will be shown in §V, it is not
difficult to identify graphs for which the model works, and to
find graphs for which it does not. Finally, we define a biased
random walk on D.
Definition 8. The absorbing discrete-time Markov chain
(DTMC) W = (W (t), t ∈ N) on D with bias parameter β ≥ 0
has states D, partitioned into absorbing states D = {dmax}
and transient states D = D \ dmax. The δ × δ transition
probability matrix PW is given in Def. 7. The absorption times
from each initial transient state are TW = (TW (l), l ∈ D),
with TW (l) = min{t ∈ N : W (t) = dmaxW (0) = l}.
Recall that for large graphs (large n) it is not possible
to compute E[TV ] and Var(TV ) for TV in Def. 2 using
Cor. 1 since we cannot obtain the fundamental matrix NV =
(In−Q)−1, on account of the difficulty of inverting the (large)
n × n matrix In − Q. In contrast, for graphs with bounded
δ, it is possible to compute E[TW ] and Var(TW ) for TW
in Def. 8 using Cor. 1 since the corresponding fundamental
matrix NW = (Iδ−1 − Q)−1 is obtained by inverting the
(smaller) (δ − 1) × (δ − 1) matrix Iδ−1 − Q.
Our model's approximation TV ≈ TW lies with the required
approximation ¯PV ≈ PW mentioned above. That is, i) al-
though the biased random walk on the graph has a probability
of transitioning to nodes of each degree l ∈ D(v) from a node
v of degree k that depends upon the exact degree neighborhood
n(v) (i.e., pl(n(v))), ii) we approximate this probability using
the expectation with respect to a random degree neighborhood
N, drawn with parameters k and J(k,·). This distribution
J(k,·) is the average distribution of the number of nodes of
each degree over all degree k nodes (Prop. 1).
V. RESULTS
The experimentation framework for this work is written
using the igraph Python library [5]. We use this framework to
generate instances of a graph family, then for each graph we
measure the random times to find a maximum degree node
by i) a biased random walk (BRW), and ii) random sam-
pling. Unless noted otherwise, the mean E[TV ] and standard
deviation Std[TV ] of the absorption of a BRW on a graph is
calculated from 500 trials for each tested bias coefficient β.
We compare the empirical mean E[TV ] and empirical standard
deviation Std[TV ] measured from the BRW on the graph with
the numerical mean E[TW ] and standard deviation Std[TW ]
of a BRW on the Markov chain of the graph's degree states.
Tab. I describes the Erdos-R´enyi (ER) graphs we used.
Graph
Erdos-R´enyi (ER)
Erdos-R´enyi (ER)
Erdos-R´enyi (ER)
Parameters
p = 0.05
p = 0.0024
p = 0.002569
Size
100
1121
1011
dmax
11
10
11
TABLE I: Parameters of the graphs used in the simulations.
As currently implemented our model has a significant
computational limitation in that the approximate random walk
(W (t)) in Def. 8 requires computing the matrix PW in (11),
and each such entry PW (k, l) requires summing over all
n ∈ Nk. As the size of Nk grows exponentially in k, we
are unable to compute it for graphs with dmax > 15; this is
the motivation behind our selecting (n, p) pairs for the ER
graphs so that dmax is small. Addressing this deficiency is the
subject of our ongoing and future work.
2
A. Erdos-R´enyi (ER) Graphs
graphs with n nodes and each of the (cid:0)n
(cid:1) possible edges is
The Erdos-R´enyi graph [6] G(n, p) is a family of random
added independently with probability p. The resulting graph
has a binomial degree distribution, pD(k) ∼ bin(n − 1, p).
To investigate larger graphs we must ensure the graph is
constructed so as to have a bounded expected maximum
degree, on account of the computational limitation dmax ≤ 15
discussed above. The following proposition gives an upper
bound on the expected maximum of n iid random variables:
Proposition 3. Let (Y1, . . . , Yn) be iid with moment gen-
erating function (MGF) φ(t) = E[etY ] and let Ymax =
max(X1, . . . , Xn). Then E[Ymax] ≤ 1
t log(nφ(t)).
It may be proved by application of Jensen's inequality to
establish etE[Ymax] ≤ nφ(t).
We apply the above rule to the random degrees (D(v), v ∈
[n]) of an ER graph G(n, p), each D(v) ∼ bin(n − 1, p),
which are identically distributed, but not independent; it can be
shown that the slight dependence is inessential and the bound
applies. Recall that the binomial distribution bin(n, λ/n) can
be approximated as a Poisson distribution Po(λ) in the case
when n is large, and that the Poisson MGF is φ(t) = eλ(et−1).
Applying Prop. 3 to this case yields the following upper bound
on the max degree of an ER graph G(n, λ/n)
E[max(D(v), v ∈ [n])] ≤ log(n) − λ
W( log(n)−λ
eλ
(12)
)
where W is the Lambert W function. The value of (12) is that
it allows us to select λ(n) so that the resulting ER graph of
order n has a specified expected max degree upper bound.
However, the ER graph is known to be disconnected with
high probability when p(n) = λ/n (or smaller) for any λ,
and our random walk is only guaranteed to find a maximum
degree node for a connected graph. It is further known that
an ER graph with p(n) = λ(β)/n will have a fraction β
of the n nodes in the giant connected component where
e4 can be wired between the endpoints of the former edges
e1 and e2 without creating self loops or multiple edges such
that the assortativity of the new graph α1 is closer to αt than
α0 we add edges e3 and e4, if not we replace e1 and e2.
This procedure is repeated until the graph's assortativity is
within a suitably small interval around αt. Notice that this
procedure preserves the degree distribution of G0, since the
degree of the end points of e1 and e2 are unchanged. When the
assortativity converges, if the graph is disconnected, then for
each disconnected component a random node is selected in the
graph's giant component and wired to the smaller component,
thereby connecting the graph. The assortativity of the 1011
node graphs compared to their target assortativity is shown in
Fig. 2 (right) for both disconnected (red) and connected (blue)
graphs. From these results we infer that connecting the graphs
in this manner has little effect on their assortativity.
D. Biased Random Walks
We carried out Monte carlo simulations of 500 trials on 10
graphs with binomial degree distributions of 100 and 1011
nodes, while sweeping the bias coefficient β of the walk. We
compared BRWs with two random sampling algorithms: i)
sampling nodes without replacement, denoted 'no-r', and ii)
sampling a node and all of its neighbors without replacement,
denoted 'no-r n'. The rationale for these two forms of sampling
is in the interest in making a fair comparison in the absorbtion
time between the BRW and a random sample. The BRW
algorithm presumes at each step that the search is able to not
only view the degree of the current node but also the degree of
all neighbors of that node. Thus any comparison between the
performance of, say, k steps of the BRW and k nodes sampled
without replacement is unfair to random sampling, since the
latter does not see as many nodes as the former. The second
sampling scheme, where at each step we sample a node and
see its degree as well as the degrees of its neighbors, offers a
more balanced comparison with the performance of the BRW.
For our comparison between the BRW and random sampling
we investigated nine sample values of the target assortativity
αt, namely the nine values {−1.0,−0.75, . . . , 0.75, 1.0}, and
we used both 100 node and 1000 node ER graphs. For each
target αt, and each graph rewired to that αt, we swept the
biased coefficient β over the range [0, 8].
The results for αt ∈ {+0.5,−0.5} are shown in Fig. 3 (for
n = 1000) and Fig. 4 (for n = 100). Several points bear
mention. First, for target assortativity αt = +0.5 (both for
n ≈ 1000 and n ≈ 100) there exist optimized β∗(αt) for
which BRW outperforms random sampling of a node and its
neighbor degrees, and there exist non-optimal values of β for
which random sampling outperforms the non-optimized BRW.
Second, for target assortativity αt = −0.5 we see that the
BRW is inferior to random sampling a node and its neighbor
degrees for all values of β. Third, the case of n ≈ 1000 and
α = +0.5 shows there exists a non-trivial value of β∗ for this
α, meaning the optimal value is not at either endpoint of the β
interval of [0, 8]. This suggests that although BRWs can yield
superior search times compared with sampling neighborhoods,
Fig. 2: Left: average absorbtion time, E[T ] (solid lines), for original
graph (blue, via simulation) and model (green, via (3)), with E[T ]±
Std[T ] (dashed lines). Right: output assortativity α as a function of
input target assortativity αt in the random rewiring algorithm.
λ(β) = − log(1 − β)/β [7]. Hence the trade-off we face in
generating ER graphs is between a large fraction β in the
giant connected component vs. a bounded max degree. For
n = 1090 and λ = 2.8 (i.e., p(n) = λ/n = 0.0024569), the
expected max degree upper bound in (12) is dmax ≤ 11.1041,
and β = 0.924975 obeys − log(1 − β)/β = λ, meaning
the giant connected component will contain approximately
nβ = 1008 nodes.
B. Results for ER Graphs
Fig. 2 (left) shows the expected absorbtion time E[T ] for
an ER graph using both i) the simulated biased random walk
(BRW) on the original graph E[TV ], and ii) the analytically
computed E[TW ] using the reduced state space model, both
swept over a range of bias coefficients β. The plot shows a
significant deviation between the measured quantity and the
model prediction. The failure of the model for this graph may
be accounted for by the fact that the ER graph is known to
have zero assortativity, i.e., the degrees of the two endpoints
of the graph are conditionally independent, and as such the
degree of the current node does not provide any substantial
information about the proximity of that node to higher degree
(and by extension, maximum degree) nodes. In this sense, our
approximation is shown to break down for such graphs, as a
central assumption of our model is the idea that the degree of a
node does contain information about its degree neighborhood
and all nodes of degree k have similar degree neighborhoods.
This led to a set of questions: i) is the accuracy of our
model of a BRW on a graph dependent upon the graph's
assortativity?, ii) are there graphs where BRW finds the max
degree nodes faster than random sampling?, and iii) what is
the optimal bias coefficient β∗ for a graph of assortativity α?
C. Graph Re-wiring Algorithm
Our approach in this paper is to offer some preliminary
numerical answers to these questions, using graph assorta-
tivity, denoted by α ∈ [−1, +1], as the independent control
parameter. To construct graphs with a target assortativity,
αt, we modified Brunet's rewiring algorithm for increasing
or decreasing a graph's assortativity [8]. Given an initial
graph G0 and a target assortativity αt. We calculate G0's
assortativity, α0 and choose two edges at random e1, e2. Then
we remove e1 and e2 from G0. If two new edges e3 and
012345Bias Coefficient: β10-1100101102103104Time to AbsorptionGraph WalkAnalytical Fig. 3: ER graph with ∼ 1000 nodes. Expected time to absorbtion
E[T ] (solid) and E[T ]±Std(T ) (dotted) for i) the BRW (blue) and ii)
random sampling without replacement (observing a) just the degree
of the sampled node (green) and b) degrees of node and its neighbors
(red)) versus β. αt = +0.5 (left) and αt = −0.5 (right).
Fig. 5: The optimal bias coefficient β∗ (points) and the interval
[βmin, βmax] of points for which E[T ] is within 10% of E[T ∗]
(shaded) vs. the target assortativity αt. Left: comparison of n ≈ 100
(gray) and n ≈ 1000 (blue). Right: comparison of β∗ for n = 100
for actual graph (gray) and reduced state space model (blue).
Fig. 4: Same caption as Fig. 3 but for an ER graph with ∼ 100 nodes.
Also shown is mean absorbtion time E[TW ] predicted by the model.
Fig. 6: Mean absorbtion times E[T ∗] vs. αt, where for each αt we
use the optimized value β∗(αt) from Fig. 5. Left: comparison of
n ≈ 100 (blue) and n ≈ 1000 (gray). Right: comparison of E[T ∗]
for actual graph (blue) and reduced state space model (green).
doing so requires a correctly-tuned value of β for the particular
value of α (and in this case also n). The inferiority of the BRW
for disassortative graphs (here, α = −0.5) may be on account
of the fact that BRWs on such graphs are more likely to spend
much of the search time trapped in a local minimum.
E. Optimal Bias Coefficient β
In this subsection we extend αt to include the nine values
mentioned earlier. For each αt we find β∗(αt) and E[T ∗(αt)]
using β∗(αt), and plot both these functions against αt. Be-
cause we observe that the dependence of E[T ] on β can be
somewhat flat near the optimal, meaning there is some degree
of insensitivity to the precise value of β, we actually compute
the interval [βmin(αt), βmax(αt)] containing α∗(αt), where
the interval holds all values of β for which the corresponding
value of E[T ] is within 10% of the optimal value E[T ∗].
The results are shown in Fig. 5 and Fig. 6. Several
points bear mention. First, Fig. 5 shows that
the optimal
bias coefficient β∗ for β ∈ [0, . . . , 8] tends to increase with
increasing assortativity αt of the graph. More sample graphs
for each αt, more points αt ∈ [−1, +1], and a larger search
range for β than the current [0, 8] are required to confirm
this initial observation. Second, Fig. 6 shows that for both
n ≈ 100 and n ≈ 1000 there exists an interval of α over
which optimized BRWs outperform random sampling of node
and neighbor degrees. Again, more extensive simulations are
required. However, these preliminary results suggest BRWs
are inferior to sampling for graphs with negative assortativity.
F. Biased Random Walk Model
Finally, the two right side plots in Fig. 5 and Fig. 6 include
results for both the simulations of BRWs on the graph as
well as analytical computations using the reduced state space
model. In particular, Fig. 5 shows that β∗ for minimizing
E[TV ] and β∗ for minimizing E[TW ] are not exactly equal, but
are comparable, and show the same rough increasing trend as
a function of α. Moreover, Fig. 6 shows E[T ∗
V ] is comparable
to E[T ∗
W ] for certain values of α.
VI. FUTURE WORK
The main contributions are i) a potentially useful model
for analytically computing the expected absorbtion time of a
biased random walk using a reduced state space model, and
ii) preliminary comparison of the absorbtion time between a
BRW and random sampling to find a max degree node. Our
current goal is to find an approximation for E[TW ].
REFERENCES
[1] S. Ikeda and I. Kubo, "Impact of local topological information on random
walks on finite graphs," in Proc. of the 30th Intl. Conf. on Automata,
Languages and Programming, 2003, pp. 1054–1067.
[2] C. Cooper, T. Radzik, and Y. Siantos, "A fast algorithm to find all high-
degree vertices in graphs with a power-law degree sequence," Internet
Mathematics, vol. 10, no. 1-2, pp. 137–161, 2014.
[3] A. Maiya and T. Berger-Wolf, "Benefits of bias: Towards better charac-
terization of network sampling," in Proc. of the 17th ACM SIGKDD Intl.
Conf. on Knowledge Discovery and Data Mining, 2011, pp. 105–113.
[4] J. Kemeny and J. Snell, Finite Markov Chains, 1st ed. Springer, 1983.
[5] G. Csardi and T. Nepusz, "The igraph software package for complex
network research," InterJournal, vol. Complex Systems, p. 1695, 2006.
[6] P. Erdos and A. R´enyi, "On random graphs I," Publicationes Mathemat-
icae (Debrecen), vol. 6, pp. 290–297, 1959.
012345678Bias Coefficient: β10-1100101102103Time to AbsorptionBiased WalkSampling no-rSampling no-r n012345678Bias Coefficient: β101102103104105Time to AbsorptionBiased WalkSampling no-rSampling no-r n012345678Bias Coefficient: β10-1100101102103Time to AbsorptionBiased WalkBiased Walk ModelSampling no-rSampling no-r n012345678Bias Coefficient: β10-210-1100101102103Time to AbsorptionBiased WalkBiased Walk ModelSampling no-rSampling no-r n1.00.50.00.51.0Assortativity: α012345678Optimal Bias Coefficient100 Nodes1000 Nodes1.00.50.00.51.0Assortativity: α012345678Optimal Bias CoefficientBiased WalkBiased Walk Model1.00.50.00.51.0Assortativity: α10-1100101102103104Optimal Time to AbsorptionBiased Walk: 100 NodesBiased Walk: 1000 NodesSampling 100 Nodes no-rSampling 100 Nodes no-r nSampling 1000 Nodes no-rSampling 1000 Nodes no-r n1.00.50.00.51.0Assortativity: α10-1100101102103Optimal Time to AbsorptionBiased Walk 100 nodesBiased Walk Model 100 nodesSampling 100 nodes no-rSampling 100 nodes no-r n [7] S. Janson and T. Luczak, Theory of random graphs.
John Wiley, 2000.
[8] R. Xulvi-Brunet and I. Sokolov, "Changing correlations in networks:
Assortativity and dissortativity," Acta Physica Polonica B, vol. 36, no. 5,
pp. 1431–1455, May 2005.
|
1804.04469 | 1 | 1804 | 2018-04-12T12:34:57 | Generative models for local network community detection | [
"cs.SI",
"stat.ML"
] | Local network community detection aims to find a single community in a large network, while inspecting only a small part of that network around a given seed node. This is much cheaper than finding all communities in a network. Most methods for local community detection are formulated as ad-hoc optimization problems. In this work, we instead start from a generative model for networks with community structure. By assuming that the network is uniform, we can approximate the structure of unobserved parts of the network to obtain a method for local community detection. We apply this local approximation technique to two variants of the stochastic block model. To our knowledge, this results in the first local community detection methods based on probabilistic models. Interestingly, in the limit, one of the proposed approximations corresponds to conductance, a popular metric in this field. Experiments on real and synthetic datasets show comparable or improved results compared to state-of-the-art local community detection algorithms. | cs.SI | cs |
Generative models for local network community detection
Twan van Laarhoven
Institute for Computing and Information Sciences,
Radboud University Nijmegen, The Netherlands
/ Open University, The Netherlands
Local network community detection aims to find a single community in a large network, while
inspecting only a small part of that network around a given seed node. This is much cheaper than
finding all communities in a network. Most methods for local community detection are formulated as
ad-hoc optimization problems. In this work, we instead start from a generative model for networks
with community structure. By assuming that the network is uniform, we can approximate the
structure of unobserved parts of the network to obtain a method for local community detection.
We apply this local approximation technique to two variants of the stochastic block model. To
our knowledge, this results in the first local community detection methods based on probabilistic
models. Interestingly, in the limit, one of the proposed approximations corresponds to conductance,
a popular metric in this field. Experiments on real and synthetic datasets show comparable or
improved results compared to state-of-the-art local community detection algorithms.
I.
INTRODUCTION
Networks are a convenient abstraction in many differ-
ent areas, such as social sciences, biology, and the world
wide web. A common structure in these real-world net-
works is a community, a group of nodes that are tightly
connected and often share other properties, for example,
biological function in a protein interaction network.
Imagine that you are trying to find such a community
If the network is very large, it
of nodes in a network.
becomes too expensive to look at all nodes and edges in
the network. Therefore local methods are needed. Local
community detection aims to find only one community
around a given set of seed nodes, by relying on local
computations involving only nodes relatively close to the
seed [1, 2]; in contrast to global community detection,
where all communities in a network have to be found.
For global community detection, it is possible to treat
the problem of finding all communities in a network as
a probabilistic inference problem. This puts global com-
munity detection on a solid foundation, and makes it
clear what a community is, and how these communities
manifest in the network structure.
But most algorithms for local community detection op-
erate by optimizing an ad-hoc objective function such as
conductance [2–4].
In this paper we will fill this gap, and propose a prob-
abilistic model for local community detection. Our con-
tributions can be summarized as follows:
1. We introduce an approximation technique for using
global models to perform local community detec-
tion.
2. We introduce the first method for local community
detection based on a generative model by using this
approximation.
3. We propose two algorithms for local community de-
tection, based on approximations of the stochastic
block model and of the degree-corrected stochastic
block model.
4. We provide a probabilistic interpretation of conduc-
tance, as limit behavior of the approximate degree-
corrected stochastic block model.
5. We show that the approximate stochastic block
model is a highly competitive algorithm, which out-
performs the state-of-the-art on three of five real
life benchmark datasets.
A. Related work
1. Local community detection
Local network community detection methods have
largely focused on optimizing conductance, which is a
measure of the quality of a graph cut. Empirically, con-
ductance has been shown to be a good quality metric for
communities in real-world networks [3], in the sense that
real-world communities often have a higher conductance
than other sets of nodes.
Because community detection is computationally hard
[5, 6], several different heuristics have been developed. A
common approach is to use spectral partitioning, which
involves finding the dominant eigenvector of a random
walk kernel. Andersen et al. [2] showed that communities
with a good conductance can be computed efficiently in
this way. In their method nodes are added to the commu-
nity in order of decreasing personalized pagerank score,
and the community along this 'sweep' with the highest
conductance is returned. Computing this personalized
pagerank is a global operation, but efficient local approx-
imations are possible, that only involve nodes near the
seed.
Several variants of this sweep method have been pro-
posed. Kloster and Gleich [7] propose an alternative to
the personalized pagerank score, based on the heat kernel
instead of random walks. Yang and Leskovec [3] propose
to find the first local optimum of conductance instead of
a global optimum. Other heuristics involve trying to find
multiple pagerank-like vectors and restarting the method
from different neighborhoods around the seed [4].
However, being based on graph cuts, a good conduc-
tance is often achieved by cutting a network into roughly
equal-sized parts, which is undesirable. To limit the size
of communities, a cut-off on the personalized pagerank
score can be used [2]; also, variations of the sweep meth-
ods have been proposed that stop at earlier local optima
[3].
2. Global community detection
Many different global community detection methods
have been developed for different classes of networks and
different community structures. For a complete overview,
we refer the reader to the surveys by Fortunato [5], Xie
et al. [8].
Here we focus on probabilistic models for global com-
munity detection. The simplest are the stochastic block
models [9, 10], which partition the nodes into communi-
ties, with varying probabilities of edges. These block
models produce networks that are very different than
real networks, in particular, the distribution of degrees
is very different. To more accurately model the node
degrees, Karrer and Newman [11] have proposed the
degree-corrected stochastic block model (DC-SBM). In
this model, they add an extra parameter to each node,
which controls the likelihood of edges to that node, and
hence the node's degree. This extra complexity comes at
the cost of making the model more difficult to fit, and
so degree correction might not be appropriate for all net-
works [12].
An issue with the stochastic block model is that the
number of communities has to be fixed because with more
communities there are more parameters in the model,
which makes it impossible to compare likelihoods. The
Infinite Relational Model [13] solves this problem by as-
suming an infinite number of communities in combination
with a Chinese restaurant prior over community struc-
tures.
The model of Newman and Leicht [14] goes one step
further, and has a parameter for each combination of
node and community, indicating the likelihood of edges
from nodes in that community to a particular other node.
This is similar to models based on non-negative matrix
factorization [15, 16]. These more complex models allow
for nodes to be in more than one community.
Aside from these flat models, also hierarchical models
have been developed for probabilistic network community
detection [17, 18].
With all probabilistic models there is the question of
inference, that is, how to find the parameters or distri-
bution of parameters that accurately model the observed
data. Common approaches are to use maximum likeli-
2
hood [used by e.g. 11, 12], variational Bayes [used by 19],
and Markov chain Monte Carlo sampling [used by e.g.
13].
More recently, there has been work using loopy be-
lief propagation for inference in stochastic block models
[20, 21]. There it has also been noted that these models
exhibit a phase transition: beyond a certain point in pa-
rameter space it is impossible to recover the true commu-
nity structure, even it was drawn from the model itself.
Furthermore, this transition is sharp for large networks.
B. Problem description
Before continuing, we will formalize the problem. The
network of interest is represented as an unweighted undi-
rected graph without self-loops. Let N be the number
of nodes in this graph, and M the number of edges.
The graph can be represented by an adjacency matrix
A, where aij = aji = 1 if there is an edge between
nodes i and j, and aij = 0 otherwise. Furthermore
(cid:80)N
(cid:80)N
j=1 aij = 2M .
i=1
The local community detection problem is now to find
the community cs that contains a given seed node s, while
inspecting only nodes and edges in or near that commu-
nity. We will only concern ourselves with a single seed
node in this paper, but this is not essential.
When working in a probabilistic setting, the goal
becomes to find the most likely community that con-
P(cs A, s). Even computing
tains the seed, argmaxcs
this probability entails a marginalization over all the
other clusters in the graph. So as a first simplifica-
tion we will instead search for the most likely clustering,
P(C A, s), and report the community cs in
argmaxC
this clustering.
We can assume that the seed is chosen independently
from the graph, and also independently from the clus-
tering, so we have that P(C A, s) ∝ P(C)P(A C). To
find the community containing the seed we only need to
maximize this quantity, and then to find the community
cs ∈ C that contains the seed.
II. THE STOCHASTIC BLOCK MODEL
We first model the community structure of all nodes
in the graph using the stochastic block model [11].
In
this global model, each node i is in exactly one commu-
nity, which we denote as ci, so the communities form a
partition of the set of nodes. The edges of the graph are
generated independently based on these communities. In
the standard stochastic block model the probability of an
edge between nodes i and j is P(aij) = πcicj where the π
are parameters.
For simplicity, we only consider two different values for
π, πc,c = λin for edges inside clusters and πcc(cid:48) = λout for
edges between different clusters c (cid:54)= c(cid:48), so
(cid:40)
P(aij) = πcicj =
λin
λout
if ci = cj
if ci (cid:54)= cj.
Since we do not know the value of these parameters λin
and λout, we put a conjugate Beta prior on them,
λin, λout ∼ β(α+, α−).
In this simpler variant of the stochastic block model,
the number of parameters of the model does not depend
on the number of communities. Hence, in contrast to
most other work on stochastic block models, we do not
have to fix the number of communities. Instead, we use
a prior over partitions, which allows for varying number
of communities of varying sizes.
It is well known that community sizes in real life net-
works follow a power law distribution [22]. Hence we
adopt the prior
(γ − 1)c−γ,
(1)
P(C) ∝(cid:88)
c∈C
where C ranges over all possible partitions of the set of N
nodes. Note that the particular choice of prior distribu-
tion is not critical to the rest of this work, and for other
applications other priors might make sense. In particular,
a common alternative choice is the Chinese Restaurant
Process.
A.
Inference
(cid:105)
(cid:105)
,
−
out
In the basic stochastic block model
P(A C) = Eλin,λout
πaij
cicj
(1 − πcicj )1−aij
(cid:104)(cid:89)
i<j
= Eλin,λout
α+
λin
(cid:104)
in =(cid:80)
in (1 − λin)α
aij, α−
−
in λout
in =(cid:80)
α+
out(1 − λout)α
i<j,ci=cj
(1−aij) and
where α+
out are the corresponding sums over ci (cid:54)= cj.
out and α−
α+
This likelihood has the same shape as the beta distri-
bution, so we can calculate the expectation exactly, to
get
i<j,ci=cj
P(A C) =
B(α+ + α+
in, α− + α−
in)B(α+ + α+
B(α+, α−)2
out, α− + α−
out)
.
(2)
Multiplying this by the prior on clusterings (1) gives the
posterior probability P(CA) up to a normalizing con-
stant.
B. Local approximation
The likelihood in equation (2) is still a function of the
entire clustering, so to find the most likely cluster con-
taining the seed we would need to consider clusterings of
3
the entire graph. To obtain a local model we make an
approximation based on the assumption that the graph
is uniform: all clusters of the graph are similar to each
other.
We make this idea concrete by assuming that all clus-
ters in the graph have approximately the same volume,
the same size, and the same fraction of within community
edges. Now, if the community containing the seed has n
nodes, while the graph has N nodes, this means that
there are approximately k = N/n communities that are
all similar to cs. Furthermore, suppose that this commu-
nity has w within community edges, then the parameters
for the stochastic block model can be approximated as
α+
in = kw,
in = kn(n − 1)/2 − α+
α−
in,
out = M − α+
α+
in,
out = N (N − 1)/2 − α+
α−
in − α−
in − α+
out.
With these quantities we can use equation (2) to approx-
imate the likelihood of the network given the community
that contains the seed. And hence also approximate the
posterior probability of a community given the network,
PSBM(cs, C A) = (γ − 1)kn−kγ
B(α+ + α+
in, α− + α−
in)B(α+ + α+
B(α+, α−)2
out, α− + α−
out)
.
(3)
Note that instead of taking k = N/n, we might rea-
son that, if the volume of the community containing the
seed is v and the graph has M edges, that there are ap-
proximately k = 2M/v communities. This is, in general,
a different estimate. For the stochastic block model, it
makes more sense to use k = N/n, because then there is
no dependence on the volume of the community.
III. DEGREE-CORRECTED BLOCK MODEL
The degree distribution of the stochastic block model
is not very realistic, because nodes inside a cluster have
similar degrees. In many real-world networks, there are
hub nodes, which have a much higher than average de-
gree; as well as leaf nodes with a very low degree. To
accurately model these phenomena Karrer and Newman
[11] have proposed the degree-corrected stochastic block
model (DC-SBM). In this model, they assign an extra
parameter di to each node, which controls the likelihood
of edges to that node, and hence the node's degree.
We can then model the edges as being drawn from a
Poisson distribution with mean didjπcicj ,
aij ∼ P (didjπcicj ) for all i < j.
Note that we use a Poisson distribution, which allows for
weighted edges with weight larger than 1, instead of the
Bernoulli distribution, because the mean might be larger
than 1.
We again place conjugate priors on all parameters,
which in this case follow a gamma distribution,
λin, λout ∼ Γ(α, θ)
di ∼ Γ(α, θ) for all i.
A.
Inference
For this degree-corrected model the likelihood of the
network given the clustering depends on parameters d
and λ. It is not possible to integrate over these parame-
ters analytically, and other authors have therefore chosen
to maximize over the parameters instead [11, 12]. Here
we use a variational approximation,
log P(A C) ≥ L(A, C)
= Ed,λ∼Q
(cid:2)log P(A, d, λ C) − log Q(d, λ)(cid:3).
(4)
tion Q(d, λ) = Qλin (λin)Qλout (λout)(cid:81)N
As is standard, we take Q to be the factorized distribu-
i=1 Qdi(di), where
each component has a gamma distribution, Qdi(di) =
Γ(di; αdi, θdi), Qλin (λin) = Γ(λin; αin, θin). This gives us
the following variational lower bound
L(A, C) =
(cid:1)
i<j
(cid:88)
(cid:0)aij log(didjπcicj ) − didjπcicj
− N(cid:88)
DKL(αdi, θdiα, θ)
− DKL(αin, θinα, θ)
i=1
− DKL(αout, θoutα, θ).
(5)
Where DKL(α, θα(cid:48), θ(cid:48)) is the Kullback-Leibler diver-
gence between two gamma distributions. And we have
assumed that aij ∈ {0, 1}, which implies that aij! = 1.
We find that the parameters that maximize L(A, C)
αdi = α − 1 +
aij
(cid:16)
θdi =
θ−1 +
πci,cj αdj θdj
(cid:17)−1
are
and
αin = α − 1 +
θin =
θ−1 +
(cid:16)
i<j,ci=cj
aij
αdiθdiαdj θdj
(cid:17)−1
,
i<j,ci=cj
and similarly for αout and θout. There is a mutual de-
pendence between these variables, so in practice we use
several iterations of the above equations to find a good
approximation of the parameters.
j=1
N(cid:88)
(cid:88)
(cid:88)
(cid:88)
j(cid:54)=i
i∈cs
(cid:88)
N(cid:88)
(cid:88)
i=1
i∈cs
v =
M =
K 2 =
(cid:16)
4
The variational approximation gives a lower bound to
the log-likelihood P(A C). In contrast, maximum like-
lihood would give an upper bound. This upper bound is
similar to L(A, C), but it does not include the Kullback-
Leibler terms. For large networks the first term of
L(A, C) dominates, and so the variational lower bound,
the true likelihood and the maximum likelihood upper
bound will all be close.
B. Local approximation
As before, we will make a local approximation of L,
which depends only on the community that contains the
seed.
It will be convenient to define
αdi = w + n(α − 1),
αdi = 2M + N (α − 1), and
αdi
2.
First of all, we can approximate θdi by changing the
sum over all other nodes j (cid:54)= i to a sum over all nodes j.
Then θdi becomes the same for all nodes in the same com-
munity, and under the assumption that all communities
are the same, θdi = θd is the same for all nodes,
θd =
θ−1 + θdλinv + θdλout( M − v)
(cid:17)−1
.
Furthermore, because αdi does not depend on the clus-
tering, we can make the following approximation L of L,
L(A, C) ≈ L(A, c) = 2M log(θd)+kw(ψ(αin)+log(θin))
+ (M − kw)(ψ(αout) + log(θout))
2αinθin − ( M 2 − kv2)θd
− k(v2 − K 2)θd
2αoutθout
+ N α log θd − M θd/θ
− DKL(αin, θinα, θ) − DKL(αout, θoutα, θ) + κ,
(6)
where κ is a constant that depends only on the network
and on the priors.
The likelihood of the degree-corrected model is based
on the degrees of nodes and the volume of communities.
So in contrast to the previous section, here it makes sense
to estimate the number of communities as k = 2M/v
instead of N/n.
As before, we multiply this approximate likelihood by
the prior, which we approximate as P(C) ≈ (γ−1)kn−kγ,
to obtain the posterior
PDCBM(cs, C A) = e L(A,c)(γ − 1)kn−kγ.
IV. LIMITING BEHAVIOR
The premise of local community detection is to find a
community without considering the entire graph, or even
a significant portion of the graph. This is only possible if
the community is small compared to the graph. We can
take this assumption one step further, and consider what
happens if the graph becomes infinitely large compared
to the cluster.
Therefore we take the limit of the approximate likeli-
hood as N → ∞, assuming that the average degree M/N
remains constant. With the stochastic block model from
Section II B we get that
log PSBM(cs, C A)
N log N
=
w
n
.
lim
N→∞
With the degree-corrected model we obtain
2 log PDCBM(cs, C A)
N log N
=
w
v
− 1,
lim
N→∞
which is exactly equal to the negation of conductance.
In other words, under this model, and in the limit of an
infinitely large graph, the a posteriori most likely cluster
corresponds to the cluster of minimum conductance.
Note that conductance has a global optimum with a
large community that contains all nodes, since in that
case w = v. Even in the non-limiting case, as the net-
work becomes larger, so does the optimal community.
And for very large networks it becomes impossible to re-
cover small communities. This phenomenon is called the
resolution limit [23], and is shared by many network com-
munity detection methods. To avoid the resolution limit,
a parameter must be introduced into the objective func-
tion, for instance by replacing the graph size or graph
volume [24]. In our case, we could take N as a formal
parameter, instead of using the actual number of nodes
in the network (keeping the average degree M/N fixed).
In this way, the search for a community is in effect per-
formed in a subnetwork of a given size.
V. EXPERIMENTS
In this section, we experimentally evaluate the pro-
posed models and approximations. We use the following
experimental protocol:
1. pick a random community from the set of all com-
munities.
2. pick a random seed from this community.
3. run the method(s) with this seed.
4. compare the recovered community to the true one
using the F1 score. For sets of nodes c, d the F1
score amounts to F1(c, d) = 2c∩d/(c+d),which
is 1 if the communities are identical, and 0 if they
5
are disjoint. We exclude the seed from this com-
parison since it always occurs in both communities,
and we would otherwise see a good F1 score for the
trivial community containing only the seed.
A. Methods
We compare three classes of methods.
1. Global generative models
We optimize the likelihood of the global clustering
models with a Louvain method [25]. This yields a par-
tition of the nodes.
In this partition, there is always
a single community that contains the seed. We denote
this method as gSBM (global stochastic block model)
and gDCBM (global degree-corrected block model). We
use uninformative priors for all parameters, β(1, 1) in the
stochastic block model and Γ(1, 1) in the degree-corrected
model. For the power law prior on community sizes we
use γ = 2.
Note that it is somewhat unfair to compare local and
global models. A global method has access to more infor-
mation. The goal of local community detection is not to
outperform global methods, but rather to achieve com-
parable results faster while looking at only a small part
of the network.
2. Local approximations
We have implemented a simple greedy algorithm to
optimize P(C A) for the stochastic and degree-corrected
block models. The algorithm starts from the community
{s} that contains only the seed. Then we consider all
neighboring nodes of the current community in a random
order, and for each node we add it to the community if
doing so would improve the approximate likelihood. This
optimization procedure is then repeated, until none of
the neighboring nodes are added. We further restart this
search 10 times, and pick the community with the highest
approximate likelihood. We denote this method as aSBM
(local approximate stochastic block model) and aDCBM
(local approximate degree-corrected block model).
Each iteration of this greedy optimization procedure
takes time proportional to the volume of the retrieved
community, and the total number of iterations is bounded
by the diameter of the community D, which is very small
in practice. This makes the total runtime O(vD).
We also consider a variant of aDCBM with an explicit
parameter N , as discussed in Section IV. We report re-
sults with N = 1000, and set M to the average node
degree times N (aDCBM1k). The supplementary mate-
rial includes results for different values of N .
3. State-of-the-art methods for local community detection
• PPR. The algorithm by Andersen et al. [2] based
on the Personalized Page Rank graph diffusion.
We use the implementation included with the HK
method.
• HK. The algorithm by Kloster and Gleich [7], using
a Heat Kernel.Code is available at https://www.
cs.purdue.edu/homes/dgleich/codes/hkgrow.
• YL. The algorithm by Yang and Leskovec [3] with
conductance as scoring function. This method uses
a different stopping condition compared to PPR,
selecting a local optimum of conductance, instead
of searching for a more global optimum. This intro-
duces a bias towards finding smaller communities.
• LEMON. The Local Expansion via Minimum One
Instead of con-
Norm algorithm by Li et al. [4].
sidering a single probability vector as in HK or
YL , this method uses the space spanned by several
short random walks. Communities are found by
solving an l1-penalized linear programming prob-
lem. The algorithm includes a number of heuris-
tic post-processing steps. Code is available at
https://github.com/yixuanli/lemon.
B. Artificial datasets
We first look at artificial datasets, by using the
LFR benchmark [26] to generate networks with a
known community structure. We used the parame-
ter settings N=5000 k=10 maxk=50 t1=2 t2=1 minc=20
maxc=100, which means that the graph has 5000 nodes,
and between 20 and 100 communities, each with between
10 and 50 nodes. We vary the mixing parameter (mu),
which determines what fraction of the edges are between
different communities. More mixing makes the problem
harder.
The LFR benchmark is very similar to the degree-
corrected block model. The differences are that node de-
grees follow a power law distribution in the LFR model,
while we used a gamma distribution, and that in the LFR
benchmark the edges are not completely independent be-
cause the degree of each node must match a previously
drawn value. Nevertheless, we expect the DCBM to give
a good fit to these networks.
The results of these experiments are shown in Table II.
Here we see that the global degree-corrected model per-
forms better than the simple stochastic block model.
This is not surprising, since the LFR benchmark has
nodes with varying degrees. These results carry over to
the local approximations, which do perform significantly
worse than the global models on these datasets. Out
of the local methods, the aDCBM and LEMON models
achieve the best results.
6
Table I. Overview of the SNAP datasets used in the experi-
ments.
Dataset
#node
#edge #comm
c
334863
317080
1134890
925872
Amazon
1049866
DBLP
Youtube
2987624
LiveJournal 3997962 34681189
Orkut
3072441 117185083
151037 19.4
13477 53.4
8385 13.5
287512 22.3
6288363 14.2
C. Real-world networks
We use five social and information network datasets
with ground-truth from the SNAP collection [27]. These
datasets are summarized in Table I. We consider all avail-
able ground-truth communities with at least 3 nodes. All
experiments were performed on a random subsample of
1000 communities.
Yang and Leskovec [3] also defined a set of top 5000
communities for each dataset. These are communities
with a high combined score for several community good-
ness metrics, among which is conductance. We therefore
believe that communities in this set are biased to be more
easy to recover by optimizing conductance.
In addition to the SNAP datasets, we also include
the Flickr social network [28]. As well as some classical
datasets with known communities: Zachary's karate club
[29]; Football: A network of American college football
games [30]; Political books: A network of books about
US politics [31]; and Political blogs: Hyperlinks between
weblogs on US politics [32]. These datasets might not be
very well suited for local community detection since they
have very few communities.
We see in Table II that, while on the artificial bench-
mark networks the global model significantly outper-
forms the local approximation, on the real-world net-
works this is not the case. We believe that this is be-
cause the ground-truth communities on these networks
are much smaller, and the considered local methods tend
to find smaller communities.
Additionally, the simple stochastic block model outper-
forms the degree-corrected model on all SNAP datasets
except for the Amazon dataset. We found this surpris-
ing, because all these datasets do have nodes with widely
varying degrees. However, the number of within commu-
nity edges varies much less. For instance a node in the
√
DBLP dataset with degree di will have on the order of
di within community edges, which means that the truth
is in between the aSBM (which assumes O(1) edges) and
aDCBM models (which assumes O(di) edges).
An issue with the local approximation is that nodes
inside communities are not representative of the entire
network. For instance, on the Youtube network the aver-
age node degree is 5.3, while the average degree of nodes
that are inside at least one community is 33.7.
When using an explicit N = 1000, the results on the
SNAP networks improve, again because the ground-truth
communities on these networks tend to be small. For the
three largest datasets, different values of N have a large
influence on the size of the recovered community. This is
likely due to the fact that it is possible to find communi-
ties at all scales in these networks. See the supplementary
material for the results for different values of N .
The results using the top 5000 communities are much
better, which is not surprising, since these top communi-
ties were selected to be easier to find. The trend between
the different methods is similar, with the aSBM and aD-
CBM methods performing best in most cases.
Surprisingly, the local approximations outperform the
global community detection methods on most of the real-
world datasets. The likely reason is that the ground-
truth communities in the SNAP datasets are relatively
small. And because of the greedy optimization strategy,
the local methods tend to find smaller communities. The
global methods, in contrast, find larger communities that
better fit the model, but which likely combine several
ground-truth communities. This means that the local
methods achieve a much better precision at the cost of a
somewhat lower recall, resulting in an overall higher F1
score.
Results on the smaller networks are mixed. All these
datasets, except for football have very large clusters for
their size, which are easier to recover with the HK and
PPR methods. The aDCBM method with N = 1000
is also sometimes better able to find good communities
on these datasets, because those networks have fewer
than 1000 nodes, so increasing N increases the size of
the found community.
We were unable to run LEMON on the large SNAP
datasets due to its memory usage.
VI. DISCUSSION
The local approximations are based on the assumption
that all communities are alike. We needed to make this
assumption to be able to say something about the global
properties of the network given only a single community.
In real-world networks there is often a large variation in
the size of the communities. Our approximation might
work well if the community is close to the average size,
but for very large or very small communities it is not
accurate. Better approximations might be possible by
finding more than one community (but still a small sub-
set of the network), or by modeling the distribution of
community sizes.
When we are interested in local community structure,
it would seem to make sense to consider models that only
7
have this local structure. For instance, models with a
single community that stands apart from a background.
But to fit such a model to an observed graph we would
also need a good model for the background, and so we
should also model the structure in the background. In
other words, we also need to find communities in the rest
of the graph, and the method would not be local. If we
were to instead use a background without further struc-
ture, then the closest fit to the observed network will be
obtained by using the structure of the single community
to explain the largest variances in the entire network, so
the obtained 'community' would cover roughly half of the
nodes. Our approach of assuming that the background is
similar to the community containing the seed is a good
compromise, as illustrated by the experiments.
In this work we maximize over clusters. It would be
interesting and useful to estimate the marginals instead.
That is, the probability that a node i is inside the same
cluster as the seed, conditioned on the graph. While
Variational Bayes gives a decent approximation to the
log-likelihood,
it does not give approximations to the
marginals. Indeed, trying to use a variational bound to
marginalize over all but one of the cluster membership
indicators leads to an objective that is identical to L, ex-
cept for a relatively small entropy term. It remains to be
seen if other inference methods can be used to estimate
the marginals, and thus to in some sense find all possible
communities containing a given seed.
The assumption used to derive the local approxima-
tions is not particular to the stochastic block models that
we have used here, and the same technique can be used
for other global community detection methods that are
based on a partition of the nodes, such as the models of
Kemp et al. [13] or Newman and Leicht [14]. However,
it is often assumed that in practice some nodes can be-
long to more than one community. There exist several
global models that include overlapping communities [see
e.g. 16, 33]. In the derivation we used the fact that if all
communities are identical and each node is in exactly one
community, then there are N/n communities. But when
nodes can be in more than one community, it is no longer
clear how many communities there are. We would need
a reliable estimate of the average number of communities
that cover a node. It therefore remains an open problem
how the local approximation can be applied to models
with overlapping clusters.
ACKNOWLEDGEMENTS
This work has been partially funded by the Nether-
lands Organization for Scientific Research (NWO) within
the EW TOP Compartiment 1 project 612.001.352.
[1] I. M. Kloumann and J. M. Kleinberg, in Proceedings
of the 20th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD '14 (ACM,
Table II. F1 score between recovered communities and ground-truth (excluding the seed node). The best result for each dataset
is indicated in bold, as are the results not significantly worse according to a paired T-test (at significance level 0.01).
Dataset
gSBM gDCBM aSBM aDCBM aDCBM-1k YL LEMON HK PPR
Global methods
Local methods
8
1.000
0.999
1.000
0.998
0.997
0.958
0.990
0.920
0.911
0.756
0.433 0.426
0.613
0.583
0.534
0.466
0.368
0.258
0.911
0.812
0.800
0.659
0.458
0.138
0.330
0.287
0.040
0.041
0.010
0.245
0.220
0.054
0.025
0.007
0.614
0.334
0.085
0.354
0.195
0.395 0.431
0.406 0.349
0.099 0.071
0.072 0.043
0.020
0.014
0.844 0.903
0.647 0.567
0.140
0.195
0.666 0.715
0.334 0.312
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
0.792
Amazon (top 5000)
DBLP (top 5000)
0.442
Youtube (top 5000)
0.083
LiveJournal (top 5000) 0.507
Orkut (top 5000)
0.233
Karate
Football
Pol.Blogs
Pol.Books
Flickr
0.165
0.101
0.790 0.817
0.192
0.039
0.429
0.274
0.204
0.090
0.379
0.727
0.103
0.295
0.164
0.448
0.682
0.040
0.243
0.050
0.895
0.799
0.726
0.529
0.322
0.093
0.447
0.344
0.084
0.052
0.016
0.895
0.571
0.172
0.672
0.119
0.740
0.769
0.090
0.451
0.066
0.307
0.274
0.168
0.121
0.085
0.069
0.381
0.245
0.082
0.054
0.034
0.780
0.419
0.241
0.521
0.097
0.562
0.784
0.015
0.175
0.012
0.925
0.859
0.587
0.533
0.427
0.279
0.229
0.240
0.075
-
-
0.397
0.339
0.098
-
-
0.683
0.322
0.151
0.605
0.025
0.881 0.352
0.090 0.136
0.039 0.040
0.039 0.040
0.039 0.041
0.037 0.039
0.221 0.119
0.199 0.194
0.031 0.052
0.028 0.035
0.032 0.019
0.709 0.527
0.342 0.329
0.067 0.116
0.569 0.478
0.312 0.260
0.799 0.908
0.452 0.260
0.661 0.535
0.629 0.653
0.054 0.118
New York, NY, USA, 2014) pp. 1366–1375.
[2] R. Andersen, F. Chung, and K. Lang, in Foundations of
Computer Science, 2006. FOCS'06. 47th Annual IEEE
Symposium on (IEEE, 2006) pp. 475–486.
[3] J. Yang and J. Leskovec, in Proceedings of the ACM
SIGKDD Workshop on Mining Data Semantics, MDS '12
(ACM, 2012) pp. 3:1–3:8.
[4] Y. Li, K. He, D. Bindel, and J. E. Hopcroft, in Pro-
ceedings of the 24th International Conference on World
Wide Web (International World Wide Web Conferences
Steering Committee, 2015) pp. 658–668.
[5] S. Fortunato, Physics Reports 486, 75 (2010).
[6] J. Shi and J. Malik (IEEE Computer Society, Washing-
[12] X. Yan, C. Shalizi, J. E. Jensen, F. Krzakala, C. Moore,
L. Zdeborov´a, P. Zhang, and Y. Zhu, Journal of Statis-
tical Mechanics: Theory and Experiment 2014, P05007
(2014).
[13] C. Kemp, J. B. Tenenbaum, T. L. Griffiths, T. Yamada,
and N. Ueda, in Proceedings of the 21st National Con-
ference on Artificial Intelligence - Volume 1 , AAAI'06
(AAAI Press, 2006) pp. 381–388.
[14] M. E. J. Newman and E. A. Leicht, Proceedings of
the National Academy of Sciences 104, 9564 (2007),
http://www.pnas.org/content/104/23/9564.full.pdf.
[15] I. Psorakis, S. Roberts, M. Ebden, and B. Sheldon, Phys-
ical Review E 83, 066114+ (2011).
ton, DC, USA, 2000) pp. 888–905.
[16] B. Ball, B. Karrer, and M. E. J. Newman, Physical Re-
[7] K. Kloster and D. F. Gleich, in Proceedings of the 20th
ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD '14 (ACM, New York,
NY, USA, 2014) pp. 1386–1395.
[8] J. Xie, S. Kelley, and B. K. Szymanski, ACM Comput.
Surv. 45, 43:1 (2013).
view E 84, 036103 (2011).
[17] H. Zhang, C. L. Giles, H. C. Foley, and J. Yen, in Pro-
ceedings of the 22Nd National Conference on Artificial
Intelligence - Volume 1 , AAAI'07 (AAAI Press, 2007)
pp. 663–668.
[18] C. Blundell and Y. W. Teh, in Advances in Neural Infor-
[9] S. L. Paul W. Holland, Kathryn Blackmond Laksey, So-
mation Processing Systems (2013).
cial Networks 5, 109 (1983).
[19] J. M. Hofman and C. H. Wiggins, Phys Rev Lett 100,
[10] K. F. Carolyn J. Anderson, Stanley Wasserman, Social
258701 (2008), 18643711[pmid].
Networks 14, 137 (1992).
[20] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´a,
[11] B. Karrer and M. E. J. Newman, Phys. Rev. E 83, 016107
Phys. Rev. E 84, 066106 (2011).
(2011).
9
[21] P. Zhang, C. Moore, and M. E. J. Newman, Phys. Rev.
(2005) pp. 36–43.
E 93, 012303 (2016).
[22] G. Palla, I. Derenyi, I. Farkas, and T. Vicsek, Nature
435, 814 (2005).
[23] S. Fortunato and M. Barth´elemy, Proceedings of the Na-
tional Academy of Sciences of the United States of Amer-
ica 104, 36 (2007).
[24] J. Reichardt and S. Bornholdt, Physical Review Letters
93, 218701 (2004).
[25] V. D. Blondel, J.-L. Guillaume, R. Lambiotte,
and
E. Lefebvre, Journal of Statistical Mechanics: Theory
and Experiment 2008, P10008 (2008), arXiv:0803.0476.
[26] A. Lancichinetti, S. Fortunato, and F. Radicchi, Physical
Review E 78, 046110 (2008).
[27] J. Leskovec and A. Krevl, "SNAP Datasets: Stan-
ford large network dataset collection," http://snap.
stanford.edu/data (2014).
[28] X. Wang, L. Tang, H. Liu, and L. Wang, Knowledge and
(2012), 10.1007/s10115-
Information Systems (KAIS)
012-0555-0.
[29] W. W. Zachary, Journal of Anthropological Research 33,
452 (1977).
[30] M. Girvan and M. E. J. Newman, Proceedings of the
National Academy of Sciences of the United States of
America 99, 7821 (2002).
[31] V. Krebs, "New political patterns," Editorial (2004).
[32] L. Adamic and N. Glance, in LinkKDD '05: Proceed-
ings of the 3rd international workshop on Link discovery
[33] A. McDaid and N. Hurley, in Proceedings of the 2010
International Conference on Advances in Social Networks
Analysis and Mining, ASONAM '10 (IEEE Computer
Society, Washington, DC, USA, 2010) pp. 112–119.
Appendix A: Additional results
The tables and figures below include additional results:
• Table III: Statistics on the additional datasets.
• Table IV: Mean size of the recovered community.
• Table V: Mean precision of the recovered commu-
nity.
• Table VI: Mean recall of the recovered community.
• Table VII: Runtime of the community detection al-
gorithms.
• Fig. 1: F1 score on SNAP datasets as a function of
the N parameter.
• Fig. 2: Mean size of the recovered community as a
function of N parameter.
10
Table III. Overview of the additional datasets.
Dataset
#node
#edge #comm
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
5000
5000
5000
5000
5000
5000
25111
25124
25125
25127
25118
25127
101
101
101
101
101
101
925872
334863
1049866
317080
1134890
2987624
3997962 34681189
3072441 117185083
151037
13477
8385
287512
6288363
c
49.5
49.5
49.5
49.5
49.5
49.5
19.4
53.4
13.5
22.3
14.2
Amazon (top 5000)
925872
DBLP (top 5000)
1049866
Youtube (top 5000)
2987624
LiveJournal (top 5000) 3997962 34681189
Orkut (top 5000)
3072441 117185083
334863
317080
1134890
Karate
Football
Pol.Blogs
Pol.Books
Flickr
34
115
1490
105
35313
78
613
16715
441
3017530
5000
5000
5000
5000
5000
13.5
22.4
14.6
27.8
215.7
2
12
2
3
17.0
9.6
745.0
35.0
171 4336.1
Figure 1. F1 score as a function of the N parameter with the aDCBM method.
10110210310410500.10.20.30.40.50.6NF1DBLPAmazonYoutubeLiveJournalOrkut Table IV. Mean size of the recovered communities.
Global methods
Local methods
Dataset
gSBM gDCBM aSBM aDCBM aDCBM-1k YL LEMON
HK
PPR
11
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
Amazon (top 5000)
DBLP (top 5000)
Youtube (top 5000)
LiveJournal (top 5000)
Orkut (top 5000)
Karate
Football
Pol.Blogs
Pol.Books
Flickr
48.9
48.7
46.2
42.9
30.5
16.3
23.9
27.6
1283.5
375.0
2089.9
15.4
24.8
434.0
98.5
973.5
2.8
8.9
100.7
10.7
2796.9
49.0
49.0
49.0
48.9
45.9
20.7
20.5
19.7
18.1
16.1
13.4
10.5
33.6
39.8
38.0
253.2
11.1
11.2
31.1
82.1
1288.8 367.5
24.5
37.7
42.0
158.1
7.6
8.8
30.6
23.7
1272.3 181.5
2.0
9.8
18.0
21.1
4.6
9.5
45.1
12.3
608.5 1054.0
45.7
42.1
40.9
30.8
20.6
13.6
16.3
21.3
78.6
107.6
283.8
10.4
15.6
42.6
48.7
219.9
6.2
10.2
20.3
9.1
222.5
44.2 9.4
39.7 8.4
32.4 5.9
21.9 5.2
14.3 4.8
9.9 4.9
11.5 6.4
20.2 6.0
31.8 9.3
89.5 10.8
163.0 11.1
9.6 7.1
12.3 6.2
26.3 6.9
36.6 13.0
202.6 9.1
15.3 8.8
10.8 8.8
97.9 7.2
25.2 6.7
436.9 12.9
50.2 159.0 1518.9
49.2 2264.6 2155.1
43.4 2410.0 2366.2
42.9 2398.3 2323.2
49.5 2412.0 2294.6
56.8 2412.8 2269.3
88.8 20819.9
41.9
46.2
55.0 24495.0
53.2 147.9 20955.5
- 153.2 3428.7
- 212.0 1634.0
28.3 4828.6
40.0
46.8 9506.8
40.0
55.3 113.9 10867.7
- 100.0
549.8
- 264.8 1607.4
16.7
40.5
17.1
15.0
44.9
56.5
68.2 492.7 1051.1
43.3
53.4
75.4 174.2 1158.1
49.3
Figure 2. Mean community size as a function of the N parameter with the aDCBM method.
101102103104105101102103NnDBLPAmazonYoutubeLiveJournalOrkut Table V. Mean precision of the recovered communities.
Global methods
Local methods
Dataset
gSBM gDCBM aSBM aDCBM aDCBM-1k YL LEMON HK PPR
12
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
1.000 1.000
1.000 1.000
0.997
0.971
0.988
0.951
0.890 0.915
0.647
0.577
0.256
0.231
0.023
0.028
0.005
0.924
0.992
0.827
0.982
0.826
0.945
0.725
0.898
0.800
0.562
0.644 0.236
0.385
0.376
0.497 0.346
0.082 0.058
0.055 0.033
0.012
0.008
0.977 0.949
0.740 0.548
0.122
0.149
0.696 0.641
0.377 0.299
0.175
0.158
0.035
0.016
0.004
0.546
0.233
0.052
0.275
0.136
0.743
Amazon (top 5000)
DBLP (top 5000)
0.343
Youtube (top 5000)
0.052
LiveJournal (top 5000) 0.405
Orkut (top 5000)
0.173
Karate
Football
Pol.Blogs
Pol.Books
Flickr
0.607
0.427
0.825 0.818
0.751
0.603
0.625
0.610
0.242
0.271
0.958 0.961
0.755
0.686
0.806 0.732
0.641
0.624
0.395 0.098
0.919
0.840
0.792
0.651
0.468
0.225
0.409
0.378
0.066
0.039
0.010
0.959
0.594
0.139
0.623
0.140
0.872
0.768
0.670
0.616
0.271
0.929
0.823
0.669
0.569
0.465
0.387
0.422
0.372
0.086
0.059
0.032
0.928
0.544
0.226
0.592
0.235
0.928
0.870
0.690
0.607
0.446
0.266
0.157
0.167
0.047
-
-
0.317
0.229
0.062
-
-
0.862 0.330
0.069 0.114
0.020 0.020
0.020 0.021
0.020 0.021
0.019 0.020
0.166 0.088
0.161 0.153
0.024 0.037
0.021 0.025
0.020 0.012
0.649 0.469
0.274 0.260
0.045 0.080
0.479 0.400
0.278 0.207
0.732
0.204
0.813 0.909
0.913
0.828
0.371 0.154
0.776 0.795 0.765 0.433
0.699
0.587 0.586
0.201 0.222
0.125
0.596
0.294
13
Table VI. Mean recall of the recovered communities.
Global methods
Local methods
Dataset
gSBM gDCBM aSBM aDCBM aDCBM-1k YL LEMON HK PPR
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
0.999
0.997
0.957
0.907
0.706
0.362
0.808
0.543
0.445
0.346
0.454
0.967
Amazon (top 5000)
0.811
DBLP (top 5000)
Youtube (top 5000)
0.669
LiveJournal (top 5000) 0.932
Orkut (top 5000)
0.663
Karate
Football
Pol.Blogs
Pol.Books
Flickr
0.101
0.777
0.129
0.196
0.259
1.000
1.000
0.997
0.993
0.914
0.366
0.831
0.593
0.267
0.345
0.445
0.972
0.856
0.605
0.938
0.683
0.060
0.822
0.020
0.364
0.089
0.453
0.425
0.380
0.322
0.248
0.168
0.575
0.421
0.300
0.286
0.225
0.792
0.649
0.421
0.707
0.396
0.241
0.710
0.055
0.198
0.127
0.909
0.811
0.794
0.647
0.438
0.130
0.750
0.495
0.181
0.172
0.204
0.901
0.737
0.483
0.892
0.463
0.317
0.679
0.023
0.171
0.049
0.892
0.792
0.713
0.510
0.297
0.071
0.689
0.416
0.252
0.260
0.257
0.883
0.667
0.420
0.803
0.175
0.721
0.781
0.069
0.417
0.046
0.232
0.218
0.114
0.077
0.052
0.041
0.467
0.223
0.112
0.081
0.056
0.763
0.396
0.364
0.554
0.095
0.446
0.758
0.008
0.112
0.012
0.940
0.865
0.561
0.503
0.439
0.330
0.913
0.701
0.436
-
-
0.943
0.897
0.714
-
-
0.641
0.894
0.084
0.670
0.022
0.995 1.000
1.000 1.000
1.000 1.000
1.000 1.000
1.000 0.999
0.971 0.964
0.755 0.985
0.483 0.853
0.117 0.757
0.193 0.661
0.432 0.690
0.977 0.997
0.723 0.939
0.272 0.826
0.910 0.962
0.530 0.711
0.794 0.913
0.921 0.923
0.593 0.707
0.760 0.819
0.042 0.125
Table VII. Mean runtime in seconds per community. Note that for the global methods a partition of the nodes is computed
only once.
14
Dataset
gSBM gDCBM aSBM aDCBM aDCBM-1k
YL LEMON
HK PPR
Global methods
Local methods
LFR (mu=0.1)
LFR (mu=0.2)
LFR (mu=0.3)
LFR (mu=0.4)
LFR (mu=0.5)
LFR (mu=0.6)
Amazon
DBLP
Youtube
LiveJournal
Orkut
0.017
0.017
0.018
0.018
0.017
0.018
0.036
0.032
0.059
0.130
0.095
0.034
Amazon (top 5000)
DBLP (top 5000)
0.034
0.056
Youtube (top 5000)
LiveJournal (top 5000) 0.124
Orkut (top 5000)
0.101
Karate
Football
Pol.Blogs
Pol.Books
Flickr
0.016
0.016
0.016
0.016
0.020
0.017 0.024
0.017 0.026
0.018 0.029
0.018 0.031
0.018 0.032
0.018 0.030
0.035 0.022
0.034 0.026
0.055 0.854
0.116 0.726
0.095 7.551
0.032 0.020
0.036 0.022
0.056 0.707
0.118 0.111
0.088 1.960
0.016 0.017
0.016 0.017
0.016 0.044
0.016 0.018
0.019 3.701
0.024
0.028
0.032
0.032
0.033
0.032
0.023
0.027
1.138
0.684
4.923
0.018
0.023
0.398
0.095
1.774
0.015
0.016
0.074
0.016
4.015
0.030
0.033
0.035
0.035
0.035
0.035
0.128
0.149
0.069
0.128
0.139
0.132
4.418
0.029
5.799
0.035
0.200
18.975
0.508 189.899
3.487 1632.087
2.629
0.027
5.637
0.028
0.146
18.295
0.095 101.696
2.144 1542.724
0.019
0.020
0.069
0.021
2.017
0.009
0.019
0.055
0.011
25.639
0.113 0.047 1.012
0.122 0.061 1.095
0.047 0.034 0.649
0.131 0.068 1.091
0.132 0.069 1.123
0.133 0.072 1.116
1.564 0.031 2.803
1.757 0.039 1.405
4.959 0.074 0.514
- 0.115 0.312
- 0.212 0.488
1.559 0.022 1.643
1.736 0.036 1.319
4.855 0.078 0.520
- 0.101 0.446
- 0.228 0.602
0.008 0.007 0.017
0.012 0.006 0.046
0.038 0.015 0.365
0.008 0.006 0.040
11.929 0.045 0.133
|
1610.09516 | 1 | 1610 | 2016-10-29T14:30:57 | Finding Street Gang Members on Twitter | [
"cs.SI",
"cs.CL",
"cs.CY",
"cs.IR"
] | Most street gang members use Twitter to intimidate others, to present outrageous images and statements to the world, and to share recent illegal activities. Their tweets may thus be useful to law enforcement agencies to discover clues about recent crimes or to anticipate ones that may occur. Finding these posts, however, requires a method to discover gang member Twitter profiles. This is a challenging task since gang members represent a very small population of the 320 million Twitter users. This paper studies the problem of automatically finding gang members on Twitter. It outlines a process to curate one of the largest sets of verifiable gang member profiles that have ever been studied. A review of these profiles establishes differences in the language, images, YouTube links, and emojis gang members use compared to the rest of the Twitter population. Features from this review are used to train a series of supervised classifiers. Our classifier achieves a promising F1 score with a low false positive rate. | cs.SI | cs | Finding Street Gang Members on Twitter
Lakshika Balasuriya
Kno.e.sis Center
Wright State University
Dayton OH, USA
[email protected]
Sanjaya Wijeratne
Kno.e.sis Center
Derek Doran
Kno.e.sis Center
Amit Sheth
Kno.e.sis Center
Wright State University
Wright State University
Wright State University
Dayton OH, USA
[email protected]
Dayton OH, USA
[email protected]
Dayton OH, USA
[email protected]
6
1
0
2
t
c
O
9
2
]
I
S
.
s
c
[
1
v
6
1
5
9
0
.
0
1
6
1
:
v
i
X
r
a
Abstract-Most street gang members use Twitter to intimidate
others, to present outrageous images and statements to the world,
and to share recent illegal activities. Their tweets may thus be
useful to law enforcement agencies to discover clues about recent
crimes or to anticipate ones that may occur. Finding these posts,
however, requires a method to discover gang member Twitter
profiles. This is a challenging task since gang members represent
a very small population of the 320 million Twitter users. This
paper studies the problem of automatically finding gang members
on Twitter. It outlines a process to curate one of the largest sets
of verifiable gang member profiles that have ever been studied.
A review of these profiles establishes differences in the language,
images, YouTube links, and emojis gang members use compared
to the rest of the Twitter population. Features from this review
are used to train a series of supervised classifiers. Our classifier
achieves a promising F1 score with a low false positive rate.
Keywords-Street Gangs, Twitter Profile Identification, Gang
Activity Understanding, Social Media Analysis
I.
INTRODUCTION AND MOTIVATION
The crime and violence street gangs introduce into neigh-
borhoods is a growing epidemic in cities around the world1.
Today, over 1.23 million people in the United States are
members of a street gang [1], [2], which is a coalition of peers,
united by mutual interests, with identifiable leadership and
internal organization, who act collectively to conduct illegal ac-
tivity and to control a territory, facility, or enterprise [3]. They
promote criminal activities such as drug trafficking, assault,
robbery, and threatening or intimidating a neighborhood [2].
Moreover, data from the Centers for Disease Control in the
United States suggests that the victims of at least 1.3% of all
gang-related2 homicides are merely innocent bystanders who
live in gang occupied neighborhoods [4].
Street gang members have established online presences co-
inciding with their physical occupation of neighborhoods. The
National Gang Threat Assessment Report confirms that at least
tens of thousands of gang members are using social networking
websites such as Twitter and video sharing websites such as
YouTube in their daily life [1]. They are very active online; the
2007 National Assessment Center's survey of gang members
found that 25% of individuals in gangs use the Internet for at
least 4 hours a week [5]. Gang members typically use social
networking sites and social media to develop online respect
for their street gang [6] and to post intimidating, threatening
images or videos [7]. This "Cyber-" or "Internet banging" [8]
behavior is precipitated by the fact that an increasing number
1http://goo.gl/OjWeYf
2The terms 'gang' and 'street gang' are used interchangeably in this paper.
of young members of the society are joining gangs [9], and
these young members have become enamored with technology
and with the notion of sharing information quickly and publicly
through social media3. Stronger police surveillance in the
physical spaces where gangs congregate further encourages
gang members to seek out virtual spaces such as social media
to express their affiliation, to sell drugs, and to celebrate their
illegal activities [10].
Gang members are able to post publicly on Twitter without
fear of consequences because there are few tools law enforce-
ment can use to surveil this medium [11]. Police departments
across the United States instead rely on manual processes to
search social media for gang member profiles and to study
their posts. For example, the New York City police department
employs over 300 detectives to combat teen violence triggered
by insults, dares, and threats exchanged on social media,
and the Toronto police department teaches officers about the
use of social media in investigations [12]. Officer training is
broadly limited to understanding policies on using Twitter in
investigations and best practices for data storage [13]. The
safety and security of city neighborhoods can thus be improved
if law enforcement were equipped with intelligent tools to
study social media for gang activity.
The need for better tools for law enforcement cannot be un-
derscored enough. Recent news reports have shown that many
incidents involving gangs start on Twitter, escalate over time,
and lead to an offline event that could have been prevented
by an early warning. For example, the media reported on a
possible connection between the death of a teenage rapper
from Illinois and the final set of tweets he posted. One of
his last tweets linked to a video of him shouting vulgar words
at a rival gang member who, in return, replied "I'ma kill you"
on social media4. In a following tweet, the teenage rapper
posted "im on 069", revealing his location, and was shot dead
soon after that post. Subsequent investigation revealed that the
rivalry leading to his death began and was carried out entirely
on social media. Other reporting has revealed how innocent
bystanders have also become targets in online fights, leaving
everyone in a neighborhood at risk5.
This paper investigates whether gang member profiles can
be identified automatically on Twitter, which can enable better
surveillance of gang members on social media. Classifying
Twitter profiles into particular types of users has been done
in other contexts [14], [15], [16], but gang member profiles
3http://www.hhs.gov/ash/oah/news/e-updates/eupdate-nov-2013.html
4http://www.wired.com/2013/09/gangs-of-social-media/
5https://goo.gl/75U3ME
Fig. 1: Twitter profile descriptions of known gang members.
Pursuant to an IRB governing human subject research, we are
prohibited from revealing personally identifiable information
in this paper. We only report Twitter handles that have already
been revealed in widely reported publications and were not
collected by the research team for this work.
pose unique challenges. For example, many Twitter profile
classifiers search for contextual clues in tweets and profile
descriptions [17], but gang member profiles use a rapidly
changing lexicon of keywords and phrases that often have only
a local, geographic context. This is illustrated in Figure 1,
which shows the Twitter profile descriptions of two verified
deceased gang members. The profile of @OsoArrogantJoJo
provides evidence that he belongs to a rival gang of the Black
Disciples by #BDK, a hashtag that is only known to those
involved with gang culture in Chicago. @PappyNotPapi's
profile mentions #PBG and our investigations revealed that this
hashtag is newly founded and stands for the Pooh Bear Gang,
a gang that was formerly known as the Insane Cutthroat Gang-
sters. Given the very local, rapidly changing lexicon of gang
members on social media, building a database of keywords,
phrases, and other identifiers to find gang members nationally
is not feasible. Instead, this study proposes heterogeneous sets
of features derived not only from profile and tweet text but also
from the emoji usage, profile images, and links to YouTube
videos reflecting their music culture. A large set of gang
member profiles, obtained through a careful data collection
process, is compared against non-gang member profiles to find
contrasting features. Experimental results show that using these
sets of features, we can build a classifier that has a low false
positive rate and a promising F 1-score of 0.7755.
This paper is organized as follows. Section II discusses
the related literature and positions how this work differs from
other related works. Section III discusses the data collection,
manual feature selection and our approach to identify gang
member profiles. Section IV gives a detailed explanation for
evaluation of the proposed method and the results in detail.
Section V concludes the work reported while discussing the
future work planned.
II. RELATED WORK
Gang violence is a well studied social science topic dating
back to 1927 [18]. However,
the notions of "Cyber-" or
"Internet banging", which is defined as "the phenomenon of
gang affiliates using social media sites to trade insults or make
violent threats that lead to homicide or victimization" [8], was
only recently introduced [19], [11]. Patton et al. introduced the
concept of "Internet banging" and studied how social media is
now being used as a tool for gang self-promotion and as a way
Fig. 2: Gang member dataset creation.
for gang members to gain and maintain street credibility [8].
They also discussed the relationship between gang-related
crime and hip-hop culture, giving examples on how hip-hop
music shared on social media websites targeted at harassing
rival gang members often ended up in real-world collisions
among those gangs. Decker et al. and Patton et al. have also
reported that street gangs perform Internet banging with social
media posts of videos depicting their illegal behaviors, threats
to rival gangs, and firearms [20], [21].
The ability to take action on these discoveries is limited by
the tools available to discover gang members on social media
and to analyze the content they post [19]. Recent attempts
to improve our abilities include a proposed architecture for
a surveillance system that can learn the structure, function,
and operation of gangs through what
they post on social
media [11]. However, the architecture requires a set of gang
member profiles for input, thus assuming that they have already
been discovered. Patton et al. [21] devised a method to
automatically collect tweets from a group of gang members
operating in Detroit, MI. However, their approach required the
profile names of the gang members to be known beforehand,
and data collection was localized to a single city in the country.
This work builds upon existing methods to automatically
discover gang member profiles on Twitter. This type of user
profile classification problem has been explored in a diverse
set of applications such as political affiliation [14], ethnic-
ity [14], gender [16], predicting brand loyalty [14], and user
occupations [17]. However, these approaches may utilize an
abundance of positive examples in their training data, and only
rely on a single feature type (typically, tweet text). Whereas
most profile classifiers focus on a single type of feature (e.g.
profile text), we consider the use of a variety of feature types,
including emoji, YouTube links, and photo features.
III. DISCOVERING GANG MEMBER PROFILES
This section discusses the methodology we followed to
study and classify the Twitter profiles of gang members auto-
matically. It includes a semi-automatic data collection process
to discover a large set of verifiable gang member profiles, an
evaluation of the tweets of gang and non-gang member posts
to identify promising features, and the deployment of multiple
supervised learning algorithms to perform the classification.
A. Data collection
Discovering gang member profiles on Twitter to build train-
ing and testing datasets is a challenging task. Past strategies
1. Seed Term Discovery2. Gang Affiliated Rappers' Twitter Profile DiscoverySearch Twitter API3. Manual Verification Of Twitter Profiles5. Using Followers and Followees to discover more profiles4. Using Retweets to discover more profilesGang MemberDatasetYESDiscardNO Method
Seed term discovery
Gang Affiliated Rappers
Retweets, Followers & Followees
Total
Number of Profiles
280
22
98
400
TABLE I: Number of gang member profiles captured.
to find these profiles were to search for keywords, phrases,
and events that are known to be related to gang activity in a
particular city a priori [11], [21]. However, such approaches are
unlikely to yield adequate data to train an automatic classifier
since gang members from different geographic locations and
cultures use local languages, location-specific hashtags, and
share information related to activities in a local region [11].
Such region-specific tweets and profiles may be used to train
a classifier to find gang members within a small region but
not across the Twitterverse. To overcome these limitations, we
adopted a semi-automatic workflow, illustrated in Figure 2, to
build a dataset of gang member profiles suitable for training a
classifier. The steps of the workflow are:
1. Seed Term Discovery: Following the success of identifying
gang member profiles from Chicago [11], we began our data
collection with discovering universal
terms used by gang
members. We first searched for profiles with hashtags for
Chicago gangs noted in [11], namely #BDK (Black Disciple
Killers) and #GDK (Gangster Disciples Killers). Those pro-
files were analyzed and manually verified as explained in
Step 3. Analysis of these profiles identified a small set of
hashtags they all use in their profile descriptions. Searching
Twitter profiles using those hashtags, we observed that gang
members across the U.S. use them, thus we consider those
terms to be location neutral. For example, gang members
post #FreeDaGuys in their profile to support their fellow
members who are in jail, #RIPDaGuys to convey the grieving
for fallen gang members, and #FuckDaOpps to show their
hatred towards police officers. We used these terms as key-
words to discover Twitter profiles irrespective of geographical
location. We used the Followerwonk Web service API6 and
Twitter REST API7 to search Twitter profile descriptions by
keywords #FreeDaGuys, #FreeMyNigga, #RIPDaGuys,
and #FuckDaOpps. Since there are different informal ways
people spell a word in social media, we also considered
variations on the spelling of each keyword; for example,
for #FreeDaGuys, we searched both #FreeDaGuys, and
#FreeTheGuys.
2. Gang Affiliated Rappers' Twitter Profile Discovery:
Finding profiles by a small set of keywords is unlikely to
yield sufficient data. Thus, we sought additional gang member
profiles with an observation from Patton et al. [8] that the
influence of hip-hop music and culture on offline gang member
activities can also be seen in their social media posts. We
thus also consider the influence of hip-hop culture on Twitter
by exploring the Twitter network of known gangster rappers
who were murdered in 2015 due to gang-related incidents8.
We searched for these rapper profiles on Twitter and manually
checked that the rapper was affiliated to a gang.
3. Manual verification of Twitter profiles: We verified each
6https://moz.com/followerwonk/bio
7https://dev.twitter.com/rest/public
8http://www.hipwiki.com/List+of+Rappers+Murdered+in+2015
profile discovered manually by examining the profile picture,
profile background image, recent tweets, and recent pictures
posted by a user. During these checks, we searched for terms,
activities, and symbols that we believed could be associated
with a gang. For example, profiles whose image or background
included guns in a threatening way, stacks of money, showing
gang hand signs and gestures, and humans holding or posing
with a gun, appeared likely to be from a gang member.
Such images were often identified in profiles of users who
submitted tweets that contain messages of support or sadness
for prisoners or recently fallen gang members, or used a high
volume of threatening and intimidating slang language. Only
profiles where the images, words, and tweets all suggested
gang affiliation were labeled as gang affiliates and added to
our dataset. Although this manual verification does have a
degree of subjectivity, in practice, the images and words used
by gang members on social media are so pronounced that we
believe any reasonable analyst would agree that they are gang
members. We found that not all the profiles collected belonged
to gang members; we observed relatives and followers of gang
members posting the same hashtags as in Step 1 to convey
similar feelings in their profile descriptions.
4. Using Retweets to discover more profiles: From the set
of verified profiles, we explored their retweet and follower
networks as a way to expand the dataset. We first considered
authors of tweets which were retweeted by a gang member
in our seed set. In Twitter, "retweeting" is a mechanism by
which a user can share someone else's tweet to their follower
audience. Assuming that a user only retweets things that
they believe or their audience would be interested in, it may
be reasonable to assume that gang members would only be
interested in sharing what other gang members have to say,
and hence, the authors of gang members' retweets could also
be gang members.
5. Using Followers and Followees to discover more profiles:
We analyzed followers and followees of our seed gang member
profiles to find more gang member profiles. A Twitter user
can follow other Twitter users so that the individual will be
subscribed to their tweets as a follower and they will be able
to start a private conversation by sending direct messages
to the individual. Motivated by the sociological concept of
homophily, which claims that individuals have a tendency to
associate and bond with similar others9, we hypothesized that
the followers and followees of Twitter profiles from the seed
set may also be gang members. Manual verification of Twitter
profiles collected from retweets, followers, and followees of
gang members showed that a majority of those profiles are non-
gang members who are either family members, hip-hop artists,
women or profiles with pornographic content. To ensure that
our dataset is not biased towards a specific gang or geographic
location, only a limited number of profiles were collected via
retweets, followers and followees.
Table I summarizes the number of profiles manually verified
as gang members from Twitter profiles collected in step 1, 2,
4 and 5. Altogether we collected 400 gang member's Twitter
profiles. This is a large number compared to previous studies
of gang member activities on social media that curated a
maximum of 91 profiles [11]. Moreover, we believe the profiles
9http://aris.ss.uci.edu/∼lin/52.pdf
(a) Gang members.
(b) Non-gang members.
Fig. 3: Comparison of words used in tweets.
collected represent a diverse set of gang members that are not
biased toward a particular geographic area or lingo as our data
collection process used location-independent terms proven to
be used by gang members when they express themselves.
B. Data analysis
We next explore differences between gang and non-gang
member Twitter usage to find promising features for classify-
ing profiles. For this purpose, profiles of non-gang members
were collected from the Twitter Streaming API10. We collected
a random sample of tweets and the profiles of the users
who authored the tweets in the random sample. We manually
verified that all Twitter profiles collected in this approach
belong to non-gang members. The profiles selected were then
filtered by location to remove non-U.S. profiles by reverse
geo-coding the location stated in their profile description by
the Google Maps API11. Profiles with location descriptions
that were unspecified or did not relate to a location in the
U.S. were discarded. We collected 2,000 non-gang member
profiles in this manner. In addition, we added 865 manually
verified non-gang member profiles collected using the location
neutral keywords discussed in Section III. Introducing these
profiles, which have some characteristics of gang members
(such as cursing frequently or cursing at law enforcement) but
are not, captures local languages used by family/friends of
gang members and ordinary people in a neighborhood where
gangs operate.
With the Twitter REST API12, we collected the maximum
number of most recent tweets that can be retrieved (3,200)
along with profile descriptions and images (profile and cover
photos) of every gang and non-gang member profile. The
resulting dataset consists of 400 gang member Twitter profiles
and 2,865 non-gang member Twitter profiles. The dataset
has a total of 821,412 tweets from gang member profiles
and 7,238,758 tweets from non-gang member profiles. Prior
to analyzing any text content, we removed all of the seed
words used to find gang member profiles, all stop words, and
performed stemming across all tweets and profile descriptions.
10https://dev.twitter.com/streaming/overview
11https://developers.google.com/maps/
12https://dev.twitter.com/rest/public
Fig. 4: Word usage in profile descriptions: gang vs non-gang.
1) Tweet text: Figure 3 summarizes the words seen most
often in the gang and non-gang members' tweets as clouds.
They show a clear difference in language. For example, we
note that gang members more frequently use curse words in
comparison to ordinary users. Although cursing is frequent in
tweets, they represent just 1.15% of all words used [22]. In
contrast, we found 5.72% of all words posted by gang member
accounts to be classified as a curse word, which is nearly five
times more than the average curse word usage on Twitter. The
clouds also reflect the fact that gang members often talk about
drugs and money with terms such as smoke, high, hit, and
money, while ordinary users hardly speak about finances and
drugs. We also noticed that gang members talk about material
things with terms such as got, money, make, real, need whereas
ordinary users tend to vocalize their feelings with terms such
as new, like, love, know, want, look, make, us. These differences
make it clear that the individual words used by gang and
non-gang members will be relevant features for gang profile
classification.
2) Twitter Profile Description: On Twitter, a user can give
a self-description as a part of the user's profile. A comparison
of the top 10 words in gang members' and non-gang mem-
bers' Twitter profile descriptions is shown in Figure 4. The
first 10 words are the most frequently used words in non-
gang members' profiles and the latter 10 words are the most
frequently used words in gang members' profiles. Word com-
parison shows that gang members prefer to use curse words
(nigga, fuck, shit) in their profile descriptions while non-gang
members use words related to their feelings or interests (love,
life, live, music, book). The terms rip and free which appear
in approximately 12% of all gang member Twitter profiles,
suggest that gang members use their profile descriptions as a
space to grieve for their fallen or incarcerated gang members.
The term gang in gang members' profile descriptions suggest
that gang members like to self-identify themselves on Twitter.
Such lexical features may therefore be of great importance
for automatically identifying gang member profiles. We take
counts of unigrams from gang and non-gang members' Twitter
profile descriptions as classification features.
Fig. 6: Examples for gang members' tweets with emojis.
emoji in an 'emoji chain'. Figure 6 presents some prototypical
'chaining' of emojis used by gang members. The chains may
reflect their anger at law enforcement officers, as a cop emoji
is often followed up with the emoji of a weapon, bomb, or
explosion. We found that 32.25% of gang members in our
dataset have chained together the police and the pistol emoji,
compared to just 1.14% of non-gang members. Moreover, only
1.71% of non-gang members have used the hundred points
emoji and pistol emoji together in tweets while 53% of gang
members have used them. A variety of the angry face emoji
such as devil face emoji and imp emoji were also common
in gang member tweets. The frequency of each emoji symbol
used across the set of user's tweets are thus considered as
features for our classifier.
5) Profile image: In our profile verification process, we
observed that most gang member profiles portray a context
representative of gang culture. Some examples of these profile
pictures are shown in Figure 7, where the user holds or points
weapons, is seen in a group fashion which displays a gangster
culture, or is showing off graffiti, hand signs, tattoos and
bulk cash. Descriptions of these images may thus empower
our classifier. Thus, we translated profile images into features
with the Clarifai web service14. Clarifai offers a free API to
query a deep learning system that tags images with a set of
scored keywords that reflect what is seen in the image. We
tagged the profile image and cover image for each profile
using 20 tags identified by Clarifai. Figure 8 offers the 20
most often used tags applied to gang and non-gang member
profiles. Since we take all the tags returned for an image, we
see common words such as people and adult coming up in the
top 20 tag set. However, gang member profile images were
assigned unique tags such as trigger, bullet, worship while
non-gang images were uniquely tagged with beach, seashore,
dawn, wildlife, sand, pet. The set of tags returned by Clarifai
were thus considered as features for the classifier.
C. Learning algorithms
The unigrams of tweets, profile text, and linked YouTube
video descriptions and comments, along with the distribution
of emoji symbols and the profile image tags were used to
train four different classification models: a Naive Bayes net, a
Logistic Regression, a Random Forest, and a Support Vector
Machine (SVM). These four models were chosen because
they are known to perform well over text features, which is
the dominant type of feature considered. The performance of
the models are empirically compared to determine the most
suitable classification technique for this problem. Data for
the models are represented as a vector of term frequencies
where the terms were collected from one or more feature sets
described above.
Fig. 5: Emoji usage distribution: gang vs non-gang.
3) Music interests: It has been recognized that music is
a key cultural component in an urban lifestyle and that gang
members often want to emulate the scenarios and activities the
music conveys [8]. Our analysis confirms that the influence
of gangster rap is expressed in gang members' Twitter posts.
We found that 51.25% of the gang members collected have a
tweet that links to a YouTube video. Following these links, a
simple keyword search for the terms gangsta and hip-hop
in the YouTube video description found that 76.58% of the
shared links are related to hip-hop music, gangster rap, and
the culture that surrounds this music genre. Moreover, this
high proportion is not driven by a small number of profiles
that prolifically share YouTube links; eight YouTube links are
shared on average by a gang member.
Recognizing the frequency with which gang members post
YouTube links on gangster rap and hip-hop, we consider the
YouTube videos posted in a user's tweets as features for the
classifier. In particular, for each YouTube video tweeted, we
used the YouTube API13 to retrieve the video's description
and its comments. Further analysis of YouTube data showed
a difference between terms in gang members' YouTube data
and non-gang members' YouTube data. For example, the top
5 terms (after stemming and stop word removal) used in
YouTube videos shared by gang members are shit, like, nigga,
fuck, lil while like, love, peopl, song, get are the top 5 terms in
non-gang member video data. To represent a user profile based
on their music interests, we generated a bag of words from the
video descriptions and comments from all shared videos.
4) Emoji: Motivated by recent work involving the use of
emojis by gang members [23], we also studied if and how gang
and non-gang members use emoji symbols in their tweets. Our
analysis found that gang members have a penchant for using
just a small set of emoji symbols that convey their anger and
violent behavior through their tweets. Figure 5 illustrates the
emoji distribution for the top 20 most frequent emojis used by
gang member profiles in our dataset. The fuel pump emoji was
the most frequently used emoji by the gang members, which
is often used in the context of selling or consuming marijuana.
The pistol emoji is the second most frequent in our dataset,
which is often used with the guardsman emoji or the police cop
13https://developers.google.com/youtube/
14http://www.clarifai.com/
DONT EVEN ASK EM WHO DEY WIT JUS BLOW EM FACES I LOST MY BRO 2 DESE STREETS NOW IM FUCKED UP #SHITREPEATF**K YOUR BLING BLING DEY GOT MY BROTHERS IN CHAINS #FREEXXXX #FREEXXXXX #FTP Fig. 7: Sample gang member profile images.
IV. EVALUATION
We next evaluate the performance of classifiers that use
the above features to discover gang member profiles on
Twitter. For this purpose, we use the training set discussed
in Section III with 400 gang member profiles (the 'posi-
tive'/'gang' class) and 2,865 non-gang member profiles (the
'negative'/'non-gang' class). We trained and evaluated the
performance of the classifiers mentioned in Section III-C
under a 10-fold cross validation scheme. For each of the
four learning algorithms, we consider variations involving only
tweet text, emoji, profile, image, or music interest (YouTube
comments and video description) features, and a final variant
that considers all types of features together. The classifiers that
use a single feature type were intended to help us study the
quality of their predictive power by itself. When building these
single-feature classifiers, we filtered the training dataset based
on the availability of the single feature type in the training
data. For example, we only used the twitter profiles that had
at least a single emoji in their tweets to train classifiers that
consider emoji features. We found 3,085 such profiles out of
the 3,265 profiles in the training set. When all feature types
were considered, we developed two different models:
1) Model(1): This model is trained with all profiles in
the training set.
2) Model(2): This model is trained with profiles that
contain every feature type.
Because a Twitter profile may not have every feature type,
Model(1) represents a practical scenario where not every
Twitter profile contains every type of feature. In this model,
the non-occurrence of a feature is represented by 'zeroing out'
the feature value during model training. Model(2) represents
the ideal scenario where all profiles contain every feature type.
For this model, we used 1,358 training instances (42% of all
training instances), out of which 172 were gang members (43%
of all gang members) and 1,186 were non-gang members (41%
of all non-gang members). We used version 0.17.1 of scikit-
learn15 machine learning library to implement the classifiers.
For each 10-fold cross validation experiment, we report
three evaluation metrics for the 'gang' and 'non-gang' classes,
namely, the Precision = tp/(tp + f p), Recall = tp/(tp + f n),
15http://scikit-learn.org/stable/index.html
Fig. 8: Image tags distribution: gang vs non-gang.
and F 1-score = 2 ∗ (P recision ∗ Recall)/(P recision +
Recall), where tp is the number of true positives, f p is the
number of false positives, tn is the number of true negatives,
and f n is the number of false negatives. We report these
metrics for the positive 'gang' and negative 'non-gang' classes
separately because of class imbalance in our dataset.
A. Experimental results
Table II presents the average precision, recall, and F 1-score
over the 10 folds for the single-feature and combined feature
classifiers. The table includes, in braces ('{ }'), the number
of gang and non-gang profiles that contain a particular feature
type, and hence the number of profiles used for the 10-fold
cross validation. It is reasonable to expect that any Twitter
profile is not that of a gang member, predicting a Twitter user
as a non-gang member is much easier than predicting a Twitter
user as a gang member. Moreover false positive classifications
of the 'gang' class may be detrimental to law enforcement
investigations, which may go awry as they surveil an innocent
person based on the classifier's suggestion. We thus believe
that a small false positive rate of the 'gang' class to be an
especially important evaluation metric. We say that a classifier
is 'ideal' if it demonstrates high precision, recall, and F 1-score
for the 'gang' class while performing well on the 'non-gang'
class as well.
The best performing classifier that considers single features
is a Random Forest model over tweet features (T), with a rea-
sonable F 1-score of 0.7229 for the 'gang' class. It also features
the highest F 1-score for the 'non-gang' class (0.9671). Its
strong performance is intuitive given the striking differences in
language as shown in Figure 3 and discussed in Section III-B1.
We also noted that music features offer promising results, with
an F 1-score of 0.6505 with a Naive Bayes classifier, as well
as emoji features with an F 1-score of 0.6067 also achieved by
a Naive Bayes classifier. However, the use of profile data and
image tags by themselves yield relatively poor F 1-scores no
matter which classifier considered. There may be two reasons
for this despite the differences we observed in Section III-B.
Features
Total Number of Profiles
{#Gang : #Non-Gang}
Classifier
Tweets (T)
Emojis (E)
3,265 {400 : 2,865}
3,085 {396 : 2,689}
Profile data (P)
2,996 {378 : 2,618}
Image tags (I)
2,910 {357 : 2,553}
Music interest (Y)
1,630 {196 : 1,434}
Model(1) {T + E + P + I + Y}
3,265 {400 : 2,865}
Model(2) {T + E + P + I + Y}
1,358 {172 : 1,186}
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Naive Bayes
Logistic Regression
Random Forest
SVM
Precision
0.4354
0.6760
0.8433
0.6301
0.4934
0.6867
0.7279
0.4527
0.6000
0.8015
0.5719
0.7501
0.2692
0.4832
0.4131
0.3889
0.5865
0.7101
0.8403
0.6232
0.3718
0.7250
0.8792
0.6442
0.4405
0.7588
0.8961
0.7185
Gang
Recall
0.9558
0.6623
0.6401
0.6545
0.7989
0.3995
0.5079
0.5642
0.243
0.2160
0.1441
0.2225
0.6973
0.1853
0.1512
0.1454
0.7424
0.5447
0.3953
0.6067
0.9387
0.6880
0.6374
0.6791
0.9386
0.7396
0.6994
0.7394
Results
F 1-score
0.5970
0.6666
0.7229
0.6388
0.6067
0.4969
0.5931
0.4955
0.464
0.3362
0.2239
0.3394
0.3851
0.2624
0.2147
0.205
0.6505
0.6110
0.5277
0.6072
0.5312
0.7038
0.7364
0.6583
0.5926
0.7433
0.7755
0.7213
Precision
0.9929
0.9529
0.9517
0.9514
0.9676
0.9164
0.9292
0.9329
0.8765
0.8974
0.8886
0.8978
0.9458
0.8950
0.8911
0.8898
0.9632
0.9395
0.9232
0.9463
0.9889
0.9564
0.9507
0.9546
0.9889
0.9639
0.9575
0.9638
Non-Gang
Recall
0.8278
0.9544
0.9832
0.9442
0.8785
0.9733
0.9721
0.8953
1.0000
0.9924
0.9859
0.9897
0.7357
0.9722
0.9731
0.9679
0.9297
0.9679
0.9895
0.9476
0.7791
0.9637
0.9881
0.9469
0.8254
0.9662
0.9873
0.9586
F 1-score
0.9028
0.9536
0.9671
0.9477
0.9207
0.9438
0.9500
0.9133
0.9341
0.9424
0.9346
0.9414
0.8271
0.9318
0.9300
0.9270
0.9460
0.9534
0.9550
0.9467
0.8715
0.9599
0.9690
0.9506
0.8991
0.9649
0.9720
0.9610
TABLE II: Classification results based on 10-fold cross validation.
First, these two feature types did not generate a large number
of specific features for learning. For example, descriptions are
limited to just 160 characters per profile, leading to a limited
number of unigrams (in our dataset, 10 on average) that can be
used to train the classifiers. Second, the profile images were
tagged by a third party Web service which is not specifically
designed to identify gang hand signs, drugs and guns, which
are often shared by gang members. This led to a small set of
image tags in their profiles that were fairly generic, i.e., the
image tags in Figure 8 such as 'people', 'man', and 'adult'.
Combining these diverse sets of features into a single
classifier yields even better results. Our results for Model(1)
show that the Random Forest achieves the highest F 1-scores
for both 'gang' (0.7364) and 'non-gang' (0.9690) classes and
yields the best precision of 0.8792, which corresponds to a low
false positive rate when labeling a profile as a gang member.
Despite the fact that it has lower positive recall compared
to the second best performing classifier (a Random Forest
trained over only tweet text features (T)), for this problem
setting, we should be willing to increase the chance that a gang
member will go unclassified if it means reducing the chance
of applying a 'gang' label to a non-gang member. When we
tested Model(2), a Random Forrest classifier achieved an F 1-
score of 0.7755 (improvement of 7.28% with respect to the
best performing single feature type classifier (T)) for 'gang'
class with a precision of 0.8961 (improvement of 6.26% with
respect to (T)) and a recall of 0.6994 (improvement of 9.26%
with respect to (T)). Model(2) thus outperforms Model(1), and
we expect its performance to improve with the availability of
more training data with all feature types.
B. Evaluation Over Unseen Profiles
We also tested the trained classifiers using a set of Twitter
profiles from a separate data collection process that may
emulate the classifier's operation in a real-time setting. For this
experiment, we captured real-time tweets from Los Angeles,
CA16 and from ten South Side, Chicago neighborhoods that
are known for gang-related activities [11] using the Twitter
streaming API. We consider these areas with known gang
presence on social media to ensure that some positive profiles
would appear in our test set. We ultimately collected 24,162
Twitter profiles: 15,662 from Los Angeles, and 8,500 from
Chicago. We populated data for each profile by using the 3,200
most recent tweets (the maximum that can be collected from
Twitter's API) for each profile. Since the 24,162 profiles are
far too many to label manually, we qualitatively study those
profiles the classifier placed into the 'gang' class.
We used the training dataset to train our best performing
random forest classifier (which use all feature types) and tested
it on the test dataset. We then analyzed the Twitter profiles
that our classifier labeled as belonging to the 'gang' class.
Each of those profiles had several features which overlap with
gang members such as displaying hand signs and weapons in
their profile images or in videos posted by them, gang names
or gang-related hashtags in their profile descriptions, frequent
use of curse words, and the use of terms such as "my homie"
to refer to self-identified gang members. Representative tweets
extracted from those profiles are depicted in Figure 9. The most
frequent words found in tweets from those profiles were shit,
nigga, got, bitch, go, fuck etc. and their user profiles had terms
such as free, artist, shit, fuck, freedagang, and ripthefallen.
They had frequently used emojis such as face with tears of joy,
hundred points symbol, fire, skull, money bag, and pistol. For
some profiles, it was less obvious that the classifier correctly
identified a gang member. Such profiles used the same emojis
and curse words commonly found in gang members profiles,
but their profile picture and tweet content was not indicative of
a gang affiliation. In conclusion, we find that in a real-time-
like setting, the classifier to be able to extract profiles with
features that strongly suggest gang affiliation. Of course, these
16http://isithackday.com/geoplanet-explorer/index.php?woeid=2442047
Fig. 9: Sample tweets from identified gang members.
profiles demand further investigation and extensive evidence
from other sources in order to draw a concrete conclusion,
especially in the context of a law enforcement investigation.
We refrain from reporting any profile names or specific details
about the profiles labeled as a 'gang' member to comply with
the applicable IRB governing this human subject research.
V. CONCLUSION AND FUTURE WORK
This paper presented an approach to address the problem
of automatically identifying gang member profiles on Twitter.
Despite the challenges in developing such automated systems,
mainly due to difficulties in finding online gang member
profiles for developing training datasets, we proposed an
approach that uses features extracted from textual descriptions,
emojis, images and videos shared on Twitter (textual features
extracted from images, and videos). Exploratory analysis of
these types of features revealed interesting, and sometimes
striking differences in the ways gang and non-gang members
use Twitter. Classifiers trained over features that highlight these
differences, were evaluated under 10-fold cross validation. Our
best classifier achieved a promising F 1-score of 0.7755 over
the 'gang' profiles when all types of features were considered.
Future work will strengthen our training dataset by in-
cluding more gang member Twitter profiles by searching for
more location-independent keywords. We also plan to develop
our own image classification system specifically designed to
classify images found on gang member profiles. We would
also like to experiment with building dictionaries that contain
gang names to understand whether "having a gang name in
the profile description" as a feature can improve our results.
Finally, we would also like to study how can we further
improve our classifier models using word embeddings [24] and
social networks of known gang members.
VI. ACKNOWLEDGEMENT
We are thankful to Uday Kiran Yeda for helping us with
data collection. We acknowledge partial support from the
National Science Foundation (NSF) award: CNS-1513721:
"Context-Aware Harassment Detection on Social Media", Na-
tional Institutes of Health (NIH) award: MH105384-01A1:
"Modeling Social Behavior for Healthcare Utilization in De-
pression" and Grant No. 2014-PS-PSN-00006 awarded by the
Bureau of Justice Assistance. The Bureau of Justice Assistance
is a component of the U.S. Department of Justice's Office
of Justice Programs, which also includes the Bureau of Jus-
tice Statistics, the National Institute of Justice, the Office of
Juvenile Justice and Delinquency Prevention, the Office for
Victims of Crime, and the SMART Office. Points of view or
opinions in this document are those of the authors and do not
necessarily represent the official position or policies of the U.S.
Department of Justice, NSF or NIH.
REFERENCES
[1] 2011 National Gang Threat Assessment Issued Emerging Trends, 2011.
[2] National Gang Report, National Gang Intelligence Center, 2013.
[3] W. B. Miller, Crime by youth gangs and groups in the United States.
US Department of Justice, Office of Justice Programs, Office of Juvenile
Justice and Delinquency Prevention Washington, DC, 1992.
"Gang homicides-five us cities, 2003-2008." Morbidity and mortality
weekly report, vol. 61, no. 3, pp. 46–51, 2012.
[4]
[5] Survey of Gang Members Online Habits and Participation (2007)
Survey results reported at the i-SAFE Annual Internet Safety Education
Review Meeting Carlsbad, California. National Assessment Center.
J. E. King, C. E. Walpole, and K. Lamon, "Surf and turf wars online-
growing implications of internet gang violence," Journal of Adolescent
Health, vol. 41, no. 6, Supplement, pp. S66 – S68, 2007.
[6]
[7] D. C. Pyrooz, S. H. Decker, and R. K. M. Jr., "Criminal and routine
activities in online settings: Gangs, offenders, and the internet," Justice
Quarterly, vol. 32, no. 3, pp. 471–499, 2015.
[8] D. U. Patton, R. D. Eschmann, and D. A. Butler, "Internet banging:
New trends in social media, gang violence, masculinity and hip hop,"
Computers in Human Behavior, vol. 29, no. 5, pp. A54 – A59, 2013.
J. C. Howell, "Gang prevention: An overview of research and programs.
juvenile justice bulletin." Office of Juvenile Justice and Delinquency
Prevention, 2010.
[9]
[10] M. Ito, S. Baumer, M. Bittanti, R. Cody, B. Herr-Stephenson, H. A.
Horst, P. G. Lange, D. Mahendran, K. Z. Mart´ınez, C. Pascoe et al.,
"Hanging out, messing around, and geeking out," Digital media, 2010.
[11] S. Wijeratne, D. Doran, A. Sheth, and J. L. Dustin, "Analyzing the social
media footprint of street gangs," in IEEE International Conference on
Intelligence and Security Informatics (ISI), 2015, May 2015, pp. 91–96.
[12] P. E. R. Forum, "Social media and tactical considerations for law
enforcement," United States Office of Community Oriented Policing
Services and United States Department of Justice, Tech. Rep., 2013.
J. Brunty, L. Miller, and K. Helenek, Social media investigation for law
enforcement. Routledge, 2014.
[13]
[14] M. Pennacchiotti and A.-M. Popescu, "A machine learning approach to
twitter user classification," 2011.
[15] R. Tinati, L. Carr, W. Hall, and J. Bentwood, "Identifying communicator
roles in twitter," in Proceedings of the 21st International Conference on
World Wide Web, ser. WWW '12 Companion. New York, NY, USA:
ACM, 2012, pp. 1161–1168.
[16] W. Liu and D. Ruths, "Whats in a name? using first names as features
for gender inference in twitter," 2013.
[17] H. Purohit, A. Dow, O. Alonso, L. Duan, and K. Haas, "User taglines:
Alternative presentations of expertise and interest in social media," in
2012 International Conference on Social Informatics (Social Informat-
ics), Washington, D.C., USA, December 14-16, 2012, pp. 236–243.
[18] F. M. Thrasher, The gang: A study of 1,313 gangs in Chicago.
University of Chicago Press, 1963.
[19] D. U. Patton, J. S. Hong, M. Ranney, S. Patel, C. Kelley, R. Eschmann,
and T. Washington, "Social media as a vector for youth violence: A
review of the literature," Computers in Human Behavior, 2014.
[20] S. Decker and D. Pyrooz, "Leaving the gang: Logging off and moving
on. council on foreign relations," 2011.
[21] D. U. Patton, "Gang violence, crime, and substance use on twitter:
A snapshot of gang communications in detroit," Society for Social
Work and Research 19th Annual Conference: The Social and Behavioral
Importance of Increased Longevity, jan 2015.
[22] W. Wang, L. Chen, K. Thirunarayan, and A. P. Sheth, "Cursing
in english on twitter," in Proceedings of the 17th ACM Conference
on Computer Supported Cooperative Work & Social Computing, ser.
CSCW '14. New York, NY, USA: ACM, 2014, pp. 415–425.
[23] D. U. Patton, J. Lane, P. Leonard, J. Macbeth, and J. R. Smith-Lee,
"Gang violence on the digital street: Case study of a south side chicago
gang members twitter communication," New Media & Society, 2016.
[24] S. Wijeratne, L. Balasuriya, D. Doran, and A. Sheth, "Word embed-
dings to enhance twitter gang member profile identification," in IJCAI
Workshop on Semantic Machine Learning (SML 2016). New York
City, NY: CEUR-WS, 07/2016 2016.
WHOLE LOTTA GOIN ON CPDK DEM BITCHESBITCH WE TAKIN GLOKS WE AIN BUY'N NUN F**K FEDS TOOK ALL DA WISE GUYS OUT THE HOOD! |
1703.02852 | 1 | 1703 | 2017-02-20T20:19:32 | Introduction to a Temporal Graph Benchmark | [
"cs.SI",
"physics.soc-ph"
] | A temporal graph is a data structure, consisting of nodes and edges in which the edges are associated with time labels. To analyze the temporal graph, the first step is to find a proper graph dataset/benchmark. While many temporal graph datasets exist online, none could be found that used the interval labels in which each edge is associated with a starting and ending time. Therefore we create a temporal graph data based on Wikipedia reference graph for temporal analysis. This report aims to provide more details of this graph benchmark to those who are interested in using it. | cs.SI | cs | Introduction to a Temporal Graph Benchmark
Wouter Ligtenberg and Yulong Pei
Department of Mathematics and Computer Science
TU Eindhoven, The Netherlands
[email protected]
Abstract. A temporal graph is a data structure, consisting of nodes and
edges in which the edges are associated with time labels. To analyze the
temporal graph, the first step is to find a proper graph dataset/benchmark.
While many temporal graph datasets exist online, none could be found
that used the interval labels in which each edge is associated with a
starting and ending time. Therefore we create a temporal graph data
based on Wikipedia reference graph for temporal analysis. This report
aims to provide more details of this graph benchmark to those who are
interested in using it.
Keywords: Temporal Graph, Graph Analysis, Graph Benchmark
1
Introduction
A temporal graph is a data structure, consisting of nodes and edges in which
the edges are associated with time labels. While many temporal graph datasets
exist online, none could be found that used the interval labels in which each
edge is associated with a starting and ending time. For this reason we generated
several synthetic datasets and modified two existing datasets, i.e., Wikipedia
reference graph [2] and the Facebook message graph, for analysis. In this report,
we introduce the creation of the Wikipedia reference graph and analyze some
properties of this graph from both static and temporal perspectives. This report
aims to provide more details of this graph benchmark to those who are interested
in using it and also serves as the introduction to the dataset used in the master
thesis of Wouter Ligtenberg [1].
2 Creation of the Temporal Graph Benchmark
This dataset was created from the Wikipedia refence set [2]. In this dataset
edges represent Wikipedia articles and every edge represents a reference from one
article to another. The edges have 2 values, one with the time stamp of the edge
creation or the edge removal, and one indicating if it was added or removed. We
transformed this dataset into an interval set in which we connected the creation
of an edge to the removal of an edge to create an edge with an interval. In this
dataset we only included edges that were first added and later removed. If an
7
1
0
2
b
e
F
0
2
]
I
S
.
s
c
[
1
v
2
5
8
2
0
.
3
0
7
1
:
v
i
X
r
a
2
Introduction to a Temporal Graph Benchmark
edge is added but never removed we ignore it. Since this is the first real interval
set we contacted both Konect1 and Icon2 to submit this dataset to their index
for further research. Both instances accepted the dataset to their database.
In details, this graph dataset contains the evolution of hyperlinks between
articles of the Dutch version of Wikipedia. Each node is an article. Each edge is
a hyperlink from one page to another (directed). In the dataset, each edge has a
start and an end time, which allows us to look at the network in different points
in time. The network under study has 67,8907 vertices and 472,9035 edges. The
first edge is created on Tuesday August 28, 2001, and the last edge removal takes
place on Sunday July 10, 2011. So, our data roughly has a time span of 10 years.
3 Properties of the Temporal Graph Benchmark
In this section, we introduce the properties of this temporal graph benchmark
in two different perspectives, i.e., static and temporal. For the static graph anal-
ysis, we ignore the time stamps and view this dataset as a static graph. Then
we calculate three graph measures on this graph including degree, clustering
coefficient and PageRank.
3.1 Static Graph Analysis
Degree The average degree, average in-degree and average out-degree charac-
terize the network in a way that it gives an indication of how many edges there
are connected on average to a node. We also calculate the max degree as the
comparison. The results are shown in Table 1.
The degree distribution using a logarithmic scale for both the degree fre-
quencies as the number of nodes including the degree distribution, in-degree
distribution and out-degree distribution, are shown in Figure 1. The curves in-
dicate that these degree distributions follow the power law distribution.
average
Degree
10.62
In-degree
Out-degree
5.31
5.31
max 12298 (n=394) 10521 (n=394) 4302 (n=306188)
Table 1. Degree statistics of the static graph.
Clustering coefficient How much nodes cluster together and form neighbor-
hoods can be measured using the clustering coefficient. This measure quantifies
the proportion of a node's neighbors to which the node is connected through
1 http://konect.uni-koblenz.de/
2 https://icon.colorado.edu
Introduction to a Temporal Graph Benchmark
3
Fig. 1. Degree distributions of the static graph.
triangles (three-node cycles). The clustering coefficient Ci of node i is defined
as:
Ci =
2ti
ki(ki − 1)
,
(1)
where ti is the number of triangles (three-node cycles) through i and and ki is
the degree of node i. Ci = 0 if none of the neighbors of a vertex are connected,
and Ci = 1 if all of the neighbors are connected. Remark that the graph is
considered as if it was undirected.
In this graph dataset, the average clustering coefficient is 0.114. For 28638
nodes Ci = 1, which means that for 4.22% of the nodes all neighbors are con-
nected. On the other hand, there are also 57.40% of the nodes have Ci = 0. This
means that these nodes are not part of any triangle at all.
PageRank This measure ranks nodes by their importance. The PageRank value
of a node is based on the nodes linking to and from that node. PageRank is
mainly used by Google to rank websites by calculating the importance each
website, but can also be used for other purposes such as rating Wikipedia articles.
To visualize the results a boxplot was plotted, which is shown in Figure 2. The
figure clearly shows that only a few pages are classified as important, i.e. that
have a high PageRank. While a lot of nodes/pages have relative small PageRank,
i.e., less importance.
3.2 Temporal Graph Analysis
We use the snapshot-based method to analyze the temporal graph. Therefore the
first step is choose the time window to divide the temporal graph into several
snapshots. The time between the creation of the first edge, and the removal
of the last edge covers a time period of 311,268,466 seconds, which is almost
10 years (9.87 years). For simplicity, we use the time window of 1 year in the
temporal analysis to divide the temporal graph into 10 snapshots.
4
Introduction to a Temporal Graph Benchmark
Fig. 2. PageRank of the static graph.
Degree In Figure 3 the degree distributions per snapshot are visualized. Each
year has its own color, ranging from blue (year 1) to red (year 10). We see a
similar pattern for degree, in-degree and out-degree. In the first years, the graph
shows a fairly large deviation from the power law. As the years pass, the graph
develops to fit the power line better. The years 5 to 8 show a similar graph as
we have seen in the static analysis of the graph. The last two years again start
to deviate from this diagonal, although still keeping a fairly diagonal line. This
behavior can be partly explained by the fact that the graph is growing at the
beginning, and getting smaller again at the end. In all snapshots it seems to
approximately hold that this network thus is a scale-free network. Moreover, the
change of degrees in different graph snapshots is shown in Figure 4.
Fig. 3. Degree distributions of the temporal graph in different graph snapshots.
Clustering coefficient Figure 5 shows the change of cluster coefficient of the
temporal graph in different snapshots and the clustering does change over time.
It is with relatively small steps, but we can conclude that the graph shows a
downward trend. This trend might indicate that more Wikipedia pages tend
Introduction to a Temporal Graph Benchmark
5
Fig. 4. Change of degrees in different graph snapshots.
to link to pages outside their general topic, which would lower the clustering
coefficient.
Fig. 5. Change of cluster coefficient in different graph snapshots.
PageRank We use the PageRank to analyze the importance of nodes change
over time. To visualize the results, we pick the top 30 nodes that scored the
highest on PageRank and the results are shown in Figure 6. From the results,
there is not a single most important node over the complete time period. Times
change, inventions are done and new articles are written about new subjects.
This is something one would expect in a data file like this, and thus is confirmed
by the data.
Acknowledgments
Section 3, i.e., the analysis of graph properties, is from the advanced homework
of 2IID0 - Web Analytics course. We thank Evertjan Peer, Hilde Weerts, Jasper
Adegeest and Gerson Foks (students of Group 12 in 2IID0) for their efforts in
the network analysis.
6
Introduction to a Temporal Graph Benchmark
Fig. 6. Top nodes with highest PageRank in different graph snapshots.
References
1. Wouter Ligtenberg. Tink, a temporal graph analytics library for Apache Flink. Mas-
ter Thesis, 2017.
2. Julia Preusse, J´erome Kunegis, Matthias Thimm, Steffen Staab, and Thomas Got-
tron. Structural dynamics of knowledge networks. In ICWSM, 2013.
|
1701.07490 | 1 | 1701 | 2017-01-17T18:52:22 | What Are People Tweeting about Zika? An Exploratory Study Concerning Symptoms, Treatment, Transmission, and Prevention | [
"cs.SI"
] | The purpose of this study was to do a dataset distribution analysis, a classification performance analysis, and a topical analysis concerning what people are tweeting about four disease characteristics: symptoms, transmission, prevention, and treatment. A combination of natural language processing and machine learning techniques were used to determine what people are tweeting about Zika. Specifically, a two-stage classifier system was built to find relevant tweets on Zika, and then categorize these into the four disease categories. Tweets in each disease category were then examined using latent dirichlet allocation (LDA) to determine the five main tweet topics for each disease characteristic. Results 1,234,605 tweets were collected. Tweets by males and females were similar (28% and 23% respectively). The classifier performed well on the training and test data for relevancy (F=0.87 and 0.99 respectively) and disease characteristics (F=0.79 and 0.90 respectively). Five topics for each category were found and discussed with a focus on the symptoms category. Through this process, we demonstrate how misinformation can be discovered so that public health officials can respond to the tweets with misinformation. | cs.SI | cs | Paper type: Original Paper
What Are People Tweeting about Zika? An Exploratory
Study Concerning Symptoms, Treatment, Transmission, and
Prevention
Corresponding Author
Michele Miller MS
Wright State University
Biological Sciences
3640 Colonel Glenn Hwy, Dayton, OH 45431
Phone (574)261-3969
Fax (937)775-3320
[email protected]
Dr. Tanvi Banerjee PhD
Wright State University
Computer Science and Engineering
3640 Colonel Glenn Hwy, Dayton, OH 45431
Roopteja Muppalla
Wright State University
Computer Science and Engineering
3640 Colonel Glenn Hwy, Dayton, OH 45431
Dr. William Romine PhD
Wright State University
Biological Sciences
3640 Colonel Glenn Hwy, Dayton, OH 45431
Dr. Amit Sheth PhD
Wright State University
Computer Science and Engineering
3640 Colonel Glenn Hwy, Dayton, OH 45431
Abstract.
Background So far an exploratory study concerning what people are tweeting about Zika has not
been performed.
Objective The purpose of this study was to do a dataset distribution analysis, a classification
performance analysis, and a topical analysis concerning what people are tweeting about four
disease characteristics: symptoms, transmission, prevention, and treatment.
Methods A combination of natural language processing and machine learning techniques were
used to determine what people are tweeting about Zika. Specifically, a two-stage classifier
system was built to find relevant tweets on Zika, and then categorize these into the four disease
categories. Tweets in each disease category were then examined using latent dirichlet allocation
(LDA) to determine the five main tweet topics for each disease characteristic.
Results 1,234,605 tweets were collected. Tweets by males and females were similar (28% and
23% respectively). The classifier performed well on the training and test data for relevancy
(F=0.87 and 0.99 respectively) and disease characteristics (F=0.79 and 0.90 respectively). Five
topics for each category were found and discussed with a focus on the symptoms category.
Conclusion Through this process, we demonstrate how misinformation can be discovered so that
public health officials can respond to the tweets with misinformation.
Keywords: Zika; Epidemiology; Social Media; Machine Learning; Topic Modeling; Big Data
Analysis
Introduction
the globe. It was shown that the spread of misinformation led to deaths due to improper practice
of appropriate preventative measures [1].
The 2014/2015 Ebola outbreak caused fear and misinformation to spread wildly across
Experts at the Center for Disease Control (CDC) and the World Health Organization
(WHO) admit that they botched the response for Ebola by not responding to the threat sooner
[2]. One year after the Ebola outbreak ended, the Zika outbreak started and also caused fear and
misinformation to spread. In the recent years, citizen sensing has picked up greatly with the rise
of mobile device popularity, as well as with the rise in social media sites such as Facebook and
Twitter. The idea with citizen sensing is that citizens play the role of sensors in the environment
[3]; providing information regarding healthcare issues such as disease outbreaks in case of Ebola
and Zika [4].
Big social data eliminate the time lag caused by traditional survey based methods,
allowing for studying public opinions on issues while addressing privacy concerns of users by
studying collective public behavior on specific issues. In particular, public opinion mining has
been studied in the past for exploration of public views on important social issues such as
gender-based violence [5], as well as to mine health related beliefs [6-7].
With respect to Zika, Twitter has served as a source of misinformation. To counter, the
CDC has been responding with correct information. For example, one user tweeted "Apparently
Florida is immune to the Zika virus" while the CDC has tweeted about Zika in Florida several
times including this tweet "Updated: CDC travel and testing recommendations for Miami-Dade
county b/c of continued local #Zika transmission". Another common misconception is found in
this re-tweet "RT @user: I saw the Zika virus has made its way to Houston. It's really only bad
for you if you are pregnant .. ". However, while Zika is typically mild to unnoticeable in adults,
it can cause health issues in some adults as the CDC explains in this tweet "Symptoms of
Guillain-Barre syndrome include weakness in arms and legs. GBS (Guillain-Barre syndrome) is
linked w/ #Zika. Learn more. bit.ly/2dqL7Dv". One more common misconception is that "It
won't do any good 2 attempt 2 prevent Zika Virus; instead, devote all funds 2 researching a
vaccine, or anecdote". Firstly, assuming that the tweet meant "antidote" instead of "anecdote,"
the statement is still incorrect since antidotes are a substance that stops the harmful effects of
poison. This tweet puts forth a fatalistic attitude towards Zika prevention in its indication that
there is nothing citizens can do to prevent Zika. The CDC responds to this by posting several
tweets about what the public can do to prevent Zika infection such as "Treating your clothing &
gear w/ permethrin can help prevent mosquito bites. Learn more ways to #ZapZika" and
In this exploratory study, a combination of natural language processing and machine
"Prevent #Zika spread after travel. Use condoms: 6 months after travel for men, 8 weeks for
women".
learning techniques were used to determine what information about Zika symptoms,
transmission, prevention, and treatment people were discussing using tweets. Specifically, a two-
stage classifier system was built that was used to find relevant tweets on Zika, and then
categorize those into four disease categories: symptoms, transmission, prevention, and treatment
(Figure 1). This information could then be used by health professionals such as people at the
CDC and WHO to know what information the public does and does not know about Zika. Such a
system may help inform them on what they need to include in messages to the public, as well as
target specific user groups to prevent the spread of misinformation.
Figure 1. Block diagram of the pragmatic function-oriented content retrieval using a hierarchical
supervised classification technique, followed by deeper analysis for characteristics of disease
content.
Zika
Zika belongs to the Flaviviridae family of viruses [8]. This family contains a few
arboviruses that are very important clinically, such as dengue virus, yellow fever virus, West
Nile Virus, tick-borne encephalitis virus, and Japanese encephalitis virus [8]. Prior to 2007 there
had only been 14 confirmed cases of Zika; however, many people do not even realize they are
sick from Zika, let alone need to go to the hospital and death due to Zika is extremely rare [9].
The Zika virus usually only causes mild symptoms such as a headache, rash, fever,
conjunctivitis, and joint pain which can last from a few days to a week after being infected.
These symptoms are similar to Dengue and Chikungunya, which are also spread by the same
mosquito as Zika. The incubation period is unknown at this time but suspected to be a few days
to a week. The virus typically remains in the blood for about a week, but has been found later in
some people. People at risk of getting Zika include people that travel or live in an area where
Zika has been found and has been bit by a mosquito, and partners of men that have traveled to an
area with the Zika virus. In the next several months the countries with active cases is expected to
increase [9].
Complications such as Guillain-Barré syndrome, microcephaly, and maculopathy may
occur in infants [8]. In Brazil, the number of infants with microcephaly increased 20 times after
the start of the Zika virus epidemic. Fetuses and infants have also been found to have cerebral
calcification and retinal abnormalities such as macular pigment mottling, macular neuroretinal
atrophy, optic nerve hypoplasia, chorioretinal atrophy, and foveal reflex loss. Since this is the
first outbreak of Zika associated with these defects, management is still an important challenge.
There are three main ways to get Zika: (i) being bitten by an infected Aedes mosquito, (ii)
through sexual contact, and (iii) from mother to fetus [9]. The Aedes mosquito aggressively bites
during the daytime, but can also bite at night. Mosquitos get the virus by feeding on someone
with the virus and then spread it by feeding on other people. An infected mother can pass the
virus to a newborn during pregnancy. It has been shown that the virus can somehow cross the
placenta. As of yet, there is no evidence of a mother passing the virus to her newborn through
breastfeeding. There have been cases of an infected man spreading the virus to his partner during
intercourse. It has been found that the virus persists in semen longer than blood and that the virus
can be spread before, during, and after, the man has symptoms. There have been no confirmed
cases of someone in the US getting Zika after receiving a blood transfusion as of February, 1,
2016. However, there have been reports of people in Brazil getting Zika from a blood transfusion
[9].
A person is most likely protected from future infections once already infected [9]. There
is currently no medicine or vaccine to treat the Zika virus. Experts suggest rest, plenty of fluids,
and acetaminophen or paracetamol for fever and pain. They say to avoid aspirin and other non-
steroidal anti-inflammatory drugs until it is confirmed to be Zika and not Dengue. Sexual
transmission can be prevented by abstaining or using condoms. Infection by mosquito bite can be
prevented by wearing long-sleeved shirts and long pants, staying in places with air conditioning,
staying in places that have door and window screens, sleeping under a mosquito net, and using
insect repellants. People returning from places with Zika should prevent being bitten by a
mosquito for three weeks to prevent the spread of the virus to uninfected mosquitoes [9].
Related Works
A study about misinformation about Ebola on Twitter found that a lot of misinformation
about cures for Ebola was circulating around Twitter [10]. They found that 44% (248) of the
tweets about Ebola were retweeted at least once. Of those 248 tweets, 38.3% were scientifically
accurate while 58.9% were inaccurate. They also found that most of the tweets containing
misinformation were never corrected.
Alvaro et al. [11] used a Twitter streaming application program interface (API) to obtain
a random sample of tweets collected over a 12-month period about first-hand experience with
cognitive enhancers or selective serotonin reuptake inhibitors. Tweets were categorized by drug
and an upper bound of 300 tweets per drug was set to collect a balanced sample with a total of
1548 tweets actually used for the study. One hundred tweets were annotated and used as the
ground truth to evaluate other annotators from crowd sourcing. In order to verify agreement
between layman annotators and experienced annotators, Cohen's Kappa and Fleiss's Kappa were
computed for different questions such as whether the tweet is in English, whether the tweet
contains drug mention, etc. These scores are then ranked using Spearman's Rho and Kendall's
tau with values of 0.471 and 0.352 respectively. N-grams, hashtags, URLs, were extracted and
classifiers were built. Bayesian Generalized Linear Modeling was found to be the best technique
for interpretation of the data in this study.
One study sought to understand how people responded to the Ebola outbreak on social
media, what type of messages people post, what factors affect these reactions, and patterns
within these reactions [4]. At first 2 billion tweets were collected using the keyword 'Ebola'
from 90 different languages. After removing duplicates and the languages with less than 100
tweets, the final tweet count is 1,167,539 tweets from 35 languages, out of which 569,888 are
geotagged. For the non-geotagged tweets, users' profile locations on Twitter were used to get the
location information that was then analyzed. In order to investigate the spatio-temporal
properties about the information propagation from the central cities (the cities where the
outbreak is reported, in this case, New York and Dallas), they considered the tweets (46,598)
which are over 2 days from the starting day of the outbreak in these cities. Results also showed
that the first reported incident had more impact and received more attention than any other
showing that people pay more attention and react more since it is a new topic. Finally, it showed
the significance of social ties in propagating the information by analyzing how the Twitter
followers post/retweet based on their followee's tweets.
Another study used Twitter to track 300 million tweets about the spread of influenza from
November 2008 to June 2010 [12]. Tweets were annotated as being positive or negative by
annotators, with positive being tweets where the person with the flu tweeted or someone around
a person with the flu tweeted. All other tweets were considered negative. This positive and
negative annotations were considered their influenza corpus. Using a support vector machine
(SVM) based classifier, they eliminated tweets that did not actually mention flu patients based on
their developed influenza corpus. The tweets were divided into four seasons: winter 2008,
summer 2009, winter 2009, and summer 2010 for the test-set. Annotations and estimated values
were compared using the Pearson correlation. They found their method performed well during
non-news periods (0.89 correlation) but did not perform well during excessive news periods, like
during the H1N1 flu epidemic. While these different studies highlighted the utility of using
social media to monitor people's thoughts regarding a specific disease outbreak, they did not
explore the specific themes within a disease.
In this study we conduct an exploratory analysis of finding the subcategories of
discussion topics from Zika related tweets; specifically in four key characteristics of Zika:
symptoms, transmission, treatment, and prevention. Using the system described in Figure 1, we
address the following research questions:
R1: Dataset Distribution Analysis: What proportion of male and female users tweeted about
Zika, what were the polarities of the tweets by male and female users, and what were the
proportion of tweets that discussed topics related to the different disease characteristics -
symptoms, transmission, treatment, and prevention?
R2: Classification Performance Analysis: What was the agreement among the annotators who
labeled a sample of the data that was used as ground truth in this study, what was the
classification performance to detect the tweets relevant to Zika, and how well were the classifiers
able to detect between tweets on the different disease characteristics?
R3: Topical Analysis: What were the main discussion topics in each of these characteristics that
can further inform members at CDC on the biggest concerns or misconceptions regarding Zika
virus?
Methods
Data Collection
Tweets are collected between the time 2016-02-24 and 2016-04-27 for a total of
1,234,605 tweets using Twitris 2.0 [13]. The keywords used to collect the tweets were "Zika,"
"Zika virus," "Zika treatment," and "Zika virus treatment." One thousand four hundred and
sixty-seven random tweets were put in the following categories: relevant or non-relevant, and
then if they were relevant they were further categorized as symptoms, treatment, transmission,
and prevention. These four categories were used because they are characteristics of disease used
in any medical journal and by the CDC and WHO and are what is important for scientists and
medical professionals to know about a disease.
Labeling Process/ Data Annotation
One-thousand four hundred and sixty-seven tweets were annotated as relevant or not, and
then by the four subcategories if they were relevant, by three microbiology and immunology
experts. Inter-rater reliability was found using Fleiss Kappa [14].
Preprocessing
As the initial step, the data were further preprocessed to remove the URL, screen handles
(@username), retweet indicators, and non-ascii characters. Data were further normalized by
removing capital letters, numbers, punctuations and whitespaces from the tweets. Terms were
filtered out to remove single characters like 'd', 'e' which do not convey any meaning about the
topics in the corpus and top words like "and", "so", etc were removed For the classification
stage, each tweet is represented as a feature vector of the words present in the tweet using
unigrams.
Classification
Supervised classification techniques including the decision tree (J48), multinomial naive
bayes (MNB), Bayesian networks (Bayes Net), SMO (sequential minimal optimization) using
SVM, Adaboost, as well as bagging or bootstrapping (Bagging) techniques were implemented on
the Zika dataset for a) classifying relevant tweets on zika, and b) if relevant, further categorizing
it into the disease characteristics. Supervised techniques rely on labeled data, in this case tweets,
that are manually labeled as relevant to Zika virus, as well as the category it belongs to: Zika
symptoms, Zika treatment, Zika transmission, and Zika prevention. They "learn" the nature of
the tweets in the different groups and subgroups.
The performance of each classifier was assessed using 10-fold cross validation, which is
a commonly used method for the evaluation of classification algorithms that diminishes the bias
in the estimation of classifier performance [15]. This approach uses the entire dataset for both
training and testing, and is especially useful when the manually labeled data set is relatively
small. In 10 fold cross-validation, the manually labeled data set is randomly partitioned into 10
equal-sized subsets. The cross-validation process is then repeated 10 times (the folds). Each time,
a single subset is retained as the validation data for testing the model, and the remaining 9
subsamples are used as training data. The 10 results from the folds are then averaged to produce
a single estimation. The study reports the average of the precision, recall, F-scores, and area
under the curve (AUC) calculated by the system. Precision is defined as the number of correctly
classified positive examples divided by the number of examples labeled by the system as
positive. Recall, also referred to as sensitivity, is defined as the number of correctly classified
positive examples divided by the number of positive examples in the manually coded gold
standard data. An F-score is a combination (harmonic mean) of precision and recall measures
[15]. The area under the curve also evaluates the tradeoff between precision and recall using
different threshold values that affect the sensitivity as well as the specificity of the different
algorithms. For the ideal classifier the AUC value would be 1, and the AUC for a random
predictor would be 0.5, so most classifiers have a value between 0.5 and 1.
Topic Modeling
Studies such as Hong [16] have shown the utility of using traditional topic modeling
methods like LDA for grouping the themes occurring in short text documents. The basic idea in
LDA is that documents (tweets in this case) are represented as random mixtures over hidden
topics, where each topic is characterized by a distribution over words that occur the most
frequently within that topic [17]. As an example, if person A makes a decision to sit at a
particular table T at a conference, the probability of A sitting at T is proportional to the other
people sitting at T. Also, as more people join the conference, the probabilities for choosing tables
converge. In LDAs the people are the words, and the tables are the topics present in a given
dataset. In this study, we use topic modeling for finding the underlying topics in each of the four
disease characteristics to find out more about the important issues in each of these categories.
Results
Dataset Distribution (addressing R1):
Overall, 42% of tweets contained a retweet and 85% contained a URL. Tweets by gender
were found using the twitter usernames using the genderize API [18]. According to genderize,
twenty-eight percent of the tweets were by males, twenty-three percent by females and 41% were
by unknown gender.
[19] (Table 1).
Table 1. Polarity and proportion of tweets divided in the gender categories.
The polarity of the individual tweets were also found using the sentiment package in R
Category All tweets
Male tweets
Female tweets
Positive
313,742 (24%) 90,928 (26%)
70,171 (25%)
Negative 783,327 (59%) 220,359 (63%) 180,276 (63%)
Neutral
137,536 (10%) 40,166 (11%)
33,760 (12%)
The polarity of the tweets between males and females were similar. Although a majority
of the tweets are categorized as having negative polarity, the percent of positive tweets was
higher than expected. Some examples of tweets that were classified as positive were "Case
report: assoc btw #Zika/teratogenicity strengthened & evidence shows impact on fetus may take
time to manifest", "RT @NEJM: At recent int'l meeting about #Zika, experts exchanged insights,
identified knowledge gaps, and agreed on a plan", and "91,387 Cases of Zika Confirmed in
Brazil This Year: Brazil has confirmed 91,387 cases of…". Words such as "strengthened",
"agreed", "confirmed" may be why some tweets are being classified as positive.
There was a class imbalance in the categories. Since there is no treatment for Zika, not
many people tweeted about it. Transmission and prevention tweets were high since they were the
most talked about topics concerning Zika. Sample tweets from these different categories as
provided in Table 3.
s
t
e
e
w
T
f
o
#
120000
100000
80000
60000
40000
20000
0
Symptoms
Treatment
Transmission
Prevention
Disease Characteristic
Figure 2. Number of tweets in each disease categorization after running all tweets (1.2 million
tweets) through the best classification model.
Classification Performance (Addressing R2)
In the first stage of the categorization process for the ground truth tweets, tweets were
first classified as being relevant or not relevant to zika (Table 2).
Table 2. Sample tweets relevant and not relevant to zika.
Classification Sample Tweets
Relevant
I guess we can now add Zika to the list of STDs
Millions of GM mosquitoes to fight Zika virus in Caymans
Not Relevant After violent crime, shoddy construction and police brutality in Brazil, Zika
seems minor.
#MoreTrustedThanHillary going to Brazil during Zika virus season
Tweets that were relevant were then categorized as being about symptoms, treatment,
transmission, or prevention (Table 3).
Table 3. Sample tweets from the four categories of relevant tweets.
Classification Sample Tweets
Symptoms WHO sees scientific consensus on Zika virus as cause for disorders
Puerto Rico Reports First Zika-Related Death Amid Outbreak
Treatment
Healthcare providers: See CDC guidelines for caring for pregnant women w/
possible #Zika exposure.
Zika drug breakthrough may be the first step towards treatment.
Transmission Zika virus strain responsible for the outbreaks in Brazil has been detected in
Africa
Zika threatens TWO BILLION people across the world: New maps reveal
where virus is likely.
Prevention
#CDC officials say men who have the #Zika virus should wait 6 months
before trying to conceive. #health
Senate Nears Deal for at Least $1.1 Billion to Fight Zika Virus
One-thousand four hundred and sixty-seven tweets were manually labeled to train the
classifiers and evaluate their performance. Figure 2 provides the distribution of the relevant
tweets in the four categories. As seen from Figure 1, the distribution of the labeled gold standard
dataset was similar to the distribution of the large data corpus.
400
350
300
250
200
150
100
50
0
s
t
e
e
w
T
f
o
#
Symptoms
Treatment
Transmission
Prevention
Disease Characteristics
Figure 3. Number of tweets from the labeled dataset for each of the four categories of disease
characteristics.
Inter-rater Reliability
Fleiss's Kappa was calculated for each category to check inter-rater reliability between
three labelers with domain expertise in microbiology and immunology. Kappa values ranged
from 0.62 to 0.93 for both the relevancy and the four disease characteristics. This indicates
substantial to almost perfect agreement among the raters [20]. Once the inter-rater reliability was
found to be substantial, a model needed to be built based on the gold standard dataset.
Classification and Performance
The table below gives the performance of different classifiers on 1,467 preprocessed
Twitter data to find the relevancy of the tweet towards Zika (Table 4). Unigram features were
extracted from the texts using the Weka toolbox [21]. For this dataset, the classifiers perform
fairly well with AUC values ranging from 0.78 to 0.94. MNB outperforms other classifiers using
both the metrics F Measure (0.86) and AUC (0.94) (Table 4). MNB classifiers perform better for
data sets that have a large variance in document length (in this case, the length of the tweets) by
incorporating the evidence of each appearing word into its model [22].
Table 4. Different classifier performances for detecting relevant tweets using J48, MNB,
Bayesian networks (Bayes Net), SMO using SVM, as well as bagging or bootstrapping
(Bagging) techniques.
Classifier
TP
FP
Precision Recall F Measure AUC
J48
0.821 0.390 0.812
0.821 0.815
0.784
MNB (bayes) 0.880 0.368 0.881
0.880 0.868
0.943
Bayes Net
0.832 0.479 0.821
0.832 0.812
0.837
SMO
0.895 0.252 0.892
0.895 0.892
0.822
Bagging
0.857 0.411 0.852
0.857 0.843
0.877
The tables below detail the actual and predicted results of MNB using the confusion
matrix (Table 5). The class imbalance was affecting the classifier performance. Although the
AUC value was high (0.94), the classifier predicted a tweet was relevant more often than not
relevant since ~80% of the tweets belong to the relevant category. This also affected the false
positive rate which was much higher than the false negative rate as seen in Table 5.
Table 5. Confusion Matrix using the best classifier MNB.
Predicted
1 (relevant) 0 (Not)
Actual
1 (relevant) 1116
0 (Not)
155
21
175
The table below gives the performance of different classifiers on 1,135 preprocessed
twitter data to find the categorical classification (symptoms, treatment, transmission, and
prevention) of the tweets (Table 6). Again, the classifiers performed fairly well with AUC values
ranging from 0.83 to 0.94. With this dataset, MNB outperforms other classifiers again.
Table 6. Different classifier performances for detecting the four disease categories within the
relevant tweets using J48, MNB, Bayesian networks (Bayes Net), SMO using SVM, as well as
bagging or bootstrapping (Bagging) techniques.
Classifier
TP
FP
Precision Recall F Measure AUC
J48
0.694 0.122 0.702
0.694 0.695
0.838
MNB
0.784 0.084 0.787
0.784 0.785
0.940
Bayes Net
0.697 0.121 0.729
0.697 0.702
0.885
SMO (SVM) 0.775 0.088 0.780
0.775 0.777
0.877
Bagging
0.727 0.112 0.741
0.727 0.730
0.901
In order to further understand the best classifier's performance a confusion matrix was
created. The tables below detail the actual and predicted results using the MNB classifier (Table
7). Here, again the proportion of categories in the gold standard dataset affects the classifier
predictions. The treatment was predicted the least number of times (169 out of 1135). However,
the diagonal values (True positives) were higher than the misclassification values which accounts
for the high AUC value. That said, there was also a noticeable overlap between Transmission
and Prevention (42 tweets belonging to Prevention categorized as Transmission and 41 tweets
belonging to Transmission categorized as Prevention). The reason these overlap may be due to
the fact that the words mosquito and sex was used to describe how Zika is transmitted and how
to prevent transmission. Also, they are closely linked in that prevention cannot occur unless the
mode of transmission is known.
Table 7. Confusion Matrix using the best classifier MNB.
Predicted
Actual
Symptoms Treatment Transmission Prevention
Symptoms
205
Treatment
11
Transmission 20
Prevention
20
4
146
4
15
45
6
264
42
8
29
41
275
Based on the above results, the two-staged classifier system was able to have a high
precision, as well as a high recall performance in categorizing the tweets into relevant and not
relevant, and further classifying the tweets into the four disease categories. Once the
performance of the model based on the gold standard data set was confirmed to have high
precision and recall, the model needed to be tested on a new set of tweets.
Confusion Matrix Error Analysis
Similar to the study by Jiang [23], an error analysis was performed on the large test
dataset by randomly selecting 530 tweets, manually labeling them, and comparing the results
using the best performing MNB classifiers for identifying relevancy (Table 8), and then further
categorizing the relevant tweets into the four subcategories (Table 9). Reevaluating the
performance using the test set enables the unbiased assessment of the generalization error of the
final chosen model. Here, 530 tweets were further annotated and tested the MNB classifiers from
the earlier section. High precision and recall values were obtained for the relevance classifier
with Precision =0.99 and Recall =0.99. Hence, the F measure was also 0.99 (harmonic mean of
precision and recall). This high performance of the classifier indicates that the gold standard
dataset was a good representation of the distribution of the tweets in the large data corpus.
Moreover, the dataset was significantly less noisy when comparing the performance of these
models with other domain areas such as gender based violence [5], drug-related tweets [24] and
religion [25].
Table 8. Relevant tweet error analysis.
Annotator
Not-relevant
Relevant
MNB
Not-relevant
Relevant
23
7
30
Table 9. Disease characteristics error analysis results.
7
493
500
30
500
530
Annotator
Symptoms Treatment Transmission Prevention Total
Symptoms
88
MNB
Treatment
2
Transmission 13
Prevention
5
0
29
1
0
5
3
129
10
Total
108
30
147
1
6
25
152
184
94
40
168
167
469
Even though the classes were unbalanced, high precision and recall values were still
obtained (Table 10). An overall high F measure of 0.9 was obtained. This further indicated the
gold standard dataset was a good representation of the tweets, as well as the disease categories in
the larger corpus and that the dataset was not very noisy.
Table 10. Precision recall and F measure for each of the four disease characteristics.
Category Symptoms Treatment Transmission Prevention Average
precision 0.98
recall
0.81
0.97
0.97
0.86
0.88
0.94
0.83
0.94
0.87
F
0.89
0.97
0.87
0.88
0.90
The error analysis indicated that the classifiers performed well with the unseen test data
and were generalized enough to work with the large dataset. The dataset was further examined
with a focus on the insights provided in the disease categories. More specifically, the topics
discussed on Twitter in the symptoms category was examined to discover the latent semantic
topics discussed.
Topical Analysis (Addressing R3):
Topic Modeling
Below the results of LDA are discussed for each of the four disease characteristics
(Tables 13-14). Based on heuristical results, the number of topics was restricted to 5 while
discussing the topics for each category. Topic modeling results are shared here [26] for the
research community to examine the outcome of using topic modeling, as well as the overlap
among the topics generated. First the results for the three categories prevention, transmission and
treatment will be discussed. Then a more detailed analysis of the topic modeling results for the
symptoms category will be discussed along with the misinformation tweeted by users in that
domain.
Prevention, Transmission and Treatment
Table 11 provides the topics for the three categories along with the sample tweets in each
topic.
Prevention: Within the prevention topics, topic #1 was about the need to control/prevent
spread, topic #2 was about the need for money to combat mosquitoes and research treatments,
topic #3 was about ways to actually prevent spread, topic #4 was about the bill to get funds, and
topic #5 was research (Table 11). These topics were not surprising considering all the discussion
about how to prevent Zika, the need for funding to prevent Zika, and the research required to
find a cure for Zika since it is an emerging disease. There is also a need to better understand Zika
virus, the disease it causes, and ways to combat it [27].
Transmission: In transmission, there was a strong overlap in topics #1 (vector i.e.
mosquitoes for Zika) and #4 (disease spread) that highlight the overlap between the mosquito
spread, and the disease in general. Looking at the tweets, both highlight the concerns and risks
associated with zika spread. Another topic (#2) was sexual which is another mode of
transmission besides through mosquitoes. The next topic (#3) was infants, who are most affected
by this epidemic due to the risk of microcephaly. The final topic (#5) was sports since the tweets
were collected during baseball season just before the olympics and many athletes were
concerned about getting affected with Zika while competing in Rio, 2016.
Treatment: There was slight overlap between topics #1 (lack of treatment) and #3
(vaccine development) for treatment primarily due to the large co-occurrence of the word
vaccine in both these topics. Blood testing (#4) was another major topic since some people got
infected with Zika after receiving a blood transfusion. Since no treatment exists, a lot of research
is focused on developing a drug for Zika, which is why test development (#5) was the final topic.
Table 11. Prevention, transmission, and treatment topic modelling results.
Disease
Characteristic
Topic
Sample Tweets for each Topic
Prevention
(#1) Control
RT @DrFriedenCDC: A2. The best way to prevent
#Zika & other diseases spread by mosquitoes is to
protect yourself from mosquito bites. #Reut
(#2) Money
Need
#healthy Congress has not yet acted on Obama's $2
billion in emergency funding for Zika, submitted in
February
(#3) Prevention RT @bmj_latest: Couples at risk from exposure to Zika
virus should consider delaying pregnancy, says
@CDCgov
(#4) Bill
https://t.co/Ke12LOdypf Senate Approves $1.1 Billion
In Funding To Fight The Zika Virus #NYCnowApp
(#5) Research
Florida is among those at greatest risk for Zika.
@FLGovScott's sweeping abortion bill blocks scientists'
access to conduct research
Transmission
(#1) Vectors
(mosquitoes)
This map shows the Northeast is at risk for Zika
mosquitos this summer
(#2) Sexual
@user1 First Sexually Transmitted Case Of Zika Virus
In U.S. Confirmed
(#3) Infants
CDC reports 157 cases of U.S. pregnant women infected
with Zika virus.
(#4) Spread
Zika strain from Americas outbreak spreads in Africa for
first time: WHO (Update)
(#5) Sports
MLB moves games from Puerto Rico due to Zika
concerns....uh..what about the Olympics?? Can't be good.
Treatment
(#1) Lack of
Treatment
RT @DrFriedenCDC: Much is still unknown about
#Zika and there is no current medicine for treatment or
vaccine to prevent the virus.
(#2) Zika Test
Rapid Zika Test Is Introduced by Researchers The test,
done with a piece of paper that changes color if the virus
…
(#3) Vaccine
Development
Researchers discover structure of Zika virus, a key
discovery in development of antiviral treatments and
vaccines
(#4) Blood Test Experimental blood test for Zika screening approved
(#5) Test
Development
New mouse model leads way for #Zika drug, vaccine
tests
Symptoms: In the topic model results for symptoms, topics #1 (zika effects), #2 (brain
defects) and #4 (zika scarier than thought) were well separated while topics #3 (confirmation of
defects) and #5 (initial reports) overlap significantly (Figure 4 & Table 12). The topics are
described in Table 16. Topics #3 and #5 overlap for symptoms because a lot of the initial reports
were about health official confirming defects. Topic #3 was more about the defects that were
confirmed while topic #5 focused on where reports came from.
Figure 4. LDA of symptoms.
All three topics (#1, #3, #5) were closely related in that they are about the defects caused
by Zika, but point to slightly different concepts (Table 12). For example, microcephaly is not the
only defect, there is also Guillain-Barre which would be topic #1. Topic #2 focuses on
microcephaly because that was a major topic of discussion as seen in these tweets: "RT
@USATODAYhealth: Zika affects babies even in later stages of pregnancy. Microcephaly seen
in babies from moms infected in 6th month"and "Zika Virus May Cause Microcephaly by
Hijacking Human Immune MoleculeFetal brain model provides first clues on how Z…"Topic #3
contains tweets when the defects were confirmed to be caused by Zika and not something else:
"Zika linked to fetal brain damage: Finnish study: infectious Zika virus from fetal tissue in cell
culture.The virus", and "Enough conspiracy theories; nature is nasty enough: U.S. health
officials confirm Zika cause of severe birth defects".
Topic #4 for symptoms was primarily just people discussing a British Broadcasting
Corporation article [28] on how more birth defects have been linked to Zika and that the virus
was expected to travel further than initially thought, leading to experts saying Zika is scarier than
they first thought. The statement of Zika being "scarier than we first thought" by the CDC was a
big topic on twitter "CDC says zika virus scarier than thought as US prepares for outbreak: On
Monday, the U.S. Centers for Disease…", and this also affected the U.S political environment:
"#2016elections U.S. Officials Warn Zika Scarier Than Initially Thought: By Timothy Gardner
and Jeff Mason WA…", , This discussion lead to some more tweets about the danger of Zika
virus: "The Edge: Zika Is Now Even More Terrifying", "Zika virus 'shrinks brains' in tests", and
"#Zika Survivor Says 'I Could Feel My Skin Shrinking' CBS Boston". These tweets demonstrate
how a statement by the CDC can be spread and how users can tweak the wording of these CDC
statements to generate more concern than the actual impact of the disease. Finally, topic #5
includes tweets about initial reports of Zika outbreaks and deaths.
Misinformation: Within symptoms, several tweets in topic #1 were calling Zika a hoax,
"Zika HOAX exposed by South American doctors: Brain deformations caused by larvicide
chemical", "The Zika Virus is a hoax! It is like calling the common cold an epidemic. It's what
they put in the drinking water.", and "CDC likely fabricating link between Zika virus and
microcephaly cases". However the CDC has stated multiple times that Zika and microcephaly
are definitely linked "CDC: Zika definitely causes severe birth defects", and"Here's a #Zika
basic: Zika infection during pregnancy can cause some severe birth defects". Some of the people
saying Zika is a hoax are misunderstanding this quote from the CDC "People usually don't get
sick enough to go to the hospital, and they very rarely die of Zika. For this reason, many people
might not realize they have been infected. Once a person has been infected, he or she is likely
protected from future infections". This statement is true for the majority of healthy adults.
However, for infants it can cause microcephaly and in some cases Guillain-Barre syndrome in
healthy adults "Symptoms of Guillain-Barre syndrome include weaknesses in arms & legs. GBS
is linked w/ #Zika". There also has already been a couple of deaths due to Zika as was detailed in
topic #5 in the LDA. The CDC has also been directly answering people's questions about Zika
on twitter. One user tweeted at the CDC "Why is #Zika of particular concern to women who are
pregnant or considering becoming pregnant? #WellnessWed #ZikaVirus" to which the CDC
responded "Zika infection in pregnancy can cause microcephaly and other severe brain defects.
http://1.usa.gov/1Pf79sK #WellnessWed". This shows that while some misinformation is still
getting tweeted, the CDC is working to get the correct information out there. This is useful
because it shows the CDC could potentially target specific user groups directly through the topic
modelling approach, and respond to users within a topic group with a similar response that can
allow the information to spread to a larger population with relatively lesser effort.
Table 12. Symptoms topic modelling results.
Topic
Words
Tweets
(#1) Zika
Effects
(#2) Brain
Defects
infect, babies, mosquito,
cause, microcephaly,
symptom, pregnancy
RT @USATODAYhealth: Zika affects babies
even in later stages of pregnancy. Microcephaly
seen in babies from moms infected in 6th month
brain, link, studies,
microcephaly, baby,
disorder, cause, damage,
infect, fetal
'Zika Virus May Cause Microcephaly by
Hijacking Human Immune MoleculeFetal brain
model provides first clues on how Z…
(#3)
Confirmed
Defects
defect, cause, birth,
confirm, health, severe,
link, official
Enough conspiracy theories; nature is nasty
enough: U.S. health officials confirm Zika cause
of severe birth defects
(#4) Scarier
Than Thought
scarier, than, thought, us,
official, health, CDC,
warn, learn, first
(#5) Initial
Reports
first, report, death, case,
puerto, confirm, rico,
cause, colombia, defect
#breakingnews Zika Virus 'Scarier Than We
First Thought,' Warn US Health Officials
Colombia Reports First Cases of Microcephaly
Linked to Zika Virus - Sun Jan 09 15:13:20 EST
In this section, the topic modeling results generate insightful results that allow
researchers to understand the citizens' concerns, as well as the misinformation spread. According
to the theory of LDA, each topic represents certain common properties, which reflects the pattern
in the tweets. Finding out the exact meanings of the topics requires additional information and
domain knowledge. We see that for each of the disease characteristics, the discovered topics can
be easily interpreted by using some domain specific knowledge.
Discussion
Classification Analysis
One of the interesting findings of our analysis was the fact that the Multivariate Naive
Bayes classifier or the MNB outperformed the other more popular classifiers in text analytics:
random forest (J48) and SVM. According to one study [29], this has to do with the class
imbalance issue in our dataset: both for the relevancy case, as well as for the disease categories
case. This also highlights the possible orthogonality of the features used in our study: the
unigrams. Specifically in this dataset, measuring the likelihood of the features in a given class
independently outperforms other complex models such as J48 and SVM. This possibly also
relates to the fact that the data are less noisy since they have been evaluated by expert annotators.
Naive bayes is one of the simplest classification models available to us, but it is nonetheless
among the most effective for this dataset. This result is non-intuitive, but is not surprising when
we consider that using text for classification is relatively imprecise compared to other types of
data. In datasets with large amounts of error, simpler models are less likely to overfit the data.
Hence we recommend that future research on text analytics begin with Naive Bayes and then
proceed to using more complex models to see if these actually improve classification accuracy.
Annotation Observations
One major issue annotating tweets was what to do about news tweets like this one: "Your
Wednesday Briefing: Bernie Sanders, Hillary Clinton, Zika Virus: Here's what you need to know
to start …". The issue was that this does give information about Zika in that it tells what news
stations were talking about it and what else was going on at the same time as the Zika outbreak.
However, the tweet itself does not give any information about Zika symptoms, treatment,
transmission, and prevention. The researchers decided to code these tweets as relevant because
they were about Zika, but not include them in the disease characteristics annotations since they
do not have any information about Zikas disease characteristics. Sporting events in general were
also included because they could be sources of transmission from athletes and fans not taking
proper precautions.
Topic Modeling
example, control, money need, prevention, bill, and research were all major topics of prevention
discussions. This indicates the classification model accurately labelled tweets. It also indicates
tweets about major topics were collected and accurately reflected in our topic model. Also, while
all four disease characteristics are important, symptoms was discussed in detail because the
researchers felt it included the most important information for public health officials to know
especially once the misconceptions/ misinformation about Zika being a hoax was found.
Categorizing the symptoms into the different topics using topic modeling also allowed us to get
deeper into the themes within the symptoms category that can allow a more targeted interaction
with agencies like CDC and specific users to provide interventions for misinformation spread.
Limitations
All of the topics under the different disease characteristics fit the characteristic. For
While we have clearly provided an exploratory framework in mining the different
characteristics of Zika, we point out the limitations we face in our dataset, and the use of social
media.
Language constraint: We have restricted our study to English language tweets that
certainly limit the strength of our study. This is more critical to address given that South
American countries were initially affected by Zika. This also restricts our analysis of measuring
disease outbreak which is why we refrained from doing so in our study.
Keyword constraint: As described in the Data collection section, we used the keywords
Zika, Zika virus, Zika treatment, and Zika virus treatment in our study. One interesting
observation here is that although the keyword "treatment" was part of the crawling process, the
treatment subcategory was still the smallest class in the distribution of the dataset (see Figures 2
and 3). Clearly, this certainly affected our dataset. This also would overlook tweets that referred
to the disease in a different name or talked about the disease without using the word Zika.
Gender & Polarity constraint: Only 49% were labeled by the gender API using the
profile name (Table 1). Similarly, around 10% of the tweets were not labeled on their polarity.
This needs to be addressed moving forward with this study by creating a customized gender
recognition tool using machine learning specifically for twitter data.
Conclusion and Future Work
This is one of the first studies to report successful creation of an automated content
classification tool to analyze Zika–related tweets; specifically in the area of epidemiology. Such
a system will help advance the field's technological and methodological capabilities to harness
social media sources for disease surveillance research. Future studies will include creating an
automated technique to detect misinformation using tweets to allow for well-targeted, timely
interventions. Such a platform will generate data on emerging temporal trends for more timely
interventions and policy responses to misinformation on Zika studies leveraging information
sources including blogs, news articles, as well as social media data.
Acknowledgements
We would like to thank Dr. Rua, Scott Holdgreve, Ryan, Becker, Amber Todd, and Sampath
Gogineni for their help.
Abbreviations
LDA: Latent dirichlet allocation
CDC: Center for Disease Control
WHO: World Health Organization
GBS: Guillain-Barre syndrome
DC: Disease classifier
RC: Relevancy classifier
API: Application program interface
SVM: Support vector machine
J48: Decision tree
MNB: Multinomial naive bayes
SMO: Sequential minimal optimization
AUC: Area under the curve
Works Cited
1. Nanlong, MT. 2014. Nigeria: Ebola - Two die after drinking salt water in Jos. Vanguard.
http://www.webcitation.org/6mhjzekwJ
2. Centers for Disease Control and Prevention (CDC). 2016. Transcript for CDC telebriefing:
Zika summit press conference. Retrieved from: http://www.webcitation.org/6mhnFbn5e
3. Berg, N. 2013. How citizens have become sensors. Retrieved from:
http://www.webcitation.org/6mhnJVcJ8
4. Tran, T, & Lee, K. Understanding citizen reactions and Ebola-related information
propagation on social media. Presented at: International Conference on Advances in Social
Networks Analysis and Mining; 2016; Retrieved from:
http://www.webcitation.org/6mhnQkdoi
5. Purohit, H, Banerjee, T, Hampton, A, Shalin, VL, Bhandutia, N, & Sheth. AP. Gender-based
violence in 140 characters or fewer: a #BigData case study of Twitter. First Monday 2016; 21(1).
6. Paul, MJ, & Dredze, M. You are what you Tweet: Analyzing Twitter for public health.
Presented at: ICWSM; 2011;20:265-272.
7. Bhattacharya, S, Tran, H, & Srinivasan, P. Discovering Health Beliefs in Twitter. Presented
at: AAAI-2012 Fall Symposium on Information Retrieval and Knowledge Discovery in
Biomedical Text 2012. Washington, DC.
8. Wong, SS, Poon, RW, & Wong, SC. Zika virus infection-the next wave after Dengue?
Journal of the Formosan Medical Association 2016;115(4):226-242.
9. Center for Disease Control (CDC). 2016. Zika Virus. Retrieved from:
http://www.webcitation.org/6mhnTZk4b
10. Oyeyemi, SO, Gabarron, E, Wynn, R. Ebola, Twitter, and misinformation: a
dangerous combination? BMJ 2014;349. doi: http://dx.doi.org/10.1136/bmj.g6178
11. Alvaro, N, Conway, M, Doan, S, Lofi, C, Overington, J, & Collier, N. Crowdsourcing
Twitter annotations to identify first-hand experiences of prescription drug use. Journal of
Biomedical Informatics 2015;58:280-287.
12. Aramaki, E, Maskawa, S, & Morita, M. Twitter catches the flu: Detecting influenza
epidemics using Twitter. Presented at: 2011 Conference on Empirical Methods in Natural
Language Processing; 2011; 1568-1576.
13. Jadhav, A, Purohit, H, Kapanipathi, P, Ananthram, P, Ranabahu, A, Nguyen, V, Mendes, P,
Smith, AG, Cooney, M, & Sheth, A. Twitris 2.0: Semantically empowered system for
understanding perceptions from social data. Presented at: Semantic Web Application Challenge
at ISWC 2010; Shanghai, China.
14. McHugh, ML. Interrater reliability: the kappa statistic. Lessons in Biostatistics
2012;22(3):276-282.
15. Hastie, T, Tibshirani, R, & Friedman, J. The elements of statistical learning: Data mining,
inference, and prediction. New York: Springer; 2009.
16. Hong, L, & Davison, BD. Empirical study of topic modeling in twitter. Presented at:
Proceedings of the first workshop on social media analytics; 2010; 80-88.
17. Blei, DM, Ng, AY, & Jordan, MI. Latent dirichlet allocation. The Journal of Machine
Learning Research 2003;3:993–1022.
18. http://www.webcitation.org/6mhnfgwtR
19. http://www.webcitation.org/6mhnirAFy
20. Landis, JR, & Koch, GG. The measurement of observer agreement for categorical data.
Biometrics 1977;33(1):159–174.
21. Frank, E., Hall, M. A., & Witten I. H. 2016. The WEKA Workbench. Online Appendix for
"Data mining: Practical machine learning tools and techniques", Morgan Kaufmann, Fourth
Edition.
22. McCallum, A, & Nigam, K. A comparison of event models for naive bayes test
classification. Presented at: AAAI-98 workshop on learning for text categorization 1998;752:41-
48.
23. Jiang, L, Yu, M, Zhou, M, Liu, X, & Zhao, T. (2011). Target-dependent twitter sentiment
classification. Presented at: Proceedings of the 49th Annual Meeting of the Association for
Computational Linguistics: Human Language Technologies;1:151-160.
24. Daniulaityte, R, Chen, L, Lamy, FR, Carlson, RG, Thirunarayan, K, & Sheth, A. "When
'bad' is 'good'": Identifying personal communication and sentiment in drug-related tweets.
JMIR Public Health Surveillance 2016;2(2):162.
25. Chen, L, Weber, I, & Okulicz-Kozaryn, A. U.S. Religious landscape on Twitter. Presented
at: Proceedings of the 6th International Conference on Social Informatics 2014.
26. http://www.webcitation.org/6mhnmw2I3
27. National Institute of Allergy and Infectious Disease (NIAID). Zika virus. Retrieved from:
http://www.webcitation.org/6mhnqsP3D
28. British Broadcasting Corporation (BBC). 2016. Zika virus 'scarier than thought' says US.
Retrieved from: http://www.webcitation.org/6mhp7CZwI
29. Frank, E, and Bouckaert, R. Naive bayes for text classification with unbalanced classes.
Presented at: Proceedings of PKDD 2006;503–510.
|
1801.01239 | 2 | 1801 | 2018-08-20T04:51:24 | How to Beat Science and Influence People: Policy Makers and Propaganda in Epistemic Networks | [
"cs.SI",
"physics.soc-ph"
] | In their recent book Merchants of Doubt [New York:Bloomsbury 2010], Naomi Oreskes and Erik Conway describe the "tobacco strategy", which was used by the tobacco industry to influence policy makers regarding the health risks of tobacco products. The strategy involved two parts, consisting of (1) promoting and sharing independent research supporting the industry's preferred position and (2) funding additional research, but selectively publishing the results. We introduce a model of the Tobacco Strategy, and use it to argue that both prongs of the strategy can be extremely effective--even when policy makers rationally update on all evidence available to them. As we elaborate, this model helps illustrate the conditions under which the Tobacco Strategy is particularly successful. In addition, we show how journalists engaged in "fair" reporting can inadvertently mimic the effects of industry on public belief. | cs.SI | cs |
How to Beat Science and Influence People:
Policy Makers and Propaganda in Epistemic Networks
James Owen Weatherall, Cailin O'Connor
Department of Logic and Philosophy of Science
University of California, Irvine
Justin P. Bruner
Department of Theoretical Philosophy
University of Groningen
Abstract
In their recent book Merchants of Doubt [New York:Bloomsbury 2010], Naomi Oreskes and
Erik Conway describe the "tobacco strategy", which was used by the tobacco industry to
influence policy makers regarding the health risks of tobacco products. The strategy involved
two parts, consisting of (1) promoting and sharing independent research supporting the
industry's preferred position and (2) funding additional research, but selectively publishing
the results. We introduce a model of the Tobacco Strategy, and use it to argue that both
prongs of the strategy can be extremely effective -- even when policy makers rationally update
on all evidence available to them. As we elaborate, this model helps illustrate the conditions
under which the Tobacco Strategy is particularly successful.
In addition, we show how
journalists engaged in 'fair' reporting can inadvertently mimic the effects of industry on
public belief.
1. Introduction
In December 1952, Readers Digest -- the publication with the largest circulation in the world
at the time -- published an article with the title "Cancer by the Carton" (Norr [1952]).1
Coming on the heels of a period during which smoking rates had risen significantly, the
article brought the growing body of evidence that cigarette smoking caused cancer to the
attention of a wide audience. Six months later, a group of doctors at Sloan-Kettering
Memorial Hospital completed a study in which they demonstrated that mice painted with
Email addresses: [email protected] (James Owen Weatherall), [email protected] (Cailin O'Connor),
[email protected] (Justin P. Bruner)
1The history presented in these opening paragraphs draws heavily on the account given by Oreskes and
Conway [2010, Ch. 1]. Note that the chronology we describe differs slightly from theirs, as they report
that the Reader's Digest article was published in the wake of the 1953 Wynder article, whereas in fact it
appeared almost a year earlier. See also the discussion in O'Connor and Weatherall [2019].
Draft of August 21, 2018
PLEASE CITE THE PUBLISHED VERSION IF AVAILABLE!
cigarette tar developed malignant carcinomas (Wynder et al. [1953]). To a public already
primed by the Reader's Digest article, this paper provided a direct and visceral link between
a known by-product of smoking and fatal cancer. It produced a media frenzy, with articles
reporting the result in a number of national newspapers and magazines.
The tobacco industry panicked. The heads of six major U.S. tobacco companies con-
tracted with the New York public relations firm Hill and Knowlton to combat claims that
their product had serious adverse health effects. At a meeting in December of 1953, they
agreed on a novel strategy: they would fight science with science. In 1954 they hired a promi-
nent geneticist, C. C. Little, to run the newly create Tobacco Industry Research Committee
(TIRC). The goal of this committee was ostensibly to support and promote research on the
health effects of tobacco. But in fact, it was a propaganda machine whose principal activity
was to find, fund, and promote scientific research that contradicted the growing consensus
that smoking kills. The goal was to create a false sense of uncertainty and controversy, so
as to delay policy change in the service of public health.
As historians Oreskes and Conway [2010] painstakingly document, drawing on documents
released as part of a series of lawsuits in the 1990s, the tobacco industry initiated this
strategy -- which Oreskes and Conway dub the "Tobacco Strategy" -- despite the fact that
their own internal research had already established a link between smoking and various
health risks.2 The Tobacco Strategy had multiple parts. One part, which we will call selective
sharing, consisted of identifying and promoting research produced independently of industry
that happened to support the industry's preferred position. For instance, in 1954 the TIRC
distributed a pamphlet entitled "A Scientific Perspective on the Cigarette Controversy" to
nearly 200,000 doctors, journalists, and policy-makers, in which they emphasized favorable
research and questioned results supporting the contrary view. A second part of the Tobacco
Strategy, which we will call biased production, consisted of funding their own research, either
through in-house scientists or by offering research grants to scientists pursuing research
activities that the industry supported, and then publishing the results of that research only
if it supported industry views.
Oreskes and Conway [2010] make a compelling case both that the Tobacco Strategy was
self-consciously adopted by industry to influence public opinion, and that those goals were
apparently achieved. They also argue that the Tobacco Strategy succeeded, in the sense
of playing a causal role in realizing these ends. But the historical analysis they provide is
necessarily in situ, and the details are complex. It is difficult to clearly distinguish the causal
role played by the TIRC's activities from other cultural and psychological considerations
that may have influenced public opinion and slowed policy-change, such as the facts that
smoking was culturally ingrained, including among doctors and legislators, and that the
tobacco industry made significant political contributions. One might worry that it was over-
determined that cultural and political change regarding tobacco would be slow, so that it is
not clear that the Tobacco Strategy played a significant role.
In this paper, we attempt to strengthen the causal link between the Tobacco Strategy and
public opinion by proposing and analyzing an agent-based model of the Tobacco Strategy.
2For other discussions of this strategy, see for instance Michaels and Monforton [2005]; Michaels [2008].
2
Our model is based on the network epistemology framework introduced by Bala and Goyal
[1998], later adapted to philosophy of science by Zollman [2007], and used more recently
by Holman and Bruner [2015, 2017] to study industry effects on science.3 This approach
allows us to test counterfactual effects of industry propaganda on public belief, free from
compounding factors.4 As we will argue, our model provides evidence that the Tobacco
Strategy can, in principle, be very effective. We also provide important insight into how
(and when) it works. As we will argue, the Tobacco Strategy trades on the fact that
empirical research into difficult, probabilistic problems will yield results that, individually,
support more than one position. It is only by viewing a maximal, unbiased sample of the
output of a research community that one can hope to come to hold reliable beliefs about the
problem at hand. This means that even subtle manipulations, not in the scientific record
itself, but in the sample of that record that one is exposed to, can have disproportionate
impact on one's beliefs5.
An important take-away from these arguments is that what may seem like weak or
subtle interventions are often most effective for the would-be propagandist. In particular,
outright scientific fraud -- intentional publication of incorrect, fabricated, or purposely mis-
leading results -- is not only unnecessary to influence public opinion on topics of scientific
inquiry, it is also riskier and often less effective than other forms of manipulation. Biased
production, which does not involve fabricating results, is a successful strategy for mislead-
ing the public. And in many cases, biased production is itself less effective than selective
sharing -- even though it involves more direct and explicit interventions on the production of
science. These observations support claims by Holman and Bruner [2017] that more subtle
industry influences -- like their "industrial selection", which we will discuss at length in the
conclusion of this paper -- can be very effective. In fact, as we will also argue in what follows,
even interventions with no malicious intent, such as those of journalists who aim to be "fair"
in their reporting on controversial science, can unwittingly mimic the effects of an active
propagandist.
Before proceeding, we wish to emphasize what is at stake in this discussion. Recognizing
how the Tobacco Strategy works is essential for understanding the relationship between
science, public policy, and democracy. There is now a large literature in philosophy of science
3Similar methods are used by O'Connor and Weatherall [2019], O'Connor and Weatherall [2017], and
Weatherall and O'Connor [2018] to discuss the propagation of false beliefs in social networks more generally.
4 To be clear, we take this ability to intervene in our model to be the key aspect that helps us strengthen
the causal link between the Tobacco Strategy and public policy related to tobacco products. The model
allows us to test where and how the presence of a propagandist can influence public belief. It also lets us
explore the effects of a propagandist in isolation from other potential causal factors. That said, we do not
claim that the factors we consider here are the only relevant causal factors, nor that other models could
not be used to identify other counterfactual relationships and possible causal factors. For more on the
interventionist account of causation that is lurking in the background here, see Woodward [2003]; see also
note 8 for more on the relationship between the model we discuss and "how-possibly" modeling.
5This conclusion is in line with prior work in social epistemology exploring how subjects come to form
incorrect beliefs on the basis of selective reporting (for an overview, see Romero [2016]). We demonstrate how
the public can be systematically mislead even in those circumstances where all experiments are published
and scientists exhibit no bias in favor of industry.
3
on what it means for science to be suitably "democratic," in the sense of reflecting the needs
and values of the members of a society affected by scientific results.6 This literature focuses
on how democratic ideals should shape scientific research. But the other direction, of how to
best understand the role that scientific knowledge plays in democratic institutions, is equally
important -- and requires us to reckon with the fact that there are powerful political and
economic interests that wish to manipulate and interfere with the distribution of scientific
knowledge. Understanding the mechanisms by which public understanding of science can fail
is crucial to developing institutions and interventions which might be effective in defanging
the Tobacco Strategy, and, in the case of 'fair' reporting, improving journalistic practice to
protect public belief.
The remainder of the paper will be structured as follows. In the next section we will
introduce the model we consider in the paper. Sections 3 and 4 will present results concerning
selective sharing and biased production, respectively. Section 5 will describe results from
a slightly modified model in which we consider ways in which journalistic practice with no
malicious intent can simulate the actions of a propagandist. We will conclude with some
reflections on the results, including a discussion of possible interventions.
2. Epistemic Network Models
Bala and Goyal [1998] introduce a model where individuals in a community try to choose
which of two theories, or practices, is the better one. In order to decide, they make use
evidence from two sources. First, each individual tests the theory they favor, and second,
they share their results with those in their social network. Over time, they update their
beliefs about the theories based on this evidence, until, in many cases, the whole group comes
to believe that one or the other is better. Zollman [2007, 2010] shows how simulations of
this model can be applied to understand scientific discovery and consensus. Under his
interpretation, each individual in the model is a scientist testing a hypothesis. For instance,
suppose a new treatment for cancer is introduced. How do researchers go about deciding
whether to adopt the new treatment or stick with the old one?
The model starts with a network, defined by a set of scientists and a set of fixed links
representing pathways of communication between them. Inquiry is modeled using what is
called a two-armed bandit problem. Each scientist is aware that there are two actions they
can perform. Action A is well understood, and is known to be successful 50% of the time.
Action B, on the other hand, is less well understood, and the agents are uncertain whether
it is more or less successful than A. In particular, the instantiation of the model we use here
assumes that B is a bit better than A, so that its chance of success is pB = .5+.7 Scientists,
however, do not know whether it is a bit better, or a bit worse (success rate .5−). Their aim
is then to discover which of these possibilities is the better one, and so adopt the superior
6See, for instance, Longino [1990, 2002], Kitcher [2003, 2011], Douglas [2009], Elliott [2011], and references
therein.
7The payoff of success for both actions is the same, and failure yields no payoff, so action B has a strictly
higher expected payoff.
4
action. In what follows, we will refer to "theory A" as the view that action A is preferable,
and "theory B" as the view that action B is preferable.
At the beginning of a simulation, scientists are given random credences about which
theory is better, represented by numbers between 0 and 1. (For example, a scientist with
credence .7 believes there is a 70% chance that theory B is right.) Each round, scientists
apply their preferred theory to the world. Agents with a credence < .5 believe that theory
A is probably correct and employ it. Agents with credence > .5, on the other hand, use
theory B. This might correspond to doctors who approve of a new treatment deciding to try
it on patients, while others stick with a more conservative approach.
In each round, scientists observe how their choices worked, and change their beliefs in
light of new evidence. In particular, it is assumed that they test their theory some number
of times, n. They then update their credence using Bayes' rule based on what they, and
all their network neighbors, have observed. Simulations stop when either all agents believe
in theory A -- and so do not learn anything new about theory B -- or when all agents have
reached some certainty threshold (say .99 credence) that B is correct, and so are exceedingly
unlikely to end up changing their minds. Both outcomes occur in simulations. Sometimes
initial beliefs and/or strings of data supporting theory A are enough to create a mistaken
consensus around the worse theory, and sometimes the community gets enough good data
to settle on the better one.
As noted in the introduction, we are particularly interested in the role of industry, and
propagandists in general, in shaping policy based on science, and in influencing popular
belief. For this reason, our models follow the framework developed by Bala and Goyal
[1998] and Zollman [2007], but with a few, key alterations. First, we introduce a new set of
agents, which we will call policy makers. These agents have beliefs, and update these in light
of evidence, but do not gather evidence themselves. Each one listens to some, or all, of the
scientists in the network, but does not influence them or anyone else. We assume that at the
beginning of each simulation, policy makers are skeptical about the new, uncertain theory,
in the sense that their initial credences are sampled from a uniform distribution supported
on the interval (0, .5).
We also introduce one further agent that we call the propagandist. This agent observes
all the scientists, and communicates with all the policy makers. Their goal is to convince
the policy makers to prefer the worse theory A, by sharing evidence that suggests action A
is preferable to action B. We consider two treatments, corresponding to different strategies
by which the propagandist may do this. The idea is to represent an individual, or set of
individuals, who are motivated to use all their resources to publicize data that is ultimately
misinformation. The actions of the propagandist capture various aspects of the Tobacco
Strategy: by sharing a biased sample of evidence with policy makers (and the public) they
seek to divide public opinion from scientific consensus.8
8 It is common in discussions of idealized models of epistemic communities -- and philosophical discussion
of modeling in the social sciences generally -- to suppose that such models provide, at best, "how-possibly"
information about a target phenomenon. (See Frey and Seselja [2018] for a discussion of this view.) We
wish to explicitly reject this attitude, at least with regard to the inferences we draw from the present
5
3. Selective Sharing
In the first treatment, which we call selective sharing, the propagandist adopts the following
strategy: in each round of the simulation, they search the network of scientists for results
that tend to favor theory A, and then they share only those results with the policy makers.
The policy makers then update their credence in light of both the evidence they receive from
the scientists to whom they are connected and the evidence received from the propagandist,
which includes a subset of the results produced by the scientists.9 In other words, each policy
maker receives, and updates on, a biased sample of evidence including a disproportionate
number of results favoring theory A. However, all of the results the policy makers receive are
"real" in the sense that they are produced by members of the scientific community, without
bias or influence by the propagandist or anyone else. In addition, the propagandist does
nothing to influence the scientific community; they only seek to shape policy maker beliefs.10
We analyze this model by varying parameter values including the number of scientists
in the network, K, the number of draws each round, n, the probability that action B is
successful, pB = .5 + , and the number of scientists each policy maker listens to, k.11 (Each
policy maker could connect to one scientist, or to two, or to three, etc.) We consider models
where the network of scientists is either a cycle, where each scientist is connected to two
peers in a ring structure, or a complete network, where every scientist communicates with
every other. Figure 1 shows these two structures. They differ in that it will tend to take
longer for the cycle to converge to a consensus, especially for large communities, because
fewer scientists share data with each other. This network structure affects our results, as
we will make clear, mainly by influencing the amount of time it takes the communities of
scientists to reach consensus. All reported results were averaged over 1k runs of simulation
model. On our view, there is ample and convincing historical evidence that the tobacco industry has
intervened to influence public beliefs using strategies very similar to the ones discussed here. The present
model supplements this historical evidence by allowing us to ask counterfactual questions concerning how
precisely this kind of intervention works. For this reason, we adopt the position that our model provides
"how-potentially" information concerning the causal role of this sort of intervention. What we mean by
this is that the model is realistic enough to explore processes that have the potential to really occur. In
this sense, we take this model to lie somewhere on a spectrum between "how-possibly" and "how-actually"
(i.e., fully realistic) models. Exploring how-potentially models can direct further inquiry into whether at
these processes are really at play, or (as in this case) supplement existing empirical work on these processes
(Rosenstock et al. [2017]). See also 4. We are grateful to an anonymous referee for pressing us to clarify
our stance on these issues, and for emphasizing that these models lie on a spectrum (a point with which we
agree).
9As we discuss below, this model involves some double-counting of results favorable to theory A, as a
given policy maker may receive the same evidence from both a scientist and from the propagandist.
10In this way, the model we consider differs significantly from that of Holman and Bruner [2015], who
consider an industry scientist who influences other scientific beliefs. We will return to this relationship in the
next section, as the treatment discussed there is somewhat closer to that discussed by Holman and Bruner.
11We consider subsets of the following parameter values: population size K = 4, 6, 10, 20, 50, number of
draws n = 1, 2, 5, 10, 20, 50, 100, 500, 1000, success of better theory pB = .51, .55, .575, .6, .7, and number of
scientists listened to by each policy maker ranges from 1 to all of them. Our choices were influenced, to
some degree, by tractability considerations; simulations with low pB and many scientists, in particular, took
longer to run.
6
that terminated when either correct or incorrect consensus was reached.12
Figure 1: Cycle and complete network structures.
We find that the presence of a single propagandist who communicates only actual findings
of (ideal, unbiased) epistemically motivated scientists can have a startling influence on the
beliefs of policy makers. No one in this scenario is doing bad science. No one is falsifying
data, or even publishing it selectively. The propagandist communicates only real scientific
findings. Nonetheless, under many scenarios we find that while the community of scientists
converges on true beliefs about the world, the policy makers reach near certainty in the false
belief. This point will become clearer as we share results in this section and the next -- for
a wide range of parameter values, public belief is highly accurate when no propagandist is
at play, but as inaccurate as it could possibly be under the influence of the propagandist.
We find that the likelihood of this sort of disagreement between science and public belief
depends on features of the community and the problem studied. First, unsurprisingly, when
policy makers listen to more scientists they are harder to deceive. Figure 2 shows the
average credence that policy makers have in the correct theory at the point where scientists
reach a correct consensus, which we understand as a state in which all members of the
scientific community have credence greater than .99. The x-axis tracks the proportion of
scientists in the entire community that each policy maker consults. As we can see, without
a propagandist, public beliefs lag behind scientific beliefs slightly, though the more scientists
are consulted, the better the public does. With a propagandist, the public is convinced that
the false theory is preferable, though if they are connected to more scientists, this effect is
ameliorated.13 In related work, Holman and Bruner [2015] model the effects of industrial
propagandists who share biased data with scientists and observe that it is best for scientists
to be highly connected to each other in such cases.14
12We ignore outcomes where scientists do not converge to the correct belief. (In these cases, policy makers
also tend to accept the worse theory.) This means that our results are averaged over fewer than 1k rounds.
In the worst case scenario, though, this only meant ignoring 10% of data points.
13These results are a cycle network with K = 20, n = 10, and pB = .55. This trend occurred across all
parameter values, as did the other trends reported unless otherwise specified.
14This shows how the suggestion from Zollman [2007, 2010] that less connected scientific communities are
more successful at reaching the truth does not hold when industry attempts to sway consensus.
7
Figure 2: Policy maker beliefs as influenced by their connections to scientists.
It is also harder to deceive policy makers when the size of the effect -- i.e., the difference
in expected payoff between action A and B -- is larger. As pB (the probability that the bet-
ter action pays off) increases, so that the difference between the better and worse theories
is more stark, the policy makers are more likely to learn the truth. This is because there
are fewer studies that happen to support the worse theory, and thus less material for the
propagandist to publicize. Figure 3 shows how as the difference between the success rates of
the two theories increases, policy makers tend to end up holding the true belief regardless of
propaganda.15 This observation is especially relevant to the biomedical sciences, because hu-
man bodies are highly variable, which means that data is often highly equivocal. We should
expect studies of human subjects to provide plenty of material for potential propagandists
looking to obscure true evidential trends.
The number of draws each round, n, also has a strong effect on the ability of the propa-
gandist to mislead. This parameter may be thought of as the sample size for each experiment
performed by the scientists. For fixed pB (i.e., fixed effect size), it is roughly analogous to
the statistical power of each scientist's experiments: the higher it is, the less likely that a
study will erroneously support theory A.16 Thus scientists who run the equivalent of well-
15Results are for K = 20 scientists in a cycle, each policy maker listening to 50% of scientists (k = 10),
and n = 10. Again the trend is stable across parameter values.
16More precisely, if we take the null hypothesis to be that that PB < PA, and the alternative hypothesis to
be that PB > PA, then for virtually any hypothesis test one considers, if we fix the effect size, = PB − .5,
increasing n will increase statistical power in the sense that it will increase the probability that the false null
hypothesis is rejected in favor of the true alternative hypothesis. Of course, our agents are Bayesians, and
so frequentist notions such as power are not strictly relevant -- the agents update their beliefs in light of any
and all evidence that they receive. Still, thinking of n as roughly tracking power is useful, here, because it
gives an intuitive sense of why small n studies are generally better for the propagandist.
8
Figure 3: Policy maker beliefs as influenced by the difficulty of the scientific problem.
powered studies and produce fewer spurious results also produce less for the propagandist
to work with. Figure 4 shows the final average beliefs of policy makers with and without a
propagandist as n increases.17 Note that the x-axis is on a log scale to make the trend more
visible. For high enough n, the propagandist can do almost nothing since they very rarely
have the spurious results necessary to mislead the public.
In these cases, policy makers
figure out the truth despite the propagandist's efforts.18
One might think that increasing the number of scientists in a network, K, holding all
else fixed, would decrease the effectiveness of the propagandist. More scientists means more
data, and more epistemically motivated minds searching for truth. But this does not occur
in general in these models. Assume that policy makers listen to some fixed number of
scientists, k. This is not an unreasonable assumption given limits on the amount of time
and energy policy makers or the public devote to science. In such cases, as the total number
of scientists increases, there is no benefit to policy makers, since they continue to listen to k
scientists. On the other hand, the increased number of scientists increases the total number
of spurious results that the propagandist can then share with policy makers, often worsening
policy maker belief.19 In the next section, we will engage in some further discussion of the
17This figure is for K = 10 scientists in a complete network, each policy maker listening to 90% of
scientists, and pB = .55.
18In an interesting connection, Holman and Bruner [2015] find that when a propagandist shares biased
data with scientists, and when these scientists are able to ignore data that is statistically unlikely given their
credences, scientists do better when n is large. This is for a different reason than we describe. When n is
large in their models, it is easier to detect a propagandist and ignore them. If n is small, more results are
statistically likely given some agent's credence.
19This is not always the case since as we outline in the Appendix the size of the community also influences
the speed of convergence to scientific consensus, which can influence policy maker beliefs.
9
Figure 4: Policy maker beliefs as influenced by the amount of data gathered by scientists.
effects that the size of the scientific community has on policy maker beliefs.
The final consideration that influences the success of the propagandist has to do with how
much time it takes for the scientific community to converge to consensus. The convergence
time depends on various factors. For instance, the size of a network generally influences
how long it takes for all scientists to reach agreement. (In a sparsely connected network,
more scientists means a longer time until each one is convinced.) In addition, as mentioned
above, for fixed population size the cycle and complete network arrangements differ in that
cycle networks take longer to converge. This is because scientists listen to fewer of their
colleagues, meaning that disagreements will tend to persist as actors in different parts of the
network gather different sorts of data.
Previous results from Zollman [2007, 2010] have shown that in many cases factors that
slow convergence can be beneficial for an epistemic community. These factors lead to tran-
sient diversity of opinion, which is important for preventing scientists from settling too
quickly on a theory that might be wrong. Heavily connected communities, on the other
hand, are easily swayed by strings of misleading evidence that are widely shared.20 But
our results show that other considerations can complicate this issue. When industrial actors
have a stake in a community's findings, taking a long time to settle on consensus can provide
more opportunity for the propagandist to find and promote results that went the wrong way
under conditions where spurious results are likely.
To try and better understand the relationship between time to convergence of the sci-
entific network and the effectiveness of the propagandist, we derive semi-analytic results
20Rosenstock et al. [2017] show that this will only occur for the region of parameter space in which the
problem of selecting the better theory is hard (due to there being fewer scientists, less difference between
theories, or lower n).
10
predicting expected outcomes in this model. In general, we can think of each policy maker's
belief as influenced by two separate factors. One comes from the community of scientists,
which, on average, will tend to pull credences in theory B up. The other comes from the
propagandist and will tend to pull credence in theory B down.
In other words, a policy
maker's belief in round t + 1 of simulation on average can be represented by:
B(t+1) = YP × YS × Bt,
where Bt is their prior belief, and YP and YS represent the average effects of the propagandist
and scientists. In the appendix, we derive expressions for these quantities and show that it
will always be the case that YP < 1 and YS > 1.
We can thus think of the dynamics in this model as a sort of tug of war between science
and industry over public belief. In general, for any set of parameter values in our model,
and some state of scientist/policy maker beliefs, one side will be winning because either
YP × YS < 1 or YP × YS > 1.21 This means that when the propagandist is winning on
average, they do better during long periods of scientific debate. In such cases, policy makers
do better when scientists quickly reach consensus. Conversely, when the YP × YS > 1
more evidence will tend to to lead the policy makers towards the true belief -- with the
propagandist's main effect being to slow this progress.22
For an example of the effects of time to convergence on belief, consider figure 5, which
shows the beliefs of policy makers consulting cycle and complete networks. We show results
in a scenario that favors true belief -- pB = .7 and n = 20 -- and also one that disfavors it --
pB = .55 and n = 5. In the first case, the slower cycle network gives policy makers time to
learn the correct theory from scientists, while relatively unencumbered by the propagandist.
In the second case, the faster complete network gives the propagandist less time to share
spurious data, and so is better for policy makers.
To be clear, if we look again at the simulation results we have presented, we can observe
that, since policy makers, on average, start simulation at credence .25, any outcome where
their final beliefs are < .25 is one where the propagandist is pulling harder than the scientists,
and vice versa. If we imagined that the process kept going -- scientists kept performing action
B, and the propagandist kept publicizing spurious results -- we would expect policy makers
to become completely convinced in either A or B eventually, with the relative values of YS
and YP determining which one. But the community can end up at a better or worse state
depending how quickly this happens. The most important take-away here is that there are
reasons to think that a long argumentative process between scientists might actually be
a bad thing when public belief is on the line. Thus we find a curious trade-off with the
Zollman effect: at least in some cases, maximizing convergence time is truth conducive for
the scientists, but not for the public.
21The situation in which the two influences exactly cancel is a knife's edge case.
22One interesting observation is that the size of the scientific network, besides influencing time to conver-
gence, also influences the rate at which policy makers change their beliefs. Adding more scientists increases
the total amount of data both coming directly from scientists and coming from the propagandist. When
the scientists are winning the tug of war, this increase benefits the policy makers. When the propagandist
is winning, this simply increases the rate at which policy makers decrease their credence in PB.
11
Figure 5: Influence of scientific network structure on policy maker beliefs.
In the next section, we will move on to discuss the second part of the Tobacco Strategy --
where industry funds their own scientists and disseminates the resulting research. Before
doing so, we mention a few potential concerns about the models just discussed, and robust-
ness checks meant to alleviate those concerns. First, as noted above, data in our model can
be "double counted" in the sense that a policy maker might update their belief on data from
a scientist and then update on the same evidence as it comes from the propagandist. We
take this to be a feature, rather than a bug, because it captures how the prevalence and
repetition of evidence can increase its impact. Nonetheless, we also ran the same simulations
where data could not be double counted this way. Results were qualitatively similar, though
the strength of the propagandist was diminished, particularly for policy makers connected to
many scientists. We also varied the pattern with which policy makers connected to scientists
to ensure that this did not have a significant effect on the results.
4. Biased Production
As mentioned in the introduction, another part of the Tobacco Strategy involved producing
industry-funded data that could be used to counter the research coming from the scientific
community.23 Our second treatment concerns this aspect of the strategy. To model this, we
23There is a close connection here to the work of Holman and Bruner [2015], who consider biased agents
in a scientific community. Like our "biased production" propagandist, their "intransigently biased agent"
(IBA) produces and shares evidence that tends to favor the wrong theory. However, in their case the IBA
draws from a biased arm relative to the other scientists in the community, whereas, as we describe, our
propagandist performs the same actions as the scientists but selectively publishes results. In addition, they
focus on the effect an IBA might have on beliefs within a scientific community, showing how industry can
shape scientific consensus itself, rather than public belief in particular.
12
now assume that the propagandist has their own resources that they can allocate towards
gathering data about the uncertain theory B. After gathering this data, the propagandist
then shares only those studies that spuriously indicated that theory A was the better one.
Notice that we do not allow the propagandist to do 'bad science' in the sense of falsifying
data, or to use biased methodologies. Their data gathering process is, in fact, identical
to scientists attempting to find the truth. The only intervention we allow on the scientific
process is one that is currently widely practiced (and not always considered particularly
unethical): selectively choosing not to publish some results.
We varied several new parameter values in this treatment. First, we varied the total
amount of resources available to the propagandist. (Think of this as the total number of
pulls from a bandit arm the propagandist can perform.) We also varied how the propagandist
chose to divide up these resources into studies. Did they run many low powered studies, or
just a few high powered ones?24
In this scenario, holding other parameters fixed, more scientists are unequivocally ben-
eficial to policy makers' beliefs. Scientists no longer unintentionally provide grist for the
propagandist, and, instead, produce and publicize data that is more likely to point in the
right direction.
It is also consistently better, as in the last section, for policy makers to
consult more scientists. The more data they receive from epistemically motivated actors,
the less the impact of the propagandists' spurious studies on their beliefs. And, again, the
better the successful theory, the lower the chance that the propagandist generates enough
spurious results to influence policy makers.
Unsurprisingly, more resources are better for the propagandist. If they have the where-
withal to run more studies, they have a greater chance of generating results supporting
theory A. Suppose each scientist has the resources to collect 10 data points per trial. As-
sume that the propagandist runs trials of the same size, but can run a different number of
these depending on their level of resources. As figure 6 shows, as this number increases, the
likelihood that the policy makers believe the correct theory goes down.25
Suppose that we hold fixed the resources a propagandist has at their disposal. The last
question we ask is: given this, what is the best way for the propagandist to allocate these
resources to different studies? The answer is that it is always better for them to break
their data up into the smallest possible chunks. Suppose the propagandist has resources
to collect 1000 data points. In the most extreme case, we could imagine a scenario where
the propagandist runs 1000 studies with one draw each. They could then report only those
where action B failed (on average this would be (.5−)∗1000 data points), making it look to
policy makers as if action B was never successful in these experiments. At the other extreme,
the propagandist would run a single experiment with 1000 data points. The chances that
24We ran simulations for a subset of the following parameter values: population size = 4, 6, 10, 20, number
of draws n = 1, 2, 10, 25, 50, 100, success of better theory pB = .55, .6, .7., and number of scientists consulted,
k, ranged from 1 to all of them. Propagandist total resources varied from equal to that of one scientist, to
equal to those of the whole scientific community. Propagandist n ranged between 1 and their total possible
pulls.
25These results are for 10 scientists, n = 10 for scientists, pB = .55, and policy makers listening to half of
scientists.
13
Figure 6: Propagandists with more resources are better able to mislead policy makers.
this experiment would support the better theory B are quite high, leaving the propagandist
with nothing to disseminate. In general, the propagandist should choose the smallest n they
can get away with.
Figure 7 shows this effect. For propagandists with resources to gather 50 data points,
If
the smaller the studies they run, the worse the eventual beliefs of the policy makers.
propagandists run a few, good studies, they are unable to mislead anyone.26
This observation actually has an analog in the previous set of results on propagandists
who selectively share others' data rather than gathering their own. In that model, given a set
of available resources for a community, it is better to use these to run fewer, better studies
than to run many less well-powered ones. We can observe this by comparing how policy
makers do given communities that have a finite amount of resources distributed differently.
Figure 8 demonstrates this. The scientific network, in this figure, has the resources to gather
a total of 100 data points each round. This is divided up by the number of scientists involved
in exploring the phenomenon so that, for example, 10 scientists gather 10 data points each
or two scientists gather 50 data points each. Without a propagandist policy makers develop
successful beliefs regardless of how resources are allocated. As the individual studies get
smaller, though, the propagandist is increasingly able to take advantage of the stochasticity
involved in gathering data to forward their agenda. Small studies mean confused policy
makers.
26These results are for 10 scientists, n = 10 (for scientists), policy makers listening to all scientists, and
pB = .55.
14
Figure 7: Propagandists who use their resources to perform small studies are better at misleading policy
makers.
5. Journalists as Unwitting Propagandists
In the two treatments we have discussed so far, the propagandist has adopted strategies
intended to influence the policy makers' beliefs. In both cases, we found that the propagan-
dist can succeed, at least under some circumstances. The basic mechanism at work in both
strategies is that the propagandist endeavors to bias the total body of evidence that the
policy makers see. They do not do this through fraud, or even by sampling from a different
distribution than the other scientists (as in Holman and Bruner [2015]). Instead, they do it
by choosing results so that the total evidence available, even though it was all drawn from
a fixed distribution, does not reflect the statistics of that distribution. What these results
show is that in many cases, particularly when evidence is relatively difficult to get or equiv-
ocal, it is essential that policy makers (and the public) have access to a complete, unbiased
sampling of studies. This suggests that curation, or even partial reporting, of evidence can
lead to epistemically poor results.
But propagandists are not the only agents who regularly curate scientific results. Non-
malicious actors can also play this role. For instance, journalists often select scientific studies
to feature in articles. Few journalists are experts in the relevant science, and in any case, the
goal is rarely to give a uniform and even-handed characterization of all available evidence.
To the contrary, journalistic practices often encourage, or even require, journalists to share
either only the most striking, surprising, or novel studies; or else, when they choose to
share a study on a controversial topic, to endeavor to be "fair" by sharing research that
supports both sides.
Indeed, the U.S. Federal Communications Commission (FCC) had
an official policy, from 1949 to 1987, that required broadcast news to present controversial
In practice, this meant giving roughly
topics of public interest in a "balanced" manner.
15
Figure 8: Policy makers do better when fixed resources are allocated to a few, good studies rather than
many with small samples.
equal time to multiple perspectives. The policy was known as the "Fairness Doctrine".
Oreskes and Conway [2010] argue that this Doctrine became a tool of the tobacco industry's
propagandists, who used it to insist that pro-industry science get publicized.
To study the effects of the Fairness Doctrine, we consider the following modification of
the models we have considered thus far.27 The network structures remain the same, but
in place of the propagandist, we have a "journalist", who like the propagandist can see all
of the scientists' results and can communicate with all of the policy makers. In this case,
we suppose that the policy makers receive all of their information from the journalist, and
consider different ways in which the journalist might share that data.
Suppose, for instance, that the journalist -- abiding by the Fairness Doctrine -- shares two
results from the network of scientists each round, one of which supports theory B, and one of
which supports theory A. More precisely: we suppose that each round, the journalist chooses
randomly among all of the results that support theory A and all that support theory B. In
rounds where no results support one of the two theories, the journalist shares the result they
shared the previous round (capturing the idea that sometimes journalists need to search past
data to find something to adequately represent the "other side"), unless there has never been
a result favoring a given theory, in which case the journalist shares only one result. What
we find is that this addition artificially slows the policy-makers' acceptance of theory B.
27We also considered a version of the model in which the journalist only published results that were
sufficiently "surprising" in the sense that the journalist judged them to have had a low probability of
occurring. For some parameter values, this sort of journalistic practice led policy makers to accept the
wrong theory. But we deemed the results to be ambiguous, because they depended sensitively on the value
of n chosen.
16
To show this, it is most informative to compare the policy makers' beliefs under the
action of this journalist (Fair) with what one would find if (a) the journalist shared two
randomly selected pieces of data each round (Random) and (b) the journalist simply shared
all available evidence each round (All). We find that (on average across runs) the Fairness
Doctrine never improves the epistemic performance of the policy makers, and for many
parameter values, policy makers receiving evidence from a journalist abiding by the Fairness
Doctrine end up with beliefs substantially less accurate than under conditions (a) or (b).
In all cases we consider, the policy makers' beliefs still tend towards the truth (whenever
the scientists' beliefs do), but they do so much more slowly.28 In this sense, then, the "fair"
journalist mimics the propagandist under the conditions where they are losing the tug-of-war
for policy makers' beliefs, but nonetheless slow the policy makers' progress.
Figure 9 shows average policy maker beliefs when scientists reach consensus on the true
belief for complete networks under different parameter values.29 Data points are for simula-
tions where journalists share Fair, Random, or All data as described above. The trend line
shows the average beliefs across parameters under each condition. As is evident, when jour-
nalists share all the data, policy maker beliefs are very accurate. With less data available --
two random sets of data -- their convergence to true beliefs is slowed. Under fair reporting,
policy makers converge to true beliefs much more slowly, since they see a result that supports
theory A every time they see a result supporting B. The strength of this effect is dependent
on parameter values, but, as mentioned, fair reporting only worsens, and never improves,
policy maker beliefs in comparison to reporting randomly or reporting all data.
Journalists are not the only actors who perform data curation. University faculty, when
teaching courses, may endeavor to give a panoptic and even-handed view of a field, but they
will not be able to share all available relevant data with a class. Literature reviews and
survey articles also curate data, commenting on its quality and selecting what seems most
relevant or significant. The mechanisms by which researchers and university instructors
select material to present are arguably motivated by epistemic considerations -- they seek
to give a compact overview that supports the views they believe are best supported, given
the total evidence that they have seen -- but it is hard to see how they can do this without
using their own beliefs -- which may be influenced by various factors -- to guide the choices.
And so, it is not clear that even this sort of curation is certain to accurately track the
statistics of the total available evidence. In this way, even independent researchers, in their
teaching and writing, can unwittingly act as propagandists. In addition, it has been widely
observed that scientists regularly fail to publish negative or null results. This 'file drawer'
effect biases the pool of evidence available to the entire community, and has already been
28When policy makers see one result supporting B and one supporting A each round, on average the result
supporting B is still stronger since this is in fact the better arm. So they learn the truth eventually.
29We chose a representative set of data points including all combinations of K = 6, 10, 20, pB = .55, .6, .7,
and n = 5, 10, 20. As in the previous models, initial policy maker beliefs were skeptical. Complete networks
demonstrated the effect more strongly because cycle networks are slow to converge and thus there is more
time for policy makers to learn true beliefs even under 'fair' reporting. But even with cycle networks the
effects of fair reporting were always neutral or negative.
17
Figure 9: 'Fair' reporting slows policy makers from reaching accurate beliefs.
shown to potentially have negative effects on scientific beliefs (Romero [2016]).30
6. Conclusion
We have now considered two aspects of the Tobacco Strategy as described by Oreskes and
Conway. We take the results we have presented to provide strong evidence that the Tobacco
Strategy can work, in the sense that biasing the total evidence presented to a group of agents
can lead those agents to converge to the incorrect belief. Of course, modeling alone cannot
establish that this did happen in the case of tobacco industry propaganda efforts. But when
you combine the modeling results we have presented with the careful historical analysis and
arguments offered by Oreskes and Conway [2010], there are strong reasons to believe that
the Tobacco Strategy was an important causal factor in the long delay between when a
link between cigarettes and cancer was established by the medical research community, and
large-scale changes in regulation and public attitudes. Moreover, we take it that if these
strategies have succeed in the past, they are very likely to succeed now and in the future.31
30See also Ioannidis [2005].
31Returning one more to the themes of notes 4 and 8: we emphasize that, at very least, these models
provide "how-possibly" information: the processes we decribe with these models can possibly occur. Fur-
thermore, as we argue here, the theoretical and historical knowledge developed by Oreskes and Conway
increases our confidence that these processes have been important ones in the real world. As noted, we
cannot be sure that this is the case. But we do not take this to be a special problem for models of this sort.
With almost any sort of scientific exploration, including experimental and other empirical work, conclusions
are drawn from data via theoretically supported inference. (See Longino and Doell [1983] for more on this
point.) In this case, we draw (conservative) inferences based on a combination of modeling and historical
work.
18
What is perhaps most interesting about the results we have presented is not that they
show what can work, but rather the insight they provide into how those strategies work.
As we have emphasized above, in our discussion on unwitting propagandists, the basic
mechanism at play in our models is that the propagandist biases the total evidence on
which the policy-makers update their beliefs. The fact that each result that is shared is, in
some sense, "real" ends up being irrelevant, because it is the statistical properties of the total
available evidence that matter. For this reason, we see that producing results favorable to
industry -- say, through industry funding -- is not necessary to manufacture favorable public
beliefs, at least in cases where the evidence is probabilistic and spurious results are possible.
On this point: one might have expected that actually producing biased science would
have a stronger influence on public opinion than merely sharing others' results. But when
one compares the two treatments we consider above, there are strong senses in which the less
invasive, more subtle strategy of selective sharing is more effective than biased production,
all things considered, particularly when the scientific community is large and the problem
at hand is difficult (or the power of generic experiments is low). The reason is that such a
scientific community will produce, on its own, plenty of research that, taken without proper
context, lends credence to falsehoods. Merely sharing this already-available evidence is cost-
effective, and much less risky than producing one's own research -- which, after all, can cost
a great deal to produce, fail to generate the desired results, and ultimately be disqualified
or ignored because of the association with industry. From this point of view, producing
industry science (or even outright fraud) is simply not worth it. In many cases, another
more effective, less expensive, and more difficult to detect strategy is available.
Indeed, if industrial interests do wish to spend their money on science, perhaps rather
than producing their own results and publishing them selectively, they would do better to
adopt a different strategy, which Holman and Bruner [2017] have called "industrial selec-
tion". The idea behind industrial selection is that there are many experimental protocols,
methodologies, and even research questions that scientists may adopt, often for their own
reasons. Some protocols may tend to produce more industry-friendly results than others.32
Industrial selection involves identifying methods already present among the community of
scientists that tend to favor one's preferred outcome, and then funding scientists who al-
ready use those methods. This extra funding works to amplify these scientists' results, by
allowing them to publish more, perform higher powered studies, and train more students
who will go on to produce consonant results. Indeed, there is a strong potential feedback
effect between industrial selection and selective sharing: by increasing the number of am-
bient industry-favorable results in a scientific literature, and then further amplifying those
results by sharing them selectively, propagandists can have an even stronger effect.
It is worth emphasizing that industrial selection does not require some scientists to have
adopted bad or unreliable research methods; it may also just be that they have focused on
32Holman and Bruner [2017] consider, for instance, the case of a class of drugs intended to stop heart
failure (Moore [1995]). In that case, studies that used reducing arrhythmia as a proxy for preventing heart
failure found that the drugs were highly successful, whereas studies that directly measured death rates from
heart failure later showed that the drugs were substantially increasing the likelihood of death.
Industry
swayed scientific practice by funding the former type of study, but not the latter.
19
research topics that promote an industrial agenda. For instance, Oreskes and Conway [2010,
p. 17] describe how the Tobacco Industry Research Council supported the work of noted
asbestos researcher Wilhelm C. Hueper of the National Cancer Institute. Hueper's research
was not methodologically flawed; to the contrary, it very convincingly established the (real)
link between lung cancer and asbestos. But for the tobacco industry, it was valuable to have
evidence that other environmental factors played a role in lung cancer, since that could be
used to obfuscate the role their own product played. Taken together, then, selective sharing
and industrial selection suggest those interested in dampening industry influence should
turn their attention to the more subtle ways propagandists can come to bias epistemic
communities. This will to a certain extent require philosophers of science to reorient their
focus, as extant work tends to address explicit, direct ways industry can impact inquiry.
It is natural to ask, given the past success of the Tobacco Strategy and the results we
have described, whether there are ways in which policy makers or members of the public
can protect themselves against a propagandist using the methods described. Of course,
one natural suggestion is to stop listening to propagandists. But this is difficult to do, in
large part because propagandists pose as (or, in some cases, are) credentialed scientists.33
Another possibility, perhaps more realistic, is to gather evidence from as many scientists as
possible: as we discuss above, policy makers connected to more scientists tend to do better,
because the propagandist has to produce more evidence to effectively influence such policy
makers' beliefs.34 Connecting to many scientists may also allow policy makers to to identify
propagandists on statistical grounds:
indeed, as Holman and Bruner [2015] point out, in
some cases it may be possible to identify biased agents by comparing the statistics of their
results with those of other agents in their network. Finally, and along similar lines, it might
also be possible, at least in some circumstances, for policy makers to adjust their updating
rule to accommodate the possibility that some of the evidence they receive is exaggerated,
filtered, or otherwise intended to mislead.35 On the other hand, as O'Connor and Weatherall
[2017] show, some apparently reasonable heuristics for updating belief in light of unreliable
sources can lead to bad epistemic outcomes, including stable polarization. We think that
it would be valuable, in future work, to study how these latter two reactions might be
implemented, and how and when they would be effective.
We will conclude by noting a few ways in which the incentive structure of science con-
tributes to the success of the would-be propagandist, particularly when employing the se-
lective sharing strategy. As we noted above, selective sharing is most effective when (a) the
problem is difficult (in the sense that pB is close to .5); (b) the number of data points per
scientist per round is small; (c) there are many scientists working on a given problem; and
33Indeed, perhaps the most striking message of Oreskes and Conway [2010] is that the most effective
propagandists have been prestigious scientists, often coming from different fields, with a political ax to
grind.
34We note that connecting to more scientists was particularly effective in the selective sharing treatment
when "double-counting" was eliminated, suggesting that policy makers would do well to attend to cases
where a small number of studies are used repeatedly to prove a point, when other studies are also available.
35We are grateful to David Manheim [2018] for emphasizing this latter possibility, which is also quite close
to what Holman and Bruner [2015] propose.
20
(d) policy makers are connected to few scientists. This situation interacts with how scien-
tists are rewarded for their work in complex, troubling ways.36 For instance, although there
are some incentives for scientists to make their work accessible and to try to communicate
it with the public and policy makers, it continues to be the case that academic scientists are
evaluated on the basis of their research productivity -- and there are sometimes reputational
hazards associated with "popularizing" too much. This sort of effect can limit the number
of scientists to whom policy makers are connected.
But a more troubling way in which propaganda and incentives intersect concerns the role
of low-powered studies. Scientists are strongly encouraged to publish research articles, and to
do so often. Moreover, publishing positive results is strongly preferred to null results. Finally,
scientists work with limited funds that they must use to meet many demands. All of these
considerations lead to strong incentives to run studies with relatively low power (particularly,
fewer data points) whenever possible (Smaldino and McElreath [2016]).37 Of course, most
scientific fields recognize that, all else being equal, higher powered studies are more rigorous,
more reliable, and make for all-around better science. But there is little incentive, and
often few resources, for individual scientists to produce consistently high quality studies.
And as we have seen, low powered studies are more likely to spuriously support erroneous
conclusions. In cases where industry is not involved, this is not as worrying: spurious results
can be detected via meta-analysis. But in the presence of a propagandist, they become a
powerful tool in shaping public belief.
Discussions of low-powered science often focus on their role in controversies about par-
ticular results, in the context of the so-called "replication crisis" in the behavioral sciences
(and especially psychology) (Open Science Collaboration [2015]). And the regular publica-
tion and acceptance of ultimately incorrect results is certainly worrying. But the arguments
and results we have given here suggest that even more is at stake: the same sorts of con-
ditions that led to a replication crisis in the first place also provide extra fodder for indus-
trial propaganda. This makes reshaping scientists' incentives to publish many, low-powered
"least-publishable units" crucial to foiling propagandists. For instance, scientists should be
granted more credit for statistically stronger results -- even in cases where they happen to be
null. Scientists should not be judged on the number of publications they produce, but on the
quality of the research in those articles, so that a single publication with a very high powered
study, even if it shows a null result, should be "worth" more than many articles showing
surprising and sexy results, but with low power. Similar reasoning suggests that, given some
fixed financial resources, funding bodies should allocate those resources to a few very high-
36Previous authors have assumed that scientists are motivated by a quest for credit, in the same way
'Credit economy' models have been used, for example, to show why
normal people seek money or goods.
scientists might divide labor (Kitcher [1990]) or commit fraud (Bright [2017]).
37The ubiquity of low-powered studies in some fields, such as neuroscience (Button et al. [2013]; Szucs
and Ioannidis [2017]), ecology (Lemoine et al. [2016]), and evolutionary biology (Hersch and Phillips [2004]),
for instance, has been remarked upon and widely discussed within those fields. Smaldino and McElreath
[2016] perform a meta-analysis of work in the behavioral sciences over the last 60 years and argue that
despite repeated calls to increase power, little change has occurred. In addition, they show why scientific
communities might move towards producing low quality, low effort studies.
21
powered studies, rather than splitting it up into many small grants with only enough funding
to run lower-powered studies. Or a more democratic alternative might involve scientists who
work together to generate and combine smaller studies before publishing, making it more
difficult for the propagandist to break off individual, spurious results.
Appendix A. Semi-analytic results
Here we describe a semi-analytic treatment of the expected progress of the models presented
in section 3.
Consider a population of K scientists each performing n trials each round, solving a
problem with pB = .5 + . Consider a policy maker connected to k scientists, with credence
Bt at time t in the true proposition that arm B is advantageous. This observer will update
her beliefs by conditionalization in light of two sources of evidence, from the scientists she
is connected to and from the propagandist. Thus, at the end of the round, we will have:
B(t+1) = YP × YS × Bt,
(A.1)
where YS represents the contribution from scientists and Yp represents the contribution from
the propagandist. What are these contributions on average?
In a generic round of play, the average result for each scientist performing the informative
action should be E = n∗pB, i.e., the expectation value for n draws from a binomial distribu-
tion with probability pB. We estimate Ys by supposing that the policy maker receives, from
each scientist performing action B that they are connected to, the average (i.e., expected)
result, and update. This yields:
(cid:18) P (n ∗ (.5 + )H)
P (n ∗ (.5 + ))
(cid:19)r∗k
YS =
(A.2)
where r ≤ 1 is the fraction of scientists in the population performing the informative action,
so r ∗ k ≤ k approximates the number of scientists connected to the observer who are
performing the informative action. Here H is the policy maker's hypothesis about the
success rate of arm B. When multiplied by the current belief, Bt, as in equation A.1, this
is just Bayes rule for the expected evidence. Notice that this equation does not account for
the fact that B will change as an agent updates on different sets of evidence within a time
step. We should think of this approximation as most useful for cases where B is changing
slowly.
One can show by direct computation that YS reduces to:
(cid:18)
YS =
1
(cid:19)r∗k
where
X =
Bt + (1 − Bt)X
(cid:18)1 − 2
(cid:19)2n
,
(A.3)
.
1 + 2
22
Observe that X < 1, and so for any value of Bt, YS > 1.
It follows that updating on
evidence from the scientists will, on average, increase credence that arm B is preferable.
Moreover, one can see that as n and each increase, X will decrease, and thus YS will
increase. Likewise, YS will increase as r, the number of scientists testing the preferred arm,
increases, and as k, the number of scientists a policy maker listens to, increases. Of course,
in the actual model stochastic effects can defeat this; we are describing average behavior.
Note that as Bt decreases, YS also becomes larger, so as observers become more skeptical,
evidence from scientists will become more powerful. Importantly this means that a set of
parameter values alone will not determine the average behavior of the model: the belief
state of a policy maker Bt and the belief states of the scientists, as represented by r, will
influence which way policy maker beliefs trend.
Now consider YP . Here we consider all (and only) spurious results produced by the
population of scientists each round. First, observe that we can calculate a rate at which
these will occur if all scientists perform the informative action, given by z = P (x < n/2)×K,
where x is the number of times arm B pays off. This expression is just the probability of
getting a spurious result -- one that suggests PB < .5 -- times the number of scientists in the
population. For instance, for even n, we have:
P (x < n/2) =
n!
i!(n − i)!
(PB)i(1 − PB)n−i
(A.4)
This is the sum of the probabilities of drawing each possible x < n/2 from a binomial
distribution with n draws and PB.
We can also calculate an expectation value for such spurious draws, E, which is given
by (again for even n):
E =
1
P (x < n/2)
n/2−1(cid:88)
i=0
in!
i!(n − i)!
(PB)i(1 − PB)n−i
(A.5)
This equation multiplies the value of each possible spurious draw, i, by the probability
of that draw, conditional on the assumption that the draw will be spurious.
This leads to an estimate for YP of:
YP =
Bt + (1 − Bt) X
,
(A.6)
where now
X =
Once again, r is the fraction of the population of scientists performing the informative action,
while z, remember, is the rate of spurious results. Observe that, by definition E < n/2, and
so whatever else is the case (2 E − n) < 1, so X > 1, which means YP is always less than 1.
Hence the propagandist will always tend to decrease the observer's confidence in the true
23
n/2−1(cid:88)
i=0
(cid:18)
(cid:19)r∗z
(cid:19)2 E−n
1
(cid:18)1 − 2
1 + 2
proposition, and this effect will be stronger when more scientists take the informative action
and the rate of spurious results is higher (more on this shortly).
Beyond these observations, it is harder to analyze the dependence of YP on other pa-
rameters, on account of the complexity of the function E. An analytic simplification of
the expression for E is apparently not available, but a numerical analysis reveals that for
small , E is very well approximated for sufficiently large n by a linear function in n, of
the form E = −.62 + 0.47n. (This expression is determined by taking a linear best fit in
Mathematica.) Thus we have:
(cid:18)1 − 2
(cid:19)−.06n−1.24
X =
1 + 2
(A.7)
We see that as becomes larger, X becomes smaller (i.e., approaches 1), and thus YP ap-
proaches 1, so that the propagandist is less effective. Notice though that as n increases in
this equation, X becomes larger, thus YP gets smaller, so the propagandist gets more effec-
tive. Thus n influences YP in two, countervailing ways. A large n makes z, the probability
of a spurious result, smaller. On the other hand, when spurious results do occur, they are
In addition, X tends to grow slowly in n and whether the
more convincing for large n.
propagandist ends up swaying the observer depends on the relative scaling of YP and YS
with n, which in turn depends on . We also have that, as Bt approaches 1, YP approaches
1, so that the propagandist becomes less effective. (Thus, both the propagandist and the
scientists are more effective when policy makers have low credence in arm B.)
A final consideration concerns z, the rate of spurious results. Clearly as z increases,
the propagandist becomes more effective. This will happen as K, the number of scientists,
increases. (Though if policy makers listen to all the scientists, k will increase as K does,
strengthening YS as well.) The rate z will also increase as r, the fraction of scientists
performing the informative action, increases, though an increase in r will increase both YS
and YP . There is also a dependence in z on P (x < n/2).
It is difficult to get a good
first order approximation for this function, but by studying plots one can characterize some
qualitative features. In general, for fixed n, increasing decreases P (x < n/2), and thus
decreases z, making the propagandist less successful.
For sufficiently large n, we find that as n increases, P (x < n/2) decreases, again making
the propagandist less effective. This is what one would expect: the probability of spurious
results decreases as n increases. But what is surprising is that for small n, there is a regime
in which increasing n actually increases P (x < n/2). For sufficiently large (say > .1 or so),
this regime is a small part of n-space. But for small (say, ≈ .001), P (x < n/2) actually
increases with n for all values of n that we study. So increasing n under these circumstances
makes the propagandist more effective. Again, an increase in n also makes scientists more
effective, so the overall influence of n will depend on whether it increases YP or YS to a
greater degree.
24
Acknowledgments
This material is based upon work supported by the National Science Foundation under
grant number 1535139 (O'Connor). The majority of the work on this project was completed
while O'Connor and Weatherall were Visiting Fellows at the Australian National University;
we are grateful to the Research School for Social Science and the School of Philosophy for
support. We are also grateful to audiences at the ANU, University of Groningen, UC Irvine,
UC Merced, UC San Diego, and the 2018 PPE Society Meeting for helpful feedback on talks
related to this work, and to Jeff Barrett, Craig Callender, Bennett Holman, David Manheim,
Aydin Mohseni, Mike Schneider, and Chris Smeenk for helpful discussions.
References
Bala, V. and Goyal, S. [1998]: 'Learning from neighbors', Review of Economic Studies, 65(3), pp. 595 -- 621.
Bright, L. K. [2017]: 'On fraud', Philosophical Studies, 174(2), pp. 291 -- 310.
Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J. and Munaf`o,
M. R. [2013]: 'Power failure: why small sample size undermines the reliability of neuroscience', Nature
Reviews Neuroscience, 14(5), pp. 365 -- 376.
Douglas, H. [2009]: Science, policy, and the value-free ideal , Pittsburgh, PA: University of Pittsburgh Press.
Elliott, K. C. [2011]: Is a Little Pollution Good For You? , Oxford, UK: Oxford University Press.
Frey, D. and Seselja, D. [2018]: 'What Is the Epistemic Function of Highly Idealized Agent-Based Models
of Scientific Inquiry?', Philosophy of the Social Sciences.
Hersch, E. I. and Phillips, P. C. [2004]:
'Power and potential bias in field studies of natural selection',
Evolution; International Journal of Organic Evolution, 58(3), pp. 479 -- 485.
Holman, B. and Bruner, J. [2017]: 'Experimentation by industrial selection', Forthcoming in Philosophy of
Science.
Holman, B. and Bruner, J. P. [2015]: 'The problem of intransigently biased agents', Philosophy of Science,
82(5), pp. 956 -- 968.
Ioannidis, J. P. [2005]: 'Why most published research findings are false', PLoS medicine, 2(8), pp. e124.
Kitcher, P. [1990]: 'The division of cognitive labor', The journal of philosophy, 87(1), pp. 5 -- 22.
Kitcher, P. [2003]: Science, truth, and democracy, New York, NY: Oxford University Press.
Kitcher, P. [2011]: Science in a democratic society, Amherst, NY: Prometheus Books.
Lemoine, N. P., Hoffman, A., Felton, A. J., Baur, L., Chaves, F., Gray, J., Yu, Q. and Smith, M. D. [2016]:
'Underappreciated problems of low replication in ecological field studies', Ecology, 97(10), pp. 2554 -- 2561.
'Body, bias, and behavior: A comparative analysis of reasoning in two
Longino, H. and Doell, R. [1983]:
areas of biological science', Signs: Journal of Women in Culture and Society, 9(2), pp. 206 -- 227.
Longino, H. E. [1990]: Science as social knowledge: Values and objectivity in scientific inquiry, Princeton,
NJ: Princeton University Press.
Longino, H. E. [2002]: The fate of knowledge, Princeton, NJ: Princeton University Press.
Manheim, D.
"How to Beat
'Comment
[2018]:
on
Science
and
Influence People"',
Https://medium.com/@davidmanheim/comment-on-how-to-beat-science-and-influence-people-https-
arxiv-org-abs-1801-01239-5e96f0ad49bd.
Michaels, D. [2008]: Doubt is Their Product, Oxford, UK: Oxford University Press.
Michaels, D. and Monforton, C. [2005]: 'Manufacturing uncertainty: contested science and the protection
of the public's health and environment', American journal of public health, 95(S1), pp. S39 -- S48.
Moore, T. J. [1995]: Deadly medicine: why tens of thousands of heart patients died in America's worst drug
disaster , New York, NY: Simon & Schuster.
Norr, R. [1952]: 'Cancer by the carton', Reader's Digest, 61(368), pp. 7 -- 8.
O'Connor, C. and Weatherall, J. O. [2017]: 'Scientific Polarization', arXiv:1712.04561 [cs.SI].
25
O'Connor, C. and Weatherall, J. O. [2019]: The Misinformation Age: How False Beliefs Spread , New Haven,
CT: Yale University Press.
Open Science Collaboration [2015]:
'Estimating the reproducibility of psychological science', Science,
349(6251), pp. aac4716.
¡http://science.sciencemag.org/content/349/6251/aac4716¿
Oreskes, N. and Conway, E. M. [2010]: Merchants of Doubt, New York, NY: Bloomsbury Press.
Romero, F. [2016]: 'Can the behavioral sciences self-correct? A social epistemic study', Studies in History
and Philosophy of Science Part A, 60, pp. 55 -- 69.
Rosenstock, S., Bruner, J. and O'Connor, C. [2017]: 'In Epistemic Networks, Is Less Really More?', Philos-
ophy of Science, 84(2), pp. 234 -- 252.
Smaldino, P. E. and McElreath, R. [2016]: 'The natural selection of bad science', Open Science, 3(9), pp.
160384.
Szucs, D. and Ioannidis, J. P. A. [2017]: 'Empirical assessment of published effect sizes and power in the
recent cognitive neuroscience and psychology literature', PLOS Biology, 15(3), pp. e2000797.
¡http://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.2000797¿
Weatherall, J. O. and O'Connor, C.
[2018]:
'Conformity in Scientific Networks',
arXiv:1803.09905
[physics.soc-ph].
Woodward, J. [2003]: Making things happen: A theory of causal explanation, Oxford university press.
Wynder, E. L., Graham, E. A. and Croninger, A. B. [1953]: 'Experimental production of carcinoma with
cigarette tar', Cancer Research, 13(12), pp. 855 -- 864.
Zollman, K. J. [2007]:
74(5), pp. 574 -- 587.
'The communication structure of epistemic communities', Philosophy of science,
Zollman, K. J. [2010]: 'The epistemic benefit of transient diversity', Erkenntnis, 72(1), pp. 17.
26
|
1607.08673 | 2 | 1607 | 2016-10-29T23:29:48 | Measuring and Modeling Bipartite Graphs with Community Structure | [
"cs.SI",
"physics.soc-ph"
] | Network science is a powerful tool for analyzing complex systems in fields ranging from sociology to engineering to biology. This paper is focused on generative models of large-scale bipartite graphs, also known as two-way graphs or two-mode networks. We propose two generative models that can be easily tuned to reproduce the characteristics of real-world networks, not just qualitatively, but quantitatively. The characteristics we consider are the degree distributions and the metamorphosis coefficient. The metamorphosis coefficient, a bipartite analogue of the clustering coefficient, is the proportion of length-three paths that participate in length-four cycles. Having a high metamorphosis coefficient is a necessary condition for close-knit community structure. We define edge, node, and degreewise metamorphosis coefficients, enabling a more detailed understanding of the bipartite connectivity that is not explained by degree distribution alone. Our first model, bipartite Chung-Lu (CL), is able to reproduce real-world degree distributions, and our second model, bipartite block two-level Erd\"os-R\'enyi (BTER), reproduces both the degree distributions as well as the degreewise metamorphosis coefficients. We demonstrate the effectiveness of these models on several real-world data sets. | cs.SI | cs |
1 of 23
Measuring and Modeling Bipartite Graphs with Community Structure
Department of Mathematics, University of California, San Diego, CA
SINAN AKSOY
Sandia National Laboratories, Livermore, CA. Corresponding author: [email protected]
TAMARA G. KOLDA
AND
ALI PINAR
Sandia National Laboratories, Livermore, CA
[Received on 1 November 2016]
Network science is a powerful tool for analyzing complex systems in fields ranging from sociology to
engineering to biology. This paper is focused on generative models of large-scale bipartite graphs, also
known as two-way graphs or two-mode networks. We propose two generative models that can be easily
tuned to reproduce the characteristics of real-world networks, not just qualitatively, but quantitatively.
The characteristics we consider are the degree distributions and the metamorphosis coefficient. The
metamorphosis coefficient, a bipartite analogue of the clustering coefficient, is the proportion of length-
three paths that participate in length-four cycles. Having a high metamorphosis coefficient is a necessary
condition for close-knit community structure. We define edge, node, and degreewise metamorphosis
coefficients, enabling a more detailed understanding of the bipartite connectivity that is not explained
by degree distribution alone. Our first model, bipartite Chung-Lu (CL), is able to reproduce real-world
degree distributions, and our second model, bipartite block two-level Erdos-R´enyi (BTER), reproduces
both the degree distributions as well as the degreewise metamorphosis coefficients. We demonstrate the
effectiveness of these models on several real-world data sets.
Keywords: bipartite generative graph model, two-way graph model, two-mode network, metamorphosis
coefficient, bipartite clustering coefficient, complex networks
1. Introduction
Network science is a powerful tool for analyzing complex systems in fields ranging from sociology to
engineering to biology. The ability to develop realistic models of the networks is needed for several
reasons. Pragmatically, we need to generate artificial networks to facilitate sharing of realistic network
data while respecting concerns about privacy and security of data. More generally, generative models
enable unlimited network data generation for computational analysis, e.g., varying the characteristics
of the graph to test a graph algorithm under different scenarios. Ultimately, we hope to understand
the underlying nature of complex systems, and modeling them mathematically is a way to test our
understanding.
This paper is focused on generative models of bipartite graphs, also known as two-way graphs or
two-mode networks. Many real-world systems are naturally expressed as bipartite graphs. The defining
characteristic of a bipartite graph is that its vertices are divided into two partitions, U and V , such that
2 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
edges, E, only connect vertices across the two partitions, i.e.,
G = (U,V,E) with U ∩V = /0
and E ⊆ U ⊗V.
Examples of bipartite graphs include author-paper networks, user-product purchase histories, user-song
play lists, actor-movie connections, document-keyword mappings, and so on. Hypergraphs can be rep-
resented as bipartite graphs in the sense of an incidence graph: the nodes and hyperedges are represented
by U and V , respectively, and edge (i, j) exists if node i is in hyperedge j. Bipartite graphs have been
widely studied; see [9, 13, 14, 20, 33, 36] and references therein. An example bipartite graph is shown
in Fig. 1.
u1
u2
u3
v1
v2
v3
v4
FIG. 1: Bipartite graph
We propose a generative model that can be easily tuned to reproduce the characteristics of real-world
networks, not just qualitatively, but quantitatively. The measurements we consider in this paper are the
degree distributions and the bipartite analog of the clustering coefficient. Of course, there are many
other measurements that we could consider, and we come back to those shortly. The degree distribution
is the number (or proportion) of nodes of degree d for each d = 1,2, . . .. A bipartite graph has two
degree distributions, one for each vertex partition. As we see in the results in Section 4.3, these two
distributions may be quite different from one another, in part because the size (and, consequently, the
average degree) for each partition may be quite different. The clustering coefficient of a one-way graph,
introduced by Holland and Leinhardt [16], is the probability that a two-path (or wedge) participates in a
three-cycle (or triangle); i.e.,
c =
3n(cid:52)
n∧ =
3× (total number of triangles)
total number of wedges
.
One characteristic of a bipartite graph is that is has no odd-length cycles; hence, it cannot have a triangle.
Robins and Alexander [33] propose a bipartite clustering coefficient that is the probability that a bipartite
three-path (or caterpillar) participates in a bipartite four-cycle (or butterfly), i.e.,
c =
4n
n =
4× (total number of butterflies)
total number of caterpillars
.
(1.1)
We call c in (1.1) the metamorphosis coefficient in reference to the notion of how often caterpillars
become butterflies. Opsahl [30] says "it could be considered a measure of reinforcement between two
individuals rather than clustering of a group of individuals." The multiplier of four in the numerator is
because every butterfly contains four distinct caterpillars, just as every triangle contains three distinct
wedges; see Fig. 2. A high metamorphosis coefficient in a bipartite graph is indicative of greater com-
munity structure in the graph, analogous to the role of a high clustering coefficient in a one-way graph.
MEASURING AND MODELING BIPARTITE GRAPHS
3 of 23
The metamorphosis coefficient plays an important role in communities because any tight-knit commu-
nity will have a high metamorphosis coefficient. In Section 5.1, we extend (1.1) to define edge, node,
and degreewise metamorphosis coefficients, enabling a more detailed understanding of the bipartite
community structure.
(a) Caterpillars
(b) Butterfly
FIG. 2: A butterfly contains four distinct caterpillars
As mentioned above, there are numerous more complex metrics that can be used in the analysis of
bipartite graphs; see, e.g., [20, 33]. Examples of such metrics include graph diameter (i.e., the maxi-
mum distance between any pair of nodes), singular values of the adjacency matrix, centrality measures,
joint degree distributions (i.e., the proportion of nodes of degree d that connect to nodes of degree d(cid:48)
for all possible pairs (d,d(cid:48))), assortativity (i.e., degree correlations), subgraph census (i.e., counts of all
patterns of three nodes, four nodes, and so on). Moreover, we can consider the analysis of the product
graphs that are produced by considering just one vertex partition and connecting any two nodes that
share a common neighbor in the bipartite graph. Because these quantities are often expensive to com-
pute for large graphs, we do not consider them in this paper. Instead, we consider the simpler metrics
of degree distribution and metamorphosis coefficients here for reasons analogous to those considered
by [20, 35]. The degree distributions are interesting because many large real-world graphs (including
the ones we consider here) do not have the Poisson distributions that occur if edges are distributed at
random; instead, they have heavy-tailed distributions (hypothesized to be power law [2] or log-normal
[18]). The metamorphosis coefficient is interesting because random graphs that have the same degree
distribution as a real-world graph tend to have a much lower metamorphosis coefficients, as we show
in Section 4.4. This is a well-known phenomena for the clustering coefficient of one-way graphs [28].
We contend that matching metamorphosis coefficients is critical for capturing social or, more generally,
interconnectedness behavior in networks.
With the goals of capturing both degree distribution and our newly-defined degreewise metamorpho-
sis coefficients, we develop two different models. The first is a straightforward extension of Chung-Lu
(CL) [1, 6, 7], which is very closely related to the configuration model. Our experimental results show
that this model is effective at reproducing the degree distributions. However, the bipartite CL graphs do
not produce the same metamorphosis coefficients as observed in real-world networks. Therefore, the
second model we propose is a bipartite extension of the Block Two-Level Erdos-R´enyi (BTER) [18, 35]
model. The BTER model is a good starting point because it reproduces both the degree distribution and
degreewise clustering coefficients of a given network. To do so, it groups nodes into Erdos-R´enyi (ER)
subgraphs, called affinity blocks, that are highly connected and so produce high clustering coefficients.
We propose a bipartite BTER that reproduces both the degree distributions and the degreewise metamor-
phosis coefficients. This extension is not straightforward since the affinity block concept does not carry
over easily to the bipartite case, so we develop a new method for creating the blocks. Computational
results for bipartite BTER show that it achieves our goals of matching both the degree distributions and
the degreewise metamorphosis coefficients.
4 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
Before we continue, we briefly survey related work. As mentioned above, the proposed bipartite
CL model is very similar to bipartite configuration models. Both Newman, Strogatz, and Watts [29]
and Guillaume and Latapy [14] propose bipartite configuration models that sample node degrees from
a distribution (one per partition), create stubs for each node equal to its degree, and then match the
stubs. Our bipartite CL model is very similar, and we discuss the connections further in Section 4.2.
On the theoretical side, Kannan, Tetali, and Vempala [17] consider the convergence of a Markov chain
rewiring algorithm for generating a bipartite graph and show that their process is rapidly mixing for
regular bipartite graphs, i.e., graphs where all the node degrees are the same. Later, this result was
extended to graphs in which only one partition is required to be regular [22]. In statistics, researchers
consider the problem of generating binary contingency tables with given row and columns sums. This
problem is equivalent to specifying the degrees in a bipartite graph. In [4, 5], the focus is on an algorithm
that is guaranteed to produce a specified row and column sums, assuming its realizable. The proposed
algorithm is not considered practical but rather of theoretical interest. As for one-way graphs, there are
also incremental growth models in which the output graph is iteratively constructed by adding vertices
and edges according to some rule. The methods are sequential since the connections created for new
nodes usually depend on the state of the graph at the iteration in which the nodes are created. Guillaume
and Latapy [14] analyze a bipartite growing model in which new vertices are linked via a preferential
attachment process. Other work has considered a similar idea except that one vertex partition remains
fixed while the other grows [12, 32]. Lastly, we note a variety of application-specific generative bipar-
tite graph models have been introduced to model specific networks, including pollination networks in
ecology [10, 34], and protein-domain networks in biology [26].
2. Data sets
We test our methods on publicly-available real-world datasets, whose properties are summarized in
Table 1. Their degree distributions are shown in Figs. 3 to 5. CondMat represents an author-paper
network from arXiv preprints in condensed matter physics from 1995–99 [27]; this has mostly been used
in the context of the coauthorship graph, but here we consider the underlying data [31]. The majority
of authors have only a single paper, whereas the most prolific author has 116 papers. Conversely, the
most coauthors on a single paper is 18, and the most likely scenario is for a paper to have 2 or 3 authors.
IMDB links movies and the actors that appeared in them [19, 20], as collected from the Internet Movie
Database. The busiest actor was in 294 movies; conversely, the largest production had 646 actors. Flickr
[24, 25] is an online photo-sharing site, and the network represents group membership of various users.
The most connected user is in 2186 groups, whereas the largest group has 34989 members. MovieLens
[8, 15] is a very famous dataset that links movies and their reviewers/critics. The most-reviewed movie
had 34864 reviews, and the most active critic reviewed 7359 movies (orange). This dataset apparently
excludes critics with less than 20 reviews. The MillionSong [3, 21, 39] dataset connects users and the
songs they played. The dataset only includes listeners of ten songs or more. The widest-ranging user
listened to 4,400 distinct songs. The top song was played by 110,479 distinct listeners. The Peer2Peer
dataset [19, 20] links users (peers) and the files they uploaded or downloaded. The busiest user touched
19,496 files. On the other hand, the most popular file only had 3396 downloads. LiveJournal [24, 25]
represents user-group memberships from a blogging site. The most engaged user is in 300 groups, which
appears to be the maximum allowed, since there are five persons in that category. The largest group has
over one million members.
MEASURING AND MODELING BIPARTITE GRAPHS
5 of 23
Table 1: Real-world bipartite graphs
Name
CondMat [27, 31]
IMDB [19, 20]
Flickr [24, 25]
MovieLens [8, 15]
MillionSong [3, 21, 39]
Peer2Peer [19, 20]
LiveJournal [24, 25]
Partition 1
16,726 authors
127,823 actors
1,728,701 users
65,133 movies
1,019,318 users
1,986,588 peers
5,284,451 users
Partition 2
22,016 papers
383,640 movies
103,648 groups
71,567 critics
384,546 songs
5,380,546 files
7,489,296 groups
Edges
58,595
1,470,418
8,545,307
10,000,054
48,373,586
55,829,392
112,307,385
3. Notation
We set up the notation for one-way and two-way graphs. We assume all graphs are simple, meaning that
there are no multiple edges. In the one-way case, we let n = V and index nodes in V by i, j ∈ {1, . . . ,n}.
In the two-way case, we let nu = U and nv = V denote the sizes of partitions one (left) and two
(right), respectively. We use i ∈ {1, . . . ,nu} and j ∈ {1, . . . ,nv} to index nodes in partition one and
two, respectively. Without loss of generality, indexing by i assumes partition one and likewise for j and
partition two. For instance, if i = j = 2 in the two-way case, although vertices i and j have the same
vertex label of 2, they belong to different partitions and are thus two distinct vertices.
One-Way Graph
G = (V,E)
Vertices : V
Edges : E ⊆ V ⊗V
# Vertices : n = V
# Edges : m = E
# Wedges : n∧
# Triangles : n(cid:52)
Clust. Coeff. : c = 3n(cid:52)/n∧
Vertex Index : i ∈ {1, . . . ,n}
Degree of i : di = { j ∈ V (i, j) ∈ E }
Table 2: Notation
Two-Way Graph
G = (U,V,E)
Vtx. Partition 1 : U
Vtx. Partition 2 : V
Edges : E ⊆ U ⊗V
# Vertices in U : nu = U
# Vertices in V : nv = V
# Edges : m = E
# Caterpillars : n
# Butterflies : n
Meta. Coeff. : c = 4n /n
Index in U : i ∈ {1, . . . ,nu}
Index in V : j ∈ {1, . . . ,nv}
Degree of i : du
Degree of j : dv
i = { j ∈ V (i, j) ∈ E }
j = {i ∈ U (i, j) ∈ E }
4. Fast Bipartite Chung-Lu Model
We adapt the Chung-Lu generative model [1, 6, 7] to bipartite graphs and demonstrate its ability to
reproduce bipartite degree distributions. We follow the notation described in Section 3.
6 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
4.1 Chung-Lu for One-way Graphs
Consider a one-way graph G = (V,E). The Chung-Lu (CL) model attempts to match the desired degrees
{d1, . . . ,dn} where di denotes the desired degree of vertex i. The model generates a random graph on n
vertices such that the probability that vertex i is adjacent to j is given by
Pr((i, j) ∈ E) =
did j
2m
n
∑
1
2
where m =
di = desired number of edges.
tation of the CL model on n vertices flips a coin for each of the(cid:0)n
To ensure the quantities are all true probabilities, we assume di (cid:54) √
i=1
(cid:1) = Ω (n2) possible edges. Many
2m for all i. A classical implemen-
2
real-world graphs are large and sparse, i.e., the number of edges m = O(n). For this reason, we favor
"fast" CL that flips only 2m coins [11, 18]. We explain the fast method below in the context of two-way
graphs.
4.2 Chung-Lu for Two-way Graphs
Consider the bipartite graph G = (U,V,E). Here we have separate desired degrees for the vertices in U
and V , denoted
{du
i }nu
i=1
and
{dv
j }nv
j=1 ,
respectively. Necessarily, the sums of the degrees in each partition must be equal to each other and to
the number of edges, i.e.,
m =
du
i =
nu
∑
i=1
nv
∑
j=1
dv
j .
Hence, the bipartite CL model generates a random bipartite graph on nu vertices in the first partition and
nv vertices in the second partition such that
i dv
du
Pr((i, j) ∈ E) =
j
m .
i (cid:54) √
m for all i and dv
j (cid:54) √
(4.1)
To ensure these are true probabilities, we assume du
m for all j. A naıve
implementation of bipartite Chung-Lu would flip a coin for all nunv possible edges. Instead, we adopt
the "fast" approach as follows. Rather than flipping a coin for every possible edge, we instead randomly
choose two endpoints for every expected edge. Since the graph is sparse, we may assume that m (cid:28) nunv,
so this approach requires many fewer random samples. For each of the m edges, we choose endpoints
in U and V proportional to
dv
j
m ,
respectively. The probability that edge (i, j) exists is then given by
and Pr( j) =
Pr(i) =
du
i
m
Pr((i, j) ∈ E) = m· Pr(i)· Pr( j) =
du
i dv
j
m ,
which is the same as (4.1). Although both implementations of CL yield identical expected degrees, a key
distinction is that multiple edges between the same pair of vertices are possible in fast CL. In practice,
for large graphs with heavy-tailed degree distributions, this rarely presents a problem. The fast bipartite
CL algorithm is described in Algorithm 1.
MEASURING AND MODELING BIPARTITE GRAPHS
7 of 23
Algorithm 1 Fast Bipartite CL
i }, {dv
j })
i
m ← ∑i du
E ← /0
for k = 1, . . . ,m do
1: procedure FBCL({du
2:
3:
4:
5:
6:
7:
8:
9: return E
10: end procedure
end for
Randomly select i ∈ U proportional to du
Randomly select j ∈ V proportional to dv
E ← E ∪ (i, j)
i /m
j /m
(cid:46) Duplicate edges discarded
As mentioned in the introduction, a closely-related approach is the bipartite configuration model
[14, 29]. These differ in some details; for instance, in the CL model, the degree distribution is not
specified exactly but rather each degree is sampled from a distribution. However, if we ignore that
detail, we may consider that each node i ∈ U has du
j stubs,
and the stubs from partition one are randomly connected to the stubs in partition two. As with fast
bipartite CL, we discard any repeated edges. In all three cases, bipartite CL, fast bipartite CL, and the
bipartite configuration models, the expected degree of a vertex is the same.
i stubs and likewise each node j ∈ V has dv
4.3 Experimental Results
We generate random graphs using CL and the degree distributions of the graphs described in Section 2.
The degree distributions are shown in Figs. 3 to 5. The degree distribution of the original graph is shown
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
103
101
10−1
Original
CL
100
101
Degree
103
101
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
10−1
100
Original
CL
101
Degree
(a) Mode 1: CondMat (Authors/Papers)
(b) Mode 2: CondMat (Authors/Papers)
FIG. 3: Mode 2: CondMat (Authors/Papers)
in blue, and the degree distribution of the graph generated by bipartite CL is shown in orange. These
are binned degree distributions, as advocated in [23]. We use powers of two for the bin borders, so the
x-coordinate 2k corresponds to the bin from [2k,2k+1). The y-coordinate is the average value for that
bin, including zero values, so the y-coordinate can be less than one. Overall, the degree distributions are
very close, especially for IMDB (Figs. 4a and 4b), Flickr (Figs. 4c and 4d), and LiveJournal (Figs. 5e
8 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
103
100
Original
CL
10−3
100
101
Degree
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
101
10−3
Original
CL
102
100
101
102
Degree
(a) Mode 1: IMDB (Actors/Movies)
(b) Mode 2: IMDB (Actors/Movies)
105
101
10−3
102
10−1
10−4
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
100
10−5
Original
CL
Original
CL
100
101
102
Degree
103
100
101
103
104
102
Degree
(c) Mode 1: Flickr (Users/Groups)
(d) Mode 2: Flickr (Users/Groups)
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
103
100
10−3
Original
CL
101
102
Degree
103
Original
CL
100
101
103
104
102
Degree
(e) Mode 1: MovieLens (Movies/Critics)
(f) Mode 2: MovieLens (Movies/Critics)
FIG. 4: Mode 2: MovieLens (Movies/Critics)
MEASURING AND MODELING BIPARTITE GRAPHS
9 of 23
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
104
100
Original
CL
101
10−4
100
105
100
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
Original
CL
102
Degree
103
100
101
103
102
Degree
104
105
(a) Mode 1: MillionSong (Users/Songs)
(b) Mode 2: MillionSong (Users/Songs)
105
100
Original
CL
101
10−5
100
103
104
102
Degree
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
101
10−3
Original
CL
100
101
102
Degree
103
(c) Mode 1: Peer2Peer (Peers/Files)
(d) Mode 2: Peer2Peer (Peers/Files)
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
104
102
100
Original
CL
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
107
100
10−7
Original
CL
100
101
Degree
102
100
102
104
106
Degree
(e) Mode 1: LiveJournal (Users/Groups)
(f) Mode 2: LiveJournal (Users/Groups)
FIG. 5: Mode 2: LiveJournal (Users/Groups)
10 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
and 5f). For CondMat (Figs. 3a and 3b), the distribution of degrees on the author nodes is a close match,
but there is some trouble matching the paper degree distribution. The model slightly overestimates at
the higher end of the degree scale and underestimates at the lower end. This is largely due to the small
size of the graph and the very small distribution (maximum of 18 authors for a paper). For MovieLens
(Figs. 4e and 4f), the model generates a few "critic" nodes of degree less than 20, even though no nodes
exist in the true degree distribution. A similar phenomena occurs for the "user" nodes in MillionSong.
In general, the CL model cannot handle gaps in the degree distribution because of Poisson distributions
in its expectations. In Peer2Peer (Figs. 5c and 5d), the number of degree-1 peer nodes is underestimated,
for reasons described in [11].
Shortcomings of bipartite CL
4.4
Overall, if we provide the degree distributions of real-world graph, the fast bipartite CL model generates
a random graph whose degree distribution closely matches the degree distribution of the original graph.
However, the graphs generated by CL typically have many fewer butterflies than the original graphs.
Table 3 shows that the number of butterflies from bipartite CL is smaller than the original graph in every
case, and sometimes by one or more orders of magnitude (CondMat, IMDB, Peer2Peer). The number of
caterpillars for the original and the generated graphs on the other hand, is closer, so the metamorphosis of
the original graph is much higher than for the generated graph with low butterfly counts. This indicates
that the bipartite CL model is omitting some important structure.
Note that the butterfly structure is what underlies the cohesive, close-knit structure in many real-
world graphs. Just as a triangle can be considered as the smallest unit of cohesion on one-way graphs,
butterflies can be considered as the smallest unit of cohesion in bipartite graphs. Conversely, without
butterflies, a bipartite graph will not have a community structure. This is our motivation for considering
a more complex model in the next section.
Table 3: Properties of the original and the BCL graphs.
Graph
Size
CondMat-Orig
CondMat-CL
IMDB-Orig
IMDB-CL
Flickr-Orig
Flickr-CL
MovieLens-Orig
MovieLens-CL
MillionSong-Orig
MillionSong-CL
Peer2Peer-Orig
Peer2Peer-CL
LiveJournal-Orig
LiveJournal-CL
nu
1.67e4
1.67e4
1.28e5
1.28e5
1.73e6
1.73e6
6.51e4
6.51e4
1.02e6
1.02e6
1.99e6
1.99e6
5.28e6
5.28e6
nv
2.20e4
2.20e4
3.84e5
3.84e5
1.04e5
1.04e5
7.16e4
7.16e4
3.85e5
3.85e5
5.38e6
5.38e6
7.49e6
7.49e6
Edges
m
5.86e4
5.86e4
1.47e6
1.47e6
8.55e6
8.39e6
1.00e7
8.78e6
4.84e7
4.81e7
5.58e7
5.58e7
1.12e8
1.11e8
Cats.
n
1.24e6
2.22e6
8.56e8
1.11e9
2.57e12
2.20e12
2.46e13
1.37e13
2.21e13
2.59e13
8.18e11
1.20e12
3.36e14
3.31e14
Buts.
n
7.05e4
3.57e2
3.50e6
1.41e5
3.53e10
1.52e10
1.20e12
5.34e11
2.15e11
6.74e10
3.80e9
1.14e8
3.30e12
2.04e12
Meta.
c
2.28e−1
6.43e−4
1.64e−2
5.10e−4
5.49e−2
2.78e−2
1.95e−1
1.56e−1
3.89e−2
1.04e−2
1.86e−2
3.79e−4
3.92e−2
2.47e−2
MEASURING AND MODELING BIPARTITE GRAPHS
11 of 23
5. Bipartite BTER Model
In the BTER model for one-way graphs, the goal is to match both the degree distribution as well as
the degreewise clustering coefficients [18, 35]. Our goal here is to extend those notions to the bipartite
case. We use (1.1) as the bipartite definition of the clustering coefficient, though other metrics exist as
surveyed by Latapy, Magnien, and Vecchio [20] and Opsahl [30].
5.1 Degreewise Metamorphosis Coefficient
BTER matches the degreewise clustering coefficient, so we need a similar measure for bipartite graphs.
We describe the degreewise metamorphosis coefficient, which provides a more nuanced measurement
of bipartite community structure than the metamorphosis coefficient. To the best of our knowledge, this
idea has not been proposed before.
We define the metamorphosis of an edge (i, j) as
n(i, j)
n(i, j)
0
c(i, j) =
=
number of butterflies containing (i, j)
number of caterpillars centered at (i, j)
if n(i, j) > 0,
if n(i, j) = 0.
(5.1)
We know the number of caterpillars centered at (i, j) immediately from the degrees of its endpoints, i.e.,
(5.2)
From this, we define the metamorphosis coefficients of vertices i ∈ U and j ∈ V as the mean value over
all edges incident to the vertex:
i − 1)(dv
n(i, j) = (du
j − 1).
cu
i =
1
du
i
∑
(i, j)∈E
c(i, j)
and
cv
j =
1
dv
j
∑
(i, j)∈E
c(i, j).
(5.3)
We may consider (5.3) to be the bipartite analogue of the clustering coefficient of a vertex, as introduced
by Watts and Strogatz [38]. Finally, we can define the per-degree metamorphosis coefficients to be
cu
d =
1
Ud ∑
i∈Ud
cu
i
and
cv
d =
1
Vd ∑
j∈Vd
cv
j,
(5.4)
where Ud and Vd denote the subsets of degree-d nodes, i.e.,
Ud = {i ∈ U du
i = d}
and Vd = { j ∈ V dv
j = d} .
Degreewise metamorphosis coefficients control for the effects of both vertex mode and vertex degree
on bipartite clustering. Accordingly, this metric may reveal insights otherwise lost in metrics based only
on the ratio of total butterfly to caterpillar counts. To illustrate, consider the CondMat author-paper
network, whose degreewise metamorphosis coefficients are shown by the blue line in Figs. 6c and 6d.
The degreewise metamorphosis coefficients are also binned in the same way that we binned the data for
the degree distributions: We use powers of two for the bins, so the x-coordinate 2k corresponds to the
bin from [2k,2k+1). The y-coordinate is the average value for that bin. For the binning, we define cu
d = 0
for any degree such that Ud = 0 (i.e., when there are no nodes of degree d), and likewise for cv
c. We
see that degreewise metamorphosis coefficients in the author mode (Fig. 6c) are higher for low degrees
than for high degrees, meaning that authors with fewer papers tended to have higher proportion of repeat
collaborations than authors with many papers. Conversely, the paper mode (Fig. 6d) shows that authors
of papers with few authors tend to have more repeats of the same author set.
12 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
5.2 Affinity Blocks
In BTER for one-way graphs, dense ER subgraphs are key to producing triangles. For bipartite BTER,
we will use dense bipartite ER subgraphs to produce butterflies. We refer to these dense ER subgraphs
as affinity blocks.
In (one-way) BTER, an affinity block ideally consists of a set of d + 1 vertices of degree d. The
connectivity of each block is computed according to the degree-d clustering coefficient. The bipartite
affinity blocks are similar in spirit, but the bipartite nature of the graph raises issues that require an
entirely new approach to the block construction.
The first key difference is that each affinity block in bipartite BTER consists of two sets of vertices,
one from each partition. While each partition set in an affinity block ideally contains vertices of the
same degree, the degrees do not necessarily match between partition sets. Indeed, in many bipartite
graphs, one partition set may have a very different range of degrees than the other, so attempting to
create blocks that match inter-partition degree is not a realistic goal.
Consequently, a second key difference concerns how we determine the sizes of the partition sets for
the affinity blocks. In the one-way BTER method, the size of each block only depends on the degree
of the vertices in the block. In the two-way case, the sizes of each partition set within a block is more
complicated.
To work out the calculations of sizes and connectivity for the blocks, we consider building a single
affinity block denoted by G = ( U, V , E). Without loss of generality, we assume all nodes in U want
degree du and all nodes in V want degree dv. Note that these degrees are with respect to the entire
graph, not the subgraph. We further assume all nodes in U want metamorphosis coefficient cu and
likewise for V and cv. When matching a real-world graph, we choose the degreewise metamorphosis
coefficients corresponding to the target degrees as defined in (5.4), i.e.,
cu = c du
and
cv = c dv.
The goal is to determine the sizes nu = U and nv = V and the connectivity, ρ, which is the probability
of an edge between a node in nu and a vertex in nv, and thus the number of edges, E
For i ∈ U, we can compute its expected metamorphosis coefficient as follows. By definition (5.3),
we have
cu
i =
1
du
i
∑
(i, j)∈E
c(i, j) =
1
du
i
c(i, j) + ∑
(i, j)∈E\ E
c(i, j)
∑
(i, j)∈ E
c(i, j).
(5.5)
du
i
The last step comes from the assumption that nearly all butterflies in the larger graph come from the
affinity blocks. Using definition (5.1), we can rewrite (5.5) as
cu
i =
1
du
i
∑
(i, j)∈ E
n(i, j)
n(i, j)
=
1
du
i
∑
(i, j)∈ E
n(i, j)
i − 1)(dv
(du
j − 1)
.
(5.6)
We have assumed that the degrees and clustering coefficients within G are constant, so (5.6) becomes
cu =
1
du( du − 1)( dv − 1) ∑
(i, j)∈ E
n(i, j) =
2ni
du( du − 1)( dv − 1)
.
(5.7)
The last equality uses the fact that the sum of butterflies involving edges of the form (i, j) is equal to
two times the number of butterflies involving node i since each such butterfly has two edges involving i.
∑
(i, j)∈ E
≈ 1
MEASURING AND MODELING BIPARTITE GRAPHS
13 of 23
In expectation,
(cid:18) nv
(cid:19)
= ρ4( nu − 1)
.
ni
because there are ( nu − 1) other choices for the second node in U and(cid:0) nv
2
,
(cid:1) choices for the two nodes in
(5.8)
V . Finally, ρ4 is the probability that all four edges exist. Combining (5.7) and (5.8) gives
2
cu .
=
ρ4 nv( nu − 1)( nv − 1)
du( du − 1)( dv − 1)
.
Using analogous reasoning for cv, we ultimately want
cu du( du − 1)( dv − 1)
nv( nu − 1)( nv − 1)
ρ4 =
=
cv dv( du − 1)( dv − 1)
nu( nu − 1)( nv − 1)
Ideally, we would have
nu = dv
and
nv = du,
(5.9)
(5.10)
(5.11)
but we cannot generally satisfy (5.10) and (5.11) at the same time. Therefore, we choose one of the
equalities in (5.11) and then solve for the other number of nodes and connectivity using (5.10) to get
nu = dv ⇒ nv =
nv = du ⇒ nu =
cu
cv
cv
cu
du,
ρ =
dv. ρ =
( du − 1)( cv)2
cu du − cv
( dv − 1)( cu)2
cv dv − cu
.
(5.12)
(5.13)
So that we can have the possibility of using a complete bipartite subgraph to yield metamorphosis
coefficients of one, we constrain our choices to satisfy:
nu (cid:62) dv
(5.14)
cv (cid:62) 1 and (5.13) otherwise. It is easy to see from (5.10) that
To satisfy (5.14), we choose (5.12) if cu
this has the added bonus of ensuring ρ (cid:54) 1. This logic forms the basis of building affinity blocks for
bipartite BTER.
nv (cid:62) du.
and
5.3 Bipartite BTER algorithm
As explained in Section 5.2, the affinity block building process for bipartite BTER takes into account
both the desired vertex degrees and corresponding desired per-degree metamorphosis coefficients when
setting the partition sizes and connectivity of each affinity block. This information is then used to model
each affinity block as an ER subgraph, where the preponderance of the graph butterflies are created.
Accordingly, the affinity block construction is key to matching per-degree metamorphosis coefficients.
In order to match the desired degree distribution, the remaining excess degree (i.e., edges not used in
constructing the affinity blocks) is connected via a fast bipartite CL procedure. The full bipartite BTER
algorithm is listed in Algorithm 2.
14 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
Algorithm 2 Bipartite BTER
i }, {dv
1: procedure BIBTER({du
j }, {cu
d }, {cv
d } )
We assume that degrees are sorted in increasing order
i
m ← ∑i du
E ← /0
i } ← {du
{eu
i ← min{i du
repeat
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
j } ← {dv
j }
i }, {ev
i > 1}, j ← min{ j dv
j > 1}
(cid:46) Excess degree initialization
(cid:46) Create affinity blocks until nodes exhausted
i , dv ← dv
du ← du
j
cu ← c du, cv ← c dv
if cu/ cv (cid:62) 1 then
else
nu ← dv, nv ← round
nv ← du, nu ← round
(cid:16) cu
cv du(cid:17)
, ρ ←(cid:16) ( du−1)( cv)2
, ρ ←(cid:16) ( dv−1)( cu)2
(cid:16) cv
cu dv(cid:17)
cu du− cv
cv dv− cu
(cid:17)1/4
(cid:17)1/4
end if
if i + nu − 1 (cid:54) nu and j + nv − 1 (cid:54) nv then
for j = j, j + 1, . . . , j + nv − 1 do
for i = i,i + 1, . . . ,i + nu − 1 do
r ← U(0,1)
if r <= ρ then
E ← E ∪ (i, j)
i = max{0,eu
eu
end if
i − 1}, ev
j = max{0,ev
j − 1}
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28: end procedure
end for
end for
end if
i ← i + nu, j ← j + nv
i } ,{ev
until i > nu or j > nv
E ← E ∪ FBCL({eu
j })
(cid:46) Create ER subgraph
(cid:46) Uniform random value in [0,1]
(cid:46) Update excess degree
MEASURING AND MODELING BIPARTITE GRAPHS
15 of 23
5.4 Experimental Results
We generate bipartite BTER graphs using the procedure described in Section 5.3. We first discuss
a single graph: CondMat. We show the resulting degree distribution and degreewise metamorphosis
coefficients in Fig. 6. The degree distributions, shown in Figs. 6a and 6b, show little difference between
bipartite BTER and CL; both are good matches to the original degree distribution. The degreewise
metamorphosis coefficients are shown in Figs. 6c and 6d. Although there is not a perfect match between
bipartite BTER and the original graph, it is much better than CL, which has almost no butterflies.
104
102
100
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
Original
CL
BTER
100
101
Degree
103
101
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
10−1
100
Original
CL
BTER
101
Degree
(a) Mode 1: CondMat (Authors/Papers)
(b) Mode 2: CondMat (Authors/Papers)
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.40
0.20
0.00
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.30
0.20
0.10
0.00
Original
CL
BTER
101
Degree
100.5
101
Degree
(c) Mode 1: CondMat (Authors/Papers)
(d) Mode 2: CondMat (Authors/Papers)
FIG. 6: Degree distribution and degreewise metamorphosis coefficients on the original CondMat graph
as well as the models generated by CL and bipartite BTER.
Summary data for all graphs is shown in Table 4. This is the same as Table 3 except that now we
have added a row for bipartite BTER. For the first four graphs, we are reporting average values over
100 experiments, and we also report the entire range of values; we do not do multiple trials for the
larger graphs because the postprocessing to count the butterflies is extremely expensive. The number
of butterflies and metamorphosis coefficients are significantly improved as compared to CL. We see
that CondMat has 70,000 butterflies, bipartite CL produces less than 400 butterflies, but bipartite BTER
produces 120,000 butterflies. The bipartite BTER number is a slight overestimate, but overall much
better than CL.
16 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
Table 4: Metamorphosis coefficients for original versus BCL graphs.
Graph
Size
nu
1.67e4
1.67e4
1.67e4
nv
2.20e4
2.20e4
2.20e4
Edges
m
5.86e4
5.86e4
6.00e4
Cats.
n
1.24e6
2.22e6
2.36e6
Buts.
n
7.05e4
3.57e2
1.10e5
[1.67 – 1.67]
[2.20 – 2.20]
[6.00 – 6.01]
[2.31 – 2.42]
[1.01 – 1.16]
1.28e5
1.28e5
1.28e5
3.84e5
3.84e5
3.84e5
1.47e6
1.47e6
1.47e6
[1.28 – 1.28]
[3.84 – 3.84]
[1.47 – 1.47]
1.73e6
1.73e6
3.96e5
1.04e5
1.04e5
1.04e5
8.55e6
8.39e6
8.34e6
[3.96 – 3.96]
[1.04 – 1.04]
[8.34 – 8.34]
6.51e4
6.51e4
1.07e4
7.16e4
7.16e4
6.99e4
1.00e7
8.78e6
8.52e6
[1.07 – 1.07]
[6.99 – 6.99]
[8.51 – 8.52]
1.02e6
1.02e6
1.02e6
1.99e6
1.99e6
1.99e6
5.28e6
5.28e6
3.20e6
3.85e5
3.85e5
3.85e5
5.38e6
5.38e6
5.38e6
7.49e6
7.49e6
7.49e6
4.84e7
4.81e7
4.80e7
5.58e7
5.58e7
5.60e7
1.12e8
1.11e8
1.10e8
8.56e8
1.11e9
1.35e9
[1.35 – 1.36]
2.57e12
2.20e12
2.36e12
[2.35 – 2.37]
2.46e13
1.37e13
1.23e13
[1.23 – 1.23]
2.21e13
2.59e13
2.93e13
8.18e11
1.20e12
1.66e12
3.36e14
3.31e14
3.50e14
3.50e6
1.41e5
6.77e6
[6.62 – 6.91]
3.53e10
1.52e10
4.25e10
[4.23 – 4.27]
1.20e12
5.34e11
1.08e12
[1.07 – 1.08]
2.15e11
6.74e10
1.90e11
3.80e9
1.14e8
6.71e9
3.30e12
2.04e12
4.61e12
CondMat-Orig
CondMat-CL
CondMat-BTER
BTER 100 Trial Range
IMDB-Orig
IMDB-CL
IMDB-BTER
BTER 100 Trial Range
Flickr-Orig
Flickr-CL
Flickr-BTER
BTER 100 Trial Range
MovieLens-Orig
MovieLens-CL
MovieLens-BTER
BTER 100 Trial Range
MillionSong-Orig
MillionSong-CL
MillionSong-BTER
Peer2Peer-Orig
Peer2Peer-CL
Peer2Peer-BTER
LiveJournal-Orig
LiveJournal-CL
LiveJournal-BTER
Meta.
c
2.28e−1
6.43e−4
1.86e−1
[1.73 – 1.99]
1.64e−2
5.10e−4
2.00e−2
[1.95 – 2.04]
5.49e−2
2.78e−2
7.21e−2
[7.19 – 7.23]
1.95e−1
1.56e−1
3.51e−1
[3.50 – 3.51]
3.89e−2
1.04e−2
2.59e−2
1.86e−2
3.79e−4
1.61e−2
3.92e−2
2.47e−2
5.26e−2
MEASURING AND MODELING BIPARTITE GRAPHS
17 of 23
The degreewise metamorphosis coefficients for the remaining graphs are shown in Figs. 7 and 8.
Even in cases where bipartite CL's overall numbers are close to the original graph as shown in Table 4,
the degreewise metamorphosis coefficients are essentially zero. Bipartite BTER corrects this problem
and gets metamorphosis coefficients that are much closer to what we see in the original graph. In some
cases like MovieLens (Fig. 7e), the CL metamorphosis coefficients are much higher than zero, which
is the result of the overall high density of the graph. Indeed, it has been proven (see Lemma 1 in [37])
that under mild assumptions, any bipartite graph with m edges and partition sizes nu and nv contains at
least on the order of ( nu
nu+nv )4 many butterflies. Thus, the presence of a certain minimum number
butterflies in bipartite graphs of sufficient density is inevitable; nevertheless, the original graph still has
higher values that are matched better by bipartite BTER. MovieLens also has an unusual profile in Mode
2 (see Fig. 7f), which is really just an artifact of the data collection. Nevertheless, bipartite BTER is able
to obtain a reasonable approximation. For completeness, the degree distributions for bipartite CL and
BTER as compared to the original graphs are shown in Figs. 9 and 10. There is little difference between
bipartite BTER and CL in terms of the degree distribution.
nv )2 · ( m
6. Conclusions
We have considered the problem of how to generate realistic bipartite graphs to reproduce the charac-
teristics of large real-world networks. Our first model, bipartite CL, accurately reproduces the degree
distribution. Our second model, bipartite BTER, goes further to capture the degreewise metamorphosis
coefficients. High coefficients are indicative of a relatively large number of butterflies indicating cohe-
sion or community structure, which is rare in sparse random graphs unless there are some behaviors that
go beyond just the degree structure. Creating realistic graph models leads to some hypotheses about the
ways the graphs were formed. In the cases where CL greatly underestimated the number of butterflies
(CondMat, IMDB, and Peer2Peer), we can surmise that there is some significant community-like behav-
ior. This is easy to see for authorship of papers (CondMat) and actors appearing in movies (IMBD). For
the Peer2Peer network, we might hypothesize that some tight-knit peer groups are sharing many files
between themselves. The other graphs have a smaller difference between the number of butterflies pro-
duced by CL and the real-world graph, indicating that there is much less community structure. In fact,
some graphs are so dense that CL has a nonzero metamorphosis coefficient, especially MovieLens.
Although these models match real-world observations in some aspects, more study is needed to
understand their limitations. How well do these models reproduce other metrics such as graph diameter,
singular values of the adjacency matrix, centrality measures, joint degree distributions, assortativity,
subgraph census, and so on? Can we find evidence that affinity blocks exist in real-world graphs?
We have created non-overlapping blocks; is that realistic? We have not mentioned time-evolution, but
certainly all networks are evolving in time and we need models that capture such changes. These are
but a few topics for future investigation.
Funding
This work was supported in part by the Defense Advanced Research Project Agency (DARPA). San-
dia National Laboratories is a multi-program laboratory managed and operated by Sandia Corpora-
tion, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy's
National Nuclear Security Administration under contract DE-AC04-94AL85000.
18 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.08
0.06
0.04
0.02
0.00
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.04
0.02
0.00
Original
CL
BTER
101
102
Degree
101
102
Degree
(a) Mode 1: IMDB (Actors/Movies)
(b) Mode 2: IMDB (Actors/Movies)
0.15
0.10
0.05
0.00
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.10
0.05
0.00
Original
CL
BTER
100
101
102
Degree
103
100
101
103
102
Degree
104
(c) Mode 1: Flickr (Users/Groups)
(d) Mode 2: Flickr (Users/Groups)
0.40
0.20
0.00
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.40
0.20
0.00
Original
CL
BTER
100
101
103
102
Degree
104
100
101
102
Degree
103
(e) Mode 1: MovieLens (Movies/Critics)
(f) Mode 2: MovieLens (Movies/Critics)
FIG. 7: Mode 2: MovieLens (Movies/Critics)
MEASURING AND MODELING BIPARTITE GRAPHS
19 of 23
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.05
0.00
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.08
0.06
0.04
0.02
0.00
Original
CL
BTER
100
101
102
Degree
103
100
101
102
103
Degree
104
105
(a) Mode 1: MillionSong (Users/Songs)
(b) Mode 2: MillionSong (Users/Songs)
0.08
0.06
0.04
0.02
0.00
0.10
0.05
0.00
Original
CL
BTER
100
101
103
104
102
Degree
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.10
0.05
0.00
100
Original
CL
BTER
101
102
Degree
103
(c) Mode 1: Peer2Peer (Peers/Files)
(d) Mode 2: Peer2Peer (Peers/Files)
Original
CL
BTER
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
s
i
s
o
h
p
r
o
m
a
t
e
M
0.15
0.10
0.05
0.00
101
102
100
102
104
106
Degree
Degree
(e) Mode 1: LiveJournal (Users/Groups)
(f) Mode 2: LiveJournal (Users/Groups)
FIG. 8: Mode 2: LiveJournal (Users/Groups)
20 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
103
100
10−3
100
Original
CL
BTER
101
Degree
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
101
10−3
Original
CL
BTER
102
100
101
102
Degree
(a) Mode 1: IMDB (Actors/Movies)
(b) Mode 2: IMDB (Actors/Movies)
105
101
10−3
102
10−1
10−4
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
100
10−5
Original
CL
BTER
Original
CL
BTER
100
101
102
Degree
103
100
101
103
104
102
Degree
(c) Mode 1: Flickr (Users/Groups)
(d) Mode 2: Flickr (Users/Groups)
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
103
100
10−3
Original
CL
BTER
101
102
Degree
103
Original
CL
BTER
100
101
103
104
102
Degree
(e) Mode 1: MovieLens (Movies/Critics)
(f) Mode 2: MovieLens (Movies/Critics)
FIG. 9: Degree distributions illustrating the original (blue), fast BCL (orange), and bipartite BTER
(green). The data is log-binned.
MEASURING AND MODELING BIPARTITE GRAPHS
21 of 23
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
104
100
10−4
Original
CL
BTER
105
100
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
Original
CL
BTER
100
101
102
Degree
103
100
101
103
102
Degree
104
105
(a) Mode 1: MillionSong (Users/Songs)
(b) Mode 2: MillionSong (Users/Songs)
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
100
10−5
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
105
101
10−3
Original
CL
BTER
100
101
102
Degree
103
Original
CL
BTER
100
101
103
104
102
Degree
(c) Mode 1: Peer2Peer (Peers/Files)
(d) Mode 2: Peer2Peer (Peers/Files)
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
104
102
100
Original
CL
BTER
)
n
i
B
r
e
p
n
a
e
M
(
t
n
u
o
C
107
100
10−7
Original
CL
BTER
100
101
Degree
102
100
102
104
106
Degree
(e) Mode 1: LiveJournal (Users/Groups)
(f) Mode 2: LiveJournal (Users/Groups)
FIG. 10: Degree distributions illustrating the original (blue), fast BCL (orange), and bipartite BTER
(green). The data is log-binned.
22 of 23
S. AKSOY, T. G. KOLDA, AND A. PINAR
Acknowledgments
We thank our colleagues for helpful discussions in the course of this work, especially Grey Ballard,
Kenny Chowdhary, Danny Dunlavy, Kevin Matulef, and Michael Wolf. We also thank the two anony-
mous referees for insightful comments that improved the work considerably.
REFERENCES
1. Aiello, W., Chung, F. & Lu, L. (2001) A random graph model for power law graphs. Experimental Mathemat-
ics, 10, 53–66.
2. Barab´asi, A.-L. & Albert, R. (1999) Emergence of scaling in random networks. Science, 286, 509–512.
3. Bertin-Mahieux, T., Ellis, D. P. W., Whitman, B. & Lamere, P. (2011) The Million Song Dataset. In ISMIR
2011: Proc. 12th Intl. Conf. on Music Information Retrieval. University of Miami.
4. Bez´akov´a, I., Bhatnagar, N. & Vigoda, E. (2006) Sampling binary contingency tables with a greedy start.
Random Structures and Algorithms, 30(1-2), 168205.
5. Chen, Y., Diaconis, P., Holmes, S. P. & Liu, J. S. (2005) Sequential Monte Carlo Methods for Statistical
Analysis of Tables. Journal of the American Statistical Association, 100(469), 109120.
6. Chung, F. & Lu, L. (2002a) The average distances in random graphs with given expected degrees. Proceedings
of the National Academy of Sciences, 99(25), 15879–15882.
7. Chung, F. & Lu, L. (2002b) Connected components in random graphs with given degree sequences. Annals of
Combinatorics, 6, 125–145.
sity of Minnesota.
8. Davies, R. (2009) MovieLens 10M. GroupLens, Department of Computer Science and Engineering, Univer-
9. Dhillon, I. S. (2001) Co-clustering documents and words using bipartite spectral graph partitioning. In KDD
'01: Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data
mining, pages 269–274, New York, NY, USA. ACM.
10. Dormann, C. F., Frund, J., Bluthgen, N. & Gruber, B. (2009) Indices, Graphs and Null Models: Analyzing
Bipartite Ecological Networks. The Open Ecology Journal, 2(1), 724.
11. Durak, N., Kolda, T. G., Pinar, A. & Seshadhri, C. (2013) A Scalable Null Model for Directed Graphs Match-
ing All Degree Distributions: In, Out, and Reciprocal. In NSW 2013: Proceedings of IEEE 2013 2nd Interna-
tional Network Science Workshop, pages 23–30.
12. Ganguly, N., Ghosh, S., Krueger, T. & Srivastava, A. (2012) Degree distributions of evolving alphabetic
bipartite networks and their projections. Theoretical Computer Science, 466, 2036.
13. Guillaume, J.-L. & Latapy, M. (2004) Bipartite structure of all complex networks. Information Processing
14. Guillaume, J.-L. & Latapy, M. (2006) Bipartite Graphs as Models of Complex Networks. Physica A, 371,
15. Harper, F. M. & Konstan, J. A. (2015) The MovieLens Datasets. ACM Transactions on Interactive Intelligent
Letters, 90(5), 215221.
795–813.
SystemsMar, 5(4), 1–19.
16. Holland, P. W. & Leinhardt, S. (1970) A Method for Detecting Structure in Sociometric Data. American
Journal of Sociology, 76(3), 492–513.
17. Kannan, R., Tetali, P. & Vempala, S. (1999) Simple Markov-chain algorithms for generating bipartite graphs
and tournaments. Random Structures and Algorithms, 14(4), 293308.
18. Kolda, T. G., Pinar, A., Plantenga, T. & Seshadhri, C. (2014) A Scalable Generative Graph Model with
Community Structure. SIAM Journal on Scientific Computing, 36(5), C424–C452.
19. Latapy, M. (2006) Bipartite Graph Tools, Version 1.0. .
20. Latapy, M., Magnien, C. & Vecchio, N. D. (2008) Basic Notions for the Analysis of Large Two-mode Net-
works. Social Networks, 30(1), 31–48.
21. McFee, B., Bertin-Mahieux, T., Ellis, D. P. & Lanckriet, G. R. (2012) The Million Song Dataset Challenge.
In WWW'12 Companion: Proc. 21st Intl. Conf. Companion on World Wide Web, pages 909–916.
22. Mikl´os, I., Erdos, P. L. & Soukup, L. (2013) Towards random uniform sampling of bipartite graphs with given
MEASURING AND MODELING BIPARTITE GRAPHS
23 of 23
degree sequence. Electronic Journal of Combinatorics, 20(1), P16.
23. Milojevi´c, S. (2010) Power Law Distributions in Information Science: Making the Case for Logarithmic
Binning. Journal of the American Society for Information Science and Technology, 61(12), 2417–2425.
24. Mislove, A. (2015) Personal Communication. .
25. Mislove, A., Marcon, M., Gummadi, K. P., Druschel, P. & Bhattacharjee, B. (2007) Measurement and Anal-
ysis of Online Social Networks. In IMC'07: Proceedings of the 7th ACM SIGCOMM Conference on Internet
Measurement, pages 29–42, San Diego, CA.
26. Nacher, J. C., Ochiai, T., Hayashida, M. & Akutsu, T. (2009) A mathematical model for generating bipartite
graphs and its application to protein networks. Journal of Physics A: Mathematical and Theoretical, 42(48),
485005.
27. Newman, M. E. J. (2001) The Structure of Scientific Collaboration Networks. Proceedings of the National
Academy of Sciences, 98(2), 404–409.
28. Newman, M. E. J. (2003) The Structure and Function of Complex Networks. SIAM Review, 45(2), 167–256.
29. Newman, M. E. J., Strogatz, S. H. & Watts, D. J. (2001) Random graphs with arbitrary degree distributions
30. Opsahl, T. (2013) Triadic Closure in Two-mode Networks: Redefining the Global and Local Clustering Coef-
and their applications. Physical Review E, 64(2).
ficients. Social Networks, 35(2), 159–167.
31. Opsahl, T. (2015) Tnet Datasets. .
32. Peruani, F., Choudhury, M., Mukherjee, A. & Ganguly, N. (2007) Emergence of a non-scaling degree distri-
bution in bipartite networks: A numerical and analytical study. Europhysics Letters (EPL), 79(2), 28001.
33. Robins, G. & Alexander, M. (2004) Small Worlds Among Interlocking Directors: Network Structure and
Distance in Bipartite Graphs. Computational & Mathematical Organization Theory, 10(1), 69–94.
34. Saavedra, S., Reed-Tsochas, F. & Uzzi, B. (2008) A simple model of bipartite cooperation for ecological and
organizational networks. Nature, 457(7228), 463466.
35. Seshadhri, C., Kolda, T. G. & Pinar, A. (2012) Community Structure and Scale-free Collections of Erdos-
R´enyi Graphs. Physical Review E, 85(5).
36. Sun, J., Qu, H., Chakrabarti, D. & Faloutsos, C. (2005) Relevance Search and Anomaly Detection in Bipartite
Graphs. SIGKDD Explorations Newsletter, 7(2), 48–55.
37. Tait, M. & Verstraete, J. (2016) On Sets of Integers with Restrictions on Their Products. European Journal of
Combinatorics, 51, 268–274.
38. Watts, D. & Strogatz, S. (1998) Collective dynamics of 'small-world' networks. Nature, 393, 440–442.
39. The Echo Nest Taste Profile Subset,
the Official User Data Collection for the Million Song Dataset.
http://labrosa.ee.columbia.edu/millionsong/tasteprofile.
|
1712.02245 | 2 | 1712 | 2018-01-15T15:18:12 | Viral content propagation in Online Social Networks | [
"cs.SI"
] | Information flows are the result of a constant exchange in Online Social Networks (OSNs). OSN users create and share varying types of information in real-time throughout a day. Virality is introduced as a term to describe information that reaches a wide audience within a small time-frame. As a case, we measure propagation of information submitted in Reddit, identify different patterns and present a multi OSN diffusion analysis on Twitter, Facebook, and 2 hosting domains for images and multimedia, ImgUr and YouTube. Our results indicate that positive content is the most shared and presents the highest virality probability, and the overall virality probability of user created information is low. Finally, we underline the problems of limited access in OSN data. Keywords: Online Social Networks, Virality, Diffusion, Viral Content, Reddit, Twitter, Facebook, ImgUr, YouTube | cs.SI | cs |
Viral content propagation in Online Social
Networks
Giannis Haralabopoulos
University of Southampton
[email protected]
Ioannis Anagnostopoulos
University of Thessaly
[email protected]
Sherali Zeadally
University of Kentucky
[email protected]
January 17, 2018
Abstract
Information flows are the result of a constant exchange in Online Social
Networks (OSNs). OSN users create and share varying types of informa-
tion in real-time throughout a day. Virality is introduced as a term to
describe information that reaches a wide audience within a small time-
frame. As a case, we measure propagation of information submitted in
Reddit, identify different patterns and present a multi OSN diffusion anal-
ysis on Twitter, Facebook, and 2 hosting domains for images and multi-
media, ImgUr and YouTube. Our results indicate that positive content
is the most shared and presents the highest virality probability, and the
overall virality probability of user created information is low. Finally, we
underline the problems of limited access in OSN data.
Keywords: Online Social Networks, Virality, Diffusion, Viral Content,
Reddit, Twitter, Facebook, ImgUr, Youtube
1 Problem Definition
Online Social Networks (OSNs) are modern information systems that allow users
to interact with friends, acquaintances, family members, colleagues or even un-
known individuals. The information exchanged in these networks is diverse and
includes local or world news, general and scientific facts, quotes, personal pref-
erences, and so on. OSN providers focus on developing new information-sharing
schemes where users can easily disseminate information efficiently.
Information propagation and diffusion fields emphasize the temporal aspects
of information systems. The scale and usability of OSNs have introduced new
terms to describe the propagation phenomenon. Rapid sharing of information
1
on a large scale is informally known as virality. The dynamics of information
describe the properties that constitute, sustain, or modify the topology of the
OSNs based on diffusion and propagation processes such as hierarchy, network
partition, clustering, and others. User influence characterizes the capacity of
OSN nodes to be a compelling force on behaviour and opinion shaping.
Our study focuses on content originating from Reddit, mainly because this
OSN is comprised of both original and reused content, and it offers a liberal
data access policy, and an open Application Programming Interface (API) with
extensive documentation. Reddit is a social news and entertainment site pow-
ered by user generated content. Registered users submit content through a
descriptive link, subject to (up/down)votes. Users who have created a post or
commented on one, gain or lose "karma", a Reddit metric for user ranking.
This metric is calculated as, the sum of all the upvotes minus the sum of all
the downvotes a user receives. Posts that acquire a high vote ratio (positive
to negative) in a short time period after their submission, are moved to the
front page. It is apparent that Reddit community defines the popularity of the
disseminated content and determines its "success" or "failure".
Information is shared and spread among various OSNs, such as Twitter,
Facebook, and Google Plus. Information submitted in any of these OSNs affects
the others as well. Our goal is to measure this impact, along with its properties
and tempo-spatial attributes. We present the two main patterns observed in
information propagation, based on whether the content existed before or was
created for the particular OSN submission. A wider range of sub-patterns is
presented in [12].
2 Related Work
The topic of diffusion has been at the center of sociology interest for many years.
Even before the emergence of OSNs, social ties and information flows have been
studied in traditional real-life social networks.
The notion that information flows, from mass media to opinion leaders and
later on to a wider population as final consumers, was first introduced during
the middle of 40's [19]. Almost a decade later, Kaltz and Lazarsfeld revisited
the subject by proposing their theory of "Two step flow of communication" [17].
In a similar conceptualization, Granovetter and Mark suggested the analysis of
social networks, as a tool for linking micro and macro levels of sociology theory
[7]. Similar conclusions were verified in OSNs after nearly 40 years, while at the
same time, interest in social media analysis skyrocketed [2, 26].
Viral marketing was introduced in 1997 by Juvertson and Draper [15]. Porter
and Golan in [25], found that provocative content is a crucial virality factor.
Leskovec et al.
in [20], modelled outbreak detection via node selection, while
performing a two-fold evaluation of their model. Concerning word of mouth
scenarios, Allsop et al.
in [1], noted that 59% of individuals frequently share
online content. Berger and Milkman in [3], observed that positive content is
more viral than negative.
2
The authors in [5] focused on identifying the optimal network that best
describes information propagation in news media and blogs. In [14], authors
analysed word of mouth through email. Social influence modeling was studied
in [6]. Similarly, the authors in [13], investigated the social influence in meme
social graph. In [32], the authors studied virality lifespan issues in Twitter, as
also mentioned in [13]. In the same OSN (Twitter), Sakaki et al.
[29] studied
users as social sensors by monitoring information flow and dissemination during
earthquake incidents. In [18], the authors analysed the entire Twitter graph in
order to assess its topological characteristics.
In [30], Tucker measured the effectiveness of virality as a marketing tool.
While in [15] direct messages were measured. Interactions had a negative effect
on virality while sharing, engaging, provocative (as defined in [25]) or humorous
content and visual appeal, affected virality in a positive way. In [24], the authors
created a map of the so-called "life cycle" of the blog. Likewise, a virality study
in OSNs conducted by Guerini et al., showed that content is more important
to virality than the influencer itself [9]. By focusing on diffusion analysis of
Twitter, Hansen et al. in [11], observed the effect of negativity on virality. In
other OSN-related researches, the authors of [4] studied the maximization of
influence. The work described in [27] examines the passivity of Twitter users.
In [23], the authors developed an algorithm to describe influence propagation.
Rajyalakshmi et al. in [26], proposed a stochastic model for the diffusion of
several topics. The authors discovered that strong ties play a significant role
in virality, having homophily as a major contributor as observed in [18]. Micro
and macro scales of the network become relevant, as seen in [7]. Furthermore,
the authors of [18] note that, acts within groups have a global impact. Romero
et al.
in [28], analyse hashtag diffusion in Twitter and discovered that initial
adopters of hashtags were fairly important. Yang and Leskovec also studied
content exposure growth and how it fades over time in Twitter [33]. Huang
et al.
in [21] tried to effectively select a number of nodes in order to monitor
information diffusion. In a different but of related context study, Guerini et al.
in [8] noted that the dynamics of Social Networks exist in the scientific literature
and used psycholinguistic analysis to determine how abstracts affect virality.
The authors of [10], studied the dynamics of information diffusion process.
Furthermore in [2], Bakshy et al. addressed the problem of information diffusion
in OSNs. The authors ended up with the statement stated in [7, 18] that, strong
ties are more influential but weak ties are responsible for propagating novel
In [16], the authors investigated why and how some messages
information.
in OSNs become viral. Finally, Weng et al.
in [31], analysed longitudinal
microblogging data.
In this research, we aim to quantify the linkage of various OSNs based on
contextual analysis, as well as to measure the diffusion in these OSNs, thus
addressing a gap in the related literature. Based on the related works discussed,
in this work, we:
• Aim to identify patterns of propagation similar to the work described in
[14]
3
• Validate virality upon time with an easy to calculate criterion similar to
metrics described in works [3, 6, 13, 30, 9, 11]
• Expand the concept of information diffusion across multiple OSNs, unlike
the works described in [32, 18, 4, 27], and [16] that deal with the same
concept but within one OSN
3 Methodology and Dataset
For our study, we need domains that provide viewership counters in order to
monitor and examine information propagation features. ImgUr and YouTube
count views according to the users IP address, Reddit only provides the voting
count of a post. As such, the sum of negative and positive votes is used as the
Reddit views counter. We were aware that both the absence of an IP based view
counter and the method Reddit uses to fuse voting, introduce some inaccuracy
in our results. However, since our main aim is toward propagation rather than
viewership volume, such kind of inaccuracy does not affect our initial perception
of information flows.
In some preliminary tests we performed on Facebook and Google Plus, we
observed that both demonstrated a fairly low propagation even on viral content.
This is because both APIs of the above mentioned OSNs do not provide access to
private posts. Thus, only public posts were taken into consideration. Addition-
ally, the network usage of Google Plus is fairly low compared to Facebook. After
considering these factors, we decided to include only Facebook in our analysis.
Finally, in order to present our results in a consistent format between different
OSNs, we address all kinds of viewership count as "Units of Interest" (UoI).
A single Unit of Interest is equal to one content view in its domain (ImgUr or
YouTube), one vote in Reddit, one mention in Twitter or one mention/like in
Facebook.
To discard content that is not gaining attention rapidly, we employed a
simple rule (mentioned hereafter as check criterion). In this rule, we examined
if "Units of Interest" in Reddit are doubled in absolute numerical values for
every check, for the first 4 checks after the creation of the post, as required by
our virality criterion. Ou proposed criterion can be applied in real time with
only a few calculations, a key difference from existing methods . Only only a
small percentage of all posts reached high viewership counts within such a short
time span. 0.66% of the posts analyzed actually passed this UoI check criterion
in our dataset (682 out of nearly 102400 posts).
4 Results
In this section, we present and analyse the results derived from our scraping
procedure. Data is separated based on the the hosting domain. Reddit posts
include topics such as "AdviceAnimals", "Aww", "Eathpon", "Funny", "Gam-
4
ing", "Gif", "Movies", "Music", "Pics", "TIL", "Videos" and "WTF", and all
were discovered with our scraping period.
4.1 Posts and Counters
We scrapped two different subreddits, "new" and "rising". In total, we found
nearly 45000 ImgUr and YouTube posts in the "rising" category and nearly
70000 in the "new" category. More specifically, the posts of the "rising" cate-
gory revealed 40966 links to ImgUr and 3084 links to YouTube. Similarly, the
distribution for posts from the "new" category for ImgUr and YouTube was
51984 and 6366 respectively.
We also observed that the "new" subreddit content is more likely to become
viral compared to the content in "rising" subreddit. Content featured in "new"
subreddit was twice as likely to pass our UoI check, while content in "rising"
category presented a higher initial UoI count but fell short on our criterion
(which is doubling the UoI count in the first four check). Out of 40966 posts
linking to ImgUr in "rising" subreddit, only 200 satisfied our check criterion
(0.48%). As for posts linking to YouTube, out of 3084 posts in total, only 14
satisfy our check (0.45%). In contrast, in "new" subreddit, out of the 51984
posts that linked to ImgUr, 431 doubled their UoI count in 4 subsequent checks
(0.82%). Similarly, the number of posts that linked to YouTube and successfully
passed our UoI criterion, was 37 out of 6366 (0.58%).
The UoI criterion presented is not a virality validation, yet it provides strong
evidence that the posts analysed have been viewed by enough viewers, within a
small time-frame and the interest is not diminishing. Among 682 highly viewed
posts, 631 were linked to ImgUr and 51 were linked to YouTube.
As mentioned in [1, 3], positive and entertaining content was found as the
most frequently shared topic.
In our case, we could label "AdviceAnimals",
"Funny" and "WTF topics", as entertaining content. "Aww" and "Eathpon"
can be considered as containing emotive content, as described in [25]. "TIL"
is mainly informative, while "Pics" and "Videos" characterize mixed content
types. "Gaming", "Movies" and "Music" contain entertainment content.
Entertaining content and positive content are shared more frequently than
anything else. Although provocative content and controversial content can be
created within Reddit and ImgUr, they rarely get enough votes to appear on
the front page. Most of the time, they appear in the form of a "news" or
"TIL" post. The number of "gaming" posts that passed our check criterion is
descriptive of the interests of Reddit users, especially when compared to other
forms of established entertainment such as movies and music. Out of the 682
posts, 371 were mentioned in Twitter, 7 in Facebook and 172 in both OSNs.
Posts of "funny" topic are the most shared in both Twitter and Facebook,
but are not included among the top-shared topics. Additionally, almost every
category presents high Twitter and low Facebook sharing percentage (a direct
result of the Facebook API limitations). "Eathpon" posts with photographs
from around the world were 100% shared in Twitter. The majority of "Music"
and "TIL" posts are shared in both OSNs, but the number of posts in these
5
categories is low. In contrast, 50% of of viral posts from "Videos" and "Movies"
are shared on all OSNs, confirming the positive attitude of Internet users towards
multimedia content.
Topics with high number of viral posts ("AdviceAnimals", "Aww", "Funny",
"Gaming" and "Pics"), were found to have the highest sharing percentage for
both OSNs pertaining to "Pics" post from "rising" subreddit (41.38%), followed
by "Funny" posts from "new" subreddit (29.75%).
In contrast, "Aww" and
"Gaming" posts from the "new" subreddit had the lowest percentages (10.53%
and 13.04% respectively). Posts from "new" subreddit were more often shared
in Twitter compared to posts from "rising" subreddit, for every topic except
"Funny".
4.2 Online Social Network Propagation
UoI checks were simultaneously conducted on (at least) an hourly-basis for Red-
dit, Twitter, Facebook and the domain the image or video was hosted (ImgUr
or YouTube). Thus, we were able to observe the propagation of a certain Reddit
post and content in both OSNs. We wanted to measure a slow-rate information
propagation from its source to both OSNs. In fact, such a propagation pattern
was indeed found and was the most common propagation standard throughout
our analysis of the posts and topics.
Figure 1: Percent UoI Variation over time, ImgUr
The time interval between subsequent checks performed in the tested do-
mains (Reddit, ImgUr or YouTube, Twitter and Facebook) is at least one hour
and the first check of every post has zero UoI. Each post was monitored for 7
days. We found that after 14 checks, the interest (in terms of UoI variation)
declines significantly. The mean number of checks was 65 per post, while the
14th check happened within a 24-hour period from the creation of the post.
6
2468101210−1100101102ImgUrRedditTwitter Adj.Facebook Adj. All posts are filtered to include those that manage to double their UoI (in
absolute numerical values) on Reddit on an hourly basis for a 4-hour time-frame.
Consequently, every post we analysed had high Reddit UoI variance for the first
5 checks, but only a few of them had higher levels of variance after that point.
As far as our OSN data scraping is concerned, we only had access to information
from the public API of Facebook, while on Twitter we could only check messages
within a timespan of a week due to Twitter API limitations.
Figure 1 presents the UoI variation over time for ImgUr, where its UoI are
growing at a high rate concurrently with Reddit, followed by Twitter with a
relatively small level of interest, and Facebook with delayed and low interest.
Similar patterns appeared in all of the distinct categories, while in most of
the cases Facebook had nearly zero level of interest. This specific pattern was
the most common in our research and appeared in 14 out of 18 ImgUr topics.
Topics such as "AdviceAnimals", "Aww, "Funny", "Eathpon", "Gif" "Pics"
and "WTF", followed this pattern - in both "rising" and "new" subreddits.
However in both "Gif" categories, interest variation in Reddit is slightly higher
than Domain.
Figure 2: Percent UoI Variation over time, YouTube
On YouTube content, we encountered a completely different propagation
pattern, Figure 2. In most cases, domain UoI was not seriously affected by the
increased interest in OSNs and Reddit. This is mainly because shared content
in YouTube is not necessarily new, but a video that existed for a long time and
becomes mildly popular over time. In our evaluation, the videos under study
presented a very high initial UoI count which resulted in a low UoI variance.
Apart from that, we observed high UoI variance in Reddit and Twitter, low UoI
variance in Facebook, and practically zero UoI variance in the YouTube.
7
2468101210−1100101102103YouTubeRedditTwitter Adj.Facebook Adj. 5 Discussion - Future Work
Our analysis verified many observations of previous researches mainly with re-
spect to the micro and macro effects of information flows [19, 17, 16]. More
specifically, a single post (micro effect) in the parent domain connected with a
post in Reddit, starts to accumulate views (macro effect) up to a point where
the information hops to OSNs (first in Twitter and then in Facebook) flow-
ing through individuals and eventually the interest dies off. Positive content
is confirmed to be the mostly shared content in OSNs as mentioned in [25, 3].
Similarly to [13], we found out that a Reddit post is most popular within the
first few hours after its creation. This effect is also enhanced by the classification
method of Reddit, where new posts need fewer votes (compared to old posts)
to move to the first page.
As mentioned, we have observed that the most shared posts are positive and
entertaining content as observed in [1, 3]. These categories include post from
topics such as "AdviceAnimals", "Funny" and "WTF topics, while we also iden-
tified emotive posts, as described in [25] on topics such as "Aww" and "Eath-
pon". Unfortunately, not enough viral posts of such content were found, in order
to allow the identification of their propagation model. "Gaming", "Movies" and
"Music" contain user-centric and very specific entertainment content, with low
post appearance but a high viral to total posts ratio, while, "TIL" is mainly
informative and is rarely used in posts linking to either ImgUr or Youtube.
The fact that only public Facebook can be scraped is a big factor to those
particular UoI. This means that content posts and view counts on Facebook
are significantly higher. Nonetheless, the approximate propagation timeframe
should not be affected because of the ripple effect of social interactions [22]. To
summarize, the most notable results of our research are:
• Online Social Networks are connected via shared information
• Shared content is mostly positive and positive content is more viral
• Shared (user generated) content has a small lifetime
• Low percentage of content can be considered as viral
• Information flows have micro and macro effects across networks they tra-
verse
• API limitations prevent further diffusion analysis
Information starts within a domain and hops to various other media within
minutes. However, upon its propagation, information flows within the original
domain do not stop, but slow down within a small period of time. Entertainment
content and positively emotive content are the most "viral", and persistence
was only found in gaming posts. Movies related posts were the only ones that
spread nearly simultaneously to every domain and OSN. We should consider a
8
new perception of information flow, one that would accurately describe content
sharing in modern OSNs.
The research community can examine what kinds of social media platforms
play the role of information broker for other ones, as well as what kinds of in-
formation their users prefer to link across different media and why. Are some
social media not capable enough of making a post viral? Is the linkage made
intentionally, or users are biased? On the other hand, social media managers
can investigate probable common information channels across other social net-
working services, reveal groups of users with similar interests, or even compare
usage and market share (market position). Finally, one can also identify the dis-
tinct role of any OSN with respect to information propagation and information
lifespan (e.g. in contrast to Twitter users, users of Reddit use the service for not
time-sensitive posts like bookmarking), by utilizing similar research methods.
In our next step, we will consider precise diffusion analysis relative to time.
Since we have already confirmed the connection of several OSNs in this work,
in the future we will look into the time frames of each OSN diffusion. But such
analysis would require data timestamps and an algorithm designed for speed.
Unfortunately our serial algorithm was designed without any time considerations
and lacked the ability to simultaneously analyse thousands of posts. We plan to
refine our scraping and analysis method to achieve those goals. Our long-term
objective is to create a viral post tool that can analyse every OSN available.
We hope that future data from these OSNs will not be locked behind pay-walls
and data access would be unobstructed, so that every scientist will be able to
analyse these unique large-scale, content-based social networks.
Additionally, we also plan to analyse a wider range of topics (such as news
or similar, which lacked required counters to observe its diffusion) in as many
OSNs as we can. We were surprised to observe that various posts in Reddit of
News topic attracted users' attention within Reddit long before being reported
in dedicated news sites. Similarly, more sites with unrestricted APIs could be
added to our analysis tool so that we can analyse similar information diffusion
metrics. Diversity of topics and sites does not affect virality analysis but it
would however have an impact on the complexity of our algorithm, while being
able to provide more interesting results. In the future, we will investigate the
aforementioned issues with a strong focus on virality, information flows and
time.
9
References
[1] Dee T Allsop, Bryce R Bassett, and James A Hoskins. 2007. Word-of-mouth
research: principles and applications. Journal of advertising research 47, 4
(2007), 398–411.
[2] Eytan Bakshy, Itamar Rosenn, Cameron Marlow, and Lada Adamic. 2012.
The role of social networks in information diffusion. In Proceedings of the
21st international conference on World Wide Web. ACM, 519–528.
[3] Jonah Berger and Katherine L Milkman. 2012. What makes online content
viral? Journal of marketing research 49, 2 (2012), 192–205.
[4] Francesco Bonchi. 2011. Influence Propagation in Social Networks: A Data
Mining Perspective. IEEE Intelligent Informatics Bulletin 12, 1 (2011), 8–
16.
[5] Manuel Gomez Rodriguez, Jure Leskovec, and Andreas Krause. 2010. In-
ferring networks of diffusion and influence. In Proceedings of the 16th ACM
SIGKDD international conference on Knowledge discovery and data min-
ing. ACM, 1019–1028.
[6] Amit Goyal, Francesco Bonchi, and Laks VS Lakshmanan. 2010. Learning
influence probabilities in social networks. In Proceedings of the third ACM
international conference on Web search and data mining. ACM, 241–250.
[7] M Granovetter. 1973. The Strength of Weak ties American journal of
Sociology 78: 13 60-1380. Examines the role of ties within a group and
their effect on performance (1973).
[8] Marco Guerini, Alberto Pepe, and Bruno Lepri. 2012. Do linguistic style
and readability of scientific abstracts affect their virality?. In ICWSM.
[9] Marco Guerini, Carlo Strapparava, and Gozde Ozbal. 2011. Exploring Text
Virality in Social Networks.. In ICWSM.
[10] Adrien Guille and Hakim Hacid. 2012. A predictive model for the temporal
dynamics of information diffusion in online social networks. In Proceedings
of the 21st international conference on World Wide Web. ACM, 1145–1152.
[11] Lars Kai Hansen, Adam Arvidsson, Finn Arup Nielsen, Elanor Colleoni,
and Michael Etter. 2011. Good friends, bad news-affect and virality in
twitter. Future information technology (2011), 34–43.
[12] Giannis Haralabopoulos, Ioannis Anagnostopoulos, and Sherali Zeadally.
2015. Lifespan and propagation of information in On-line Social Networks:
A case study based on Reddit. Journal of network and computer applica-
tions 56, (2015), 88–100. Elsevier.
10
[13] Dino Ienco, Francesco Bonchi, and Carlos Castillo. 2010. The meme ranking
problem: Maximizing microblogging virality. In Data Mining Workshops
(ICDMW), 2010 IEEE International Conference on. IEEE, 328–335.
[14] Jos´e Luis Iribarren and Esteban Moro. 2011. Affinity paths and information
diffusion in social networks. Social networks 33, 2 (2011), 134–142.
[15] Steve Jurvetson and Tim Draper. 1997. Viral marketing: Viral marketing
phenomenon explained. DFJ Network News (1997).
[16] Aditya Karnik, Atul Saroop, and Vivek Borkar. 2013. On the diffusion of
messages in on-line social networks. Performance Evaluation 70, 4 (2013),
271–285.
[17] Elihu Katz and Paul Felix Lazarsfeld. 1966. Personal Influence, The part
played by people in the flow of mass communications. Transaction Publish-
ers.
[18] Haewoon Kwak, Changhyun Lee, Hosung Park, and Sue Moon. 2010. What
is Twitter, a social network or a news media?. In Proceedings of the 19th
international conference on World wide web. ACM, 591–600.
[19] Paul Felix Lazarsfeld, Bernard Berelson, and Hazel Gaudet. 1948. The
peoples choice: how the voter makes up his mind in a presidential campaign.
(1948).
[20] Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos,
Jeanne VanBriesen, and Natalie Glance. 2007. Cost-effective outbreak de-
tection in networks. In Proceedings of the 13th ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM, 420–429.
[21] Yongcheng Li, Shuguang Huang, Chaoran Fan, and Guozheng Yang. 2013.
The selection of information diffusion monitoring nodes in directed online
social networks. In Proceedings of the 2012 international conference on in-
formation technology and software engineering. Springer, 533–540.
[22] Norman Long. 2003. Development sociology: actor perspectives. Routledge.
[23] Michael Mathioudakis, Francesco Bonchi, Carlos Castillo, Aristides Gionis,
and Antti Ukkonen. 2011. Sparsification of influence networks. In Proceed-
ings of the 17th ACM SIGKDD international conference on Knowledge
discovery and data mining. ACM, 529–537.
[24] Karine Nahon, Jeff Hemsley, Shawn Walker, and Muzammil Hussain. 2011.
Blogs: spinning a web of virality. In Proceedings of the 2011 iConference.
ACM, 348–355.
[25] Lance Porter and Guy J Golan. 2006. From subservient chickens to brawny
men: A comparison of viral advertising to television advertising. Journal
of Interactive Advertising 6, 2 (2006), 30–38.
11
[26] S Rajyalakshmi, Amitabha Bagchi, Soham Das, and Rudra M Tripathy.
2012. Topic diffusion and emergence of virality in social networks. arXiv
preprint arXiv:1202.2215 (2012).
[27] Daniel M Romero, Wojciech Galuba, Sitaram Asur, and Bernardo A Hu-
berman. 2011. Influence and passivity in social media. In Proceedings of
the 20th international conference companion on World wide web. ACM,
113–114.
[28] Daniel M Romero, Brendan Meeder, and Jon Kleinberg. 2011. Differences
in the mechanics of information diffusion across topics:
idioms, political
hashtags, and complex contagion on twitter. In Proceedings of the 20th
international conference on World wide web. ACM, 695–704.
[29] Takeshi Sakaki, Makoto Okazaki, and Yutaka Matsuo. 2010. Earthquake
shakes Twitter users: real-time event detection by social sensors. In Pro-
ceedings of the 19th international conference on World wide web. ACM,
851–860.
[30] Catherine E Tucker. 2014. The reach and persuasiveness of viral video ads.
Marketing Science 34, 2 (2014), 281–296.
[31] Lilian Weng, Jacob Ratkiewicz, Nicola Perra, Bruno Gon¸calves, Carlos
Castillo, Francesco Bonchi, Rossano Schifanella, Filippo Menczer, and
Alessandro Flammini. 2013. The role of information diffusion in the evolu-
tion of social networks. In Proceedings of the 19th ACM SIGKDD interna-
tional conference on Knowledge discovery and data mining. ACM, 356–364.
[32] Jiang Yang and Scott Counts. 2010. Predicting the Speed, Scale, and Range
of Information Diffusion in Twitter. ICWSM 10 (2010), 355–358.
[33] Jaewon Yang and Jure Leskovec. 2011. Patterns of temporal variation in
online media. In Proceedings of the fourth ACM international conference
on Web search and data mining. ACM, 177–186.
12
|
1302.4765 | 1 | 1302 | 2013-02-19T22:11:52 | Design Features for the Social Web: The Architecture of Deme | [
"cs.SI",
"cs.SE"
] | We characterize the "social Web" and argue for several features that are desirable for users of socially oriented web applications. We describe the architecture of Deme, a web content management system (WCMS) and extensible framework, and show how it implements these desired features. We then compare Deme on our desiderata with other web technologies: traditional HTML, previous open source WCMSs (illustrated by Drupal), commercial Web 2.0 applications, and open-source, object-oriented web application frameworks. The analysis suggests that a WCMS can be well suited to building social websites if it makes more of the features of object-oriented programming, such as polymorphism, and class inheritance, available to non-programmers in an accessible vocabulary. | cs.SI | cs | To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
Design Features for the Social Web:
The Architecture of Deme
Todd R. Davies and Mike D . Mintz
Symbolic Systems P rogram, S tan ford University , Stanford, California, 94305-2150 USA
[email protected] tanfo rd.edu, MikeM [email protected] rd.edu
Abstract. We characterize the “social Web” and argue for several features that are
desirable for users of socially o riented web applications. We describe the architec-
ture of Deme, a web content management sy stem (WCMS) and extensible frame-
work, and show how it implements these desired features. We then compare Deme
on our desiderata with other web technologies: traditional HTML, previou s open
source WCMS s (illustrated by Drupal), commercial Web 2.0 application s, and
open-source, object-oriented web application frameworks. The analysis suggests
that a WCMS can be well suited to building social websites if it makes more of the
features of object-o riented programm ing, such as polymorphism , and class inheri-
tance, availab le to nonprogrammers in an accessible vocabulary.
Keywords: conten t management sy stems, content management frameworks, social
web applications, user controlled data
1 Introduction
In early February of 2009, the social networking website Facebook.com changed its
Terms of Serv ice (ToS) agreement, apparently taking away users’ r ight to remove, and
thereby to revoke Facebook’s license to, their con ten t [16]. The move by Facebook
caused considerable con troversy, lead ing to a threatened lawsuit by privacy advocates ,
headline stories in major media, and an online revolt within Facebook by groups of its us-
ers. After several press commun iqués and blog posts by Facebook representatives defend-
ing the new ToS failed to convince skeptics, the company changed course and rever ted to
its previous Terms of Serv ice agreement [10].
The controversy over Facebook’s ToS showed both the g rowing importance of the
Web’s social functions, as ev idenced by the w idespread media coverage it received, and
also the widespread sensitiv ity of users to questions about who controls their data. Th is
paper exp lores how the arch itecture of a web application can reflect these and o ther goals
in the emerg ing, more socially orien ted Web. Af ter d iscussing use character istics that
distinguish social web applications, we define concep ts and desired features for the social
Web, and then describe the arch itecture of a new, socially oriented web conten t manage-
ment system (WCMS) and ex tensible framework we call Deme (pronounced “deem”),
which imp lements our approach to the challenges posed by our desiderata.
1
Todd R. Davies and Mike D. Min tz
The unusual characteristics of the Web that pose engineering challenges were enumer-
ated by Murugesan and G inige [11]. There are var ious possible approaches to web engi-
neering, includ ing model-driven approaches [e.g. 5,15], principle-based approaches such
as represen tational state transfer (REST)[8], and programming techn iques such as agile
programming [4]. Within specif ic programming language families, a notable development
has been the widespread use of web application frameworks such as Ruby on Rails . A
WCMS is a tool for building web applications without requiring website bu ilders to have
a background in computer programming.1 Our approach combines the user exper ience
goals of WCMSs with concep ts and princip les from object-orien ted web application
frameworks in an architecture built for the social Web .
2 Web 2.0 and the Social Web
The set of trends often referred to as “Web 2.0” – the second generation of web design
and developmen t – compr ise roughly the following [cf 9,14]:
− websites that act as rich Internet applications (RIAs), with substan tial and complex
processing of HTTP requests;
− heavier use of clien t-side processing, i.e. more code loaded into the browser and less
need for full page requests;
− larger websites supporting many in teracting users, such as Wik ipedia, Youtube, and
Facebook; and
− availab ility and use of new , higher- level programming tools such as web app lication
frameworks and rich Internet application frameworks.
In contrast to Web 2 .0, the earlier generation (“Web 1.0”) was based on the browser/client
requesting full HTML pages or med ia files from a server directory, thin clien ts, relatively
little in teractiv ity with users, and little server-side code. The transition between the two
was gradual, w ith some features of Web 2 .0 being present even on some ear ly websites .
Web conten t management systems pre-dated the introduction of the term “Web 2 .0” in
2004 [14], and their development has lagged beh ind the phenomena cited above. The
most popular general-purpose WCMSs today (Joomla/Mambo, Drupal, and Plone) were
all first released in 2001-’02, and since then site administrators have grav itated toward
newer versions of these tools rather than to newer WCMSs.2 The greater emphasis on
large (usually commercial) sites availab le to end users, on one hand, and developer tools
requiring programming skills, on the other , has made nonprogrammers who admin ister
websites less represented in the grow th of the Web than both end users and programmers.
But, as we will see, the characteristics of WCMSs are especially appropriate for the social
Web.
While “Web 2 .0” refers to a generation of technologies and trends that emerged in the
second decade of the Web’s existence, the term “social Web” refers to character istics of
1 See http ://www .contentmanager.eu.com/histo ry.htm.
2 See http://mameou.wordpress.com/2008/05/31/dries-buytaert-vision-of-drupal/. Wordpress, wh ich
is more specific to blog management, was introduced a few years later and also been very suc-
cessful. It could also be classified as a CMS , but is less general purpose than the other three tools
cited.
2
To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
web applications that pre-dated and will ou tlast the Web 2.0 era. The following use char-
acteristics distinguish social web applications from non-social ones [2,12]:
User-generated social content. Social web applications enable site visitors to submit
conten t that others can access, such as photos , their own profile data, links to other web-
sites, and comments on o ther users’ conten t.
Social network ing. Users of social web applications jo in together in online groups and
relationships (e.g. friends), which allow them to see iden tity-related information about the
people to whom they are connected .
Collaborat ion. Users engage in conversations , co-creation of content (e.g. on w ikis) ,
collaborative filtering , and collective action.
Cross-platform data sharing. Increasingly, shar ing content requires that a user be
able to transfer data across sites , imp lying that the site on which the remote content is to
be shared can interface correctly with the other site’s data. When the remo te data need to
be processed locally , the two sites must agree on its meaning, which is a def ining charac-
ter istic of the semantic Web [1,12].
A website need not exhibit all of the above characteristics in order to be considered so-
cial. For examp le, a newspaper b log may enable users to make commen ts, but w ith no
support for networking through member profiles or collaboration. For our purposes, the
main thing that defines the social Web is that it enables visitors, not just site admin istra-
tors, to contribu te some form of content that o ther users can access .
3 Content Management Concepts
We aim to show that our approach to content management provides advantages for the
social Web over other approaches. To do this, we will def ine a set of dimensions through
which content technologies can be compared . Consider the following concepts, with ex-
amples drawn primar ily from traditional web concep ts :
Unit. A chunk of content that can be referenced independently of other chunks, e.g. an
HTML f ile/webpage.
Subsegment. A content segment or portion of a unit; for example, a semantic element
in an HTML file.
Unit type. The classification of a unit of con ten t that def ines its subsegment structure
and constrains what v iewer code can be execu ted in order to display it, e.g. a MIME type.
Behav iors . A set of actions available to a user with respect to a given unit of content,
e.g. the GET, POST, PUT, and DELETE methods of HTTP .
Container. A data structure for grouping multiple units together , e.g . a directory on a
file server.
Type structure. The relationships between un it types , e.g. XHTML is a subtype of
XML.
Type-viewer matching . A system for specifying which v iew code to invoke for a
given unit type, e.g. the preferences-applications table in the Firefox browser.
Relation specif ier. A way to represent and disp lay a relationship between un its , e.g. a
hyperlink.
Access control. The system for specifying a user’s abilities to perform actions , such as
read and write, on a unit of content, e.g. through file system permissions.
3
Todd R. Davies and Mike D. Min tz
Addressing. The means of specifying a particu lar un it, e.g. the URL of a page.
Versioning. A system for stor ing previous versions of a unit, e.g. an arch ive of old
files.
Deletion method(s). A behav ior of a un it that results in it becoming h idden or re-
moved from use, e.g . the rm command in Un ix and the DELETE method of HTTP .
Software license. A legal agreement w ith users specify ing the rights and responsibili-
ties of bo th the user and the provider of a site’s software.
In the next section, we will argue for specific features along these dimensions that are
most compatible w ith the social Web.
4 Guiding Features for Social Web Content Management
The use characteristics of the social Web suggest preferred features from the perspective
of users (i.e. what users themselves are likely to want) . These guiding features can be de-
fined w ith respect to the conten t managemen t concep ts defined in section 3.
Units shou ld be page independent. A web page may contain many pieces of data to
which a user might want independen t access. For example, the comments at the bottom of
a blog posting should, logically , each be addressab le and includable ind ividually, as
should any subset of them. Under HTML, by con trast, the page is the unit of reference.
References to individual elemen ts are possib le through anchors or the DOM, but these are
still tied to a par ticular page. Database tables are page- independent, bu t users do not gen-
erally have page-independent access to them. WCMSs generally allow users or admin is-
trators to def ine page independent units , such as the “nodes” of Drupal,3 but commercial
Web 2.0 sites such as Youtube do not permit end users this capab ility for each conten t
unit, e.g. a par ticular comment. They requ ire a user to refer to a unit through a URL.
Subsegments should be fu lly po intab le. Units themselves can be made up of parts or
subsegmen ts. But a user or content manager in a social web application may want a sub-
segmen t to be a reference to ano ther unit, or to a subsegment within a unit. We call this
feature “full poin tability”. The “creator” subsegmen t of a content unit might poin t, for ex-
ample, to a unit represen ting a particular person. If the unit describ ing a person is div ided
into their name, email address , telephone number , etc., one of their fr iends might want to
place their email address, w ithout any of the other information about that person, in a list
that will dynamically update whenever the email address changes . Although database-
driven web applications generally allow references to o ther units, and the WCMS Drupal
supports pointing to subsegments (node fields) in its Conten t Constructor Kit module,4 in
general commercial Web 2 .0 sites do not allow the user to refer d irectly to con tent fields
or subsegmen ts.
Unit types should be polymorphic. Polymorphism refers to the ability of a unit of data
of one type to be treated as hav ing a different type [3]. Th is is important for the social
Web because, as no ted in section 2, when one person shares conten t with another person,
they may not be on the same platform , and so the code necessary to view a specific type
may not be available to every user. Po lymorphism exists for a unit when its type is a sub-
3 See http ://drupal.o rg/node/19828.
4 See http ://drupal.o rg/project/fieldreference.
4
To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
type of an understood supertype (e.g. an HTML file w ith microformat markup can be
rendered by any HTML interpreter). The Web 2.0 emphasis on both server- and clien t-
side processing can break polymorphism for social web applications, because a g iven
user’s browser may not be compatib le w ith a specific type of content. Cross-p latform data
sharing may also be impossible between differen t applications or even different installa-
tions of the same app lication.
Behaviors should be extensible. In general, a designer of a website w ill not anticipate
all the possible actions and sequences of actions a user may want to do. For examp le, a
user of a search engine may want to sort the results by a cr iterion for which there is no
widget on the site, or a reader of a message board migh t want to view all the posts by
people from a cer tain city. Extensib ility implies that the user can create new behaviors for
a given unit type. Bu t this depends on an ability to modify either the data model or the
view code. WCMSs such as Drupal provide this ability by allowing modif ications of the
open source code, and through optional contributed modules . Commercial Web 2 .0 sites ,
by contrast, often leave the user w ithout a way to add a desired behavior when the code is
not available for modif ication. For examp le, users of many v ideo shar ing sites cannot sort
videos by date.
Containers should be referen tia l. In a tree-structured file system paradigm , the con-
tainer (a folder or directory) stores a copy of each unit (a file), and every file must be
stored in a folder . As Ted Nelson has pointed out [13], th inking this way results in unnec-
essary f ile duplication that can cause incompatibilities . Referential containers , by con-
trast, store only the addresses of content units , and units are stored separately. Th is form
of container is better suited to the Web than value containers are, because differen t people
have d ifferen t ways of categorizing content, and with reference-based categor ies , they
need not in terfere w ith each o ther. Referential containers allow for a single poin t of stor-
age, rather than copies that can be updated d ifferentially and become inconsistent. Social
web applications emphasize practices such as the shar ing of tags or labels , which are ref-
erence containers .
Type structure should be inheritance hierarchical. In Drupal, conten t types such as
“article” and “even t” are def ined in a flat hierarchy , with configurable options, but w ith-
out inheritance of structure from other types.5 A site developer who uses Drupal ex-
pressed the need for type inheritance in the fo llow ing blog commen t in December 20086:
“…the structure of most of our con ten t types is similar, or close enough that much of
the template is the same for all of them . Most of the code in each temp late is repeated
from one to the next. I really w ish there was some kind of content type hierarchy or in-
heritance so those types of proper ties could be passed on to "children" con ten t types.”
The trad itional structure of web pages is very nonhierarchical. Internet media types can
associate different actions to different types of conten t in the browser, bu t they do not ex-
ist in an inheritance hierarchy. Content type inheritance, wh ich is found in only a few en-
terprise CMSs such as Alfresco and Documentum (and no t, to our knowledge, in general
use Web CMSs) has been called by one blogger “the holy grail of conten t management”.7
Type-viewer matching should be server-side specialized. The advantage of content typ-
ing is only fu lly realized if each con tent type is associated with v iew code that is tailored
5 See http ://drupal.o rg/project/inherit regarding experimental conten t type inheritance in Drupal.
6 See http ://www .yelvington.com/node/517.
7 See http ://gadgetopia.com/post/6360.
5
Todd R. Davies and Mike D. Min tz
to that type. This makes it possib le to tailor the user in terface exper ience to the content
type. WCMSs that allow content typing generally specialize the v iew by type to some ex-
tent, but when, as in Drupal, a user can create new conten t types by filling out a form, the
view code must be gener ic enough across con tent types to allow the def inition of a new
type withou t writing code for an associated viewer . This limits the ex tent to which the
viewer can be specialized . Again , web browsers trad itionally render all content as pages
of HTML, and Internet med ia/MIME types are handled differently by each browser . If a
site administrator wan ts all (or near ly all) end users to be able to view content in a spe-
cialized way, the v iew code must be def ined at the server level.
Relation specifiers should be in tegrally un itizab le. Relations between con tent units
such as hyper links have traditionally been specified within one unit, po inting to ano ther
unit. This leads to a basic structure for links that is one-way , which can make it diff icu lt
to detect incoming links. Moreover, specify ing relations with in a unit usually requ ires that
the user specifying the relation have write pr ivileges for the target unit, and that these re-
lations be visible to all users viewing the target unit, such as a web page. In social web
applications , on the other hand, users may wish to specify relations between units in a
way that is external to the related units. For examp le, a user may w ish to inser t a com-
ment at a particu lar location in a document, which will have differen t associated permis-
sions from those of the document itself. Th is is especially useful when referring to docu-
ments on other websites . A lthough even basic HTML supports linking to an external
page, th is kind of reference specifier is not in tegrated within the application, because it is
not visible when v iew ing the referenced pages. A solu tion is to allow relations between
specific locations in con ten t units to be specif ied as units themselves, w ith their own per-
missions, and an integrated tool for displaying references when view ing an item . WCMSs
such as Drupal generally support th is8 but not for the general case of relationships be-
tween locations within units,9
Access contro l should be flu id-granular . Web applications generally provide much
coarser con trol over who can view, ed it, and delete content than does an operating sys-
tem . But th is type of control is what users typically want, because each piece of conten t is
different. Moreover, the ideal pr ivilege def initions are even more complex than in operat-
ing systems , since they can be defined for an arbitrary number of groups, with compli-
cated ru les of precedence, and for each subsegment (field) o f each conten t unit. Commer-
cial websites generally g ive limited control to the user to define these permissions ,
although social networking sites such as Facebook have evolved to be fairly granular.10
Drupal embodies fluid ity through the ability to def ine an indefinite number of roles , or
packages of privileges, and its Content Creator Kit module makes f ield-level permissions
availab le as well.11 But the comb ination of fluid ity (many distinctions between adminis-
trator and user) and granularity (control over each field o f a unit) is very difficult to
achieve and generally not found in commercial Web 2.0 sites .
Addressing should be domain independen t. As much as possible, conten t should be
addressab le independen tly of its path, so that links will no t break if the content moves.
8 See http ://drupal.o rg/node/414018.
9 A well-known advocate of more flex ible reference specifiers, w ith support fo r two-way links and
deep transclusion, has been Ted Nelson [13].
10 See http ://www .allfacebook.com/2009/04/facebook-privacy -limitations/.
11 See http ://drupal.o rg/node/310 and http ://drupal.org/p roject/cck_field_perms.
6
To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
This can be imp lemen ted through redirects , but that depends on the con tent owner’s con-
trol over a domain, since URLs are tied to domain names. Commercial web applications
generally do not support domain- independen t addressing, but Drupal does support a lim-
ited version of it, internally to a site, through the node ID combined w ith the ability to
move the database to ano ther domain .
Versioning should be comprehensive. Since, on a social web site users are provid ing
conten t, they may need access to earlier versions of a unit. This is built into Drupal12 as
well as w iki sites, but is generally not available on commercial Web 2.0 sites .
Deletion methods should be user con trolled . In a social web application, a user
uploads content to the host site in lieu of placing it on their own site. For commercial
Web 2.0 sites , th is means that the user’s ab ility to remove content is limited by the tools
provided to users, and by the ToS agreement. As in the Facebook ToS controversy , this
can lead users to feel that they have lost contro l over their own data, and may pose pri-
vacy risks . A site should not unduly limit users’ ability to delete their own data, e.g. by
making true deletion impossible (as opposed to flagging the data as h idden in a database).
Drupal makes true deletion availab le to users,13 but commercial sites generally do not.
Software licenses should be free/open source. Another aspect of user contro l is the
ability to inspect and modify the code. Although this generally requires moving data to
one’s own server space, and most users w ill not want to do it, a free/open source platform
gives all users f lexib ility by enabling others to provide alternative hosting environmen ts
for their data.
5 The Deme Architecture
In this section, we describe the arch itecture of Deme,14 our new WCMS and framework
written in Django/Python, w ith a PostgreSQL database, licensed under the Affero GPLv3
license.15 Recently, the term “content management framework” has been used, somewhat
controversially, to denote “an application programming in terface for creating a custom-
ized content management system”.16 We use the term “framework” to indicate that the
system is designed to facilitate custom code developmen t. Deme attempts to make avail-
able the concep ts of ob ject-oriented programming (OOP) to end users and nonprogram-
mer website administrators, using language that we believe will be more understandable
to nonprogrammers. We def ine the terminology of Deme below with respect to concepts
familiar to a technical aud ience. Desired features from section 4 are noted in bold ita lics.
Items and item types . Un its of con ten t in Deme are stored in “items”. An item is an
instance of a par ticular “item type”. The Deme item types are inheritance hierarchical. If
the Person item type inherits from the Agent item type, then any item that is a Per-
son is also an Agent. Every item type ultimately inherits from the Item item type
12 See http ://drupal.o rg/node/70591.
13 See http ://agaric.com/note/disable-delete-regular-u sers.
14 See http ://deme.stan ford.edu.
15 See http ://www .gnu.org/licenses/agpl-3.0.htm l.
16 See fo r example
http://en.w ikipedia.org/w/index.php?title=List_of_content_management_framewo rks&old id=282
731961
7
Todd R. Davies and Mike D. Min tz
(which corresponds to the Object class in many programming languages). We allow mu l-
tip le inheritance, and use it occasionally (e.g., TextComment inher its from both Com-
ment and TextDocument). Deme items are stored in a database using object relational
mapping (ORM) 17 with mu lti-table inher itance. For example, if our item type h ierarchy is
Item -> Agent -> Person, and our items are Mike[Person] and Robot[Agent],
then there w ill be one row in the Person tab le (for Mike), two rows in the Agent table
(for Mike and Robot), and two rows in the Item tab le (for Mike and Robot). An
abridged basic v iew of the Deme item type hierarchy is shown in F igure 1.
Fig. 1. The Deme item type hierarchy (abridged basic view ).
Pieces. Every item type def ines the “pieces” (mapped to f ields/columns in the data-
base) relevan t for that type’s items, and item types inherit p ieces from their supertypes. If
Item defines the description piece, Agent defines no new pieces, and Person de-
fines the first_name piece, then every person has a description and a
first_name.
17 See Sco tt W. Ambler’s explanation ORM at
http://www .ag iledata.org/essay s/mappingObjects.htm l.
8
To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
Piece types . Every piece of an item has a type (e.g. String, Integer, and Boo-
lean). Pieces can point to other items (foreign keys in the database), and can addition-
ally specify which piece of another item is being pointed to (fully poin table) . Pointing
pieces are usefu l for defining relationsh ips between items. For example, the Item item
type has a creator piece pointing to the Agent that created it. Multiple items can
point to a common item . Pieces cannot store data structures like lists. So rather than stor-
ing differen t contact methods as pieces of each Agent, we make ContactMethod an
item type, and give it an agent_pointer piece. The contact methods for agent 123
are represented by all of the ContactMethods that have agent_pointer equal to
123.
Item IDs . The most importan t piece of an item is the id (a page-independent primary
key). Every item has a unique, immutable id. Items share the same id w ith their su-
pertype versions (so Mike’s row in the Person table has the same id as Mike’s row in
the Agent and Item tables) . Pointing pieces are references to the id of the pointed-to
item. A lthough not imp lemen ted yet, we p lan to make available (optionally) “universal
item id’s” through a reserved namespace approach like URNs or i-names , for doma in-
independen t addressability across installations of Deme. Relations between items and
pieces are shown for a portion of the item type hierarchy in F igure 2.
Other h ighligh ts of the Deme architecture include the following.
Versioning. For every item type, there is a comprehensive “old versions” tab le.
Deleting items . There are two ways to delete items: deactivate and destroy.
Decativating is recoverable (through reactivate), but destroy is not (user control).
The user interface ensures that deactivating happens before destroying.
Collect ion. An item type that represents an unordered referential set of o ther items ,
Collections use pointers from Memberships (which are items in their own right; in-
tegra l un itizability) to represen t their con ten ts, so multip le Collections can poin t to
the same contained items. Collections “directly” contain items via Memberships ,
but they also “ind irectly” contain items via chained Memberships.
Transclusion. An embedded reference from a location in a specif ic version of a
TextDocument to ano ther Item . A Transclusion [13] is a separate item (in tegra l
unitizab ility).
Comment. A unit of discussion about an item. Each Comment specif ies the com-
mented item and item_version_number. Comments can be associated with spe-
cif ic locations in a TextDocument v ia Transclusions.
Excerpt . An item that refers to a portion of another item or (in a planned future ver-
sion) an external resource, such as a webpage (fu ll poin tability).
Permissions. Permissions define what actions an arbitrary g roup of Agents (fluid ity)
can and cannot do to each item and its pieces (granu larity).
Viewer types. Deme takes advantage of the model-view-temp late architecture of
Django. A viewer is a Python class that processes browser or API requests . Each viewer
defines the item type it can accept (server side specia lization), and mu ltiple viewers can
accept the same item type. V iewers that accept an item type will also accept sub types of
that item (polymorph ism). Each viewer type def ines a set of actions , e.g. item_show .
Custom viewers and item types can define new actions (extensibility).18
18 There are many other item types and architectural featu res not discussed above. For a fu ll de-
scription, see http://deme.stanford.edu/static/docs/index.html.
9
Fig. 2. A detailed partial view of the Deme item type hierarchy. Solid connecto rs denote supertype-
subtype inheritance. Do tted connecto rs denote pointers from pieces to items.
Todd R. Davies and Mike D. Min tz
6 Comparing Deme With Other Technologies for the Social Web
Table 1 summarizes how Deme achieves the desired features for the social Web , by com-
parison with other web technologies: (a) file system-based “Web 1.0” sites (basic
HTML); (b) the w idely used WCMS Drupal; (c) commercial Web 2.0 sites such as You-
tube, Facebook, and Myspace; and (d) object-orien ted web application frameworks such
as Ruby on Rails and Django.
On three dimensions (unit type, type structure, and addressing), the only other technol-
ogy besides Deme that achieves the desired feature is OOP /web applicaton frameworks,
which require programming skill. On eigh t dimensions (unit, subsegment, behav iors, rela-
tion specif iers , access control, versioning, deletion methods, and software license) , all of
the open-source approaches (Drupal, web frameworks, and Deme) achieve the desired
features, but commercial Web 2.0 sites do not. The remaining two dimensions (con tainer
and type-viewer) are ones for which all of the technolog ies beyond basic HTML achieve
the desired feature.
The social Web is especially associated with commercial Web 2.0 sites. Bu t our analy-
sis suggests that these sites do not meet users’ needs as well as open-source technologies
that g ive more con trol to users . Previous WCMSs, represented here by Drupal, generally
exhibit more of the desired features than large commercial sites do, and they do no t re-
quire a programming background to administer them. Bu t they do not meet the desired
social web cr iteria quite as well as OOP web frameworks do. The frameworks, on the
10
To appear in 8th Int’l Wo rkshop on Web-Orien ted Software Techno logies (IWWOST 2009)
other hand, require more programming skill. Deme makes availab le powerfu l OOP con-
cepts from web frameworks to nonprogrammers for managing content, in a code base that
is built for modif ication. We also believe that the terminology used in Deme w ill make it
easier for nonprogrammers to learn than Drupal, but that remains to be tested empirically.
Table 1. Comparison o f web technologies by content management concept. Approaches that make
available the desired feature fo r each content concept are highlighted in bold.
Content
Desired
(a) File sys-
(b) Web
(c)Commer-
(d) OOP
(e) Deme
/Web app
cial Web
CMS
tem/Web
social
managment
v0.9 WCMS
2.0 sites
feature
frameworks
(Drupal)
1.0 HTML
concep t
object/row
node
item
page
photo,
file/page
unit
video, etc.
independen t
field
attribute/
piece,
fully
semantic
custom
subsegment
field
excerp t
pointable
fields
element
class
item type
polymor-
content type custom
Internet
unit type
phism
media type
types
menus
methods
actions
exten sible
HTTP
widgets
behaviors
methods
container
tags/labels
categories
collection
referential
directo ry
container
classes
class
item type
inheritance
(flat)
(flat)
MIME type
type
inheritance
hierarchy
hierarchy
/subtype
structu re
site-defined
views and
model-view
viewer
server-side
browser
type-viewer
separa tion
modules
viewer
types
specialized
application
matching
preferences
relation
relation
transclu-
integ rally
limited
one-way
relation
nodes
objects
sions,
unitizable
hyperlinks
bidirectional
specifiers
member-
links
ships
customiza-
perm ission s
fluid-
custom
ble
granular
permissions
object
(universal)
domain
permalink
identiy
item id
independen t
version
old ver-
comprehen-
none or
wiki diffs
contro l sys-
sions table
sive
tem
file edit and
deactivate,
user
file system
limited data
deletion
delete
destroy
controlled
delete
removal
methods
open
Affero
free/open-
default
usually
softwa re
source
GPLv3
source
copyright
proprietary
license
The version of Deme presented here is the latest step in a multi-year project aimed at
creating a p latform for deliberative in teractions , e.g. document-cen tered discussion [6].
Future work will involve ref ining the interface to enable easier collaboration and com-
menting. The social Web is ultimately about foster ing conversation. In the words of Cory
Doctorow [7], “Conversation is king. Content is just something to talk about.” Users are
likely to continue to wan t this conversation to ex tend to an open dialogue abou t the social
web platform itself. Even the most techn ically minded of tool providers should be pre-
admins and
roles
node ID
conten t
versioning
node delete
GPLv2
access
control
addressing
versioning
restricted
directo ries
URL
old files
11
Todd R. Davies and Mike D. Min tz
pared to justify their design and licensing cho ices to end users in relation to their needs
and desires , and to provide technology that is responsive to user demands.
Acknowledgments
We wish to thank Leo Perry, Ben Newman, Brendan O’Connor, Joseph Marrama, Jane
Huang, and Ivan Sag for helpful contribu tions to this version of Deme.
References
1. Berners-Lee, T., Hendler, J., Lassila, O . The Semantic Web. Sci. Am. 284(5), 34-43 (2001)
2. Boulos, M .N., Wheelert, S. The Emerging Web 2.0 Social Software: An Enabling Suite o f So-
ciable Technologies in Health and Health Care Education. Health Information and Libraries
Journal. 24, 2 -23 (2007)
3. Cardelli, L, Wegner, P. On Understanding Types, Data Abstraction , and Polymorphism. Com-
puting Surveys. 17(4), 471-522 (1985)
4. Cohen, D., Lindvall, M ., Costa, P. An Introduction to Agile Methods. In Advances in Comput-
ers. pp. 1-66. E lsevier Science (2004)
5. Constantine, L.L., Lockwood, L.A.D. Usage-Cen tered Engineering for Web Applications.
IEEE Software. 19(2), 42-50 (2002)
6. Davies, T., Newman, B., O'Connor, B. Tam, A, Perry , L. Document Centered Discu ssion: A
Design Pattern for Online Deliberation. In Schuler, D. (ed.). L iberating Voices: A Pattern Lan-
guage fo r Communication Revolution. pp. 384-386. MIT Press (2008)
7. Doctorow, C. D isney Exec: Piracy Is Just a Business Model. October 10 (2006),
http://www .boingboing.net/2006/10/10/disney-exec-p iracy -i.h tml, quoted in Shirky, C. Here
Comes Everybody. Penguin (2008)
8. Fielding, R.T.; Tay lor, R .N.. Principled Design of the Modern Web Arch itecture. ACM Trans-
actions on Internet Technology. 2(2), 115–150 (2002)
9. MacManus, R ., Porter, J. Web 2 .0 for Designers. Digital Web Magazine, May 4 (2005)
10. Milian, M., Facebook Backtracks on Terms of Use Update. Lo s Angeles Times. February 18
(2009). , h ttp://latimesblogs.latimes.com/technology/2009/02/facebook-tos.htm l
11. Murugesan, S ., G inige, A . Web Engineering: Introduction and Perspectives. Idea Group (2005)
12. Mikroyannidis, A. Toward a Semantic Social Web . Computer. 40(11), 113-115 (2007)
13. Nelson, T.H. Xanalogical Structure, Needed Now More than Ever: Parallel Documents, Deep
Links to Content, Deep Versioning, and Deep Re-Use. ACM Computing Surveys. 31(4 )
(1999)
14. O’Reilly , T. What Is Web 2.0? September 30 (2005),
http://www .oreillynet.com/pub/a/oreilly/tim/news/2005/09/30/what-is-web -20.html
15. Rossi, G. Urbieta, M., Ginzburg, J., Distante, D., Garrido, A. Refactoring to Rich Internet Ap-
plications: A Model-D riven Approach. In 8th In ternational Conference on Web Engineering.
pp. 1-12. IEEE (2008)
16. Walters, C . Facebook’s New Terms of Service: We Can Do Whatever We Want With Your
Content, Fo rever. The Consumerist. Feb ruary 15 (2009).
http://consumerist.com /5150175/facebooks-new -terms-o f-service-we-can -do-anyth ing-we-
want-with -your-content-fo rever
12
|
1504.05163 | 1 | 1504 | 2015-04-20T19:19:10 | Trend of Narratives in the Age of Misinformation | [
"cs.SI",
"cs.HC",
"physics.soc-ph"
] | Social media enabled a direct path from producer to consumer of contents changing the way users get informed, debate, and shape their worldviews. Such a {\em disintermediation} weakened consensus on social relevant issues in favor of rumors, mistrust, and fomented conspiracy thinking -- e.g., chem-trails inducing global warming, the link between vaccines and autism, or the New World Order conspiracy.
In this work, we study through a thorough quantitative analysis how different conspiracy topics are consumed in the Italian Facebook. By means of a semi-automatic topic extraction strategy, we show that the most discussed contents semantically refer to four specific categories: {\em environment}, {\em diet}, {\em health}, and {\em geopolitics}. We find similar patterns by comparing users activity (likes and comments) on posts belonging to different semantic categories. However, if we focus on the lifetime -- i.e., the distance in time between the first and the last comment for each user -- we notice a remarkable difference within narratives -- e.g., users polarized on geopolitics are more persistent in commenting, whereas the less persistent are those focused on diet related topics. Finally, we model users mobility across various topics finding that the more a user is active, the more he is likely to join all topics. Once inside a conspiracy narrative users tend to embrace the overall corpus. | cs.SI | cs |
Trend of Narratives in the Age of Misinformation
Alessandro Bessi1, 2(cid:89), Fabiana Zollo2,(cid:89), Michela Del Vicario2,¤a, Antonio Scala2,3‡,
Guido Caldarelli2,3,4‡, Walter Quattrociocchi2,*,
1 IUSS Pavia, Italy
2 IMT Lucca, Italy
3 ISC-CNR, Rome, Italy
4 LIMS, London, UK
* [email protected]
Abstract
Social media enabled a direct path from producer to consumer of contents changing the
way users get informed, debate, and shape their worldviews. Such a disintermediation
weakened consensus on social relevant issues in favor of rumors, mistrust, and fomented
conspiracy thinking – e.g., chem-trails inducing global warming, the link between
vaccines and autism, or the New World Order conspiracy.
In this work, we study through a thorough quantitative analysis how different
conspiracy topics are consumed in the Italian Facebook. By means of a semi-automatic
topic extraction strategy, we show that the most discussed contents semantically refer to
four specific categories: environment, diet, health, and geopolitics. We find similar
patterns by comparing users activity (likes and comments) on posts belonging to
different semantic categories. However, if we focus on the lifetime – i.e., the distance in
time between the first and the last comment for each user – we notice a remarkable
difference within narratives – e.g., users polarized on geopolitics are more persistent in
commenting, whereas the less persistent are those focused on diet related topics. Finally,
we model users mobility across various topics finding that the more a user is active, the
more he is likely to join all topics. Once inside a conspiracy narrative users tend to
embrace the overall corpus.
Introduction
According to [1], causation is bound to the way communities attempt to make sense to
events and facts. Such a phenomenon is particularly evident on the web where users,
immersed in homophile and polarized clusters, process information through a shared
system of meaning [2, 3]. Indeed, social media enabled a direct path from producers to
consumers of contents – i.e., disintermediation – changing the way users get informed,
debate, and shape their opinions [4–8]. However, confusion about causation encourages
speculation, rumors, and mistrust [9]. In 2011 a blogger claimed that Global Warming
was a fraud aimed at diminishing liberty and democratic tradition [10]. More recently,
rumors about Ebola caused disruption to health-care workers [11–13]. Such a scenario
fostered the production of an impressive amount of conspiracy-like narratives aimed at
explaining reality and its phenomena, and provide an unprecedented opportunity to
study the dynamics of narratives’ emergence, production, and popularity.
1/18
Recently, we pointed out that the more the users are exposed to unsubstantiated
rumors, the more they are likely to jump the credulity barrier [2, 3]. As pointed out
by [14], individuals can be uninformed or misinformed, and the actual means of
corrections in the diffusion and formation of biased beliefs are not effective. In
particular, in [15] online debunking campaigns have been shown to create a
reinforcement effect in usual consumers of conspiracy stories. Narratives grounded on
conspiracy theories have a social role in simplifying causation because they tend to
reduce the complexity of reality and are able at the same time to contain the
uncertainty they generate [16–18]. Conspiracy theories create a climate of
disengagement from mainstream society and from officially recommended practices [19] –
e.g. vaccinations, diet, etc.
Despite the enthusiastic rhetoric about the collective intelligence [20, 21] the role of
socio-technical systems in enforcing informed debates and their effects on the public
opinion still remain unclear. However, the World Economic Forum listed massive digital
misinformation as one of the main risks for modern society [22]. A multitude of
mechanisms animates the flow and acceptance of false rumors, which in turn create false
beliefs that are rarely corrected once adopted by an individual [23–26]. The process of
acceptance of a claim (whether documented or not) may be altered by normative social
influence or by the coherence with the system of beliefs of the individual [27, 28]. A
large body of literature addresses the study of social dynamics on socio-technical
systems from social contagion up to social reinforcement [6, 7, 29–36].
In this work we analyze a collection of conspiracy news sources in the Italian
Facebook. In particular, we focus on the emergence of conspiracy topics and the way
they are consumed. We identify pages diffusing conspiracy news – i.e. pages promoting
contents neglected by main stream media. We define the space of our investigation with
the help of Facebook groups very active in debunking conspiracy theses (Protesi di
Protesi di Complotto, Che vuol dire reale, La menzogna diventa verita e passa alla
storia). We categorize pages according to their contents and their self description.
Concerning conspiracy news pages, their self claimed mission is to inform people
about topics neglected by main stream media. Through a semi-automatic topic
extraction strategy, we show that the most discussed contents refer to four well specified
topics: environment, diet, health, and geopolitics. Such topics are consumed in a very
similar way by their respective audience – i.e, users activity in terms of likes and
comments on posts belonging to different categories are similar and resolves in
comparable information consumption patterns. Conversely, if we focus on the lifetime
–i.e., the distance in time between the first and the last comment for each user – we
notice a remarkable difference within topics. Users polarized on geopolitics subjects are
the most persistent in commenting, whereas the less persistent users are those focused
on diet narratives. Finally we analyze mobility of users across topics. Our findings show
that users can jump independently from one topic to another, and such a probability
increases with the user engagement. This work provides important insights about the
fruition of conspiracy like rumors in online social media and more generally about the
mechanisms behind misinformation diffusion.
Results and Discussion
The analysis aims at characterizing the topical space in the conspiracy corpus of the
Italian Facebook. We start our investigation by outlining the emerging topics and then
we focus on the information consumption patterns. Finally we provide a data-driven
model of users information consumption patterns.
Details about the mathematical and statistical tools as well as the data used in the
analysis are described in Methods section.
2/18
Topics extraction and validation
As a first step in our analysis we apply a semi-automatic topic extraction strategy
aimed at classifying content.
We have 205, 703 posts (98.62% of the total corpus of conspiracy posts) containing a
message – i.e. a simple text or a description of the associated photo, video, or link. We
build a Document-Term matrix (205, 703 posts × 216, 696 terms) and take all the terms
with more than 500 occurrences (1820). We then apply a supervised preprocessing in
order to identify terms related to the conspiracy storytelling. Such a supervised task is
performed by 20 volunteers separately1. The resulting set is composed by 159 terms.
Then, we derive the co-occurrence network of conspiracy terms – i.e., a graph where
nodes are conspiracy terms, edges bond two nodes if the corresponding terms are found
in the same post, and weights associated to edges indicate the number of times the two
terms appear together in the corpus. Such a graph has 159 nodes and 11, 840 edges.
Since the co-occurrence network is a dense graph, we apply a filter in order to
remove edges which are not significant. More precisely, we apply the disparity filter
algorithm [37] (see Methods section for details) to extract the network backbone
structure, thus reducing the number of edges while preserving its multi-scale nature.
The application of the filtering algorithm with a statistical significance level of α = 0.05
results in a graph with 159 nodes and 1, 126 edges. We asked to the volunteers to give a
generic class to each term. By accounting only for 90% of concordance within volunteers
tags on names, the semantics of conspiracy terms results in four main categories:
environment, health, diet, and geopolitics. In Figure 1 we show the backbone of the
co-occurrence term network, where different colors indicate nodes belonging to different
conspiracy class.
Figure 1. Backbone of conspiracy terms co-occurence network. Different
colors indicate nodes belonging to different conspiracy topics according to our manual
classification. In particular, purple nodes belong to geopolitics, red nodes to
environment, blue nodes to health, and green to diet.
1We consider as conspiracy terms only those terms labeled as conspiratorial by at least the 90% of
volunteers.
3/18
HEALTHGEOPOLITICSENVIRONMENTDIET To validate the classification, we apply three different community detection
algorithms – i.e., Walktrap [38], Multilevel [39], and Fast greedy [40] (see Methods
section for further details) – to the backbone of conspiracy terms co-occurence network.
In Figure 2 we show the classification provided by each community detection
algorithm. Multilevel and Walktrap algorithms assign each term to the same community
and their accuracy with respect to our manual classification is 100%, while the
concordance index of the Fast greedy algorithm is 88.68%.
(a) Walktrap
(b) Multilevel
(c) Fast greedy
Figure 2. Communities of conspiracy terms. Membership of conspiracy terms
according to different community detection algorithms. Purple nodes belong to
geopolitics, red nodes to environment, blue nodes to health, and green to diet.
We assign a post to a given topic according to the term in it. In case of terms
belonging to different topics, we apply the majority rule, in case of ties, the post is not
labeled. Through such a criterion we are able to label 44, 259 posts – i.e. 9, 137
environment posts, 8, 668 health posts, 3, 762 diet posts, and 22, 692 geopolitics posts.
Attention patterns
Content consumption
In order to characterize how information belonging to the different categories of
conspiracy topics are consumed, we perform a quantitative analysis on users’
interactions – i.e. likes, shares, and comments2.
In Figure 3 we show the complementary cumulative distribution functions (CCDFs)3
of the number of likes, comments, and shares received by posts belonging to different
conspiracy categories of topics. Likes, comments, and shares of posts are long-tailed
distributed and best fitted by discrete power law distributions. To further characterize
differences within the distributions, in Table 1 we summarize the estimated lower
bounds and scaling parameters for each distribution.
2Notice that each of these actions has a particular meaning [41]. A like stands for a positive feedback
to the post; a share expresses the will to increase the visibility of a given information; and a comment is
the way in which online collective debates take form around the topic promoted by posts. Comments
may contain negative or positive feedbacks with respect to the post.
3We prefer to show CCDFs rather than CDFs since the former allow to plot heavy-tailed distributions
in doubly logarithmic axes, and thus emphasize their long tail behavior.
4/18
(a)
(b)
(c)
Figure 3. Topics attention patterns. Complementary cumulative distribution
functions (CCDFs) of the number of likes (a), comments (b), and shares (c) received by
posts belonging to different conspiracy topics.
Environment
Health
Diet
Geopolitics
Likes
xmin
142
172
135
167
α
2.82
2.68
2.84
2.36
α
Comments
xmin
42
37
15
135
2.82
2.59
2.36
3.14
Shares
α
xmin
408
435
358
407
2.62
2.39
2.59
2.25
Table 1. Power law fit of conspiracy topics attention patterns. Lower bounds
and scaling parameters estimates for the distributions of the number of likes, comments,
and shares received by posts belonging to different conspiracy topics.
To analyze lifetime of posts from different categories, we compute the temporal
distance between the first and last comment for each post. In Figure 4 we show the
Kaplan-Meier estimates of survival functions (see Methods for further details) for posts
belonging to different conspiracy topics. The p-value associated to the Gehan-Wilcoxon
test (a modification of the log-rank test) is p = 0.091, which lets us conclude that there
are not significant statistical differences between the survival functions of posts
belonging to different conspiracy topics.
5/18
likesCCDFllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll10010110210310410510-510-410-310-210-1100lEnvironmentHealthDietGeopoliticscommentsCCDFlllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll10010110210310410-510-410-310-210-1100lEnvironmentHealthDietGeopoliticssharesCCDFlllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll10010110210310410510-510-410-310-210-1100lEnvironmentHealthDietGeopolitics Figure 4. Lifetime of conspiracy topics. Kaplan-Meier estimates of survival
functions of posts belonging to different conspiracy topics.
Our findings show that conspiracy topics are consumed in a similar way. In
particular, we find that survival functions of posts belonging to different conspiracy
topics do not show different statistical signatures.
Users activity
Here, we consider users’ attention patterns with respect to different conspiracy topics by
analyzing the number of likes and comments, as well as the lifetime of each user – i.e.
the temporal distance between his first comment and last comment on a post belonging
to a specific category – that can be intended as an approximation of users persistence in
online collective debating.
We consider as conspiracy users those whose liking activity on conspiracy pages is
greater than the 95% of their total liking activity – i.e. a conspiracy user left at most
5% of her likes on posts belonging to science pages. Such a method allows to identify
790, 899 conspiracy users. Moreover, we consider a user polarized towards a given
conspiracy topic if she has more than the 95% of her likes on posts belonging to that
topic. Such a criterion allows to classify 232, 505 users (29.39% of the total). Table 2
summarizes the classification task’s results. We observe that the majority of users is
concerned about conspiracy stories related to geopolitics (62.95%), whereas conspiracy
narratives about environment (18.39%) and health (12.73%) attract a smaller yet
substantial number of users, while diet (5.94%) seems to be considered a less attractive
subject. However, notice that here we are analyzing conspiracy users that have the most
of their total activity on a given topic. Indeed, Figure 3 shows that diet receives as
much attention as the other conspiracy topics, even if it is not a polarizing topic, as
shown in Table 2.
In Figure 5 we show the CCDFs of the number of likes and comments of users
polarized towards different conspiracy topics. We observe minor yet significant
6/18
t (days)S(t)0102030405000.20.40.60.81EnvironmentHealthDietGeopolitics Geopolitics
Environment
Health
Diet
Users
146, 359
42, 750
29, 587
13, 807
%
62.95
18.39
12.73
5.94
Table 2. Polarization of users towards different conspiracy topics.
differences between attention patterns of different conspiracy users. Table 3 summarizes
the estimated lower bounds and scaling parameters for each distribution. These results
show that users polarized towards different conspiracy topics consume information in a
comparable way – i.e, with some differences all are well described by a power law.
(a)
(b)
Figure 5. Users attention patterns. CCDFs of the number of likes (a) and
comments (b) by users polarized on different conspiracy topics.
Environment
Health
Diet
Geopolitics
Likes
xmin
5
5
4
6
α
4.37
2.51
5.52
3.61
Comments
xmin
α
3
3
3
6
2.49
2.56
2.94
2.88
Table 3. Powerlaw fit of conspiracy users attention patterns. Lower bounds
and scaling parameters estimates for the distributions of the number of likes and
comments left by users polarized towards different conspiracy topics.
In order to analyze the persistence of users polarized towards different conspiracy
topics, we compute the temporal distance between the first and last comment of each
user on posts belonging to the conspiracy topic on which the user is polarized. In Figure
6 we show the Kaplan-Meier estimates of survival functions (see Methods section for
further details) for conspiracy users polarized towards different topics.
functions (all p-values are less than 10−6).
The Gehan-Wilcoxon test assesses a significant difference between the four survival
7/18
likesCCDFllllllllllllllllllllllllllllllllllllllll10-510-410-310-210-1100100101102103lEnvironmentHealthDietGeopoliticscommentsCCDFlllllllllllllllllllll10-510-410-310-210-1100100101102103104lEnvironmentHealthDietGeopolitics Figure 6. Persistence of conspiracy users. Kaplan-Meier estimates of survival
functions for users polarized towards different conspiracy topics.
Summarizing, we observe minor yet significant differences in the way users polarized
towards different conspiracy topics consume information. Moreover, by focusing on the
lifetime – i.e. the temporal distance between users’ first and last comment – we find a
remarkable difference within those users. In particular, we notice that users polarized
on geopolitics subjects are the most persistent in commenting, whereas the less
persistent users are those focused on diet narratives.
Modeling user mobility
In this section we focus on users’ activity across different conspiracy topics. In Table 4
we summarize users’ behavior by showing the Pearson correlations of their liking
activity within the different categories of topics. We see meaningful correlations
between the liking activity of users across different conspiracy topics.
Envir
Health
Diet
GeoPol
Envir Health Diet GeoPol
1.00
0.64
0.65
0.48
1.00
0.68
1.00
0.61
0.78
1.00
Table 4. Mobility of users across topics. Pearson correlations coefficients of
conspiracy users’ liking activity between different categories.
We analyze the relationship between the engagement of a user – i.e. the number of
likes she left on conspiracy posts – and how her activity is distributed across categories.
Figure 7 shows that the more a conspiracy user is engaged the more his activity is
distributed across different conspiracy topics.
8/18
t (days)S(t)020040060080000.20.40.60.81EnvironmentHealthDietGeopolitics Figure 7. Engagement and mobility across topics. Light blue lines represent the
median of the likes distributions; pale blue shaded boxes represent the interquartile
range (25–75 percentile); horizontal bars represent the extremes of the distributions.
Users active on 4 topics are 15, 510; users active on 3 topics are 20, 929; users active on
2 topics are 21, 631; and users active on 1 topic are 9, 980.
By considering only users with at least 4 likes (n = 68, 050) – necessary condition to
In particular, we consider the number of topics liked by users as the ordinal
be active on the four identified topics – we can model the relationship between the
number of likes and the number of considered topics by means of a proportional odds
model (see Methods section for an extended discussion).
dependent variable, i.e. we have j = (K − 1) = 3 ordered categories: 12, 23, and 34.
We consider the number of likes left by users as the predictor of our model. Thus, we
need to estimate three intercepts and one regression coefficient. Table 5 reports details
about the performed regression.
The estimated coefficient, β, can be difficult to interpret because it is scaled in terms
of logs. Another way to interpret these kind of regression models is to convert the
coefficient into a odds ratio (see Methods section for further details). To get the odds
ratio (OR) we exponentiate the estimate:
OR = exp(β) = exp(0.1141) = 1.12
Since OR > 1, an increase in the number of likes left by a user raises her probability
to consider a greater number of topics. In particular, an increase in the number of likes
is associated with 12% times increased odds of considering a higher number of topics.
The model provides four probabilities for each user to belong to one of the four
categories. For each user we consider the category associated with the higher
probability as the predicted category. In order to evaluate the goodness of fit of the
model, we compare the predicted categories vs the real ones by means of the absolute
9/18
4321# of topics# likes100101102103 Coefficients Value
0.1141
# of likes
Std. Error
t-value
0.001016
112.3
p-value
< 10−6
Intercepts Value
-0.7602
1.0783
2.9648
12
23
34
Std. Error
0.0135
0.0126
0.0177
p-value
t-value
-56.4267 < 10−6
< 10−6
85.7607
167.4990 < 10−6
Table 5. Proportional Odds Model. Log-odds regression coefficient and intercepts
with related standard errors, t-values, and p-values. Confidence interval at 95% for the
estimated coefficient is (0.1121, 0.1161). Chi-Square test’s p-value is 1, so we do not
reject the null hypothesis (H0 : current model is good enough) and conclude that the
model is a good fit.
distance coefficient and we find that:
δ = 1 −
(cid:80)di
n(K − 1)
= 0.852,
where di = xi − yi is the absolute distance between the real and the predicted
categories, n is the total number of users, and K is the number of categories. Since the
absolute distance coefficient is close to 1, the proposed model provides a good fit for the
data.
Summarizing, the more a user is engaged in conspiracy storytelling the more her
probability to consider a higher number of different conspiracy topics. Indeed, we
deliver a data-driven model of information consumption pointing out that users
engagement on different topics is mainly driven by their overall commitment on
conspiracy storytelling and that with the increasing of the engagement they tend to
embrace the overall corpus without a specific pattern.
Conclusions
Conspiracy theories are considered to belong to false beliefs overlooking the pervasive
unintended consequences of political and social action [9, 42]. Social media fostered the
production of an impressive amount of rumors, mistrust, and conspiracy-like narratives
aimed at explaining (and oversimplifying) reality and its phenomena. Such a scenario
provides an unprecedented opportunity to study the dynamics of topics emergence,
production, and popularity. Indeed, in this work we focus on how conspiracy topics are
consumed in the Italian Facebook.
The understanding of consumption patterns behind unsubstantiated claims might
provide important insight both at the level of popularity of topics as well as to prevent
misinformation spreading. Users activity in terms of likes and comments on posts
belonging to different categories are similar and resolves in similar information
consumption patterns. Conversely, if we focus on the lifetime – i.e., the distance in time
between the first and the last comment for each user – we notice a remarkable difference
within topics. Users polarized on geopolitics subjects are the most persistent in
commenting, whereas the less persistent users are those focused on diet narratives.
Finally we focus on the mobility of users across topics. In particular we address the
patterns behind the consumption of different topics. Users can jump independently
from one topic to another, and such a probability increases with the user engagement
(number of likes on a specific topic). Each new like on the same category increases of
10/18
the 12% the probability to pass to a new one. Our work provides important insights
about the diffusion and emergence of misinformed narratives.
11/18
Acknowledgments
Funding for this work was provided by EU FET project MULTIPLEX nr. 317532 and
SIMPOL nr. 610704. The funders had no role in study design, data collection and
analysis, decision to publish, or preparation of the manuscript.
Methods
Data Collection
We define the space of our investigation with the help of some Facebook groups very
active in the debunking of conspiracy theses. The resulting dataset is composed by 39
public Italian Facebook pages.
Notice that the dataset is the same used in [2] and [43]. However, in this paper we
focus on the 39 (exhaustive set) conspiracy pages with the intent to further characterize
the attention dynamics driving the diffusion of conspiracy topics on the Italian
Facebook. We download all posts from these pages in a timespan of 4 years (2010 to
2014). In addition, we collect all the likes and comments from the posts, and we count
the number of shares. In total, we collect around 9M likes and 1M comments,
performed by about 1.2M and 280K Facebook users, respectively. In Table 6 we
summarize the details of the data collection.
Entity
Pages
Posts
Likes
Comments
Shares
Likers
Commenters
Total
39
208, 591
6, 659, 382
836, 591
16, 326, 731
864, 047
226, 534
Table 6. Breakdown of the Facebook dataset.
Preliminaries and Definitions
Bipartite Network. We consider a bipartite network whose nodes are conspiracy
posts and conspiracy terms. The presence of a term on a given post determines a link
between the term and the post. More formally, a bipartite graph is a triple
G = (A, B, E) where A = {ai i = 1, . . . , nA} and B = {bj j = 1, . . . , nB} are two
disjoint sets of nodes, and E ⊆ A × B is the set of edges – i.e. edges exist only between
nodes of the two different sets A and B. The bipartite graph G is described by the
matrix M defined as
(cid:26) 1
Mij =
if an edge exists between ai and bj
0
otherwise
Referring to A as the set of conspiracy terms, in our analysis we use the
co-occurrence matrix C A = M M T , that is the weighted adjacency matrix of the
co-occurrence of conspiracy terms on conspiracy posts. Each non-zero element of C A
corresponds to an edge among nodes ai and aj with weight P A
ij .
12/18
Disparity Filter. Disparity filter is a network reduction algorithm which extracts
the backbone structure of a weighted network, thus reducing its size without destroying
its multi-scale nature. In particular, the method introduced in [37] is based on the null
hypothesis that the normalized weights corresponding to the connections of a given
node with degree k are uniformly distributed. The disparity filter identifies which links
for each node should be preserved in the network. The null model allows such a
discrimination through the computation – for each edge of a given node – of the
probability αij that its normalized weight pij is compatible with the null hypothesis. All
the links with αij smaller than a certain significance level α reject the null hypothesis,
and can be considered as significant heterogeneities characterizing the network. The
statistically significant edges will be those whose weights satisfy the relation
(cid:90) pij
αij = 1 − (k − 1)
(1 − x)k−2dx < α,
indicating that by decreasing the significance level α we can filter out additional
links, and thus focus on more relevant edges.
0
In order to validate our manual classification
Community Detection Algorithms.
of conspiracy terms, we apply three well known community detection algorithms to the
backbone of the conspiracy terms co-occurrence network.
Walktrap [38] computes a measure of similarities between nodes based on random
walks which has several important advantages: it captures well the community structure
in a network, it can be computed efficiently, and it can be used in an agglomerative
algorithm to compute efficiently the community structure of a network. Such an
algorithm runs in time O(mn2) and space O(n2) in the worst case, and in time
O(n2 log n) and space O(n2) in most real-world cases, where n and m are respectively
the number of nodes and edges in the network.
Multilevel [39] is based on multilevel modularity optimization. Initially, each node is
Fast greedy [40] it is a hierarchical agglomeration algorithm for detecting community
assigned to a community on its own. In every step, nodes are re-assigned to
communities in a local, greedy way. Nodes are moved to the community in which they
achieve the highest modularity. Such an algorithm runs in linear time when m ∼ n,
where n and m are respectively the number of nodes and edges in the network.
structure. Its running time on a network with n nodes and m edges is O(md log n)
where d is the depth of the dendrogram describing the community structure. Many
real-world networks are sparse and hierarchical, with m ∼ n and d ∼ log n, in which
case such an algorithm runs in essentially linear time, O(n log2 n).
Kaplan-Meier estimator. Let us define a random variable T on the interval [0,∞),
indicating the time an event takes place. The cumulative distribution function (CDF),
F (t) = Pr(T ≤ t), indicates the probability that such an event takes place within a
given time t. The survival function, defined as the complementary CDF (CCDF4) of T ,
represents the probability that an event lasts beyond a given time period t. To estimate
this probability we use the Kaplan–Meier estimator [44].
Let nt denote the number of that commented before a given time step t, and let dt
denote the number of users that stop commenting precisely at t. Then, the estimated
survival probability after time t is defined as (nt − dt)/nt. Thus, if we have N
observations at times t1 ≤ t2 ≤ ··· ≤ tN , assuming that the events at times ti are jointly
4We remind that the CCDF of a random variable X is one minus the CDF, the function f (x) =
Pr(X > x).
13/18
independent, the Kaplan-Meier estimate of the survival function at time t is defined as
(cid:89)
(
ti<t
nti − dti
nti
).
S(t) =
Odds ratio. Probability and odds are two basic statistical terms to describe the
likeliness that an event will occur. Probability is defined5 as the fraction of desired
outcomes in the context of every possible outcome with a value in [0, 1], where 0 would
be an impossible event and 1 would represent an inevitable event. Conversely, odds can
assume any value in [0,∞), and they represent a ratio of desired outcomes versus
undesired outcomes. Given a desired outcome A, the relationship between the
probability P (A) that event A will occur, and its odds O(A) is
P (A) =
O(A)
1 + O(A)
and O(A) =
P (A)
1 − P (A)
.
It follows that the odds ratio (OR) of two events A and B is defined as
OR(A, B) =
O(A)
O(B)
=
P (A)
1−P (A)
P (B)
1−P (B)
P (A)[1 − P (B)]
P (B)[1 − P (A)]
.
=
Proportional Odds Model. The proportional odds model is a class of generalized
linear models used for modeling the dependence of an ordinal response on discrete or
continuous covariates.
Formally, let Y denote the response category in the range 1, . . . , K with K ≥ 2, and
let πj = Pr(Y ≤ j x) be the cumulative response probability when the covariate
assumes value x. The most general form of linear logistic model for the jth cumulative
response probability,
logit(πj) = ln
= αj + βT
j x,
(cid:19)
(cid:19)
(cid:18) πj
1 − πj
(cid:18) πj
1 − πj
is one in which both the intercept α and the regression coefficient β depend on the
category j. The proportional odds model is a linear logistic model in which the
intercepts depend on j, but the slopes are all equal, i.e.
logit(πj) = ln
= αj + βT x.
In other words, proportional odds model takes logistic regression one step further in
order to account for ordered categorical responses. For instance, in our analysis we
could have used a logistic regression model to investigate the effect of the number of
comments on the odds ratio (OR) of ”considering < 3 topics” vs ”considering ≥ 3
topics”. However, in such a case the cut-point would be arbitrary, and we could have
used a similar logistic regression model to analyze the effect on the odds ratio (OR) of
”considering < 2 topics” vs ”considering ≥ 2 topics”. In this sense, proportional odds
model averages up over all possible cut-point logistic regression models to maximize the
amount of information one can get out of the data.
5For the sake of brevity we consider only the frequentist approach.
14/18
References
1. Furedi F. Culture of Fear Revisited. Bloomsbury Publishing; 2006. Available
from: http://books.google.it/books?id=34CvAwAAQBAJ.
2. Bessi A, Coletto M, Davidescu GA, Scala A, Caldarelli G, Quattrociocchi W.
Science vs Conspiracy: collective narratives in the age of (mis)information. Plos
ONE. accepted;.
3. Mocanu D, Rossi L, Zhang Q, Karsai M, Quattrociocchi W. Collective attention
in the age of (mis)information. Computers in Human Behavior.
accepted;abs/1403.3344.
4. Brown J, Broderick AJ, Lee N. Word of mouth communication within online
communities: Conceptualizing the online social network. Journal of interactive
marketing. 2007;21(3):2–20.
5. Kahn R, Kellner D. New media and internet activism: from the ‘Battle of Seattle’
to blogging. new media and society. 2004;6(1):87–95.
6. Quattrociocchi W, Conte R, Lodi E. Opinions Manipulation: Media, Power and
Gossip. Advances in Complex Systems. 2011;14(4):567–586.
7. Quattrociocchi W, Caldarelli G, Scala A. Opinion dynamics on interacting
networks: media competition and social influence. Scientific Reports. 2014;4.
Available from: http://dx.doi.org/10.1038/srep04938.
8. Kumar R, Mahdian M, McGlohon M. Dynamics of Conversations. In:
Proceedings of the 16th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. KDD ’10. New York, NY, USA: ACM; 2010. p.
553–562.
9. Sunstein CR, Vermeule A. Conspiracy theories: Causes and cures*. Journal of
Political Philosophy. 2009;17(2):202–227.
10. The Goal Is Power: The Global Warming Conspiracy; 2011.
http://www.forbes.com/sites/charleskadlec/2011/07/25/
the-goal-is-power-the-global-warming-conspiracy/. Forbes.
11. Ebola Lessons: How Social Media Gets Infected; 2014. Available from:
http://www.informationweek.com/software/social/
-ebola-lessons-how-social-media-gets-infected/a/d-id/1307061 [cited
29.10.2014].
12. The Ebola Conspiracy Theories; 2014. Available from: http://www.nytimes.
com/2014/10/19/sunday-review/the-ebola-conspiracy-theories.html
[cited 29.10.2014].
13. The inevitable rise of Ebola conspiracy theories; 2014. Available from:
http://www.washingtonpost.com/blogs/wonkblog/wp/2014/10/13/
the-inevitable-rise-of-ebola-conspiracy-theories/ [cited 29.10.2014].
14. Kuklinski JH, Quirk PJ, Jerit J, Schwieder D, Rich RF. Misinformation and the
Currency of Democratic Citizenship. The Journal of Politics. 2000;62(3):790–816.
Available from:
http://www.journals.cambridge.org/abstract_S0022381600000335.
15/18
15. Bessi A, Caldarelli G, Del Vicario M, Scala A, Quattrociocchi W. Social
Determinants of Content Selection in the Age of (Mis)Information. In: Aiello L,
McFarland D, editors. Social Informatics. vol. 8851 of Lecture Notes in Computer
Science. Springer International Publishing; 2014. p. 259–268. Available from:
http://dx.doi.org/10.1007/978-3-319-13734-6_18.
16. Byford J. Conspiracy Theories: A Critical Introduction. Palgrave Macmillan;
2011. Available from: http://books.google.it/books?id=vV-UhrQaoecC.
17. Fine GA, Campion-Vincent V, Heath C. Rumor Mills: The Social Impact of
Rumor and Legend. Social problems and social issues. Transaction Publishers;.
Available from: http://books.google.it/books?id=dADxBwgCF5MC.
18. Hogg MA, Blaylock DL. Extremism and the Psychology of Uncertainty.
Blackwell/Claremont Applied Social Psychology Series. Wiley; 2011. Available
from: http://books.google.it/books?id=GTgBQ3TPwpAC.
19. Bauer M. Resistance to New Technology: Nuclear Power, Information Technology
and Biotechnology. Cambridge University Press; 1997. Available from:
http://books.google.it/books?id=WqlRXkxS36cC.
20. Surowiecki J. The Wisdom of Crowds: Why the Many Are Smarter Than the
Few. Abacus; 2005. Available from:
http://books.google.it/books?id=_EqBQgAACAAJ.
21. Welinder P, Branson S, Belongie S, Perona P. The Multidimensional Wisdom of
Crowds. In: NIPS; 2010. p. 2424–2432.
22. Howell L. Digital Wildfires in a Hyperconnected World. In: WEF Report 2013.
World Economic Forum; 2013. .
23. Garrett RK, Weeks BE. The promise and peril of real-time corrections to
political misperceptions. In: Proceedings of the 2013 conference on Computer
supported cooperative work. CSCW ’13. New York, NY, USA: ACM; 2013. p.
1047–1058. Available from: http://doi.acm.org/10.1145/2441776.2441895.
24. Meade M, Roediger H. Explorations in the social contagion of memory. Memory
& Cognition. 2002;30(7):995–1009. Available from:
http://dx.doi.org/10.3758/BF03194318.
25. Koriat A, Goldsmith M, Pansky A. Toward a psychology of memory accuracy.
Annu Rev Psychol. 2000;51:481–537. Available from:
http://research.haifa.ac.il/~{}iipdm/personal_homepages/Papers/
towards_a_psychology_of_memory_accuracy.pdf.
26. Ayers M, Reder L. A theoretical review of the misinformation effect: Predictions
from an activation-based memory model. Psychonomic Bulletin & Review.
1998 Mar;5(1):1–21. Available from: http://dx.doi.org/10.3758/bf03209454.
27. Zhu B, Chen C, Loftus EF, Lin C, He Q, Chen C, et al. Individual differences in
false memory from misinformation: Personality characteristics and their
interactions with cognitive abilities. Personality and Individual Differences.
2010;48(8):889 – 894. Available from:
http://www.sciencedirect.com/science/article/pii/S0191886910000784.
28. Frenda SJ, Nichols RM, Loftus EF. Current Issues and Advances in
Misinformation Research. Current Directions in Psychological Science.
2011;20:20–23.
16/18
29. Onnela JP, Reed-Tsochas F. Spontaneous emergence of social influence in online
systems. Proceedings of the National Academy of Sciences. 2010
Oct;107(43):18375–18380. Available from:
http://dx.doi.org/10.1073/pnas.0914572107.
30. Ugander J, B L, M C, Kleinberg J. Structural diversity in social contagion.
Proceedings of the National Academy of Sciences. 2012;Available from:
http://www.pnas.org/content/early/2012/03/27/1116502109.abstract.
31. Lewis K, Gonzalez M, Kaufman J. Social selection and peer influence in an
online social network. Proceedings of the National Academy of Sciences. 2012
Jan;109(1):68–72. Available from:
http://dx.doi.org/10.1073/pnas.1109739109.
32. Kleinberg J. Analysis of large-scale social and information networks.
Philosophical Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences. 2013;371.
33. Paolucci M, Eymann T, Jager W, Sabater-Mir J, Conte R, Marmo S, et al. Social
Knowledge for e-Governance: Theory and Technology of Reputation. Roma:
ISTC-CNR; 2009.
34. Quattrociocchi W, Paolucci M, Conte R. On the effects of informational cheating
on social evaluations: image and reputation through gossip. International Journal
of Knowledge and Learning. 2009;5(5/6):457–471.
35. Bond R, Fariss J, Jones J, Kramer A, Marlow C, Settle J, et al. A
61-million-person experiment in social influence and political mobilization.
Nature. 2012;489(7415):295–298.
36. Centola D. The Spread of Behavior in an Online Social Network Experiment.
Science. 2010;329(5996):1194–1197.
37. Serrano MA, Boguna M, Vespignani A. Extracting the multiscale backbone of
complex weighted networks. Proceedings of the National Academy of Sciences.
2009;106(16):6483–6488. Available from:
http://dx.doi.org/10.1073/pnas.0808904106.
38. Pons P, Latapy M. Computing Communities in Large Networks Using Random
Walks. In: Lecture Notes in Computer Science. vol. 3733; 2005. p. 284–293.
39. Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E. Fast unfolding of
communities in large networks. Journal of Statistical Mechanics: Theory and
Experiments. 2008;2008. Available from:
http://dx.doi.org/10.1088/1742-5468/2008/10/P10008.
40. Clauset A, Newman MEJ, Moore C. Finding community structure in very large
networks. Phys Rev E. 2004 Dec;70:066111. Available from:
http://link.aps.org/doi/10.1103/PhysRevE.70.066111.
41. Ellison NB, Steinfield C, Lampe C. The Benefits of Facebook ” Friends:” Social
Capital and College Students’ Use of Online Social Network Sites. Journal of
Computer-Mediated Communication. 2007 July;12(4):1143–1168.
42. Locke S. Conspiracy culture, blame culture, and rationalisation1. The
Sociological Review. 2009;57(4):567–585.
17/18
43. Bessi A, Caldarelli G, Vicario MD, Scala A, Quattrociocchi W. Social
determinants of content selection in the age of (mis)information. Proceedings of
SOCINFO 2014. 2014;abs/1409.2651. Available from:
http://arxiv.org/abs/1409.2651.
44. Kaplan EL, Meier P. Nonparametric Estimation from Incomplete Observations.
Journal of the American Statistical Association. 1958;53(282):457–481.
18/18
|
1311.7213 | 1 | 1311 | 2013-11-28T05:43:32 | Finding a Maximum Clique using Ant Colony Optimization and Particle Swarm Optimization in Social Networks | [
"cs.SI",
"cs.NE"
] | Interaction between users in online social networks plays a key role in social network analysis. One on important types of social group is full connected relation between some users, which known as clique structure. Therefore finding a maximum clique is essential for some analysis. In this paper, we proposed a new method using ant colony optimization algorithm and particle swarm optimization algorithm. In the proposed method, in order to attain better results, it is improved process of pheromone update by particle swarm optimization. Simulation results on popular standard social network benchmarks in comparison standard ant colony optimization algorithm are shown a relative enhancement of proposed algorithm. | cs.SI | cs | Finding a Maximum Clique using Ant Colony Optimization and Particle Swarm
Optimization in Social Networks
Mohammad Soleimani-Pouri
Department of Electrical, Computer
and Biomedical Engineering, Qazvin
branch, Islamic Azad University,
Qazvin, Iran
[email protected]
Alireza Rezvanian
Computer and Information
Technology Engineering Department
Amirkabir University of Technology,
Tehran, Iran
[email protected]
Mohammad Reza Meybodi
Computer and Information
Technology Engineering Department
Amirkabir University of Technology,
Tehran, Iran
[email protected]
Abstract—Interaction between users in online social networks
plays a key role in social network analysis. One on important
types of social group is full connected relation between some
users, which known as clique structure. Therefore finding a
maximum clique is essential for some analysis. In this paper,
we proposed a new method using ant colony optimization
algorithm and particle swarm optimization algorithm. In the
proposed method, in order to attain better results, it is
improved process of pheromone update by particle swarm
optimization. Simulation results on popular standard social
network benchmarks in comparison standard ant colony
optimization algorithm are shown a relative enhancement of
proposed algorithm.
Keywords- social network analysis; clique problem; ACO;
PSO.
I.
INTRODUCTION
Today online social networks are formed a new type of
life due to some facilities and services for a wide ranges of
ages such as young to old. There is no doubt about either
influence or growth of social networks. Therefore, several
many of researchers are focused on social network analysis
aspects. It seems to be useful, studying structure of relation
between users in social networks. One of the important user
group structure associated with a full connected of some
users, which known as clique structure [1]. Several
applications of finding clique are reported by researchers
such as social networks analysis [2], online shopping
recommendation [3] and evolution of social networks [12].
In literature, finding clique is categorized as NP-complete
problems in graph theory [4].
Evolutionary
stochastic
are
(EAs)
algorithms
optimization techniques based on the principles of natural
evolution which applied several NP problems [13-15] such
as clique problem. Various types of algorithms have been
presented to solve clique problem, while evolutionary
algorithms such as genetic algorithm (GA) and ant colony
optimization (ACO) have been used more. Popular algorithm
named Ant-clique algorithm, which make a maximal clique
using sequential greedy heuristics based on ant colony
optimization by adding vertices to the partial cliques
iteratively [5]. Besides, another ACO based method
hybridized by simulated annealing (SA) [6] and tabu search
[7]. Although new hybrid algorithm obtained good results,
they have a high complexity in practice.
In this study, Particle Swarm Optimization (PSO)
algorithm has been applied as the heuristic to enhance the
performance of ACO algorithm for finding a maximal clique
in social network graph. Simulation results on social network
benchmark are shown the better results in comparison with
standard ACO algorithm. In the rest of this paper, section II
and III are consisted of ACO and PSO introduction
respectively, in section IV, proposed method is discussed.
Simulation results on social networks datasets are reported in
section V.
II. ANT COLONY OPTIMIZATION (ACO)
Ant Colony optimization (ACO) algorithm works well
for solving several discrete problems. The basic algorithm of
ACO was proposed by Dorigo as a multi agent approach in
order to solve traveling salesman problem (TSP) [8]. The
main idea of ACO algorithm is inspired from the behavior of
seeking out food by colonies of ants. Ants search their
environment randomly to find food. They return some of the
food to their nest once found a food and leave pheromone in
their return path. The amount of pheromone left on their path
depends on quality and size of the food source and it
evaporates gradually. Remaining pheromones will persuade
other ants to follow the path and just after a short time,
majority of the ants will trace the shorter path which is
marked with stronger pheromone. Procedure of ACO
algorithm has been presented in Figure 1.
Procedure ACO_MetaHeuristic
while(termination_conditions)
generateSolutions()
daemonActions() {Optional}
pheromoneUpdate()
end while
end procedure
Figure 1. Pseudo-code of ACO algorithm [9]
During running of the algorithm, ants first produce
different solutions randomly
in
the main
loop after
initialization. Afterwards, the solutions are improved by
updating the pheromones and using a local search optionally.
According to the problem and graph traverse, pheromones
set on vertices or edges of graph. Traversing the edge
between vertices i and j depends on the probability of edge
which is calculated as below:
(
)
τ
α
ij
(
τ
∑
α
ij
(1)
=
)
p
k
ij
k is probability of traversing the edge between
Where, pij
α is amount of pheromone present on
vertices i and j, while τij
the above mentioned edge. An optional local search can
contribute to improvement of the results prior to updating the
pheromones. However, method of updating the pheromones
can be like this:
(
1
+ =
1
+ Δ
)
ρτ
t
ij
(2)
τ
t
ij
τ
t
ij
−
t+1 is
Where, ρ is evaporation rate of pheromone, τij
t is
amount of new pheromone for edge between i and j, τij
t
amount of current pheromone for edge between i and j, Δτij
is amount of reinforced pheromone for proper solutions
which can be calculated from the following equation.
τ
Δ
t
ij
τ
∈
good s olution
t
ij
Otherw ise
(3)
if
1
⎧
⎪
= ⎨
0
⎪⎩
III.
PARTICLE SWARM OPTIMIZATION (PSO)
Particle swarm optimization (PSO) is a population based
optimization technique developed by Eberhart and Kennedy
in 1995, which inspired from social behavior of birds
seeking for food. In this method, group of birds seek for food
in a specified space randomly. Birds follow the bird with the
shortest distance to food in order to find position of the food
[10].
Every solution in PSO algorithm which is called a
particle corresponds to a bird in their pattern of social
movement. Each particle has a value of fitness calculated by
a fitness function. A particle bears greater fitness once it is
located in a closer position to the target (food in the model of
moving birds) within search space. Moreover, any particle
represents a speed which is responsible to direct the particle.
A particle will keep going through the problem space by
following the optimum particle at current state [16].
A group of particles (solutions) are created randomly at
the beginning of particle swarm optimization algorithm and
they try to find the optimum solution through being updated
among generations. A particle is updated in each step using
the best local and global solution. The best position a particle
has ever succeeded to reach is called pbest and saved while
the best position achieved by the population of particles is
named gbest. Velocity and location of each particle will be
updated using Equations (4) and (5) in each step of
implementing the algorithm after finding the best local and
global values.
(
)
(
= wv + c r pbest - x + c r gbest - x
t
t
t
t
2 2
1 1
i
i
i
i
)
t
i
(4)
t +1
v
i
(5)
i = 1, ..., m
x
= x + v
t
t
t +1
i
i
i
Where, vi is the velocity of ith particle and xi is the current
position of it. r1 and r2 are random values in the range of
(0,1). c1 and c2 are learning parameters usually assumed
equal (c1=c2). w is inertia weight which is considered as a
constant or variable coefficient as random, linear, nonlinear
and adaptive [11]. PSO has been used in various applications
and
this research utilizes
it
to
improve amount of
pheromones.
IV. PROPOSED ALGORITHM
High complexity was a major drawback of the previous
heuristic methods for solving the clique problem since it
significantly adds to the volume of calculations. All methods
provided so far apply the following relation to calculate Δτ
although the proposed algorithm take the advantage of PSO
algorithm to improve results and reduce complexity.
1
τ+Δ
=
t
1
(6)
+
G
- best-tour )
-best
(1
This algorithm has used the hybrid of ACO and PSO
algorithms in order to find the maximal clique in a graph. For
this purpose, some ants are placed initially on the graph and
follow paths to find the maximal clique. After evaporation of
existing pheromones, proper path is updated by its amount
on the edges using particle swarm optimization algorithm.
This procedure is repeated until the optimum clique is
obtained on the desired graph. Determining the amount of
pheromone through PSO is such that the total pheromone
measured at this step and the total pheromone associated
with the best answer up to present step will be calculated
taking into account the amount of pheromones at current
step. Now, Δτ for the reinforced pheromone of the desired
clique will be calculated with PSO algorithm using the
following equation:
1τ
= Δ +
τ
+Δ
tV
t
t
t is amount of reinforcement for current
Where, Δτij
t+1 is amount of reinforcement for new
pheromone and Δτij
pheromone, while Vt gives the amount of change which can
be achieved from this equation:
V = c r (pτ - Δτ) + c r (gτ - Δτ) + c V
t +1
3
2 2
1 1
Where, Vt+1 is the new value of Vt. r1 and r2 are two
random values within range of (0,1), while c1, c2 and c3 are
learning parameters (c1=c2=c3). pτ and gτ are considered as
the pheromone correspondent with the best current clique
and the best clique found so far, respectively. In this case,
(7)
(8)
t
discovering the changed amount of pheromone will be
implemented much more intelligent.
Taking into consideration the mentioned items, general
routine of proposed the algorithm can be summarized as
below:
1.
Initialize the parameters
2. Repeat steps 3 to 5 until reaching conditions of
termination
3. Traverse the graph by ants and create sub-graphs as
clique
4. Evaluate the obtained cliques
5. Update pheromones by PSO according to equation
(8)
6. End.
V. SIMULATION RESULTS
For evaluation of the proposed method, experiments applied
on some popular social network datasets. The description of
popular social networks is listed in the table 1.
the
297
62
V
VI
III
IV
8479
I
II
TABLE I.
SOCIAL NETWORKS DESCRIPTIONS
Dataset
Nodes
Edges
78
34
Zachary's karate club
Common adjective and nouns
112
425
in "David Copperfield"
Neural
network
of
nematode C. Elegans
social network of dolphins,
Doubtful Sound, New Zealand
Pajek
network:
Erdos
collaboration network 971
Pajek
network:
Erdos
collaboration network 991
Pajek network: World Soccer,
VII
Paris 1998
VIII Pajek network: graph and
digraph glossary
Pajek
network:
Slovenian
journals 1999-2000
co-authoship of scientists in
network theory & experiments
Pajek
network:
SmaGri
citation network
interchange network,
email
Univ. of Rovira
i Virgili,
Tarragona
4919
1314
35
72
472
492
295
118
1589
IX
X
XI
1059
1133
5451
159
1417
124
823168
1190
XII
Topology of zachary’s karate club and social network of
dolphins are presented in figure 2 and 3 respectively as the
most popular of the social network datasets.
Figure 2. Zachary's karate club social network
Figure 3. Dolphins social network
The setting of parameters for experiment is listed in
below. It is noted that choosing different values for
improving the results is also possible.
In this paper has used 30 ants with ρ=0.95, Φ=0.0002,
Δτinitial=0, τmin=0.01, τmax=6.
Meanwhile, parameters of PSO were initialized as V=0,
c1=c3=0.3, and c2=1-c1. α and ρ were given the following
values based on the experiments done and t is the number of
iterations.
α
t
( )
1
⎧
⎪
2
⎪
= ⎨
3
⎪
⎪
4
⎩
≤
t
100
< ≤
t
100
< ≤
t
400
>
t
800
400
800
⎫
⎪
⎪
⎬
⎪
⎪
⎭
(9)
+
ρ
(t 1)
ϕρ
(1- ) (t)
⎧
= ⎨
ρ
(t)>0.95
0.95
if
⎩
Average of 10 independent runs with 1000 iterations for
each implementation have been listed in table 2 and table 3
for proposed method (ACO-PSO) and ACO algorithm
respectively, including the maximum (Best), average (Avg),
(10)
standard deviation (Std) and run-time of algorithm for
finding a maximum clique in each graph datasets.
TABLE II.
SIMULATION RESULTS OF ACO-PSO FOR FINDING A
MAXIMUM CLIQUE
Std
Avg
4.995
0.070
4.783
0.412
5.543
1.482
4.998
0.044
0.6673
5.848
6.011
0.819
4.118
0.322
3.985
0.121
3.185
0.299
0.371
2.253
6.409
0.765
2.241
7.997
Run-time
11.84
23.37
49.47
15.69
105.11
111.89
10.742
12.94
22.07
118.59
481.34
529.04
Best
5
5
7
5
7
7
5
4
4
3
8
12
SIMULATION RESULTS OF ACO FOR FINDING A
MAXIMUM CLIQUE
Avg
Std
4.991
0.082
4.709
0.495
5.521
1.568
0.094
4.991
8.753
0.861
5.961
0.924
3.910
0.293
3.943
0.232
0.316
3.017
2.175
0.434
6.025
0.815
2.374
7.562
Run-time
64.03
355.81
1121.71
132.12
5853.59
6281.35
55.15
153.66
438.87
2142.03
9153.65
11109.17
Best
5
5
7
5
7
7
5
4
4
3
8
12
Graph
I
II
III
IV
V
VI
VII
VIII
IX
X
XI
XII
Graph
I
II
III
IV
V
VI
VII
VIII
IX
X
XI
XII
TABLE III.
Table 2 and table 3 indicate that the proposed method
(ACO-PSO) produces better results in comparison with ACO
method since the proposed approach is an appropriate
method to update pheromones of the traversed paths for ants
in calculating the optimum clique.
VI. CONCLUSION
A new hybrid algorithm has been presented in this paper
using ACO and PSO (ACO-PSO) for finding a maximum
clique in social networks. Traditional algorithms suffered
high complexity while the hybrid proposed algorithm just
change the process of update pheromone. It has been shown
that the new algorithm was able to improve the basic ACO
algorithm as simply and quickly. Simulation results on
popular social networks datasets indicate the improved
results for proposed algorithm in comparison with the ACO
algorithm.
[4]
REFERENCES
[1] D. Easley, and J. Kleinberg, Networks, Crowds, and Markets:
Reasoning About a Highly Connected World. Cambridge University
Press, 2010.
[2] S. Sadi, S. O(cid:31)g(cid:31)u(cid:31)du(cid:31)cu(cid:31), and A. Uyar, “An efficient community
detection method using parallel clique-finding ants,” Proc. 2010 IEEE
Congress on Evolutionary Computation (CEC 2010), IEEE Press,
July 2010, pp. 1-7, doi: 10.1109/CEC.2010.5586496.
[3] Q. Yang, P. Zhou, H. Zhang, and J. Zhang, “Clique Discovery Based
on User Similarity
for Online Shopping Recommendation,”
Information Technology Journal, vol. 10, pp. 1587-1593, 2011.
J. Kleinberg, and E. Tardos, Algorithm Design. Addison Wesley,
2005.
[5] S. Fenet, and C.Solnon, “Searching for Maximum Cliques with Ant
Colony Optimization,” Applications of Evolutionary Computing,
LNCS 2611, pp. 236-245, 2003.
[6] X.Xu, J.Ma, and J. Lei, “An Improved Ant Colony Optimization for
the Maximum Clique Problem,” Proc. Third International Conference
on Natural Computation (ICNC 2007), IEEE Press, Aug. 2007, pp.
766-770, doi: 10.1109/ICNC.2007.205.
[7] M.Al-Fayoumi, S.Banerjee, and P. K. Mahanti, “Analysis of Social
Network Using Clever Ant Colony Metaphor,” World Academy of
Science, Engineering and Technology, vol. 5, pp. 970-974, 2009.
[8] E. Bonabeau, M. Dorigo, and G. Theraulaz, Swarm intelligence.
Oxford University Press, 1999.
[9] K. Socha, and M. Dorigo, “Ant colony optimization for continuous
domains,” European Journal of Operational Research, vol. 185, no. 3,
pp. 1155-1173, 2008.
[10] J. Kennedy, and R.C. Eberhart, “Particle Swarm Optimization,” Proc.
IEEE International Conference on Neural Networks (ICNN 1995),
IEEE
Press, Nov/Dec.
1995,
pp.
1942-1948,
doi:
10.1109/ICNN.1995.488968.
[11] S. Nabizadeh, K. Faez, S. Tavassoli, and A. Rezvanian, “A novel
method for multi-level image thresholding using particle swarm
Optimization algorithms,” Proc. 2010 2nd International Conference
on Computer Engineering and Technology (ICCET 2010), IEEE
Press,
Apr.
2010,
pp.
V4-271-V4-275,
doi:
10.1109/ICCET.2010.5485600.
[12] F. Yan, S. Cai, M. Zhang, G. Liu, and Z. Deng, “A clique-
superposition model for social networks,” Science China Information
Sciences, pp. 1-19, 2012. doi: 10.1007/s11432-011-4526-y, In press.
[13] A. Rezvanian, and M. R. Meybodi, “An adaptive mutation operator
for artificial immune network using learning automata in dynamic
environments,” 2010 Second World Congress on Nature and
Biologically Inspired Computing (NaBIC 2010), Dec. 2010, pp. 479-
483, doi: 10.1109/NABIC.2010.5716360.
[14] A. Rezvanian, and M. R. Meybodi, “Tracking Extrema in Dynamic
Environments Using
Immune
a Learning Automata-Based
Algorithm,” Grid and Distributed Computing, Control And
Automation, Communications in Computer and Information Science,
2010, vol. 121, pp. 216-225, doi: 10.1007/978-3-642-17625-8_22.
[15] A. Rezvanian, and M. R. Meybodi, “LACAIS: Learning Automata
Based Cooperative Artificial
Immune System
for Function
Optimization,” Contemporary Computing, Communications
in
Computer and Information Science, 2010, vol. 94, pp. 64-75, doi:
10.1007/978-3-642-14834-7_7.
[16] A. B. Hashemi, and M. R. Meybodi, “A note on the learning automata
based algorithms for adaptive parameter selection in PSO,” Applied
Soft Computing, vol. 11, no. 1, pp. 689-705, 2011.
|
1705.10122 | 1 | 1705 | 2017-05-29T11:16:50 | Reciprocity-driven Sparse Network Formation | [
"cs.SI",
"cs.GT"
] | A resource exchange network is considered, where exchanges among nodes are based on reciprocity. Peers receive from the network an amount of resources commensurate with their contribution. We assume the network is fully connected, and impose sparsity constraints on peer interactions. Finding the sparsest exchanges that achieve a desired level of reciprocity is in general NP-hard. To capture near-optimal allocations, we introduce variants of the Eisenberg-Gale convex program with sparsity penalties. We derive decentralized algorithms, whereby peers approximately compute the sparsest allocations, by reweighted l1 minimization. The algorithms implement new proportional-response dynamics, with nonlinear pricing. The trade-off between sparsity and reciprocity and the properties of graphs induced by sparse exchanges are examined. | cs.SI | cs |
Reciprocity -- driven Sparse Network Formation
1
Konstantinos P. Tsoukatos
Department of Computer Science and Engineering
Technological Education Institute of Thessaly, Greece
Abstract
A resource exchange network is considered, where exchanges among nodes are based on reciprocity.
Peers receive from the network an amount of resources commensurate with their contribution. We
assume the network is fully connected, and impose sparsity constraints on peer interactions. Finding the
sparsest exchanges that achieve a desired level of reciprocity is in general NP-hard. To capture near --
optimal allocations, we introduce variants of the Eisenberg -- Gale convex program with sparsity penalties.
We derive decentralized algorithms, whereby peers approximately compute the sparsest allocations,
by reweighted ℓ1 minimization. The algorithms implement new proportional-response dynamics, with
nonlinear pricing. The trade-off between sparsity and reciprocity and the properties of graphs induced
by sparse exchanges are examined.
Network formation, proportional-response, nonlinear pricing, sparse interactions.
Index Terms
I. INTRODUCTION
The unprecedented increase in wireless traffic poses significant challenges for mobile operators,
who face extensive infrastructure upgrades to accommodate the rising demand for data. To ease
strain on networks, a viable alternative seeks to take advantage of already deployed resources,
that presently remain underutilized. For example, in device-to-device communications, devices in
close proximity may establish either direct links, or indirect communication via wireless relays,
altogether bypassing the cellular infrastructure. In recently launched Wi-Fi internet services,
e.g., FON (fon.com), Open Garden (opengarden.com), Karma (yourkarma.com), sharing wireless
access is a prominent feature, and subscribers are rewarded for relaying each other's traffic. In
all these scenarios, it is important to design mechanisms which foster cooperation and encourage
user contribution, in ways that realize fair and efficient use of pooled resources.
September 22, 2018
DRAFT
2
In this paper, we study a network exchange model, where collaborative resource consumption
is based on reciprocation. Incentive mechanisms based on reciprocation have been proposed in
the context of peer-to-peer and user-provided networks [1], [2]. Participant nodes earn credits (or
virtual currency) for assisting other nodes in transmitting their data to the destination. Ideally,
reciprocation implies that each peer receives back from the network an amount of resources
or utility equal to what he contributed to other users. However, such perfect reciprocation may
in general not be feasible, due to constraints arising from network structure/connectivity, and
differing resource endowments possessed by nodes, also depending on their position in the
network graph. Moreover, peers can typically maintain a limited number of connections (in the
popular BitTorrent peer-to-peer protocol users upload to at most four peers). Hence, it seems
reasonable to explore situations where graphs representing exchanges of resources among peers
are in some sense sparse.
The exchange model considered in this paper builds upon the so-called linear Fisher market
in economics (see [3], [4], [2] and references), where each participant aims to receive as much
resources as possible from the market. In this model, the optimal resource allocations are captured
by a classic convex program discovered by Eisenberg and Gale in 1959 [5]. In the context
of peer-to-peer bandwidth trading, the authors in [3] proposed a simple distributed algorithm
called proportional -- response, that computes solutions to the Eisenberg -- Gale program, hence
also equilibrium allocations in the resource exchange model: At every time slot, each peer
distributes his available resource to other peers in proportion to the resources it received from
them in the previous time slot.
Here, we formulate an optimization problem that balances benefits from reciprocation with
fixed per-link costs, therefore induces peers to maintain only a few active connections. Clearly,
rational peers will not engage in an exchange if costs out-weight potential benefits. We impose
sparsity penalties on peer interactions, to reflect the fact that peers often cannot afford the cost
of establishing and maintaining a link, the associated communication overhead, etc. fixed costs,
or are simply limited by physical constraints, such as limited range of wireless devices. This
reciprocity versus sparsity optimization is solved by decentralized tit-for-tat algorithms, whereby
peers communicate bids for each other's resource, so as to approximately compute the sparsest
allocations, achieving close-to-perfect reciprocation. Our algorithms implement nonlinear pricing,
and extend the proportional -- response dynamics of [3] to reinforce interactions where large
amounts of resources are exchanged. Starting from a complete graph, the algorithms prescribe
September 22, 2018
DRAFT
3
how nodes can gradually form a network of exchanges, that progressively gets sparser. As a result,
the proposed schemes suggest a network formation model where directed graphs, representing
sparsity-constrained resource exchanges, are constructed. The graphs may manifest either direct
reciprocation between peers, i.e., both edges (i, j) and (j, i) are typically present in the network
graph, or indirect reciprocation, in which case most of the edges do not have their reverse in the
graph. We illustrate the formation of resource exchange networks by peers who achieve almost
perfect reciprocation with only a small number of connections, and discuss the properties of the
sparse graphs in several numerical examples.
From a mathematical standpoint, the sparse exchange algorithms are derived by applying
majorization-minimization [6], [7] and reweighted ℓ1 minimization [8] to a combinatorial prob-
lem, and optimize the trade-off between reciprocation and sparsity up to a local optimum. Starting
from different initial conditions, different local optima arise, corresponding to different resource
allocations and sparse exchange graphs.
II. SYSTEM MODEL AND BACKGROUND
Consider a network of N peers who exchange resources over a graph G describing connectivity.
Exchanges take place only between peers that are neighbours in the graph G. Each peer allocates
spare resources to other peers, in exchange for their resources (in the future). There exists a
single resource/commodity in the network. Peers spend their own spare resource for acquiring
resources, i.e., there is no monetary budget. Let ai be the resource endowment of peer i. Let
xij(t) be the amount of resource allocated from user j to user i at time t = 0, 1, . . .. Vector
xi denotes the allocations of peer i, and x−i denotes the allocations of all others. Each peer i
allocates the entire budget ai to his neighbours, ai = Pj6=i xji, and receives in return a total
amount ri(x−i) := Pj6=i xij of resource. We assume peers value only the resource received from
others (not their own spare resource). That is, utility is linear in the amount of received resources
Ui(ri) = ri, and each peer i = 1, . . . , N , allocates resources to solve
(PEER) max
xi≥0
ri(x−i)
subject to Xj6=i
xji = ai.
These N (PEER) problems are intertwined, because each peer's utility depends on resources
received from other peers.
Notation: Subscript in allocation xij is understood as given from j to i, similarly bij is the
bid of peer j for resource of peer i, and µij the price peer j charges to i.
September 22, 2018
DRAFT
4
a:1
d:1
a:1
d:1
a:1
d:1
b:1
c:1
b:1
c:1
b:1
c:1
complete graph
ring (cycle)
pairs
Fig. 1. Exchange network with 4 nodes. Left: Complete graph. Middle, right: Graphs with minimum number of links.
Motivation. In this paper, we consider an exchange network where the connectivity graph G
is complete. Every node can, in principle, engage in exchanges with everybody else. However,
we assume that establishing and maintaining exchange links carries a cost, so that peers tend
to limit the number of their active connections. This is often the case in practice, where peers
choose a few trading partners and avoid spreading themselves thin, so as to reduce overhead,
friction etc. costs associated with exchanges. Limits on the number of connections may also arise
due to physical or protocol constraints (e.g., in the BitTorrent peer-to-peer protocol). In a slightly
different context, reducing transaction costs is a motivation for sparse portfolio selection [9].
Here, in a similar spirit, we use a penalty term that encourages peers to form sparse connections.
Our goal is to develop a quantitative model for dynamic formation of exchange networks driven
by reciprocation, where the directed graphs representing exchanges are sparse.
Example. The example network of Figure 1 illustrates the graphs implementing sparse
exchanges. All four nodes a, b, c, d have resource endowment equal to 1. Perfect reciprocation
can be realized in infinitely many ways, across the 12 links of the complete graph (left). The
sparsest exchange graphs, where each peer gives to exactly one peer 1 unit of resource, consist
of only 4 links, arranged either in ring (middle) or pairs (right), whence we also see the sparsest
solution need not be unique. In addition, the pairs configuration of Figure 1 shows that the
sparsest solution may partition the (initially complete) graph into disconnected components.
Previous work. We recap several useful results from a large literature. Exchange network is
an instance of a linear Fisher market, for which an equivalent convex formulation was given by
Eisenberg and Gale (1959) [5]:
max
x Xi
ai log ri
subject to Xj6=i
xji = ai,
∀i.
(1)
The objective in (1) resembles the familiar proportionally-fair allocation, where each peer i
September 22, 2018
DRAFT
5
receives an amount of resources ri proportional to his contribution ai. It is also similar to
Kelly's NETWORK problem [10], if the contributions ai are viewed as payments. The receive
vector r achieving optimality in (1) is unique, however the optimal allocation x is not unique,
because the objective in (1) is not strictly concave in x. That is, the same optimal receive
vector may be realized with different allocations. The equivalence between (PEER) and (1) can
be established as follows. Let ρj be the price at which peer j "sells" its resource (although
no actual monetary payments mediate the exchange). User i, by allocating amount xji to j,
"purchases" back xij = xji/ρj. Hence, the total resource received by user i is ri = Pj6=i xji/ρj.
Therefore, to maximize utility in (PEER) user i allocates resources to (and consequently receives
resources xij > 0 from) only peers with the largest 1/ρj, i.e., the cheapest neighbors,
xij > 0
if and only if ρj = min
k∈Ni
ρk,
(2)
where Ni is the set of neighbors of i. Now, to find the prices ρ, consider the convex program
(1), relax the constraints and write the Lagrangian
L(x, ρ) = Xi
ai log ri +Xi
ρi(ai −Xj6=i
xji).
The KKT conditions at the saddle point of the Lagrangian imply that either
or
xij = 0
and
xij > 0
and
∂L
∂xij
∂L
∂xij
< 0 in which case
= 0 in which case
ai
ri
ai
ri
< ρj,
= ρj.
From the equations above we deduce that xij > 0 if and only if ρj = mink∈Ni ρk, which is
precisely condition (2). Therefore, allocations x and prices ρ solving (PEER) can be computed
through the Eisenberg -- Gale program (1).
Define the exchange ratio for each peer i as the ratio of the resources ri peer i receives, over
the ai he allocates to others. Since xij = xji/ρj, by summing over i we get aj = rj/ρj, that is
ρj =
rj
aj
,
∀j.
Hence, the exchange ratio coincides with the Lagrange multiplier/price ρ in the Eisenberg -- Gale
program (1). May use the term price and exchange ratio interchangeably.
When network connectivity is given, we summarize the following facts from [2], [3], [4], [11]:
Network graph decomposes into components, and resource exchanges take place only within each
September 22, 2018
DRAFT
6
component. Since the exchange ratio is also the price (Lagrange multiplier) at which each node
sells its resource, peers with high exchange ratio are expensive and more constrained, i.e., "poor"
and struggle to contribute more resources. At equilibrium, rational peers exchange resource only
with their minimum price (cheapest/most generous) neighbours. Exchange can be viewed as a
reverse auction: Acting as sellers, peers compete to sell their resource by lowering their prices
(raising their bids), whereas, in the role of buyers, they purchase resource from the cheapest
(highest "bang-per-buck") neighbor. Moreover, the prices of peers who exchange resources with
each other are inversely proportional. That is, the price at which a node buys resources from
peers is equal to the inverse of the price at which he sells his own resource. The upshot is that,
by measuring his own price, each peer can infer the price of the peers he exchange resources
with. In particular, inexpensive nodes (with prices smaller than one) know they interact with
expensive nodes (with price larger than one).
New connections. Previous work typically analysed exchanges in networks where connectivity
and resource endowments were a priori fixed and immutable. Then, due to the existing connec-
tivity and neighbouring node endowments, certain nodes may end up receiving significantly less
resources than what they contribute to their peers. Unless it is possible to alter either connectivity,
or node resource endowments, rational peers with exchange ratio much lower than one have little
incentive to engage in exchanges and contribute. Lack of participation will decrease the total
amount of resources contributed to the network, i.e., is detrimental to social welfare.
Rational nodes with low prices are motivated to seek out new peers, who are less expensive
than the ones they presently interact with. Neighbor selection by a Gibbs sampling algorithm
was proposed in [2]. By connecting with richer peers, inexpensive nodes receive more resources,
hence their exchange ratio (which is also their price) increases. Along the way, inexpensive peers
become more expensive, thereby also less attractive as candidates for resource exchange. In a
fully connected network, whenever perfect reciprocity is possible, this balancing act may drive
all exchange ratios (prices) to one, i.e. all nodes receive from the network an amount of resources
equal to what they contribute -- perfect reciprocation.
III. NETWORK FORMATION PROBLEMS
Starting from a complete network, we allow peers to gradually form an exchange graph
that progressively gets sparser. First, we discuss centralized optimization problems, that aim
to identify the sparsest interactions that guarantee a desired level of reciprocation. Then, we
September 22, 2018
DRAFT
7
introduce variants of the Eisenberg -- Gale program (1), where the objective is to balance benefits
from equitable allocations with fixed per-link costs, hence form only a few connections. These
formulations lead to distributed algorithms that enable peers to compute sparse exchanges in
a decentralized manner, by communicating bids for each other's resource. The algorithms are
simple and natural to understand.
A. Sparse Exchanges with Reciprocation Guarantees
Let x0 denote the pseudo-norm that counts the number of nonzero entries in x.
Problem P0. The objective is to find sparsest allocations that achieve a minimum exchange
ratio at least θ, where θ ≤ 1:
(P 0) min
x
x0
s. t. Xj6=i
xji = ai,
ri
ai
≥ θ ∀i.
(3)
This is a combinatorial problem (assuming the minimum desired level θ of reciprocity is feasible),
hence intractable. To find an approximate solution, we may replace the nonsmooth x0 norm
by a smooth proxy. A typical choice, justified by the limit
1 {x 6= 0} = lim
ǫ↓0
log(1 + x/ǫ)
log(1 + 1/ǫ)
,
is given by the logarithmic approximation ℓ(x) := Pi,j log(ǫ + xi,j), leading to the problem
min
ℓ(x)
x
s. t. Xj6=i
xji = ai,
≥ θ ∀i.
(4)
ri
ai
Hence, the combinatorial objective in (3) has been substituted by a minimization of a concave
function (4). This is again hard, and can be tackled as follows.
Problem P1. Successively minimize a linear upper bound to the logarithm in (4), given by
min
x
T ∇ℓ(x(t))
x
s. t. Xj6=i
xji = ai,
ri
ai
≥ θ ∀i,
(5)
formed around the previous solution x(t), for t = 0, 1, . . .. This amounts to solving a series of
linear programs to find a local minimum, hence can be computed efficiently. Moreover, to avoid
getting trapped in local minima, a small random perturbation may be employed.
Problem P2. Instead of bounding the logarithm in (4) with a linear upper bound, we bound
with a quadratic. More specifically, it holds that log(ǫ + x) ≤ q(x, ¯x) + k, where
q(x, ¯x) :=
x2/(2δ(ǫ + δ)),
0 ≤ x ≤ δ
x2/(2¯x(ǫ + ¯x)), x > δ,
(6)
DRAFT
September 22, 2018
8
for appropriate constant k and small δ > 0. Starting from an iterate x(t), we use (6) to bound
(4) and get a quadratic program in x:
min
x Xi,j
q(xij, xij(t)) s. t. Xj6=i
xji = ai,
ri
ai
≥ θ ∀i.
(7)
Problem (7) subsequently reduces to a linear system computing 2N multipliers from 2N linear
equations. As in (5), we apply the majorization-minimization procedure [8], hence solve a
sequence of quadratic programs to obtain the final solution (IRLS algorithm). Instead of solving
each quadratic program completely, we may run one (or a few) iterations towards solution (e.g.
fixed-point iteration) of linear system. We anchor a new upper bound to the computed allocations,
and resume iterations for the updated linear equations.
The solutions discussed above are centralized. In the following, we focus on distributed
algorithms, obtained by balancing reciprocation with a penalty that encourages sparse exchanges.
B. Eisenberg -- Gale Program with Sparsity Penalty
We consider an Eisenberg -- Gale program (1) augmented with a sparsity promoting term
(EG) max
log(ǫ + xij)
x Xi
ai log ri − cXi,j
subject to Xj6=i
xji = ai,
∀i.
(8)
Optimization (8) is nonconvex because the objective is a difference of two concave functions.
We will derive a distributed algorithm that computes an approximate solution. First, relax the
constraints, introduce the multipliers λ and write the Lagrangian
L(x, λ) = Xi
ai log ri − cXi,j
log(ǫ + xij) +Xi
λi(ai −Xj6=i
xji).
The dual optimization requires solving the relaxed primal
max
L(x, λ).
x
(9)
A local maximum for (9) can be determined in a iterative fashion using majorization-minorization.
To that end, we bound the logarithms log(ǫ + xij) using
log y ≤ log y0 +
y − y0
y0
.
(10)
DRAFT
September 22, 2018
9
Next, with a logarithmic change of variables xij := log xij, define the function
φ(x) := Xi
ai log(Xj6=i
exij ),
so that φ(x) = Pi ai log ri(x). The convexity of the log-sum-exp function [12, page 74] implies
that φ is convex in x, therefore it holds that
φ(x) ≥ φ(x(t)) + (x − x(t))T ∇φ(x(t)).
(11)
Making use of inequalities (10) and (11), we lower bound the Lagrangian L by a surrogate
function g anchored at x(t),
constructed as
L(x, λ) ≥ g(xx(t)),
g(xx(t))
:= φ(x(t)) + (x − x(t))T ∇φ(x(t))
(log(ǫ + exi,j (t)) − 1)
−cXi,j
−Xj (cid:0)λj + w(exij(t))(cid:1) exij .
The weights w above are defined as in [8] by
w(x) :=
c
ǫ + x
.
(12)
(13)
Function g(xx(t)) in (12) is concave in the transformed allocations x and easy to maximize.
Successive maximization of the lower bound (12) yields the iteration
x(t + 1) = arg max
x
g(xx(t)),
where, at each time t = 0, 1, . . ., multipliers λ are also updated to satisfy the node endowment
constraints in (8). We differentiate (12) with respect to x, transform back to the x domain, and,
after some algebra, arrive at the following solution:
Each peer i communicates at time t = 0, 1, . . . its exchange ratio
ρi(t) :=
ri(t)
ai
to other peers.
Let the bids of peer j for peer i's resource be
bij(x(t), λi) :=
xji(t)
ρj(t)
1
λi + w(xji(t))
,
September 22, 2018
(14)
(15)
DRAFT
10
where the weights w are defined in (13) and λi is the multiplier associated with the budget
constraint for peer i. In market terms, bij is the amount of resource peer i can purchase from j
at price (per unit) ρj(t)(λi + w(xji(t))) by paying xji(t). Note that pricing is nonlinear; price
per unit decreases as payment xji(t) increases, and asymptotically drops to λiρj(t) as payment
xji(t) goes to infinity.
Each peer i selects λi to exhaust the entire budget
ai = Xj6=i
bij(x(t), λi),
(16)
this can be computed by bisection search. Check that if endowment ai of peer i is large, then,
all other quantities in (16) remaining fixed, multiplier λi (which is also the price associated with
i's endowment) will be small, i.e., peer i will be less resource constrained, as expected.
Finally, peer i allocates resources to j proportionally to bids
xji(t + 1) = bij(x(t), λi)
(17)
where the bids bij are defined by (15) and multipliers λi solve (16). This is a proportional-
response with nonlinear price discrimination. We call the algorithm an EG-sparse proportional-
response (EGsPaRse).
Algorithm 1 Eisenberg-Gale Sparse Proportional Response (EGSPARSE)
1: Initialization (time t = 0): Peers allocate resources x(0) either equally or randomly.
2: repeat
3:
4:
5:
6:
Each peer i computes its exchange ratio ρi(t) (14) and communicates it to the network.
Each peer i determines the bids bij(t) for its resource from (15), where the multiplier λi
solves (16).
Each peer i allocates resources xji(t + 1) according to (17)
t ← t + 1
7: until Convergence
When there is no sparsity-promoting penalty c = 0, the recursion becomes
xij(t + 1) = aj xij(t)
ai
ri(t)(cid:30)Xk6=j
xkj(t)
ak
rk(t)
.
(18)
Updates (18) coincide with the standard proportional-response dynamics of [3]. To verify this,
observe that iteration (18) corresponds to two steps of proportional-response: In the numerator,
September 22, 2018
DRAFT
11
peer i reciprocates j by charging a constant per-unit price ri(t)/ai (linear pricing), likewise each
peer k in the denominator, and for peer j reciprocating i in the entire fraction.
In the general case c > 0, peers are required to communicate either their exchange ratio,
or the multiplier λ (which is also related to the exchange ratio). This implicitly assumes peers
declare their true ratio. In practice, peers may be unwilling to disclose their exchange ratio (due
e.g. to privacy) or strategically misreport it, to extract additional resources. Such strategic/non-
cooperative behaviour by peers who anticipate the effect of reporting their ratio may result in
loss of optimality.
C. An Alternative Formulation: SPaRse Algorithm
We next turn to an alternative formulation, which leads to an intuitively appealing algorithm.
Recall the definition of the Kullback -- Leibler divergence between two vectors,
D(u, v) := Xi
ui log
ui
vi
−Xi
(ui − vi), u, v ≥ 0.
Inspection of the Eisenberg -- Gale program (1) shows it is equivalent to minimizing the divergence
D(a, r) between allocated a and received r resources (subject to constraints). It is natural to
wonder whether we may seek to minimize D(r, a) instead of D(a, r); although divergence is
in general not symmetric. It turns out that the former optimization also captures the optimal
allocations, a result due to Shmyrev [13] (see also discussion in [4]). The key advantage of
this alternative formulation is that it nicely fits the proportional-response dynamics. Hence, we
consider a convex program equivalent to (1), obtained from minx D(r, a), together with a sparsity
penalty:
(S) min
log(ǫ + xij)
x
D(r, a) + cXi,j
subject to Xj6=i
xji = ai,
∀i.
(19)
Optimization problem (19) is nonconvex; we will derive an algorithm that computes a local
minimum using the minorization-majorization procedure [7]. The updates can be expressed in
terms of a Bregman divergence Bh, associated with the convex negative entropy function h.
Definition 1: Let ψ : X → IR be a strongly convex function on a convex set X. The Bregman
divergence Bψ : X × X → IR associated with the strongly convex function ψ is defined by
Bψ(u, v) := ψ(u) − ψ(v) − (u − v)T ∇ψ(v).
(20)
DRAFT
September 22, 2018
12
The Bregman divergence is a distance -- like function, as it satisfies Bψ(u, v) ≥ 0 for all u, v,
thanks to the convexity of ψ. For example, ψ(u) = 1
2u2 induces the usual Euclidean distance
2 u − v2, however the Bregman divergence is in general not symmetric (for more
Bψ(u, v) = 1
properties see e.g. [14]). The Bregman distance generated by the negative entropy
is the Kullback -- Leibler divergence:
h(r) := Xi
ri log ri,
r ≥ 0,
Bh(u, v) = D(u, v), u, v ≥ 0.
(21)
(22)
This particular choice of Bregman function (instead of usual Euclidean distance) is motivated
by the fact that entropy better reflects the geometry of the simplex constraints [14], [4] (so that
the latter are easily eliminated).
Let f (x) be the objective function in (19). In the majorization step, a point x(t) is used to
anchor a surrogate function g(xx(t)) which upper bounds f ,
f (x) ≤ g(xx(t)), x ≥ 0,
and is easy to minimize. Function g is chosen to be tight at x(t), i.e., f (x(t)) = g(x(t)x(t)).
In the minorization step, the upper bound is minimized with respect to x, generating a sequence
for each t = 0, 1, . . . . We form the surrogate g(xx(t)) as follows: Write the divergence D(r, a)
x(t + 1) = arg min
g(xx(t)),
x
(23)
as
D(r, a) = h(r) −Xi
ri log ai − (ri − ai).
Because of (22), negative entropy (21) satisfies
h(r) = h(r(t)) + (r − r(t))T ∇h(r(t)) + D(r, r(t)).
(24)
(25)
The divergence D(r, r(t)) in (25) is bounded using Lemma 1 (end of Section III). The logarithm
in (19) is bounded by the first-order Taylor expansion
log y ≤ log y0 +
y − y0
y0
,
(26)
DRAFT
September 22, 2018
as is customary in the reweighted ℓ1 minimization [8] framework. Inserting (24) and (25) in the
objective (19) and taking into account inequalities (26) and (32) gives
g(xx(t)) = D(x, x(t)) + h(r(t)) + (r − r(t))T ∇h(r(t))
13
log(ǫ + xij(t)) +
xij − xij(t)
ǫ + xij(t)
+cXi,j
−Xi
ri log ai − (ri − ai).
(27)
The updated allocations x(t + 1) are computed by minimizing the surrogate (27) in (23). After
some algebra, also making use of ∂D(x, x(t))/∂xij = log xij − log xij(t), we get
xij(t + 1) = xij(t)
ai
ri(t)
exp(cid:18)−
c
ǫ + xij(t)(cid:19) .
(28)
Finally, allocations (28) are normalized to satisfy the endowment constraint Pi6=j xij = aj
for each peer j. We thus arrive at a second algorithm for sparse proportional-response, where
nonlinear prices (with an exponential factor) are charged to users:
Each peer i computes the price µji(t) (per unit resource) charged to peer j 6= i at time
t = 0, 1, . . . as
µji(t) :=
ri(t)
ai
exp(cid:18)
c
ǫ + xij(t)(cid:19) .
(29)
Pricing above is nonlinear, because µji(t) depends on the amount of resource xij(t) (payment)
offered to i, inside the exponential. The higher the resource xij(t) (payment) offered by peer
j to peer i, the lower the price µij(t) (per unit resource) charged to i, and the price converges
to the exchange ratio ri(t)/ai as payment xij(t) goes to infinity. Hence, the proposed dynamics
reinforce exchanges that involve large amounts of resources.
Next, each peer i communicates to other peers j the price µji(t). Peer i computes the bids of
other peers j for i's resource as
bij(x(t)) :=
xji(t)
µij(t)
,
j 6= i.
(30)
This corresponds to the amount of resource with which j intends to reciprocate i. Alternatively,
peers can communicate directly the bids instead of prices. Bid bij is also the number of resource
units that peer i can purchase from j with total payment xji(t), at price µij(t).
Subsequently, peer i allocates his resource to peer j proportionally to the received bids,
xji(t + 1) = ai
bij(x(t))
bik(x(t))
Xk6=i
,
j 6= i.
(31)
DRAFT
September 22, 2018
14
This is a proportional-response with nonlinear price discrimination, where larger amounts of
resource are "sold" at lower per-unit price (discount). We call this algorithm a Shmyrev-sparse
proportional-response (SPaRse).
Algorithm 2 Shmyrev Sparse Proportional Response (SPARSE)
1: Initialization (time t = 0): Peers allocate resources x(0) either equally or randomly.
2: repeat
3:
4:
5:
6:
Each peer i computes the nonlinear price µji(t) (29) (per unit resource) he charges to
each peer j, and communicates µji(t) to peer j.
Each peer i determines the bids bij(t) for its resource by peer j using (30).
Each peer i allocates resources xji(t + 1) proportionally to bids (31).
t ← t + 1
7: until Convergence
Each round of SPaRse has O(N 2) computation and communication complexity. If there is no
sparsity penalty (set c = 0) we recover again recursion (18), which is the standard proportional-
response [3]. As will be seen in the numerical results of Section IV, the variant with Equal first
round allocation x(0) tends to generate graphs with mostly direct reciprocation, while Random
first round leads to indirect reciprocation, This is likely due to the fact that a random initial
allocation adds uncertainty and erases symmetry, so that it gets impossible to recover the more
orderly direct reciprocation.
The analysis above can be extended to address a slightly different model, where each peer is
constrained by the maximum number of active connections it can maintain at all time slots.
Lemma 1: For all x, y ≥ 0 it holds that
D(r(x), r(y)) ≤ D(x, y).
(32)
Proof: Inequality (32) follows from the joint convexity of the function d(x, y) = x log(x/y)
in (x, y) and Jensen's inequality [4].
IV. NUMERICAL RESULTS
We evaluate the performance of the SPaRse proportional-response algorithm in several numer-
ical examples; EGsPaRse is omitted for brevity. The examples showcase the formation of sparse
exchange graphs by peers who communicate bids/prices in a distributed manner, and compute
September 22, 2018
DRAFT
15
1: 155.81
108.9
30.31
29.93
16.98
6: 107.07
122.58
8: 86.49
107.07
56.18
57.81
5: 122.58
4: 278.09
76.46
75.33
53.52
52.76
31.78
31.37
58.51
57.66
2: 75.33
3: 52.76
7: 31.37
9: 57.66
Fig. 2. SPaRse-Equal: Exchange graph in a N = 9 node network, T = 10, 000 iterations.
allocations that achieve close to perfect reciprocation (minimum exchange ratio near one). We
discuss the influence of the initial split (SPaRse-Equal versus SPaRse-Random variants) on the
properties of the induced graphs in terms of direct/indirect reciprocation, the role of the link
cost parameter c, and the temporal effects (number of iterations T ) on the sparsity and fairness
of the resulting allocations.
The SPaRse algorithm is applied to a 9-node network, where node endowments are shown in
the circles, i.e., node 1 endowment is 155.81. After 10, 000 iterations of SPaRse-Equal (with c =
0.1, ǫ = 0.01), a graph with 16 links is generated, shown in Figure 2, together with the computed
allocations xij. The minimum exchange ratio is 0.981, and the divergence between received and
allocated resources is D(r, a) = 0.125. We see the majority of links are bidirectional: Only links
1 → 6, and 6 → 5 do not have their reverse in the graph, so among these three nodes indirect
reciprocation takes place.
We next consider a 25-node network, with sample mean node endowment ¯a = 106.22 and
standard deviation std(a) = 53.48, drawn from a lognormal log ai ∼ N (4.5, 0.25) distribution.
We compare influence on the resulting allocations of the Random and Equal initial splits.
Figure 3 shows six sample paths of the SPaRse algorithm (with c = 0.2, ǫ = 0.01); solid
red line corresponds to Equal initial split of resources, and dashed blue lines correspond to
five Random initial splits. It appears that starting with Equal allocation requires more time to
converge. We see that, in general, convergence takes place to different allocations, and slightly
different minimum exchange ratios, which are larger than 0.9, not too far from 1. The top left
plot shows that the cardinality of the final allocation is roughly the same under both Random
September 22, 2018
DRAFT
100
80
60
40
100
50
0
s
k
n
i
l
f
o
.
m
u
N
s
k
n
i
l
l
a
c
o
r
p
c
e
r
f
o
.
i
m
u
N
Equal
Random
1000
2000
iterations
1.5
1
0.5
e
c
n
e
g
r
e
v
d
L
−
K
i
3000
0
0
1
o
i
t
a
r
0.95
m
u
m
n
M
i
i
0.9
1000
2000
iterations
3000
0.85
0
1000
2000
iterations
1000
2000
iterations
16
3000
3000
Fig. 3. SPaRse convergence in a N = 25 node network: Equal (solid) vs. Random (dashed) initial allocation.
and Equal, i.e., regardless of initial conditions. More interestingly, the bottom left plot in Figure
3 suggests qualitatively different behavior of the two variants: (a) SPaRse-Equal forms a graph
that implements direct reciprocation (number of reciprocal links is almost equal to total number
of links, at about 45); while (b) SPaRse-Random generates graphs that implement indirect
reciprocation, as there are very few reciprocal links.
The impact of different random initial allocations is quantified in the the same 25-node
network, and node endowments as Figure 3. We run SPaRse-Random 1, 000 times (with c = 0.1,
ǫ = 0.01), each time with a different random split of resources in the first round. Runs are 5, 000
iterations long, by then allocations have converged. We record four performance metrics: (i)
the cardinality of the final allocation (the number of directional links in the resulting exchange
network), (ii) the reciprocity (i.e., the number of links for which their reciprocal is also in the
graph), (iii) the minimum exchange ratio over the 25 nodes, and (iv) the divergence D(r, a)
between received r and allocated a resources. Figure 4 shows histograms and mean values
(vertical black line) for all 4 metrics. We see that SPaRse-Random usually achieves a minimum
exchange ratio larger than 0.92, with sparse graphs consisting of less than 50 edges, out of
25 × 24 = 600 totally in a complete graph with 25 nodes. The top histograms (cardinality of
x, reciprocity) once more indicate that graphs generated by SPaRse-Random manifest mostly
indirect reciprocation, since (on the average) only about 4 out of the 46 links are reciprocal.
September 22, 2018
DRAFT
17
y
c
n
e
u
q
e
r
f
e
v
i
t
l
a
e
R
0.4
0.3
0.2
0.1
0
0.2
0.15
0.1
0.05
y
c
n
e
u
q
e
r
f
e
v
i
t
l
a
e
R
40
Num. of links
45
50
0
0.9 0.92 0.94 0.96 0.98
Minimum ratio
0.4
0.3
0.2
0.1
0
0
0.2
0.15
0.1
0.05
0
5
15
Num. of reciprocal links
10
0.2 0.4 0.6 0.8
1
Divergence D(ra)
1.2
Fig. 4. SPaRse-Random in N = 25 node network. Histogram of metrics over 1000 runs with random initial splits.
50
40
30
20
s
k
n
i
l
f
o
.
m
u
N
10
0
0.05
0.1
Sparsity cost c
0.4
0.3
0.2
0.1
e
c
n
e
g
r
e
v
d
i
L
−
K
0
0
0.05
0.1
Sparsity cost c
0.15
Fig. 5. SPaRse-Equal: Performance under different sparsity parameters c (N = 11 node network).
The role of sparsity parameter c is examined in Figure 5. In a N = 11 node network, five
endowment vectors are randomly drawn from the same lognormal distribution as before. For
each endowment vector we run SPaRse-Equal with different sparsity parameters c and record
the cardinality of the resulting allocation, and the divergence D(r, a), to get five cardinality
and divergence curves. The duration of each run is 104 iterations. As c decreases, the algorithm
computes more fair allocations (smaller divergence, larger minimum exchange ratio), but takes
longer to converge. Decreasing c below 0.05 (while average endowment is about 100) yields
close to zero divergence, i.e., perfect reciprocation (Figure 5, right). However, for c smaller than
0.05, and when computations stop after 104 iterations, we see that almost zero divergence is
accompanied by an increase in number of links in the graph (Figure 5, left).
The discussion above suggests that, by tuning the parameters c and ǫ, our model can generate
September 22, 2018
DRAFT
18
28.36
28.37
1: 155.81
36.72
36.72
5: 122.58
19.11
19.12
3.18
3.08
31.51
31.5
38.22
38.20
39.32
39.29
8: 86.49
0.1
0.07
2: 75.33
8.68
8.86
12.32
12.14
24.34
0.02
24.18
35.75
35.66
6: 107.07
9: 57.66
43.77
43.76
48.8
48.90
4: 278.09
31.32
31.36
7: 31.37
52.75
52.74
29.38
29.51
3: 52.76
Fig. 6. SPaRse-Equal: Exchange graph in a N = 9 node network, T = 500 iterations.
graphs with various levels of sparsity and reciprocation, which also evolve temporally as the
allocation of resources changes over the course of time. Apart from the static graphs that arise
after SPaRse converges, one may also take a snapshot of the network at some finite time, during
the transient. For example, at time t = 0 let us start with the 9-node endowments of Figure 2
and apply SPaRse in a complete graph, which sparsifies as time elapses. By stopping early after
500 iterations, we obtain the graph shown in Figure 6. This consists of 35 links (as compared
to 16 links in Figure 2), where only link 3 → 1 is not directly reciprocated (but has small
allocation 0.02). The minimum exchange ratio is 0.998, and divergence D(r, a) = 0.001, while
the respective values in Figure 2 were 0.981 and 0.125. Therefore, graph in Figure 6 is less
sparse than Figure 2, but realizes more fair exchanges. A common feature of the allocations in
both Figures 2 and 6 is that low endowment nodes apparently never exchange resources with
each other.
V. CONCLUSION
We studied a resource exchange network where exchanges among nodes are based on reci-
procity. To incorporate costs of establishing and maintaining active connections, we imposed
sparsity penalties on peer interactions. Finding the sparsest graphs that achieve a certain level of
reciprocation is in general NP-hard. We proposed decentralized algorithms, that enable peers to
approximately compute the sparsest allocations, by generalized proportional-response dynamics,
with nonlinear pricing. Numerical results illustrate the performance of the SPaRse algorithms and
September 22, 2018
DRAFT
19
the formation of exchange graphs by peers who achieve close-to-perfect reciprocation (minimum
exchange ratio near one), in a network with a limited number of active connections.
REFERENCES
[1] G. Iosifidis, L. Gao, J. Huang, and L. Tassiulas, "Incentive mechanisms for user-provided networks," IEEE Communications
Magazine, vol. 52, no. 9, pp. 20 -- 27, 2014.
[2] M. Zubeldia, A. Ferragut, and F. Paganini, "Neighbor selection for proportional fairness in P2P networks," Computer
Networks, vol. 83, no. 4, pp. 249 -- 264, 2015.
[3] F. Wu and L. Zhang, "Proportional response dynamics leads to market equilibrium," in Proceedings of the Thirty-ninth
Annual ACM Symposium on Theory of Computing, ser. STOC '07, 2007, pp. 354 -- 363.
[4] B. Birnbaum, N. R. Devanur, and L. Xiao, "Distributed algorithms via gradient descent for Fisher markets," in Proceedings
of the 12th ACM Conference on Electronic Commerce, ser. EC '11, 2011, pp. 127 -- 136.
[5] E. Eisenberg and D. Gale, "Consensus of subjective probabilities: The pari-mutuel method," Annals of Mathematical
Statistics, vol. 30, no. 1, pp. 165 -- 168, 1959.
[6] D. R. Hunter and K. Lang, "A tutorial on MM algorithms," The American Statistician, vol. 58, no. 1, pp. 30 -- 37, January
2004.
[7] Y. Sun, P. Babu, and D. P. Palomar, "Majorization-minimization algorithms in signal processing, communications, and
machine learning," IEEE Transactions on Signal Processing, vol. 65, no. 3, pp. 794 -- 816, February 2017.
[8] E. J. Cand`es, M. B. Wakin, and S. P. Boyd, "Enhancing sparsity by reweighted ℓ1 minimization," Journal of Fourier
Analysis and Applications, vol. 14, no. 5, pp. 877 -- 905, 2008.
[9] J. Brodie, I. Daubechies, C. De Mol, D. Giannone, and I. Loris, "Sparse and stable Markowitz portfolios," Proceedings of
the National Academy of Sciences, vol. 106, no. 30, pp. 12 267 -- 12 272, 2009.
[10] F. Kelly, "Charging and rate control for elastic traffic," European Transactions on Telecommunications, vol. 8, no. 1, pp.
33 -- 37, 1997.
[11] L. Georgiadis, G. Iosifidis, and L. Tassiulas, "Exchange of services in networks: Competition, cooperation, and fairness,"
in Proceedings of the 2015 ACM SIGMETRICS, ser. SIGMETRICS '15, 2015, pp. 43 -- 56.
[12] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY, USA: Cambridge University Press, 2004.
[13] V. I. Shmyrev, "An algorithm for finding equilibrium in the linear exchange model with fixed budgets," Journal of Applied
and Industrial Mathematics, vol. 3, no. 4, pp. 505 -- 518, 2009.
[14] A. Beck and M. Teboulle, "Mirror descent and nonlinear projected subgradient methods for convex optimization,"
Operations Research Letters, vol. 31, no. 3, pp. 167 -- 175, May 2003.
September 22, 2018
DRAFT
|
1411.3675 | 3 | 1411 | 2016-07-23T22:37:22 | Scalable Link Prediction in Dynamic Networks via Non-Negative Matrix Factorization | [
"cs.SI",
"cs.AI",
"cs.IR"
] | We propose a scalable temporal latent space model for link prediction in dynamic social networks, where the goal is to predict links over time based on a sequence of previous graph snapshots. The model assumes that each user lies in an unobserved latent space and interactions are more likely to form between similar users in the latent space representation. In addition, the model allows each user to gradually move its position in the latent space as the network structure evolves over time. We present a global optimization algorithm to effectively infer the temporal latent space, with a quadratic convergence rate. Two alternative optimization algorithms with local and incremental updates are also proposed, allowing the model to scale to larger networks without compromising prediction accuracy. Empirically, we demonstrate that our model, when evaluated on a number of real-world dynamic networks, significantly outperforms existing approaches for temporal link prediction in terms of both scalability and predictive power. | cs.SI | cs |
Scalable Temporal Latent Space Inference for
Link Prediction in Dynamic Social Networks
Linhong Zhu, Dong Guo, Junming Yin, Greg Ver Steeg, Aram Galstyan
1
Abstract-We propose a temporal latent space model for link prediction in dynamic social networks, where the goal is to predict
links over time based on a sequence of previous graph snapshots. The model assumes that each user lies in an unobserved
latent space, and interactions are more likely to occur between similar users in the latent space representation. In addition, the
model allows each user to gradually move its position in the latent space as the network structure evolves over time. We present
a global optimization algorithm to effectively infer the temporal latent space. Two alternative optimization algorithms with local
and incremental updates are also proposed, allowing the model to scale to larger networks without compromising prediction
accuracy. Empirically, we demonstrate that our model, when evaluated on a number of real-world dynamic networks, significantly
outperforms existing approaches for temporal link prediction in terms of both scalability and predictive power.
Index Terms-Latent Space Model, Link Prediction, Non-negative Matrix Factorization, Social Network Analysis
!
1 INTRODUCTION
Understanding and characterizing the processes driv-
ing social interactions is one of the fundamental prob-
lems in social network research. A particular instance
of this problem, known as link prediction, has recently
attracted considerable attention in various research
communities. Besides academic interest (see [15], [24]
for a survey of different methods), link prediction
has many important commercial applications, e.g.,
recommending friends in an online social network
such as Facebook and suggesting potential hires in
a professional network such as LinkedIn.
In this work we focus on the temporal link predic-
tion problem: Given a sequence of graph snapshots
··· , Gt from time 1 to t, how do we predict
G1,
links in future time t + 1? To perform link predic-
tion in a network, one needs to construct models
for link probabilities between pairs of nodes. Recent
research interests in link prediction have focused on
latent space modeling of networks. That is, given
the observed interactions between nodes, the goal
is to infer the position of each node in some latent
space, so that the probability of a link between two
nodes depends on their positions in that space. Latent
space modeling allows us to naturally incorporate the
well-known homophily effect [25] (birds of a feather
flock together). Namely, each dimension of the latent
space characterizes an unobservable homogeneous
attribute, and shared attributes tend to create a link in
a network. We illustrate this concept with an example
Linhong Zhu, Greg Ver Steeg and Aram Galstyan are with Information
Sciences Institute, University of Southern California, email: {linhong,
gregv, galstyan}@isi.edu.
Dong Guo is with the Department of Computer Science, University of
Southern California, email: [email protected].
Junming Yin is with the Department of Management Information Sys-
tems, University of Arizona, email: [email protected].
Fig. 1. An example of observed interactions among Alice,
Bob and Kevin, and their positions in a simplified one-
dimension latent space. Alice is very liberal, Bob is biased to-
wards being liberal, and Kevin is very conservative. However,
all their profile information as well as the dimension label are
unobservable.
shown in Figure 1.
Various approaches,
Example 1: Assume that we have observed some
interactions during a political campaign. An exam-
ple of a latent space is shown in Figure 1, where
most observed links were formed between users with
similar political attitudes. Based on the learned latent
space, we could then predict that Bob is more likely to
interact with Alice than Kevin in a political campaign.
including Bayesian infer-
ence [16], [45], multidimensional scaling (MDS) [32],
matrix factorization [26], [29], [11], [44], [47], [9], and
mixed-membership model [3], have been proposed
to infer the static latent space representation of the
observed network. Nevertheless, most of these studies
were focusing on static graphs, where the latent posi-
tions of the nodes are fixed. In social networks, "There
is nothing permanent except change"; the network
itself is dynamic and changing with time [14] [36]. In
addition, the latent position of each node might also
be evolving over time [32], [34], [8], [48], [46]. A naive
approach to extend static latent space modeling for
dynamic networks is to model each node as a single
latent representation and then update its position
whenever the network evolves. Unfortunately, this
approach tends to overfit on the current time step,
AliceBobKevinconservativeliberal?? 2
problem with latent space models, a significant gap
remains between theoretical capabilities and practi-
cal applications. Towards this end, in this work we
further introduce two new variants of BCGD: a local
BCGD algorithm and an incremental BCGD algo-
rithm. In a local BCGD algorithm, instead of using
all historical graph snapshots to jointly infer temporal
latent space at all timestamps, it sequentially infers
the temporal latent space at each time position with
a single graph snapshot and previous temporal la-
tent space-thus significantly reducing computational
cost. In addition, as we stated earlier, the latent po-
sitions of nodes are changing smoothly over time,
not suddenly. To make use of this temporal smooth-
ness, we develop an incremental BCGD algorithm
to adaptively infer changes in latent positions based
only on the changes in interactions. As shown in
Figure 2, when Bob receives new interactions, his
latent position in time t = 2 requires an update which,
however, can be performed efficiently leveraging his
previous latent position in time t = 1.
In summary, the contributions of this work are:
1) We propose a temporal latent space model for
dynamic networks to model the temporal link
probability of node pairs, which can be further
used to accurately recover or predict the forma-
tion of links.
2) We address algorithmic issues in learning the
temporal latent space representation of nodes
by developing the standard global BCGD algo-
rithm. We also provide a set of theoretical results
for the standard global BCGD algorithm.
3) We develop two fast BCGD algorithms: a lo-
cal BCGD algorithm and an incremental BCGD
algorithm. Furthermore, we illustrate that the
proposed incremental BCGD algorithm only re-
quires a conditional update on a small set of af-
fected nodes, and is therefore significantly faster
than all existing approaches.
4) We conduct experiments over large real-life
graphs. The experimental results show that
the proposed approaches-global BCGD, local
BCGD, and incremental BCGD-achieve good
predictive quality in both sparse and dense net-
works with average AUC scores of 0.81, 0.79
and 0.78 respectively. In addition, it takes only
less than half an hour for the incremental BCGD
algorithm to learn the temporal latent spaces of
massive networks with millions of nodes and
links.
The remainder of the paper is organized as follows:
We first introduce our problem formulation in Sec-
tion 2, and then present the proposed global BCGD
algorithm and two fast BCGD algorithms in Section 3
and Section 4 respectively. In Section 5, we conduct an
experimental study to evaluate the effectiveness and
efficiency of our approach. Related work is discussed
Fig. 2. An example of temporal
latent positions of user
Bob. With the temporal smoothness assumption, Bob is more
likely to be in the blue position than the white position in time
t = 2.
leading to abrupt transitions and poor incorporation
of historical information. Therefore, we are interested
in inferring temporal latent positions for all the nodes
in a dynamic network. The underlying problem can
be stated as follows: Given a sequence of graph snapshots
G1, ··· , Gt, how can we infer the temporal latent space so
that at each time point links are more likely to be formed
between nodes that are closer in the temporal latent space?
With the inferred temporal latent space, links can be
accurately predicted at future times.
Unfortunately, we still have a limited understand-
ing of how the latent space is evolving for temporal
social networks. In this work we propose a Temporal
Latent Space Model for Dynamic Link Prediction. Our
model is based on the intuition that nodes can move
smoothly in the latent space over time, and that
large moves are less likely [32], [46]. As illustrated in
Figure 2, in time t = 1, Bob is biased towards liberal;
it is unlikely that Bob will suddenly move to very
conservative in time t = 2. In addition to temporal
smoothness, our model imposes a constraint on the
dimensionality of the latent space and assumes that
the dimension of latent space is much smaller than the
number of nodes. With the dimensionality constraint,
the online link prediction using our model is efficient
in both computational time and storage cost (see
related work for more details). In addition, varying
the dimension of latent space offers an opportunity
to fine-tune the compromise between computational
cost and solution quality. The higher dimension leads
to a more accurate latent space representation of each
node, but also yields higher online computational
cost.
One of the most widely used methods for infer-
ring low-rank latent space in networks is via matrix
factorization with the "block-coordinate gradient de-
scent (BCGD)" algorithm [6], [38], which has been
successfully applied in community detection for static
networks [42]. However, the introduced additional
temporal smoothness term in the objective function
creates new challenges for the BCGD approach. More-
over, while there is a lot of work on studying con-
vergence of BCGD, it is unknown whether the new
objective function satisfies these criteria. We prove
that the global BCGD algorithm proposed in this work
has a quadratic convergence rate.
The global BCGD algorithm is computationally
expensive and not applicable for truly large-scale
problems. Therefore, for the temporal link prediction
Bobconservativeliberalt=1Bobt=2Bob t=1Bob t=2?Bob t=2?ObservationLatent space in Section 6. We conclude this work and present future
work in Section 7.
2 PROBLEM FORMULATION
A graph G is denoted as (V , E), where V is the set
of nodes and E ⊆ V × V is the set of (directed or
undirected) interactions. In this work we first focus
on undirected graphs, in which case its matrix rep-
resentation is symmetric. We use u and v to denote
individual nodes, and t and τ to denote timestamps.
Let Gτ = (Vτ , Eτ ) be a time-dependent network
snapshot recorded at time τ, where Vτ is the set of
nodes and Eτ ⊆ Vτ × Vτ is the set of interactions. In
addition, denote by ∆Vτ and ∆Eτ the sets of vertices
and interactions to be introduced (or removed) at time
τ, and let ∆Gτ = (∆Vτ , ∆Eτ ) denote the change in the
whole network.
Dynamic social network. A dynamic network G is a
sequence of network snapshots within a time interval
and evolving over time: G = (G1, . . . , Gt).
Temporal latent space and its model. Let Zτ be the
low-rank k-dimension temporal latent space represen-
tation for node set Vτ at time τ. For each individual u
at time τ, we use a row vector Zτ (u) to denote its tem-
poral latent space representation and a scalar Z(u, c)
to denote its position in cth dimension. Inspired by the
dynamic community model [36], [37], we impose the
following assumptions on the temporal latent space:
1) Temporal smoothness: Individuals change their
latent positions gradually over time.
2) Network embedding: Two nodes that are close
to each other in the interaction network, in terms
of the network distance, are also close to each
other in the temporal latent space in terms of
latent space distance.
3) Latent Homophily: Two members who are close
to each other in latent space interact with one
another more frequently than two faraway mem-
bers.
Within our model, we assume that the probability
of a link between two nodes depends only on their
latent positions. In principle, given a dynamic graph
G = (G1,
. . . , Gt), we need Zt+1 to predict future
graph Gt+1. However, since Zt+1 is not available, in
our model we assume it can be approximated by
Z1,··· , Zt. Therefore, given a dynamic graph G =
(G1, . . . , Gt), we infer Z1, ··· , Zt based on G1, ··· ,
Gt, use Z1, ··· , Zt to approximate Zt+1, and finally
predict Gt+1.
With the above definitions, we focus on the follow-
ing problem:
Problem 1: (Temporal Latent Space Inference.) Giv-
en a dynamic social network G = (G1, G2, . . . , Gt), we
aim to find a k-dimension latent space representation
at each timestamp Zτ that minimizes the quadratic
3
loss with temporal regularization:
t(cid:88)
(cid:88)
t(cid:88)
(cid:107)Gτ − Zτ Z T
τ (cid:107)2
arg min
Z1,··· ,Zt
subject to: ∀ u, τ, Zτ ≥ 0, Zτ (u)Zτ (u)T = 1
F + λ
τ =1
τ =1
u
(1 − Zτ (u)Zτ−1(u)T )
(1)
where λ is a regularization parameter, and the term
(1 − Zτ (u)Zτ +1(u)T ) penalizes node u for suddenly
changing its latent position. Note that when comput-
ing the quadratic loss (cid:107)Gτ − Zτ Z T
F , we ignore all of
the diagonal entries.
τ (cid:107)2
In the above model, the latent representation of each
node corresponds to a point on the surface of a unit
hypersphere. Note that this is different from mixed
membership stochastic block models [3] where nodes
are mapped onto simplex. In practice, we find that
sphere modeling gives us a clearer boundary between
linked pairs and non-linked pairs when we project
all pairs of nodes into the latent space. In addition,
we impose the constraints Zτ ≥ 0, not only because
the non-negativity establishes the duality between
our modeling and non-negative matrix factorization,
but also because it gives latent space an intuitive
parts-based interpretation, as suggested by Lee and
Seung [22]. In the facial image example [22], with
non-negative constraints, each dimension of latent
space corresponds to a part of the face, such as nose,
eye and ear; while without non-negative constraints,
each dimension of latent space corresponds to a blur
representation of the entire face, of which it is very
difficult to interpret. Similarly, in the social network
domain, with the non-negative constraints, each di-
mension of latent space corresponds to a part of users'
attributes such as current city, hometown, personality
and so on. It is more intuitive to represent each user
as an additive mixture of attributes (current city: A,
hometown: B, personality: openness to experience),
rather than represent each user as a combination of
different representatives.
Link prediction. Given that we have inferred
Z1, . . . , Zt by optimizing Eq. 1, our goal is to predict
the adjacency matrix Gt+1 at the next timestamp
t + 1. The most natural estimator is the conditional
expectation: Yt+1 = E[Gt+1 Z1,··· , Zt]. By assuming
that the temporal dynamics of latent positions is
Markovian and satisfies E[Zt+1 Z1,··· , Zt] = Zt and
t+1 Z1,··· , Zt] = Dt (a diagonal matrix), as well
cov[Z T
as Zt+1Z T
t+1 an unbiased estimate of Gt+1, we obtain
Yt+1 = E[Gt+1 Z1,··· , Zt]
t+1 Z1,··· , Zt]
= E[E[Gt+1 Z1,··· Zt, Zt+1] Z1,··· , Zt]
= E[Zt+1Z T
= cov[Z T
+ E[Zt+1 Z1 ··· Zt]E[Zt+1 Z1 ··· Zt]T
= Dt + ZtZ T
t .
t+1 Z1 ··· Zt]
However, as we ignore the diagonal entries of the
adjacency matrix (the graph does not contain self-
t
t
τ
to predict Gt+1.
loops) and because the conditional covariance matrix
Dt is assumed to be diagonal, we will use the off-
diagonal of ZtZ T
Generalization. Although in this work, we use ZtZ T
to predict Gt+1, without losing generality, the pre-
dicted graph at t + 1 can be formulated as: Yt+1 =
Φ(f (Z1,··· , Zt)), where Φ is the link function and f is
the temporal function. Learning or selecting the best
link function Φ and temporal function f is beyond
the scope of this work. For example, we could apply
nonparametric approaches [31] to automatically learn
the link function Φ. Additionally, though this work
focuses on undirected graphs, our model can be gen-
eralized to directed and weighted graphs. Specifically,
this can be done by introducing another matrix B,
which represents the weighted mapping from one
dimension to another dimension in the latent space.
With the matrix B, the term Gτ − Zτ Z T
in Eq. 1 is
now generalized to Gτ − Zτ BZ T
τ . If B is symmetric,
the underlying graph is undirected; otherwise, it is
directed.
Problem complexity. Vavasis recently proved that
non-negative matrix factorization (NMF)
is NP-
hard [39]. In addition, with the separability condition,
that is, for each dimension c in the latent space Z,
there is a node u such that Z(u, c) ≥ p and Z(u, c(cid:48)) = 0
for c(cid:48) (cid:54)= c, there exists a polynomial time exact algo-
rithm to solve the non-negative matrix factorization
problem [4]. Unfortunately, in our modeling, there is
no guarantee that the latent space Z satisfies the sep-
arability condition. In addition, even if there exists a
latent space Z that satisfies the separability condition,
Recht et al. [30] pointed out that an exact algorithm
is still too computationally expensive for large-scale
data. Therefore, in the following we focus on the
block coordinate gradient descent (BCGD) approach
to obtain an approximate solution to Problem 1.
3 A STANDARD BCGD ALGORITHM
In this section we present the details of the standard
BCGD algorithm that provide a local optimal solution
to Problem 1.
With a partial observed graph structure, the objec-
tive function in Eq. 1 can be decomposed into a linked
part and a non-linked part. That is,
(Gτ (u, v) − Zτ (u)Zτ (v)T )2
(cid:88)
u,v∈Eτ
τ =1
t(cid:88)
(cid:88)
(cid:88)
u,v(cid:54)∈Eτ
arg min
Z1,··· ,Zt
t(cid:88)
t(cid:88)
τ =1
+
(Zτ (u)Zτ (v)T )2
(2)
(1 − Zτ (u)Zτ−1(u)T )
+ λ
subject to: ∀ u, τ, Zτ ≥ 0, Zτ (u)Zτ (u)T = 1
τ =1
u
Unfortunately, the decomposed objective function
in Eq. 2 for Problem 1 is still a fourth-order poly-
nomial and non-convex. Therefore, we then adopt a
4
block coordinate descent approach to solve Eq. 2. We
update Zτ (u) for each node u at time τ by fixing both
latent positions Zτ (v) of all other nodes v at time τ
as well as all the temporal latent positions other than
at time τ. Proceeding in this way, each step of block
coordinate descent is solving a convex problem.
For each node u at time τ, we focus on optimizing
the following problem:
arg min
Zτ (u)≥0
(cid:88)
J(Zτ (u)), where J(Zτ (u)) =
(Gτ (u, v) − Zτ (u)Zτ (v)T )2 +
(cid:88)
(Zτ (u)Zτ (v)T )2
v∈N (u)
+ λ(1 − Zτ +1(u)Zτ (u)T ) + λ(1 − Zτ (u)Zτ−1(u)T )}
v(cid:54)∈N (u)
(3)
In the following, we use the projected gradient
descent algorithm to find the approximation solution
with a non-negativity constraint. With the gradient
descent optimization algorithm, for each node u at
timepoint τ, we could iteratively update Zτ (u) in each
iteration r + 1 with the following rule:
Z (r+1)
τ
(u) = Z (r)
τ (u) − η∇Zτ (u)J(Z (r)
τ (u))
√
where η is the step size.
Step size. Nesterov's gradient method [28], [13], [5]
iteratively updates the step size using the Lipschitz
constant. According to the following lemma, we show
that our step size can also be iteratively updated using
the Lipschitz constant.
Lemma 1: The gradient of the objective function
J(Zτ (u)) is Lipschitz continuous, and the Lipschitz
n2 − 2n + k, where n is the
constant L is equal to 2
number of nodes in a graph, and k is the number of
dimensions.
Proof: A detailed proof is in Supplement A. (cid:3)
With the Lipschitz constant and ∇Zτ (u)J(Z (r)
τ (u))
shown in Eq. 12 in Supplement A, each node u at
timepoint τ can be computed via the update rule
stated in the following Lemma.
Lemma 2: The latent position of each node u at
timepoint τ can be iteratively computed with the
following update rule:
Z (r+1)
τ
(u) = max((1 + 2α)Z (r)
+ 2α
Gτ (u, v)Z (r)
τ (u) + αλ(Z (r)
τ (u)Z (r)
τ (v) − 2αZ (r)
τ
T
Z (r)
τ
, 0)
τ−1(u) + Z (r)
τ +1(u))
(cid:88)
v∈N (u)
(4)
where α = ar+1+ar−1
ar+1L , N (u) denotes the set of u's
neighbors, L is the Lipschitz constant, d(u) is the
degree of node u, and
√
(cid:40)1
(5)
ar =
1+
4a2
r−1+1
2
if r = 0,
if r > 0.
Proof: A detailed proof is in Supplement B. (cid:3)
Note that when updating Zτ (u) according to
Lemma 2, it requires the latent positions of all the
neighbors of u. For nodes with extra-large degrees,
the above update will become time-consuming. One
dimension k
Algorithm 1 The global BCGD algorithm for jointly
inferring temporal latent space
Input: Graphs {G1,··· , Gt} and latent space
Output: latent space {Z1,··· , Zt} and prediction Yt+1
1: Nonnegative initial guess for {Z1,··· , Zt}
2: Repeat
3:
4:
5:
update Zτ (u) by Eq. 4
6:
normalize Zτ (u)
7: until {Z1,··· , Zt} converges.
8: return Yt+1 = Φ([Z1,··· , Zt]) and {Z1,··· , Zt}
for each time τ from 1 to t
for each u in graph
interesting direction in the future is to apply stochastic
gradient descent update rules to nodes with large
degrees. That is, instead of using the latent positions
of all of u's neighbors, each iteration randomly selects
one of u's neighbors v, and leverages v's latent posi-
tions to update the latent position of u.
We summarize our global block coordinate descent
approach to solve Eq. 2 in Algorithm 1. In the follow-
ing, we provide additional theoretical properties of
the proposed BCGD algorithm, including convergence
analysis and complexity analysis.
τ (u)(cid:107)2
F
3.1 Theoretical Analysis
Convergence rate. Since our algorithm uses Nes-
terov's gradient method to determine the step size
(see Lemma 2), we can conclude that our algorithm
achieves the convergence rate O( 1
r2 ) as stated in the
following theorem.
τ (u)−Z∗
(r+2)2
τ (u))−J(Z∗
for each node u at
τ (u)) ≤ 2L(cid:107)ZI
Theorem 1: Given the sequence generated by Al-
timepoint τ, we
gorithm 1,
have J(Z (r)
, where
Z∗
τ (u) is the optimum solution for Zτ (u) with respect
to the subproblem in Eq. 3, Zτ (u)I is the initialization
for Zτ (u), L is the Lipschitz constant in Lemma 1,
and r is the iteration number (not the total number of
iterations).
Proof: The proof is shown in Supplement C. (cid:3)
Local error bound. The solution returned by Al-
gorithm 1 is a local optimum of the objective in
Eq. 2. Unfortunately, it is not trivial to assess how
this locally optimal solution compares to the global
optimum of the objective in Eq. 2. This is because the
objective function in Eq. 2 is non-convex, and thus
arriving at the global optimum via local iterations
is not guaranteed. Only if the input matrix G is
separable or near-separable [4], [30], we are able to
achieve the global optimum or a global error bound.
Computational complexity. Table 1 summarizes the
cost of each operation in an iteration for both sparse
and dense graphs. Fortunately, Gibson et al. [12]
concluded that real-world networks are generally
globally very sparse, where most of nodes have a
relatively small number of neighbors. Therefore, for
5
TABLE 1
Time complexity analysis for Algorithm 1, n is the number of
nodes, mτ is the number of edges in graph Gτ , d(u) is the
degree of node u, k is the number of dimensions, and T is
the number of timestamps.
O((cid:80)
O((cid:80)
Sparse
Dense
τ (n + mτ )k) O(n2T k)
O(d(u)k)
O(nk)
τ (n + mτ k)
O(n2T k)
Initialize
Update Zτ (u)
Convergence
(a) BCGD
(b) Local BCGD
Fig. 3. BCGD versus local BCGD.
single iteration of BCGD should be O((cid:80)
complexity is bounded by O(Rk(cid:80)
storage complexity is bounded by O((cid:80)
such real-world sparse networks, the total cost of a
τ (n + mτ )k),
which is linear in the number of edges and nodes. Assume
that the total number of iterations is R, then the time
τ (n + mτ )). Since
in Algorithm 1, we need to store all the input matrices
and the output temporal latent space matrices, the
τ (nk + mτ )).
4 FAST BCGD ALGORITHMS
Unfortunately, the standard BCGD is very expensive
in both computation and storage since it requires all
historical graph snapshots {G1,··· , Gt} to jointly and
cyclicly update all temporal latent space {Z1, ··· , Zt}.
In this section we aim to further reduce the time
complexity of the BCGD algorithm so that it scales
with big data. In contrast to the joint inference in
BCGD, we propose a sequential inference algorithm
and an incremental inference algorithm, both of which
utilize the locality and temporal
information. The
proposed two faster BCGD algorithms are as efficient
as the fast independent inference (i.e., infer Zτ from
Gτ only) and as effective as the global joint inference
(BCGD).
As illustrated in Figure 3, we first introduce the
local BCGD algorithm, which sequentially infers each
temporal latent space Zτ from a single graph snapshot
Gτ and partial prior knowledge Zτ−1.
4.1 Local BCGD Algorithm
Specifically, at each timestamp τ, we aim to optimize
the following local objective function to compute Zτ :
(Gτ (u, v) − Zτ (u)Zτ (v)T )2
(cid:88)
arg min
Zτ
+
+ λ
(cid:88)
(cid:88)
u∈Vτ
u,v(cid:54)∈Eτ
u,v∈Eτ
(Zτ (u)Zτ (v)T )2
(1 − Zτ (u)Zτ−1(u)T )
(6)
G1G2GtZ1Z2ZtGraphLatent space......G1G2GtZ1Z2ZtGraphLatent space...... dimension k
Algorithm 2 The local BCGD algorithm for sequen-
tially inferring temporal latent space
Input: Graphs {G1,··· , Gt} and latent space
Output: Yt+1 and latent space {Z1,··· , Zt}
1: Nonnegative initial guess for Z1
2: for each τ from 1 to t
3:
4:
5:
6:
7:
8: return Yt+1 = Φ([Z1,··· , Zt]) and {Z1,··· , Zt}
Initial Zτ based on Zτ−1
repeat
update Zτ (u) by Eq. 7 and normalize it
for each u in graph Gτ
until Zτ converges.
Using the same BCGD approach, we iteratively
update the latent position Zτ (u) of each node u by
fixing the latent positions of all the other nodes. This
leads to the following update rules for Zτ (u) in the
(r + 1)th iteration:
Z (r+1)
τ
(u) = max((1 + 2α)Z (r)
τ (u) + αλZτ−1(u) + 2α
(cid:88)
v∈N (u)
τ (v) − 2αZ (r)
τ (u)Z (r)
τ
T
Z (r)
τ
, 0)
Gτ (u, v)Z (r)
where α = ar+1+ar−1
ar is defined in Lemma 2.
(7)
ar+1L , L is the Lipschitz constant, and
We summarize the proposed local BCGD algorithm
in Algorithm 2. Note that in the global BCGD algo-
rithm, for each iteration we jointly update Z1, Z2 until
Zt and then iterate back to Z1 in the next iteration.
This cyclic update schema is very expensive; while in
the local BCGD algorithm, as shown in Lines 4–7, we
sequentially compute Zτ by a single inner iteration.
That is, we iteratively update Zτ until Zτ converges
and then move to the computation of temporal la-
tent space Zτ +1. This local sequential update schema
greatly reduces the computational cost in practice, as
we analytically show in the following.
Complexity analysis. The cost of each operation in
Algorithm 2 remains the same as that of Algorithm 1,
as reported in Table 1. Thus, the total computational
τ (n + mτ )Rτ k). In practice, since we
leverage prior knowledge Zτ−1 to locally update Zτ ,
it converges much faster than the global BCGD algo-
rithm, and thus the local BCGD algorithm is more
efficient than the global BCGD algorithm, as also
verified in our experiments. In addition, in the in-
memory algorithm where we put all the data into
memory, the global BCGD algorithm requires at least
τ (nk + mτ )) memory cost; while the local BCGD
only requires storage of a single graph snapshot and
two latent space matrices representation. Therefore,
the memory cost of the local BCGD algorithm is
bounded by O(nk + maxτ mτ ).
cost is O((cid:80)
O((cid:80)
4.2 Incremental BCGD Algorithm
In this section we further study how to infer or main-
tain temporal latent space incrementally with graph
6
dimension k
for each u in graph Gτ
Algorithm 3 The incremental BCGD algorithm for
inferring temporal latent space
Input: Graphs {G1,··· , Gt} and latent space
Output: Yt+1 and latent space {Z1,··· , Zt}
01: Nonnegative initial guess for Z1
02: for each time stamp τ from 1 to t
03:
04:
05:
06:
07:
08:
09:
10:
11:
12:
update affected node set S with Alg. 4
13:
14: return Yt+1 = Φ([Z1,··· , Zt]) and {Z1,··· , Zt}
affected node set S=∆Vτ
repeat
update Zτ (u) by Eq. 7 and normalize it
until Zτ converges or S is empty.
Initialize Zτ (u) using Eq. 8
if Gτ (u) is not updated
Zτ (u) = Zτ−1(u)
for each u in S
else
changes (nodes and edges insertion and deletion).
Instead of recomputing the temporal latent space Zτ
with the entire graph snapshot Gτ and the prior
knowledge Zτ−1, our dynamic update strategy is to
adjust Zτ incrementally from Zτ−1 as the network
structure changes. Specifically, we take advantage of
what we already learned in previous snapshots, as
well as the temporal transition for inferring the dy-
namics in the current timestamp.
Overview of incremental BCGD. The overview of
this algorithm is as follows: First, we identify nodes
whose local neighborhood has changed between two
consecutive snapshots, including cases where an exist-
ing node adds or deletes links, or a node is added to or
removed from the network. With the matrix represen-
tation, we model all the four updates towards Gt in
terms of row-wise updates on matrix representation.
For example, a newly added node with degree d is
modeled as an update from a row with all-zero entries
to a row with d non-zero entries. For nodes whose
local neighborhood has not changed, we assume that
their initial temporal latent positions do not change
either. For nodes whose neighborhood has changed,
we initialize their new temporal latent positions based
on their new neighbors' latent space positions. Next,
we iteratively perform a conditioned latent position
update for each affected node (i.e., a candidate set
of nodes whose latent position might change) and an
update for the affected node set until there are no
more changes in the temporal latent position of each
node. The entire process is outlined in Algorithm 3.
Initialize updated nodes. In our algorithm, the four
update operations are handled simply by comparing
whether Gτ (u) is identical to Gτ−1(u). For each up-
dated node u, we initialize its latent position based
on the probability of seeing its updated neighbors'
latent positions. Specifically, for each node u and the
dimension c of the latent space at time τ, the position
of u in dimension c is computed using the following
equation:
Zτ (u, c) =
(cid:80)
(cid:80)
v∈N (u) Gτ (u, v)Zτ (v, c)
q∈N (u) Gτ (u, q)
(8)
The initialization of latent position for an updated
node u follows the notion of "latent homophily"
introduced earlier: The latent position of the node u is
as close as possible to those of its network neighbors.
Identifying affected nodes. Our dynamic update
strategy can be viewed as an extra conditional update
by which only nodes affected accept a new latent posi-
tion. Unfortunately, the set of affected nodes for which
the latent positions need to be updated is not limited
to only the set of recently changed nodes and/or their
network neighbors. Therefore, how could we identify
the set of affected nodes?
The overall idea of our affected nodes identification
is outlined as follows. Initially, the set of affected
nodes is identical to the set of updated nodes (Line 8
in Algorithm 3). Next, after performing each iteration
of a conditional update with Eq. 7 in Line 11 in
Algorithm 3, some old affected nodes are no longer
required to be updated since their latent positions
have converged. On the other hand, the effects of
the update could be further propagated from old
affected nodes to their neighborhood. The details of
our affected nodes update procedure are presented in
Algorithm 4.
Algorithm 4 Updating affected nodes
Input: Zτ−1, Zτ , Sold
Output: A set of affected nodes S
1: S = Sold
2: for each u in Sold
3:
S = S \ {u}
4:
5:
if Zτ (u)Zτ (w)T − Zτ−1(u)Zτ−1(w)T ≥ ζ
6:
7:
8: return S
if ∀c, Zτ (u, c) − Zτ−1(u, c) < δ
for each w ∈ N (u)
S = S ∪ {w}
τ
of Algorithm 3 is O(k(cid:80)
(cid:80)
Complexity analysis. The total computational cost
r Sr + N (Sr)), where
r denotes the iteration number of the conditional
update process outlined in Lines 9–13 in Alg. 3, Sr
denotes the set of affected nodes and N (Sr) denotes
the neighborhood of affected nodes. When τ=1, Sr is
equal to the number of nodes n, and N (Sr) is equal
to the number of edges in the first graph snapshot
m1. For the in-memory storage cost, compared with
Algorithm 2, Algorithm 3 requires additional storage
cost about ∆Gτ . Therefore, the in-memory storage
cost of Algorithm 3 is O(nk + maxτ (mτ + ∆mτ )).
5 EXPERIMENTS
5.1 Dataset and Evaluation
We use five real temporal datasets in our experiments,
which are obtained from the Koblenz Large Network
7
TABLE 2
Statistics of data sets, where volume denotes the total
number of interactions, and the number of edges denotes
the number of distinct interactions.
# nodes
# edges
volume
410
63,731
28,093
17,298
2,765
817,035
183,412
4,596,803
3,148,447
1,314,050
3,223,589
18,986,618
9,375,374
10,724,828
9,375,374
# snap-
shots
8
5
9
11
7
Data
Infection
[18]
Facebook
[40]
HepPh
[23]
DBLP [1]
YouTube
[27]
Collection [21]. The statistics of each data set are
reported in Table 2. Note that in our experimental
setting, each graph snapshot Gτ consists of all the
nodes and the interactions that appear within the time
interval [τ − 1, τ ].
Evaluation metrics. We evaluate the performance
of the proposed approaches from three aspects: the
effect of parameters, the efficiency of both offline
latent space inference and online link prediction, and
the accuracy of online link prediction. We use total
running time and memory consumption to evaluate
the efficiency of both offline latent space inference
and online link prediction. We use prediction error
to evaluate the inference accuracy. Given the train-
··· , Gt, prediction error is defined
ing graph G1,
τ−1(cid:107)F . Therefore, a smaller
as
prediction error indicates better inference accuracy.
(cid:80)t
τ =2 (cid:107)Gτ − Zτ−1Z T
1
t−1
For link prediction accuracy, we use Area Under
Curves (both Receiver Operating Characteristic (ROC)
and Precision-Recall (PR) curves), termed as AUCROC
and AUCPR. The ROC curve is created by plotting the
true positive rate vs. false positive rate; while the
PR curve is created by plotting precision vs. recall at
various threshold settings. Thus, AUC score evaluates
the overall ranking yielded by the algorithms with a
larger AUC indicating better link prediction perfor-
mance. In the following we only report the experiment
results based on AUCROC; the experiment results based
on AUCPR are reported in the Supplement.
Test pair generation. Given the training graph G1,
··· , Gt, we perform link prediction for the test graph
Gt+1. We provide two different setups to generate
test pairs: (1) all
links: In this setup, we focus on
how well the methods predict links in the test graph,
no matter whether those links are repeated links or
new links. Toward this goal, we randomly generate
100,000 linked and an equal number of non-linked
pairs from Gt+1 as test pairs. (2) new links: In this
setup, we focus on how well the methods predict the
emergence of new links and deletion of existing links.
Again, we randomly generate 10,000 linked and an
equal number of non-linked pairs from ∆Gt+1.
TABLE 4
Effect of the smoothing parameter λ
8
Prediction error by BCGDG
0
228
3436
2043
80566
195023
[0.0001, 1]
225±1.26
3312±16
1697±7.9
65878±429
161239±726
Prediction error by BCGDL
λ
Infection
Facebook
Hepph
DBLP
Youtube
λ
Infection
Facebook
Hepph
DBLP
Youtube
0
250
3659
2566
81633
198646
[0.0001, 1]
243±3.5
3485±36
2106±51
65957±18
162287±357
10
288
3410
1769
65926
162282
10
249
3525
2172
66471
163675
(a) Inferring time (s)
(b) Prediction errors
Fig. 4. Effect of dimensionality k to proposed approaches.
Comparable approaches. We compare our approaches
(standard BCGD BCGDG, local BCGD BCGDL and in-
cremental BCGD BCGDI) with a set of state-of-the-
art approaches, which are summarized in Table 3
and Section 6. For the offline latent space inference
time comparison, we compare our approaches with
global optimization approach DMMSB [10], local opti-
mization approaches NMFR [43], Hott [30], BIGCLAM
[42], PTM [45] and incremental optimization approach
LabelRT [41]. We compare online link prediction times
of our approaches to that of high-dimension latent
space approaches BIGCLAM and LabelRT, and pop-
ular graph heuristic AA [2]. For the Hott, NMFR and
DMMSB, the total running time for their approaches
is the same as that of our BCGD approaches. This is
because the online prediction time cost is proportional
to the number of dimensions k, and k is set to 20 for all
the low-dimension constraint latent space approaches
including our BCGD approaches, Hott, NMFR, PTM
and DMMSB.
Configurations. The results are reported as an average
over 10 runs. If the maximum number of iterations is
required as an input; we set it to 100. All experiments
are conducted on a single machine with 8G memory
and i7 2.7 GHZ CPU.
5.2 Effect of Parameters
This set of experiments aims to determine the impact
of parameters on the performance of the proposed
approaches and/or the comparable approaches.
The effect of temporal smootheness. While fixing
all the other parameters (e.g., k=20), we first study
the effect of the smoothing parameter λ. The effect
of parameter λ might be related to the inference
algorithms, thus we evaluate the performance of both
global BCGD and local BCGD when varying λ. We
do not plot the running time since we notice that the
running time is insensitive to the parameter λ. We
vary λ from 0.0001 to 10 with logarithmic scale and
report the prediction errors shown in Table 4. We also
report the prediction errors for λ=0. Note that when
(a) Inferring time (s)
(b) Prediction accuracy
Fig. 5. Effect of dimensionality k to other approaches on
Facebook dataset.
λ is 0, the update rules stated in Eq. 4 and Eq. 7
become identical. BCGDG differs from BCGDL since
they have different ways to proceed with initialization
and convergence, see Lines 2–7 in Algorithm 1 and 2.
Clearly, either absence of temporal smoothness (λ=0)
or strong temporal smoothness (λ=10) leads to higher
prediction errors. While λ varies from 0.0001 to 1,
there are no significant differences in terms of predic-
tion errors. Therefore, in the following experiments,
we simply fix λ as 0.0001 for BCGDG and 0.01 for
BCGDL.
The effect of dimensionality. Next, we fix all the
other parameters and analyze the effect of k. Since
the effect of number of dimensions k is lightly cor-
related with inference algorithms, here we choose
the representative local BCGD algorithm. We vary k
from 5 to 30, and report both the running time of
temporal latent space inference and prediction error in
Figure 5. The overall trends indicate that the running
time increases with number of dimensions k, while
prediction error decreases with number of dimensions
k. However, increasing number of dimensions does
not necessarily lead to the decrease of prediction
error. For example, for a Facebook dataset, when the
number of dimensions k is increased from 25 to 30,
the prediction error is increased. In order to balance
the efficiency and effectiveness, we opt to fix k=20 for
the proposed approach in all of the following experi-
ments. A nonparametric approach that automatically
selects the best dimension k will be an interesting
future direction.
In order to perform a fair comparison, we also ex-
amine the effect of k to other comparable approaches
InfectionFacebookHepPhDBLPYoutube 1101102103 5 10 15 20 25 30Running time (S)Number of dimensions k103104105 5 10 15 20 25 30Prediction errorNumber of dimensions kBCGDLPTMDMMSBNMFRHott101102103104 5 10 15 20 25 30Running time (S)Number of dimensions k 0.4 0.5 0.6 0.7 5 10 15 20 25 30AUCROCNumber of dimensions k A summary of state-of-the-art approaches, where time series denotes using a sequence of graph snapshots as
inputs, aggregated denotes using a single aggregated static graph snapshot as an input, low denotes low
dimensional latent space and high denotes without dimensionality constraint.
TABLE 3
aggregated
time-series
inference
Scalability
√
√
prediction
√
√
Accuracy
all links
new links
Latent space dimension
√
low high
√
NA
√
√
√
PTM [45], NMFR [43]
Hott [30]
LabelRT [41]
BIGCLAM [42]
AA [2]
BCGDG, DMMSB [10]
BCGDL, BCGDI
√
√
Input graph
√
√
√
√
√
√
√
√
√
√
√
NA
√
√
√
√
√
√
9
√
√
√
(a) Inferring time (s)
(b) Prediction accuracy
Fig. 6. Effect of number of snapshots to proposed and other
approaches on Facebook dataset.
(a) Inferring time (s)
(b) Prediction errors
Fig. 7. Effect of parameters ζ and δ on Facebook dataset
(Best viewed in color).
including DMMSB, PTM, NMFR, and Hott. LabelRT
and BIGCLAM are not included because their best
dimension k is automatically computed. Furthermore,
since several baselines DMMSB, PTM, NMFR are not
scalable and can not support running on the two large
datasets DBLP and YouTube, we simply choose the
relatively largest dataset, Facebook, on which all of
the approaches are able to be tested. The comparison
of running time and AUCROC on new links is reported
in Figure 5(a) and Figure 5(b) respectively. Clearly, the
overall trends of other approaches are very similar
to those of BCGDL: The inference time is growing
linearly or even quadratically with the dimensionality
k, while the prediction accuracy also improves with
k. Therefore, we use the same value of k=20 for other
comparable approaches as well.
The effect of time intervals.
In order to understand
the proposed model more deeply, we also study how
the value of time interval τ impacts prediction perfor-
mance. Specifically, we choose the Facebook dataset,
vary the value of time interval τ, and create five
different sequences of graph snapshots. As reported
in Figure 6(a) and Figure 6(b), both the inference
time and prediction accuracy increase with number
of snapshots. The inference time increases because
a larger number of snapshots mean more frequent
model updates per shorter time interval. Meanwhile,
shorter time interval leads to less change of graphs,
which indicates that it is more reliable to forecast links
at time t+1 based on the current observation at time t.
Therefore, the prediction accuracy also improves with
number of snapshots. Again, the overall trends hold
for both proposed and other approaches. Therefore,
we simply fix the number of snapshots as reported in
Table 2.
The effect of lazy update. Finally, we evaluate the
effect of two parameters ζ and δ that are related to the
incremental BCGD algorithm. Since parameters ζ and
δ are correlated with each other, we vary both of them
and report the total inference time and prediction
error on the Facebook dataset in Figure 7. The results
verify that when setting both ζ and δ to smaller
values, the predictive accuracy is improved, while
the total running time is also increased. However,
Figure 7(a) shows that when we reduce ζ and δ from
10−2 to 10−4, the running time almost remains the
same. The results verified that we can safely set ζ and
δ to relatively small values without losing too much
efficiency. We repeated this set of experiments on all
the other datasets. Then, using curve fitting to capture
how the best value of ζ and δ can be approximated
by parameters such as n, m, T , and k, we noticed
that a good value for ζ and δ can be achieved around
(cid:112)− log(1 − 1/n) and 2ζ/k. Therefore, in the follow-
(cid:112)− log(1 − 1/n) and δ as 2ζ/k.
ing experiments, without specification, we set ζ as
5.3 Evaluation of Efficiency
In this section we first compare the proposed global
BCGD BCGDG,
local BCGD BCGDL, and incremen-
tal BCGD BCGDI with other latent space inferring
BCGDGBCGDLBCGDIlabelRTDMMSB101102103104 6 12 18 24 30Inferring time (s)Number of snapshots 0.5 0.6 0.7 0.8 6 12 18 24 30AUCROCNumber of snapshots10-510-410-310-2ζ10-410-310-2δ 9.3 9.6 9.9 10.2Running time (S) 9.3 9.6 9.9 10.210-510-410-310-2ζ10-410-310-2δ 3480 3490 3500 3510 3520Prediction error 3480 3490 3500 3510 3520 (a) Inferring time (s)
(b) Memory usage (MB)
Fig. 8. Inference efficiency of BCGD algorithms
approaches in terms of inference time and mem-
ory consumption. We then subsequently compare
BCGD approaches with two auto-dimensionality la-
tent space approaches BIGCLAIM and LabelRT and
graph heuristic AA in terms of online link prediction
time and memory consumption.
Offline inference efficiency. We first report the total
running time and memory consumption of our BCGD
approaches in Figure 8. Clearly, BCGDG is neither
memory efficient nor time efficient. Both BCGDI and
BCGDL are very time and memory efficient: BCGDI is
more efficient in running time, while BCGDL is more
efficient in memory usage. This is because BCGDI
requires additional storage for graph changes ∆G;
but it adaptively updates latent space with the graph
changes and thus is very time efficient.
10
(cid:29)t PTM (cid:29)t {NMFR, BCGDG} (cid:29)t DMMSB, where
A (cid:29)t B denotes that A is faster than B on average. On
HepPh data, {NMFR, BCGDG} (cid:29)t {LabelRT, PTM} (cid:29)t
DMMSB. Clearly, our global optimization algorithm
BCGDG is at least five times faster than the other global
approach, DMMSB, and is comparable to inefficient
local approach NMFR. For small graphs, it is not
surprising that the most efficient algorithm is the
incremental approach LabelRT since it incrementally
maintains and updates latent spaces. Unfortunately,
it consumes large amounts of memory, and becomes
much slower than our approach BCGDG for graph
HepPh. In addition, all of these approaches are much
slower than our fast BCGD algorithms BCGDL and
BCGDI, each of which takes less than 400 seconds to
finish learning temporal latent space for HepPh.
(a) Inferring time (s)
(b) Memory usage (MB)
Fig. 10.
baselines Hott and BIGCLAM.
Inference efficiency comparison with scalable
We further compare our fast BCGD algorithms with
state-of-the-arts scalable approaches Hott and Big-
CLAM. BCGDL is comparable to Hott and BigCLAM.
In addition, the incremental approach BCGDI is consis-
tently more efficient than alternative approaches Hott
and BigCLAM. Because we noticed from Figure 10(b)
that the memory usage of these four approaches is
similar, we conclude that our two fast BCGD algo-
rithms are both memory- and time-efficient.
(a) Inferring time (s)
(b) Memory usage (MB)
Fig. 9. Inference efficiency comparison with DMMSB, PTM,
LabelRT and NMFR.
In the following we group other baselines into two
classes to facilitate results visualization. We first com-
pare those memory- or time-inefficient approaches
DMMSB, PTM, LabelRT and NMFR with our BCGDG
and report the results in Figure 9. Among them, PTM
and LabelRT require much larger memory consump-
tion and thus fail to process the two large graph DBLP
and YouTube due to their memory bottlenecks. NMFR
is more memory efficient than PTM and LabelRT, but
it still can not handle the two large graphs. In ad-
dition, global optimization algorithm DMMSB takes
too long to process the two large graph DBLP and
Youtube due to running time bottleneck. Figure 9(a)
shows that on Infection and Facebook data, LabelRT
(a) Prediction time (s)
(b) Memory usage (MB)
Fig. 11. Online prediction time and memory comparison for
20 thousand node pairs. Note that BCGD represents all of
the three proposed algorithms (BCGDG, BCGDL, and BCGDI)
since the online prediction time and memory consumption of
these three are the same.
Online Prediction Efficiency. We now verify that the
low-dimension latent space approach is very efficient
for online link prediction. Here we choose AA as
BCGDGBCGDLBCGDI10-1 1 10102103104InfectionFacebookHepPhDBLPYoutubeInferring time (s) 1 10102103InfectionFacebookHepPhDBLPYoutubeMemory usage (MB)BCGDGDMMSBNMFRLabelRTPTM10-1100101102103104InfectionFacebookHepPhInferring time (s) 1 10102103104InfectionFacebookHepPhMemory usage (MB)BCGDLBCGDIHottBIGCLAM10-1 1 10102103104InfectionFacebookHepPhDBLPYoutubeInferring time (s) 1 10102103InfectionFacebookHepPhDBLPYoutubeMemory usage (MB)BCGDBIGCLAMLabelRTAA10-1 1 10102103104InfectionFacebookHepPhDBLPYoutubePrediction time (s) 1 10102103InfectionFacebookHepPhDBLPYoutubeMemory usage (MB) 11
Fig. 12. All links prediction accuracy comparison.
Fig. 13. New links prediction accuracy comparison.
a representative graph heuristic approach for link
predication, not only because it achieves good accu-
racy [31], but also because it is fast to compute com-
pared to other graph heuristics [7]. However, AA is
still much slower than latent space-based approaches
in online link probability computation. As shown in
Figure 11, on the HepPh dataset, for 20,000 node pairs,
it takes more than 150 seconds to compute the AA
scores; while all the low-dimension latent space-based
approaches are able to finish online link probability
computation in less than one second. In addition,
latent space-based link prediction approaches are also
more efficient than AA in terms of memory con-
sumption, see Figure 11(b). To predict whether two
nodes are linked, latent space-based approaches only
need to read the latent positions of two nodes; while
AA requires the neighborhood information of two
nodes. Finally, Figure 11(a) also intuitively supports
the reason why we use low-dimension constraint.
With low-dimension constraint (k=20), our BCGD ap-
proach obtains at least four times speed-up than auto-
dimensionality approaches LabelRT and BigCLAM
(sometimes k can be more than two hundred) in
online link prediction.
5.4 Evaluation of Link Prediction Accuracy
In this section we quantitatively evaluate the quality
of learned latent spaces in terms of their predictive
power. We give an overall comparison of all the ap-
proaches in terms of AU CROC on predicting all links
and report the results in Figure 12. Here, Gpre denotes
the simple baseline using Gt to predict Gt+1. Clearly,
link prediction accuracy, we have {BCGDG,
for all
BCGDL, BCGDI, PTM, DMMSB} (cid:29)a {Hott, NMFR}
(cid:29)a {BIGCLAM, LabelRT, AA, Gpre}, where A (cid:29)a B
denotes that on average, A is more accurate than B
in terms of predictive power. BigCLAM and LabelRT
have poor AUC scores because they only output a
single, hard assignment of nodes to each dimension,
which provides no means to control the trade-off
between true and false positive rates. AA did not
perform well on Facebook and YouTube because it
can not capture the deleted links prediction; AA still
gives each unlinked node pairs a high score based
on their topological similarity on aggregated graph
from G1 to Gt. Our methods perform much better
than Gpre due to: 1) Zt encodes not only the link
information of Gt, but also the temporal information
from G1 to Gt; 2) Zt is a good estimate for Zt+1; 3) the
reconstruction error between ZtZ T
t and Gt reflects the
possible future changes. Given any pair (u, v), if the
reconstruction error between Zt(u)Zt(v)T and Gt(u, v)
is large, then it indicates that the proximity of the pair
(u, v) is changing over time.
We next evaluate the link prediction performance
of all the approaches on new links. The results are
plotted in Figure 13. As expected, the AUC values
are much lower for the new links. For the new links
prediction, we have {BCGDG, BCGDL, BCGDI, LabelRT,
DMMSB} (cid:29)a {Hott, NMFR} (cid:29)a {AA, BIGCLAM}. It
is not surprising that all the temporal link prediction
approaches perform better than static link prediction
approaches.
Figure 13 also reflects an interesting phenomenon:
our fast BCGD algorithms BCGDL and BCGDI not only
improve the efficiency bottleneck of latent space infer-
ence, but also achieve comparable prediction accuracy
with the global optimization approach BCGDG. For the
new links prediction, they even outperform global
approaches on both dataset Infection and YouTube.
This is because both BCGDL and BCGDI utilize local
structures and local invariance, which significantly
enhances local and temporal information encoding
in the temporal latent space. In addition, we notice
that on average, the incremental approach BCGDI is
able to obtain higher accuracy than the local approach
BCGDL. Although both BCGDL and BCGDI utilize local
invariance, BCGDI is able to exploit the local structure
and local invariance in an adaptive way with graph
changes, which leads to a better performance. Because
0.2 0.4 0.6 0.8 1InfectionFacebookHepPhAUCROCBCGDGBCGDLBCGDIPTMDMMSBNMFRHottAABIGCLAMLabelRTGpre 0.2 0.4 0.6 0.8 1DBLPYoutubeAUCROCBCGDGBCGDLBCGDIHottGpreAABIGCLAM 0.2 0.4 0.6 0.8 1InfectionFacebookHepPhAUCROCBCGDGBCGDLBCGDIDMMSBLabelRTHottNMFRPTMAABIGCLAM 0.2 0.4 0.6 0.8 1DBLPYoutubeAUCROCBCGDGBCGDLBCGDIHottAABIGCLAM BCGDI is more scalable than all the other alternative
approaches (as shown in Figure 10(a)). We conclude
that BCGDI is a practical solution for link prediction
on large-scale dynamic networks.
6 RELATED WORK
Recently, link prediction has attracted significant at-
tention, and various methods have been proposed for
different types of networks. In the following, we first
give a brief overview of related work in link predic-
tion. Particularly, we focus on two categories: graph-
based heuristics and latent space-based approaches.
Next, we present some related work in inferring latent
positions of nodes in a latent space.
6.1 Link Prediction
The link prediction problem in static networks has
been extensively studied. We refer readers to two
recent survey papers [24], [15] and the related work
section of the paper [49] for an exhaustive intro-
duction to this thriving field. Among these existing
works, one of the most popular categories of methods
is graph-based heuristics. Graph-based heuristics [24],
[2], [35], [20], [19] model the link probability be-
tween two nodes as a function of their topological
similarity. The topological similarity can be defined
based on local proximity such as common neigh-
bors [2], triad closures [35], or global proximity such
as weighted shortest-paths [20]. Here we highlight
Adamic and Adar (AA) [2] who proposed a degree-
weighted common-neighbor heuristic that works well
in collaboration networks.
Unfortunately, in online link prediction, the compu-
tation of topological similarity metrics, especially for
those metrics that require path computation or ran-
dom walk over the entire network, is time-consuming
if computed on-the-fly [7]. On the other hand, one can
precompute all pairs of topological similarity scores in
advance, leading to a quadratic storage cost. To strike
a balance between online computation cost and offline
storage cost, recent research interests in link predic-
tion have been directed towards temporal link pre-
diction with latent space modeling [16]. For example,
Fu et al. [10] extended the mixed membership block
model to allow a linear Gaussian trend in the model
parameters (DMMSB). Sewell and Chen [33] proposed
a model which embeds longitudinal network data as
trajectories in a latent Euclidean space, and the prob-
ability of link formation between two nodes depends
on their Euclidean distance and popularity. In this
work we do not take the popularity of each node into
consideration since we assume that the popularity
of each node is automatically captured by the latent
space modeling: a popular node is surrounded by
many other nodes in the latent space. Furthermore,
our model penalizes sudden and large changes in the
latent positions.
12
Dunlavy et al. [8] developed a tensor-based latent
space modeling to predict temporal links. Unfortu-
nately, in online prediction, it requires the tensor prod-
uct of vectors, which is much more computationally
expensive than the dot product of vectors in the pro-
posed approach. In addition, their approaches require
very large storage costs since they need to put the
entire tensor into memory. Recently, there have been
several more efficient approaches to conduct tensor
decomposition. For instance, Huang et al. [17] pro-
posed to conduct tensor decomposition with singular
value decomposition to find temporal communities in
a GPGPU setting.
6.2 Inferring Latent Space
Recent work has explored static latent space model-
ing for social network analysis. For example, Hoff
et al. [16] developed a class of models where the
probability of a relation between actors depends on
the positions of individuals in an unobserved social
space. They make inference for the latent space with
the maximum likelihood estimation and Bayesian pos-
terior inference. Obtaining this posterior is, however,
often infeasible in large-scale graphs. Yin et al. [45]
proposed an efficient stochastic variational inference
algorithm and a parsimonious triangular model to
infer the latent spaces of large networks (PTM). Sarkar
and Moore [32] first generalized the classical mul-
tidimensional scaling to get an initial estimate of
the positions in the latent space, and then applied
nonlinear optimization directly to obtain the latent
space representation.
Matrix factorization approaches are also applied
to observed networks to learn the low-dimension
latent space representation. Yang et al. [43] propose
a multiplicative algorithm with graph random walk
to solve the symmetric graph factorization problem
(NMFR). However,
their approach does not scale
well due to the high computation cost in each it-
eration. Additionally, the effectiveness of their ap-
proach decreases as the density of the input graph
increases. Yang and Leskovec [42] proposed a matrix
factorization approach over networks to learn the
static latent space representation by maximizing the
link likelihood (BIGCLAM). Their approach is very
scalable; nevertheless, the learned latent space is of
high dimension, which leads to expensive online link
prediction computation. HottTopixx (Hott [30]) uses
a new approach for NMF with low-rank constraint
which is highly parallel and allows it to run on
very large datasets. Unfortunately, their approach is
designed to factorize the traditional rectangle ma-
trices, and scalable approaches in symmetric graph
factorization are much less studied than rectangle
matrix factorization such as Hott. In this work we
apply symmetric matrix factorization approaches di-
rectly on the observed networks to infer low-rank
latent spaces. We propose a block coordinate gradient
descent algorithm, which is more scalable than NMFR
and performs well, regardless of the topology of input
graphs.
In addition to matrix factorization approaches, re-
cently Xie et al. [41] (LabelRT) proposed a label
propagation-based approach that incrementally de-
tects communities in dynamic networks. In this work
we propose an incremental BCGD algorithm to incre-
mentally detect latent positions at each time step with
conditional updates on a set of affected nodes.
7 CONCLUSION AND FUTURE WORK
In this work we propose a scalable approach for link
prediction with a temporal latent space model, which
assumes two nodes are more likely to be linked if
they are close to each other in their latent space.
In addition, our temporal latent space model prefers
smoothly evolving by penalizing frequent changes in
latent positions for each node. With respect to the
proposed model, we developed three different infer-
ence algorithms, BCGDG, BCGDL and BCGDI to learn
the temporal
latent space via non-negative matrix
factorization approaches. We also provide a set of
theoretical analyses characterizing their performance
guarantees. We conducted a set of experiments on
large networks with millions of nodes and links to
verify both the efficiency and predictive quality of all
the proposed approaches.
Our model still has limitations. First, the temporal
smoothness assumption may not hold in some cir-
cumstance. For example, in a road network, the latent
position of each node changes significantly from rush
hour to a non-rush hour. And even in social networks,
where temporal smoothness assumptions typically
hold, external events may cause significant shifts to
the network that reflect abrupt changes in latent node
positions. We plan to extend the proposed model to
support temporal nonlinear transitions. Second, we
plan to propose a new continuous time model that
supports continuous inputs rather than discretized
graph snapshots. Moreover, we plan to extend our
approach to generalized cases including directed and
weighted graphs. In addition to model improvement,
another interesting direction is to further improve the
efficiency. Note that in our block coordinate gradient
descent approach, the latent position update for a
node u can be conducted simultaneously with another
node v if they do not share the same neighborhood.
This property can be leveraged for very good paral-
lelization in the future.
ACKNOWLEDGMENT
We are very grateful to Prof. Dacheng Tao, Prof. Kun
Zhang, Prof. Rong Ge, Li Han and Dingxiong Deng
for their insightful discussions. This research was sup-
ported in part by DARPA grant No. W911NF1210034.
REFERENCES
13
[1]
network
2015.
"Dblp
[Online]. Available: http://konect.uni-koblenz.de/networks/
dblp coauthor
– KONECT," May
dataset
[2] L. A. Adamic and E. Adar, "Friends and neighbors on the
web," SOCIAL NETWORKS, vol. 25, pp. 211–230, 2001.
[4]
[3] E. M. Airoldi, D. M. Blei, S. E. Fienberg, and E. P. Xing,
"Mixed membership stochastic blockmodels," J. Mach. Learn.
Res., vol. 9, pp. 1981–2014, 2008.
S. Arora, R. Ge, R. Kannan, and A. Moitra, "Computing a non-
negative matrix factorization – provably," in STOC Conference,
2012, pp. 145–162.
[5] 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.
[6] M. W. Berry, M. Browne, A. N. Langville, V. P. Pauca, and
R. J. Plemmons, "Algorithms and applications for approximate
nonnegative matrix factorization," Comp. Stat. & Data Analysis,
vol. 52, no. 1, pp. 155–173, 2007.
S. Cohen and N. Cohen-Tzemach, "Implementing link-
prediction for social networks in a database system," in SIG-
MOD DBSocial Workshop, 2013, pp. 37–42.
[8] D. M. Dunlavy, T. G. Kolda, and E. Acar, "Temporal link
prediction using matrix and tensor factorizations," ACM Trans.
Knowl. Discov. Data, vol. 5, no. 2, pp. 10:1–10:27, 2011.
[9] D. Erdos, R. Gemulla, and E. Terzi, "Reconstructing graphs
from neighborhood data," ACM Trans. Knowl. Discov. Data,
vol. 8, no. 4, pp. 23:1–23:22, 2014.
[7]
[10] W. Fu, L. Song, and E. P. Xing, "Dynamic mixed membership
blockmodel for evolving networks," in ICML Conference, 2009,
pp. 329–336.
[11] S. Gao, L. Denoyer, and P. Gallinari, "Temporal link prediction
by integrating content and structure information," in CIKM
Conference, 2011, pp. 1169–1174.
[12] D. Gibson, R. Kumar, and A. Tomkins, "Discovering large
dense subgraphs in massive graphs," in VLDB Conference,
2005, pp. 721–732.
[13] N. Guan, D. Tao, Z. Luo, and B. Yuan, "Nenmf: An optimal
gradient method for nonnegative matrix factorization." IEEE
Trans. on Signal Processing, pp. 2882–2898, 2012.
[14] P. Gupta, V. Satuluri, A. Grewal, S. Gurumurthy, V. Zhabiuk,
Q. Li, and J. Lin, "Real-time twitter recommendation: Online
motif detection in large dynamic graphs," Proc. VLDB Endow.,
vol. 7, no. 13, pp. 1379–1380, 2014.
[15] M. A. Hasan and M. J. Zaki, "A survey of link prediction in
Springer
social networks," in Social Network Data Analytics.
US, 2011, pp. 243–275.
[16] P. D. Hoff, A. E. Raftery, and M. S. Handcock, "Latent space
approaches to social network analysis," Journal of the American
Statistical Association, vol. 97, 2002.
[17] F. Huang, N. U. N, M. U. Hakeem, P. Verma, and A. Anand-
kumar, "Fast detection of overlapping communities via online
tensor methods on gpus." CoRR, vol. abs/1309.0787, 2013.
[18] L. Isella, J. Stehl´e, A. Barrat, C. Cattuto, J.-F. Pinton, and
W. V. den Broeck, "What's in a crowd? analysis of face-to-face
behavioral networks," J. of Theoretical Biology, vol. 271, no. 1,
pp. 166–180, 2011.
[19] G. Jeh and J. Widom, "Simrank: a measure of structural-
context similarity," in SIGKDD Conference, 2002, pp. 538–543.
[20] L. Katz, "A new status index derived from sociometric analy-
sis," Psychometrika, vol. 18, no. 1, pp. 39–43, March 1953.
[21] J. Kunegis, "KONECT - The Koblenz Network Collection," in
Proc. Int. Web Observatory Workshop, 2013.
[22] D. D. Lee and H. S. Seung, "Learning the parts of objects by
nonnegative matrix factorization," Nature, vol. 401, no. 6755,
pp. 788–791, 1999.
[23] J. Leskovec, J. Kleinberg, and C. Faloutsos, "Graph evolution:
Densification and shrinking diameters," ACM Trans. Knowledge
Discovery from Data, vol. 1, no. 1, pp. 1–40, 2007.
[24] D. Liben-Nowell and J. Kleinberg, "The link prediction prob-
lem for social networks," in CIKM Conference, 2003, pp. 556–
559.
[25] M. Mcpherson, L. Smith-Lovin, and J. M. Cook, "Birds of
a feather: Homophily in social networks," Annual Review of
Sociology, vol. 27, pp. 415–444, 2001.
[26] A. K. Menon and C. Elkan, "Link prediction via matrix fac-
torization," in ECML/PKDD Conference, 2011, pp. 437–452.
[27] A. Mislove, "Online social networks: Measurement, analysis,
and applications to distributed information systems," Ph.D.
dissertation, Department of Computer Science, Rice Univer-
sity, 2009.
[28] Y. Nesterov, Introductory lectures on convex optimization : a basic
course. Kluwer Academic Publ., 2004.
[29] G.-J. Qi, C. C. Aggarwal, and T. Huang, "Link prediction
across networks by biased cross-network sampling," in ICDE
Conference, 2013, pp. 793–804.
[30] B. Recht, C. Re, J. A. Tropp, and V. Bittorf, "Factoring non-
negative matrices with linear programs," in NIPS Conference,
2012, pp. 1223–1231.
[31] P. Sarkar, D. Chakrabarti, and M. I. Jordan, "Nonparametric
link prediction in dynamic networks," in ICML Conference,
2012, pp. 1687–1694.
[32] P. Sarkar and A. W. Moore, "Dynamic social network analysis
using latent space models," SIGKDD Explor. Newsl., vol. 7,
no. 2, pp. 31–40, 2005.
[33] D. K. Sewell and Y. Chen, "Latent space models for dynamic
networks," Journal of the American Statistical Association, vol.
110, no. 512, pp. 1646–1657, 2015.
[34] J. Sun, C. Faloutsos, S. Papadimitriou, and P. S. Yu, "Graph-
scope: Parameter-free mining of large time-evolving graphs,"
in SIGKDD Conference, 2007, pp. 687–696.
[35] P. Symeonidis, E. Tiakas, and Y. Manolopoulos, "Transitive
node similarity for link prediction in social networks with
positive and negative links," in Proceedings of the 4th ACM
conference on Recommender systems, 2010, pp. 183–190.
[36] C. Tantipathananandh, T. Berger-Wolf, and D. Kempe, "A
framework for community identification in dynamic social
networks," in SIGKDD Conference, 2007, pp. 717–726.
[37] C. Tantipathananandh and T. Y. Berger-Wolf, "Finding com-
munities in dynamic social networks," in ICDM Conference,
2011, pp. 1236–1241.
[38] P. Tseng and S. Yun, "A coordinate gradient descent method
for nonsmooth separable minimization," Math. Program., vol.
117, no. 1-2, pp. 387–423, 2009.
[39] S. A. Vavasis, "On the complexity of nonnegative matrix
factorization," J. on Optimization, vol. 20, no. 3, pp. 1364–1377,
2009.
[40] B. Viswanath, A. Mislove, M. Cha, and K. P. Gummadi,
"On the evolution of user interaction in Facebook," in Proc.
Workshop on Online Social Networks, 2009, pp. 37–42.
[41] J. Xie, M. Chen, and B. K. Szymanski, "Labelrankt: Incremental
community detection in dynamic networks via label propaga-
tion," CoRR, vol. abs/1305.2006, 2013.
[42] J. Yang and J. Leskovec, "Overlapping community detection
at scale: A nonnegative matrix factorization approach," in
WSDM Conference, 2013, pp. 587–596.
[43] Z. Yang, T. Hao, O. Dikmen, X. Chen, and E. Oja, "Cluster-
ing by nonnegative matrix factorization using graph random
walk," in NIPS Conference, 2012, pp. 1088–1096.
[44] J. Ye, H. Cheng, Z. Zhu, and M. Chen, "Predicting positive and
negative links in signed social networks by transfer learning,"
in WWW Conference, 2013, pp. 1477–1488.
[45] J. Yin, Q. Ho, and E. P. Xing, "A scalable approach to proba-
bilistic latent space inference of large-scale networks," in NIPS
Conference, 2013, pp. 422–430.
[46] J. Zhang, C. Wang, J. Wang, and J. X. Yu, "Inferring continuous
dynamic social influence and personal preference for temporal
behavior prediction," Proc. VLDB Endow., vol. 8, no. 3, pp. 269–
280, 2014.
[47] Y. Zhang, M. Zhang, Y. Liu, S. Ma, and S. Feng, "Localized
matrix factorization for recommendation based on matrix
block diagonal forms," in WWW Conference, 2013, pp. 1511–
1520.
[48] L. Zhu, A. Galstyan, J. Cheng, and K. Lerman, "Tripartite
graph clustering for dynamic sentiment analysis on social
media," in SIGMOD Conference, 2014, pp. 1531–1542.
[49] L. Zhu and K. Lerman, "A visibility-based model for link
prediction in social media," in ASE SocialCom Conference, 2014.
14
Linhong Zhu is a computer scientist at
USC's Information Sciences Institute. Prior to
that she worked as a Scientist-I in the data
analytics department at the Institute for In-
focomm Research, Singapore. She obtained
her Ph.D. degree in computer engineering
from Nanyang Technological University, Sin-
gapore in 2011 and the BS degree from
the University of Science and Technology of
China in 2006. Her research interests are in
large-scale graph analytics with applications
to social network analysis, social media analysis, and predictive
modeling. She is a member of IEEE and ACM.
Dong Guo received the BS degree in
physics from the University of Science and
Technology of China, Hefei, China, in 2007,
and entered the Ph.D. program in computer
science at the University of Southern Califor-
nia in 2011. His doctoral projects consisted
of studying the data-driven representation
learning algorithms and applying machine
pipeline in industrial data analysis.
Junming Yin is an assistant professor in
the Department of Management Information
Systems at the University of Arizona. Prior to
that he was a Lane Fellow in the Lane Center
of Computational Biology at Carnegie Mellon
University. He received his Ph.D. from UC
Berkeley in Computer Science. His research
focus is on statistical machine learning and
its applications in business intelligence, digi-
tal marketing, healthcare systems, and com-
putational biology.
Greg Ver Steeg is a research assistant pro-
fessor in computer science at USC's Infor-
mation Sciences Institute. He received his
Ph.D. in physics from Caltech in 2009 and
since then has focused on using ideas from
information theory to understand complex
systems like human behavior, biology, and
language. His work has been recognized with
an AFOSR Young Investigator Award.
Aram Galstyan is a research director for
data science and machine learning at USC's
Information Sciences Institute and a re-
search associate professor
in the USC
Viterbi School of Engineering's Computer
Science Department. His current research
focuses on different aspects of machine
learning, with applications in computational
social sciences, forecasting, and bioinformat-
ics.
Supplemental Materials to "Scalable Temporal Latent Space
Inference for Link Prediction in Dynamic Social Networks"
A PROOF OF LEMMA 1
According to Eq. 3, we can obtain the gradient of
J(Zτ (u)) at time stamp τ
∇J(Zτ (u)) = − λ(Zτ−1(u) + Zτ +1(u))
(cid:88)
Zτ (v)T Zτ (v)
v(cid:54)∈N (u)
(Gτ (u, v) − Zτ (u)Zτ (v)T )Zτ (v)
− 2
v∈N (u)
+ 2Zτ (u)
(cid:88)
We rewrite the term(cid:80)
(cid:88)
Zτ (v)T Zτ (v) − (cid:88)
τ Zτ − (cid:88)
v∈V
=Z T
=
v∈N (u)
v∈N (u)
v(cid:54)∈N (u) Zτ (v)T Zτ (v) as:
Zτ (v)T Zτ (v) − Zτ (u)T Zτ (u)
Zτ (v)T Zτ (v) − Zτ (u)T Zτ (u)
(9)
(10)
By combining Eq. 9 and Eq. 10, we have:
∇J(Zτ (u)) = − λ(Zτ−1(u) + Zτ +1(u)) + 2Zτ (u)Z T
τ Zτ
− 2Zτ (u)Zτ (u)T Zτ (u) − 2
Gτ (u, v)Zτ (v)
(cid:88)
v∈N (u)
(11)
Since for each u ∈ V , Zτ (u)Zτ (u)T = 1, the gradient
of J(Zτ (u)) can be simplified as
∇J(Zτ (u)) = − λ(Zτ−1(u) + Zτ +1(u)) + 2Zτ (u)Z T
τ Zτ
− 2Zτ (u) − 2
Gτ (u, v)Zτ (v)
(cid:88)
v∈N (u)
(12)
For any two matrices Z1, Z2 ∈ R1×k, we replace
Zτ (u) in Eq. 12 with Z1 and Z2, we then have
2
(cid:107)∇J(Z1) − ∇J(Z2)(cid:107)2
=(cid:107)2(Z1 − Z2)Z T
=(cid:107)2(Z1 − Z2)(Z T
=4tr(((Z1 − Z2)(Z T
=4tr((Z1 − Z2)T (Z1 − Z2)(Z T
τ Zτ − (Z1 − Z2)(cid:107)2
τ Zτ − I)(cid:107)2
2,
2
τ Zτ − I))T (Z1 − Z2)(Z T
τ Zτ − I)(Z T
τ Zτ − I)),
τ Zτ − I)T ),
Since for any two positive semi-definite matrices A
and B, we have tr(AB) ≤ tr(A)tr(B). By substitut-
ing A = (Z1− Z2)T (Z1− Z2), B = (Z T
τ Zτ −
I)T , we have
τ Zτ − I)(Z T
2
(cid:107)∇J(Z1) − ∇J(Z2)(cid:107)2
=4tr((Z1 − Z2)T (Z1 − Z2)(Z T
≤4tr((Z T
=4tr((Z T
τ Zτ − I)(Z T
τ Zτ − I)(Z T
τ Zτ − I)(Z T
τ Zτ − I))T ),
τ Zτ − I)T )tr((Z1 − Z2)T (Z1 − Z2)),
τ Zτ − I)T )(cid:107)Z1 − Z2(cid:107)2
2
Let us apply the above trace inequality again and
using tr(Z T
τ Zτ ) = n, tr(I) = k, we then have
Thus, we
2(cid:112)(n2 − 2n + k)(cid:107)Z1 − Z2(cid:107)2. Therefore, ∇J(Zτ (u)) is
2(cid:112)(n2 − 2n + k). This completes the proof.
Lipschitz continuous and the Lipschitz constant is
(cid:107)∇J(Z1) − ∇J(Z2)(cid:107)2
have
≤
B PROOF OF LEMMA 2
Based on [28], [13], [5], for each node u, at time point
τ, we could iteratively update Zτ (u) in each iteration
r + 1 with the following rule:
Y (r+1) = Y (r) − η∇Zτ (u)J(Y (r))
Z (r+1)
(u) = PL(Y ) = max(Y (r+1), 0)
τ
Y (r+1) = Z (r+1)
τ
(u) +
(Z (r+1)
τ
(u) − Z (r)
τ (u))
ar − 1
ar+1
5, PL(Y )
=
non-negative
is
τ (u) (Zτ (u)I
ar
is defined in Eq.
where
the
arg minX≥0 φ(Y, X)
projection of Y , and Y (0) = Z I
the initialization for Zτ (u)).
denotes
Let us replace the ∇Zτ (u)J(Z (r)
the step size η with 1
which leads to the expression in Lemma 2.
τ (u)) with Eq. 12 and
L (L is the Lipschitz constant),
Therefore, the update rule in Lemma 2 is correct.
C PROOF OF THEOREM 1
(cid:107)Y − PL(Y )(cid:107)2
F
L
2
According to Theorem 2.2.2 in [28] or Lemma 2.3 [5],
for any matrix X ∈ R1×k
and matrix Y ∈ R1×k, we
have
J(X) ≥ J(PL(Y )) + L < PL(Y ) − Y, Y − X > +
+
Now let Y (r+1) = Z (r)(u) + ar−1
(13)
where PL(Y ) = arg minX≥0 φ(Y, X) denotes the non-
negative projection of Y , and (cid:104),(cid:105) denotes the matrix
inner product operator.
(Z (r)(u) −
Z (r−1)(u)) and PL(Y (r)) = Z (r)(u). Note that we omit-
ted the subscript τ that denotes the time to simplify
notations. By substituting X = Z (r)(u), Y = Y (r+1),
we have
J(Z (r)(u)) ≥J(Z (r+1)(u)) +
(cid:107)Y (r+1) − Z (r+1)(u)(cid:107)2
ar+1
F
+ L(cid:104)Z (r+1)(u) − Y (r+1)(u), Y (r+1) − Z (r)(u)(cid:105)
(14)
Similarly, by substituting X = Z∗(u), Y = Y (r+1), we
have
L
2
(cid:107)∇J(Z1) − ∇J(Z2)(cid:107)2
≤4tr((Z T
τ Zτ − I)(Z T
≤4(n2 − 2n + k)(cid:107)Z1 − Z2(cid:107)2
2
2
τ Zτ − I)T )(cid:107)Z1 − Z2(cid:107)2
2
∗
J(Z
(u)) ≥J(Z (r+1)(u)) +
(cid:107)Y (r+1) − Z (r+1)(u)(cid:107)2
F
+ L(cid:104)Z (r+1)(u) − Y (r+1)(u), Y (r+1) − Z
∗
L
2
(u)(cid:105)
(15)
(u)))
2
since ar+1 > 1, by multiplying both sides of Eq. 14 by
ar+1 − 1 and adding it to Eq. 15, we have
(ar+1 − 1)J(Z (r)(u)) + J(Z
, ar+1Y (r+1) − (ar+1 − 1)Z (r)(u) − Z
+ ar+1J(Z (r+1)(u)) +
(u)) ≥ L(cid:104)Z (r+1)(u) − Y (r+1),
(cid:107)Y (r+1) − Z (r+1)(u)(cid:107)2
(u)(cid:105)
Lar+1
∗
∗
F
∗
∗
r = a2
(u))) − a2
r+1(J(Z (r+1)(u)) − J(Z
F . That is, (cid:107)A− B(cid:107)2
((cid:107)ar+1Z (r+1)(u) − ar+1Y (r+1)(cid:107)2
F +(cid:107)B−C(cid:107)2
F − (cid:107)B − C(cid:107)2
F +2(cid:104)A−B, B−C(cid:105) = (cid:107)A−B+B−C(cid:107)2
(16)
r+1 − ar+1.
From Line 3 in Algorithm 1, we get a2
By using this equality and multiplying both sides of
Eq. 16, we have
r(J(Z (r)(u)) − J(Z
∗
a2
F + 2(cid:104)ar+1Z (r+1)(u)
≥ L
2
− ar+1Y (r+1), ar+1Y (r+1) − (ar+1 − 1)Z (r)(u) − Z
(u)(cid:105))
(17)
Since for any matrices A, B and C, we have (cid:107)A −
B(cid:107)2
F =
F + 2(cid:104)A− B, B− C(cid:105) = (cid:107)A−
(cid:107)A− C(cid:107)2
C(cid:107)2
F . By replacing A with ar+1Z (r+1)(u),
B with ar+1Y (r+1), C with (ar+1 − 1)Z (r)(u) + Z∗(u),
we simplify Eq. 17 as
r(J(Z (r)(u)) − J(Z
a2
≥ L
2
− (cid:107)ar+1Y (r+1) − (ar+1 − 1)Z (r)(u) − Z
((cid:107)ar+1Z (r+1)(u) − (ar+1 − 1)Z (r)(u) − Z
(u)(cid:107)2
F )
((cid:107)ar+1Z (r+1)(u) − (ar+1 − 1)Z (r)(u) − Z
∗
(u)(cid:107)2
=
− (cid:107)arZ (r)(u) − (ar − 1)Z (r−1)(u) − Z
(18)
The last equality holds due to Y (r+1) = Z (r)(u) + ar−1
(Z (r)(u) − Z (r−1)(u)).
By varying r=0 to r − 1 and summing up all these
r+1(J(Z (r+1)(u)) − J(Z
(u)(cid:107)2
(u))) − a2
(u)(cid:107)2
L
2
ar+1
∗
∗
∗
F
F
∗
∗
F
(u)))
expressions in Eq. 18, we get
(u)) − a2
(u)(cid:107)2
∗
J(Z (0)(u)) − J(Z
∗
≥ − L
(cid:107)Z (0)(u) − Z
2
F
r(J(Z (r)(u)) − J(Z
∗
(u)))
(19)
On another hand, we substitute X and Y in Eq. 13 by
X = Z∗(u) and Y = Y (0), we have
J(Z (0)(u)) − J(Z
L
2
((cid:107)Y (0) − Z
(u)) ≤
∗
F − (cid:107)Z (0)(u) − Z
(u)(cid:107)2
∗
∗
(u)(cid:107)2
F )
(20)
By combining Eq. 19 and Eq. 20, we have
r(J(Z (r)(u)) − J(Z
a2
∗
(u))) ≤ L
2
(cid:107)Y (0) − Z
∗
(u)(cid:107)2
F
(21)
By substituting Y (0) as Z I (u) and using ar ≥ (1 +
r/2(cid:112)α0/L) = (1+r/2) in Lemma 2.2.4 in [28], we get:
J(Z (r)(u)) − J(Z∗(u)) ≤ 2L(cid:107)Z I (u) − Z∗(u)(cid:107)2
F
(r + 2)2
This completes the proof for Theorem. 1.
Fig. 14. All links prediction accuracy comparison in terms
of AUC under PR curves.
D EXTRA RESULTS BASED ON AUCPR
In this set of experiments, we evaluate the link predic-
tion accuracy according to the measurement of AUCPR.
Figure 14 gives the overall comparison of different
approaches in terms of AUC under PR curves. The
results based on AUCPR are similar with those based
on AUCROC: the proposed approaches are consistently
better than other state-of-the-art approaches. We also
observed that the performance of AA based on AUCPR
is improved, possibly because AA gives high score
for node pairs that have common neighbors and zero
scores otherwise, which leads to high precision and
relatively lower recall. To the contrary, the perfor-
mance of DMMSB based on AUCPR is getting worse.
All the results indicate that the proposed approach,
especially the incremental algorithm BCGDI, performs
well in terms of link prediction task.
0.2 0.4 0.6 0.8 1InfectionFacebookHepPhAUCPRBCGDGBCGDLBCGDIPTMDMMSBNMFRHottAABIGCLAMLabelRTGpre 0.2 0.4 0.6 0.8 1DBLPYoutubeAUCPRBCGDGBCGDLBCGDIHottGpreAABIGCLAM |
1804.01576 | 1 | 1804 | 2018-03-13T22:32:16 | A Bayesian Model for False Information Belief Impact, Optimal Design, and Fake News Containment | [
"cs.SI"
] | This work is a technical approach to modeling false information nature, design, belief impact and containment in multi-agent networks. We present a Bayesian mathematical model for source information and viewer's belief, and how the former impacts the latter in a media (network) of broadcasters and viewers. Given the proposed model, we study how a particular information (true or false) can be optimally designed into a report, so that on average it conveys the most amount of the original intended information to the viewers of the network. Consequently, the model allows us to study susceptibility of a particular group of viewers to false information, as a function of statistical metrics of the their prior beliefs (e.g. bias, hesitation, open-mindedness, credibility assessment etc.). In addition, based on the same model we can study false information "containment" strategies imposed by network administrators. Specifically, we study a credibility assessment strategy, where every disseminated report must be within a certain distance of the truth. We study the trade-off between false and true information-belief convergence using this scheme which leads to ways for optimally deciding how truth sensitive an information dissemination network should operate. | cs.SI | cs | A Bayesian Model for False Information Belief
Impact, Optimal Design, and Fake News
Containment
Amin Khajehnejad
3Red Trading Group LLC., University of Michigan Ann Arbor
Email: [email protected], [email protected]
Shima Hajimirza
Texas A&M University
Email: [email protected]
8
1
0
2
r
a
M
3
1
]
I
S
.
s
c
[
1
v
6
7
5
1
0
.
4
0
8
1
:
v
i
X
r
a
Abstract-This work is a technical approach to modeling
false information nature, design, belief impact and containment
in multi-agent networks. We present a Bayesian mathematical
model for source information and viewer's belief, and how the
former impacts the latter in a media (network) of broadcast-
ers and viewers. Given the proposed model, we study how a
particular information (true or false) can be optimally designed
into a report, so that on average it conveys the most amount of
the original intended information to the viewers of the network.
Consequently, the model allows us to study susceptibility of a
particular group of viewers to false information, as a function of
statistical metrics of the their prior beliefs (e.g. bias, hesitation,
open-mindedness, credibility assessment etc.). In addition, based
on the same model we can study false information "containment"
strategies imposed by network administrators. Specifically, we
study a credibility assessment strategy, where every disseminated
report must be within a certain distance of the truth. We study the
trade-off between false and true information-belief convergence
using this scheme which leads to ways for optimally deciding
how truth sensitive an information dissemination network should
operate.
I. INTRODUCTION
Adversarial spread of false information based on exploita-
tion of viewer's naivety, gullibility and bias has been remarked
numerously over social media and news networks, most no-
tably in the midst of the most recent US presidential election
and Brexit [1]–[5]. The notion is familiarly referred to as
fake news. However, manipulative demonstration and spread of
false information is not limited to news, and indisputably exists
in other domains such as advertising [6]–[9], financial markets
(electronic trading) [10], [11], cyber-security [12], adversarial
machine learning [13], [14] and more, often referred to by
terms such as misinformation or disinformation.
When it comes to studying false information spread,
the
goals are multifaceted and significantly vague. This is due
to various reasons. Firstly, The nature of false information
is highly context-dependent and debatable. The notion of
fake news for instance is not fully resolved and well-defined
for many [15]–[17]. The practice of fake news fabrication
and broadcasting can take various forms, from an individual
creating and promoting fake social media posts, news websites
etc. [18], to manipulative journalism of well-established news
organizations doing so with the intention of appealing to a
certain group of audience [19]. False news creation and spread
false
how is
it
it
from the
is also not necessarily bound to malicious intention, and can
be a mere reflection of erroneous, exaggerative or neglectful
quoting of absorbed facts, rumors, panic epidemics and similar
phenomena. Furthermore, false information spread can affect
different systems in different ways, and consequently, admin-
istrators decide to handle it in various forms: identification,
detection, prediction, regulation, education or even ignorance
are common mechanism that are usually adopted to combat
false information [20]. Nevertheless, one needs to identify
and study the following aspects of false information in their
corresponding contexts:
1. Nature. What
information,
is
originated/designed, and what distinguishes
truth?
2. Impact. How does false information or fake news impact
beliefs of the users/viewers of a system/network, and how
does that in turn affect the dynamics of information flow and
system operation?
3. Detection. What methodologies can be used to identify
false information accurately and efficiently?
4. Spread (epidemic).
In an interactive network where
information is propagated in a distributed way by users and
outside feeds, what are the chances of epidemic false news
spread as a function of network parameters?
5. Containment and mitigation. What mechanisms can
prevent false information dissemination, or offset
it with
truth? At what costs can those strategies be implemented and,
what are the expected outcomes and damages?
Mathematical modeling of
false information is far more
challenging than semantic and social analysis, due to lack
of basic fundamentals. Yet
there are a handful of recent
works that have modeled various aspects of false information
nature and spread using technical methods. Most notably,
the majority of works have investigated fraudulent accounts
and credibility of news in online social networks such as
twitter [21]–[26]. A few other works have proposed models
for disseminating good information in social networks to
mitigate the effect of misinformation,
thereby improving
reliability [27]–[31]. Misinformation and manipulation in
financial limit order markets are probably studied far more
in-depth, especially using game-theoretic approaches (see
e.g. [32]–[35]), due to much simpler and readily available
numerical representations of information and beliefs. There
are recent works that have tried to adopt those approaches
in the context of social networks. For example, the study
of [30] offers a game-theoretic approach to truth and fake
news dissemination based on a mixed-strategy formulation of
volunteer's dilemma in an adhoc social network. However,
those studies are still preliminary and immature compared
to areas where numerical representation of information is
available.
The current work is a technical (yet different) approach to false
information design, impact and containment in multi-agent
networks. We present a Bayesian mathematical model for
information, viewer's belief, and how the former impacts the
latter in a media (network) of broadcasters and viewers. Given
the proposed model, we study how a particular information
(true or false) can be optimally designed into a report, so
that on average it conveys the most amount of the original
intended information to the viewers of the network. The model
then allows one to study susceptibility of a particular group
of viewers to false information, as a function of statistical
metrics of the their prior beliefs (e.g., bias, hesitation, open-
mindedness, etc.). In addition, based on the same model we
can study false information "containment" strategies imposed
by network administrators. Specifically, we study a credibility
assessment strategy, where every disseminated report must be
within a certain distance of the truth. We study the trade-
off between false and true information containment using this
scheme which leads to ways for optimally deciding how truth
sensitive an information dissemination network should operate.
A. Proposed Model
an
consider
We
information
network
(medium) with an information source, a transmitter (reporter),
a network administrator and multiple viewers (audience). The
schematics of such a network are illustrated in Figure 1.
dissemination
report is not permitted to be broadcast if it is too distant
from the truth xt. Therefore, the reporter's optimal design
must fall in an acceptable neighborhood of the truth which
we denote by C(xt). We assume that xs, xt and y are all real
vectors of length n. For the present, work we use the simple
euclidean distance as the measure of deviation between facts:
d(x1, x2) = (cid:107)x1 − x2(cid:107). Therefore, the authenticity filter has
the following mathematical form:
y ∈ C(xt) = {x (cid:107)x − xt(cid:107) (cid:54) }
(1)
where is a positive constant. There are N >> 0 viewers,
and every viewer i has a prior distribution pi(x) about the
truth. When a viewer receives a report y, he/she forms a
posterior distribution given the prior belief and the observation,
and consequently adopts a new belief in a Bayesian manner,
which we denote by ζi(y). The details of the viewers belief
impact
is described in the next subsection. We make the
assumption of the well-informed reporter. This means that the
reporter is fully aware of the truth xs, and also of the network
administration policy . However, the reporter is not aware of
every individual's prior distribution. Instead, he only knows
average statistics such as mean of first and seconds moments
of the viewers' priors, as will be explained the next section.
The reporter's goal is to convey the source's information to as
many viewers as possible. For a single viewer i, information
conveyance can be measured by the distance of the original
information xs and the adopted belief ζi(y):
τi,s(y) = (cid:107)xs − ζi(y)(cid:107)
(2)
For simplicity of future reference, we call this quantity the
"reporter-viewer information convergence", or in short "con-
vergence". The smaller τi,s(y), the more successful the re-
porter has been for impacting that particular viewer. However,
the statistical notion of information throughput is obtained by
averaging τi,s(y) over all viewers:
τs(y) = E(cid:107)xs − ζi(y)(cid:107)
(3)
The reporter's goal is therefore to minimize average conver-
gence τs(y) for a valid report y which is permitted to pass
through the network, i.e.
τs(y)
argmin
s.t. y ∈ C(xt)
y
(4)
Fig. 1: Schematics of an information dissemination network
with source, reporter, administrator and viewer.
The transmitter intends to disseminate in the network a piece
of information xs from the source. However, he can modify
the information content and design a report y, in order for
the viewers to accept as much of the source information as
possible. There is an authenticity filter in the network that
verifies every input report by comparing it with the truth. The
B. Bayesian Viewer
We model the viewer of the news (audience) as an agent i
that has a prior belief about the truth at a particular point
in time. We denote the probability distribution of this prior
with pi(·). Upon receiving a piece of report y, the viewer'
instantaneous belief of the truth (his point of view) is the
Maximum A Posteriori fact:
ζi(y) = argmax
x
pi(xy),
(5)
where the posterior distribution of the true news is obtained
from the Bayes rule :
pi(xy) ∝ pi(x)pi(yx)
(6)
The likelihood distribution pi(yx) is a function of
the
uncertainty and credibility of the source of news in the
viewer's mind. This is essentially an estimation of the statistics
of truth manipulation (in the mind of the viewer) that an
information source performs on the facts before disseminating
them. If in the viewer's mind, the source of the news is
very noisy, exaggerative or inaccurate, then the likelihood
distribution pi(yx) is very wide and has a large variance. If
there is a reporting bias to the journalism which is known to
the viewer, then the distribution is nonzero mean. In summary,
the viewer blends his prior opinion with what he/she observes,
but takes into account the presumed credibility of the source
(which could in all likelihood be false).
In a perfectly Gaussian world, we can assume that the prior
and likelihood are both n-dimensional Gaussian distributions:
pi(x) ∼ N (µi, Σi),
pi(xy) ∼ N (y, Σs,i),
(7)
where Σi and Σs,i are covariance matrices of the viewer's prior
and the conditional likelihood distribution, respectively. Intu-
itively, Σi is a multi-dimensional measure of viewer's intrinsic
hesitation about the state of truth at the time being, while Σs,i
measures suspicion of noisy/false reporting against the news
source. The well-informed reporter assumption means that
the reporter is not aware of individual viewer characteristics
µi, Σi, Σs,i, but has perfect knowledge of the joint distribution
of these quantities. Combining equations (5),(6),(7), we get:
s,i (x − y)
log pi(xy) ∝ −(x − µi)T Σ−1
and by maximizing this, we get that the viewers post report
belief becomes:
(x − µi) − (x − y)Σ−1
i
where
ζi(y) = Aiy + Biµi ,
Ai =(cid:0)Σ−1
Bi =(cid:0)Σ−1
(cid:1)−1
(cid:1)−1
i + Σ−1
i + Σ−1
s,i
s,i
Σ−1
s,i ,
Σ−1
i
(9)
(10)
Therefore, in the world of Gaussian news and beliefs, a viewer
adopts a linear combination of his prior belief and the viewed
report as his/her new belief.
C. Optimal Reporter
Combining equations (3),(4) and (9) together, the optimal
report y∗ is revealed to be the solution of the following
optimization:
E(cid:107)Aiy + Biµi − xs(cid:107),
y
min
s.t. (cid:107)y − xt(cid:107) (cid:54) .
(11)
(11) is a convex optimization. Therefore, using the method of
Lagrange multipliers (see e.g., [36]), y∗ is the minimizer of
the following utility function for some nonnegative λ (cid:62) 0:
L(y, λ) = E(cid:107)Aiy + Biµi − xs(cid:107)2 + λ(cid:107)y − xt(cid:107)2
Taking derivatives with respect to y, we get:
∂L(y, λ)
∂y
where:
= (A2 + λI)y + ABµ − Axs − λxt
A2 = EAT
i Ai, ABµ = EAT
i Biµ, A = EAi.
(12)
(13)
(14)
Finally, by setting (13) equal to zero, the optimal report is
obtained as:
y∗ = (A2 + λI)−1(cid:0)λxt + Axs − ABµ(cid:1) ,
(15)
where λ is the smallest nonnegative constant
that makes
(cid:107)y∗ − xt(cid:107) (cid:54) xt. The λ in (15) is a regularization factor that
forces the report to be closer to reality. Note that the well-
informed reporter assumption means that A2, A and ABµ are
computable for the reporter, hence so is the optimal report y∗.
D. Ergodic Covariances
For further simplicity of the analysis, we consider the
case of ergodic covariances, i.e., Σi = Σ and Σi,s = Σs
for all viewers, and only mean belief µi varies. Under this
assumption, it is easy to show that
y∗ = (A2 + λI)−1 (λxt + Axs − ABµ) .
(16)
When λ = 0, there is no authenticity requirement for the
report, thus the optimal report can be designed solely based on
the input source information and the viewers' belief statistics.
Specifically:
(8)
y∗ = A−1xs − A−1Bµ = xs + ΣsΣ−1 (xs − µ) .
(17)
In other words, the optimal report is the linear sum of the news
source and a skewed version of the average news surprise
(xs − µ), which can also be interpreted as exaggeration.
The optimal magnitude and rotation of such exaggeration is
determined by the relative ratio of the covariance matrices Σ
(hesitation) and Σs (source reliability). The more reliable a
source (smaller (cid:107)Σs(cid:107)∗) or the more open minded/less certain
the viewers (larger (cid:107)Σ(cid:107)∗), more exaggeration is allowed. (cid:107)X(cid:107)∗
denotes the nuclear norm of a matrix X which can be a
measure of the magnitude of a covariance matrix. However,
this also tells us that elements of the report (i.e. different
coordinates of y) can be exaggerated differently, according to
the certainty distribution. It should be noted that even in the
case of a perfectly truthful source xs = xt, the optimal report
is exaggerative and not completely honest. A credibility as-
sessment strategy or filtration policy should take that fact into
consideration. In other words, enforcing truthful reporting will
compensate the performance of an optimal reporter completely
based on the truth if (cid:107)ΣsΣ−1 (xt − µ)(cid:107) >> . The optimal
convergence metric τs(y∗) for λ = 0 in this case becomes:
τs(y∗) = E(µi − µ)T B2(µi − µ)
(18)
which is independent of the particular piece of news xs, and
only depends on the source credibility and viewers' belief
statistics.
E. Authentication Policy
Assuming that every reporter does the optimal action know-
ing the specified filter parameter , it is up to the network
administrator to set a value that promotes reporting based on
truthful sources and limits the unauthentic ones. This requires
definition of an appropriate truth dissemination utility function.
For a given source xs, policy parameter and optimal report
y∗, we denote the optimal convergence with c(xs, ), which
can be obtained from the derivations of the previous section
as:
c(xs, ) = E
(cid:107)Ay∗
+ Bµi − xs(cid:107)
i
(19)
The criteria for in filter design is intuitive: average conver-
gence for a source of false information must be significantly
smaller than that of the authentic source. In addition, should
not be not too small to limit the truthful optimal reporter.
These criteria can be represented by:
(cid:18)
(cid:12)(cid:12)(cid:12)(cid:12) (cid:107)xt − xs(cid:107) (cid:62) dmin
(cid:19)
U1() = E
xt,xs
c(xs, ) − c(xt, )
U2() =
c(xt,∞)
E
xt
E
c(xt, )
xt
(cid:62) α,
(cid:62) δ,
(20)
(21)
where δ and α are positive constants. The first equation
states that the average convergence distance for false news
information sources must be farther than that of the true
information source by a minimum margin δ. Note that the
expectation is over the distribution of all information sources,
and an information vector is considered false if it is at a
distance dmin from the truth. The second equation on the other
hand guarantees that the network is not too restrictive for the
optimal reporter of truth compared with a free network with no
authentication policy. Combining these metrics, we can define
a unified utility function, and choose the maximizer ∗:
∗ = argmax
U ()
(cid:18)
= E
xt,xs
c(xt,∞)
E
xt
E
c(xt, )
xt
+ β
c(xs, ) − c(xt, )
(cid:12)(cid:12)(cid:12)(cid:12) (cid:107)xt − xs(cid:107) (cid:62) dmin
(cid:19)
(22)
Fig. 2: Statistics of belief convergence of optimal report for
true information (blue) and random unauthentic information
(red).
II. NUMERICAL EVALUATIONS
We set up a basic numerical simulation to calculate and
illustrate the convergence metrics and utility functions of true
and unauthentic information sources based on the derivations
of the previous section. We consider a simple case of n =
2, with ergodic covariances Σ, Σs given below in (23). We
assume that the true information xt and source information xs
are both independently generated from a uniform distribution
of vectors on the unit 2-sphere, S2. The average viewers' belief
µ is also drawn from the same distribution.
(cid:20)1
0
(cid:21)
0
1
Σ =
, Σs =
(cid:20)0.5
0
(cid:21)
0
0.5
, µi ∼ N
µ, xt, xs ∼ S2,
(cid:18)
µ,
(cid:20)0.1
0
(cid:21)(cid:19)
0
0.1
(23)
t )−
In Figure 2, the average convergence for true source E(cid:107)ζ(y∗
xt(cid:107) is compared with the convergence of a random source
s) − xs(cid:107), as a function of network filter parameter .
E(cid:107)ζ(y∗
The standard deviations of the same quantities are displayed as
error areas. Since the mean belief is completely independent
of the source and truth, this figure is a representative of a
convergence trade-off in a network of indifferent audience.
Figure 3 shows convergence trade-off conditioned on the fact
that average belief is closer to the source (false) information
than the truth, i.e., (cid:107)¯µ− xs(cid:107) (cid:54) (cid:107)¯µ− xt(cid:107). We can think of this
scenario as "uneducated audience". Note that for larger , the
false source convergence error area is below the average con-
vergence of truth, meaning that the false information reporter
has a significant chance of conveying more information that
the truth.
III. CONCLUSION AND FUTURE WORK
A Bayesian model for information, viewer's belief and the
impact of the former on the latter was proposed in this work.
The model was studied from the point of an optimal reporter
intending to convey the message of a particular source to a
statistically large group of viewers, through a network that
implements a basic authentication policy. The simplifying
assumption of numerical information vectors and Gaussian
priors, posteriors and belief impact was made. The reporter
was also assumed to be well-informed, which means he has
the knowledge of truth, network policy and joint distribution
of viewers' belief statistics. Consequently, we derived the
formulation of the optimal reporter, and showed that it is a
linear combination of a portion of truth, source information
and exaggeration. In addition, knowing this optimal strategy,
we studied the optimal choice of the network authentication
policy parameter. If the policy is too restrictive, it hurdles
the performance of a well-intended reporter that only uses
exaggerative techniques to convey the truth to an uneducated
or biased audience. However, too loose of a policy grants a
large capacity for false information dissemination to false or
malicious sources. We proposed a computable utility function
that takes these factors into account, and can be optimized for
the best choice of authentication policy.
Given that this study is the first of its own kind, the proposed
model
is expected to be significantly more progressed in
the future work, in order to account for more practicalities.
First and foremost, our primary goal for a follow-up work
would be to validate the model with real-world data in various
contexts (social media, news websites and electronic trading
platforms). Secondly, we intend to incorporate less perfect
assumptions such partially informed reporters, and selective
or random authentication policies, and provide more analytical
and concrete derivations than those of the current work.
Another limitation of the current model is that it is very static.
In reality the source credibility parameters Σs,i is a function
of the long term reporting of a source, and varies through the
time. Furthermore, viewers have ways of cross-validating and
comparing information coming from different sources. Other
model parameters are also expected to be dynamic.
The present model is also passive, which means the viewers
do not spread the information, and are only receivers of it.
A very interesting study would be to adopt the current model
and study dynamics of belief evolution and propagation in
a network where viewers act as two-way agents passing and
receiving information.
Finally, the role of education can be explained and studied with
the current model to some extent. Education can be considered
as a meta-function that skews viewers beliefs towards the truth
over time. Given the importance of education in the age of
social media and abundance of information, we believe that it
is essential to provide modeling tools for studying its role in
prevention of misinformation.
Fig. 3: Statistics of belief convergence of optimal report for
true information (blue) and random unauthentic information
(red), for uneducated audience (cid:107)¯µ − xs(cid:107) (cid:54) (cid:107)¯µ − xt(cid:107).
Figure 4 shows convergence trade-off conditioned on the fact
that average belief is farther to the source (false) information
than the truth (cid:107)¯µ − xs(cid:107) (cid:62) (cid:107)¯µ − xt(cid:107).
Fig. 4: Statistics of belief convergence of optimal report for
true information (blue) and random unauthentic information
(red), for educated audience (cid:107)¯µ − xs(cid:107) (cid:62) (cid:107)¯µ − xt(cid:107).
.
Finally, an example of network utility function for β =
1.6, dmin = 1.1 as a function of is shown in Figure 5.
Fig. 5: Example of network utilities for β = 1.6, dmin = 1.1.
criminal ecosystem on twitter," in Proceedings of the 21st international
conference on World Wide Web. ACM, 2012, pp. 71–80.
[27] C. Budak, D. Agrawal, and A. El Abbadi, "Limiting the spread of mis-
information in social networks," in Proceedings of the 20th international
conference on World wide web. ACM, 2011, pp. 665–674.
[28] N. P. Nguyen, G. Yan, M. T. Thai, and S. Eidenbenz, "Containment of
misinformation spread in online social networks," in Proceedings of the
4th Annual ACM Web Science Conference. ACM, 2012, pp. 213–222.
[29] H. Zhang, M. A. Alim, X. Li, M. T. Thai, and H. T. Nguyen, "Misinfor-
mation in online social networks: Detect them all with a limited budget,"
ACM Transactions on Information Systems (TOIS), vol. 34, no. 3, p. 18,
2016.
[30] A. Ehsanfar and M. Mansouri, "Incentivizing the dissemination of truth
versus fake news in social networks," in System of Systems Engineering
Conference (SoSE), 2017 12th.
IEEE, 2017, pp. 1–6.
[31] N. A. Abdullah, D. Nishioka, Y. Tanaka, and Y. Murayama, "User's
action and decision making of retweet messages towards reducing mis-
information spread during disaster," Journal of Information Processing,
vol. 23, no. 1, pp. 31–40, 2015.
[32] A. Chakraborty and B. Yolmaz, "Manipulation in market order models,"
Journal of financial Markets, vol. 7, no. 2, pp. 187–206, 2004.
[33] Chatterjea, Arkadev, and R. A. Jarrow, "Market manipulation, price
bubbles, and a model of the us treasury securities auction market,"
Journal of Financial and Quantitative Analysis, vol. 33, no. 2, pp. 255–
289, 1998.
[34] J.-L. Vila, "Simple games of market manipulation," Economics Letters,
vol. 29, no. 1, pp. 21–26, 1989.
[35] B. Biais, P. Hillion, and C. Spatt, "Price discovery and learning during
the preopening period in the paris bourse," Journal of Political Economy,
vol. 107, no. 6, pp. 1218–1248, 1999.
[36] D. P. Bertsekas, "Constrained optimization and lagrange multiplier
methods," Academic press, 2014.
REFERENCES
[1] A. Benham, E. Edwards, B. Fractenberg, L. Gordon-Murnane, C. Het-
herington, D. A. Liptak, M. Smith, C. Thompson, and A. P. Mintz, "Web
of deceit: Misinformation and manipulation in the age of social media,"
Information Today, Inc., 2012.
[2] H. Allcott and M. Gentzkow, "Social media and fake news in the 2016
election," Journal of Economic Perspectives, vol. 31, no. 2, pp. 211–36,
2017.
[3] H. Rashkin, E. Choi, J. Y. Jang, S. Volkova, and Y. Choi, "Truth of
varying shades: Analyzing language in fake news and political fact-
checking," Proceedings of the 2017 Conference on Empirical Methods
in Natural Language Processing, pp. 2931–2937, 2017.
[4] A. Kucharski, "Post-truth: Study epidemiology of fake news," Nature,
vol. 540, no. 7634, p. 525, 2016.
[5] H. Rainie, J. Q. Anderson, and J. Albright, "The future of free speech,
trolls, anonymity and fake news online," Washington, DC: Pew Research
Center, 2017.
[6] J. C. Olson and P. A. Dover, "Cognitive effects of deceptive advertising,"
Journal of Marketing Research, pp. 29–38, 1978.
[7] G. M. Armstrong, M. N. Gurol, and F. A. Russ, "Detecting and
correcting deceptive advertising," Journal of Consumer Research, vol. 6,
no. 3, pp. 237–246, 1979.
[8] S. Romani, "Price misleading advertising: effects on trustworthiness
toward the source of information and willingness to buy," Product and
Brand Management, vol. 15, no. 2, pp. 130–138, 2006.
[9] J. Richards, Deceptive advertising: Behavioral study of a legal concept.
Routledge, 2013.
[10] E. J. Lee, K. S. Eom, and K. S. Park, "Microstructure-based manipula-
tion: Strategic behavior and performance of spoofing traders," Journal
of Financial Markets, vol. 16, no. 2, pp. 227–252, 2013.
[11] P. Kollock, "The production of trust in online markets," Advances in
group processes, vol. 16, no. 1, pp. 99–123, 1999.
[12] P. W. Singer and A. Friedman, Cybersecurity: What everyone needs to
know. Oxford University Press, 2014.
[13] L. Huang, A. D. Joseph, B. Nelson, B. I. Rubinstein, and J. D. Tygar,
"Adversarial machine learning," in 43-85. ACM, 2011.
[14] M. Barreno, B. Nelson, R. Sears, A. D. Joseph, and J. D. Tygar.,
"Can machine learning be secure?" in Proceedings of the 2006 ACM
Symposium on Information, computer and communications security,
2006, pp. 16–25.
[15] S. R. Knowlton and B. Reader, "Moral reasoning for journalists. santa
barbara," CA: Praeger, 2009.
[16] A. Day and E. Thompson, "Live from new york, it's the fake news!
saturday night live and the (non) politics of parody," Popular Commu-
nication, vol. 10, no. 1-2, pp. 170–182, 2012.
[17] L. Feldman, "The news about comedy: Young audiences, the daily show,
and evolving notions of journalism," Journalism, vol. 8, no. 4, pp. 406–
427, 2007.
[18] L. Sydell, "We tracked down a fake-news creator in the suburbs. heres
what we learned," National Public Radio 23, 2016.
[19] S. L. Borden and C. Tew, "The role of journalist and the performance
of journalism: Ethical lessons from fake news (seriously)," Journal of
Mass Media Ethics, vol. 22, no. 4, pp. 300–314, 2007.
[20] B. R. Lindsay, "Social media and disasters: Current uses, future options,
and policy considerations," 2011.
[21] C. Castillo, M. Mendoza, and B. Poblete, "Information credibility on
twitter," in Proceedings of the 20th international conference on World
wide web. ACM, 2011, pp. 675–684.
[22] A. Gupta and P. Kumaraguru, "Credibility ranking of tweets during
high impact events," in Proceedings of the 1st workshop on privacy
and security in online social media. ACM, 2012, p. 2.
[23] M. Gupta, P. Zhao, and J. Han, "Evaluating event credibility on twitter,"
in Proceedings of the 2012 SIAM International Conference on Data
Mining. SIAM, 2012, pp. 153–164.
[24] S. M. Shariff, X. Zhang, and M. Sanderson, "User perception of
Springer, 2014,
information credibility of news on twitter." in ECIR.
pp. 513–518.
[25] K. Thomas, D. McCoy, C. Grier, A. Kolcz, and V. Paxson, "Trafficking
fraudulent accounts: The role of the underground market in twitter spam
and abuse." in USENIX Security Symposium, 2013, pp. 195–210.
[26] C. Yang, R. Harkreader, J. Zhang, S. Shin, and G. Gu, "Analyzing
spammers' social networks for fun and profit: a case study of cyber
|
1709.08348 | 2 | 1709 | 2019-01-08T08:24:42 | SCube: A Tool for Segregation Discovery | [
"cs.SI",
"physics.soc-ph"
] | Segregation is the separation of social groups in the physical or in the online world. Segregation discovery consists of finding contexts of segregation. In the modern digital society, discovering segregation is challenging, due to the large amount and the variety of social data. We present a tool in support of segregation discovery from relational and graph data. The SCube system builds on attributed graph clustering and frequent itemset mining. It offers to the analyst a multi-dimensional segregation data cube for exploratory data analysis. The demonstration first guides the audience through the relevant social science concepts. Then, it focuses on scenarios around case studies of gender occupational segregation. Two real and large datasets about the boards of directors of Italian and Estonian companies will be explored in search of segregation contexts. The architecture of the SCube system and its computational efficiency challenges and solutions are discussed. | cs.SI | cs | SCube: A Tool for Segregation Discovery
Salvatore Ruggieri
Alessandro Baroni
University of Pisa, Italy
University of Pisa, Italy
[email protected]
[email protected]
9
1
0
2
n
a
J
8
]
I
S
.
s
c
[
2
v
8
4
3
8
0
.
9
0
7
1
:
v
i
X
r
a
ABSTRACT
Segregation is the separation of social groups in the physical or
in the online world. Segregation discovery consists of finding
contexts of segregation. In the modern digital society, discov-
ering segregation is challenging, due to the large amount and
the variety of social data. We present a tool in support of seg-
regation discovery from relational and graph data. The SCube
system builds on attributed graph clustering and frequent itemset
mining. It offers to the analyst a multi-dimensional segregation
data cube for exploratory data analysis. The demonstration first
guides the audience through the relevant social science concepts.
Then, it focuses on scenarios around case studies of gender oc-
cupational segregation. Two real and large datasets about the
boards of directors of Italian and Estonian companies will be
explored in search of segregation contexts. The architecture of
the SCube system and its computational efficiency challenges
and solutions are discussed.
1 SOCIAL SEGREGATION
Ethical issues in data and knowledge management are gaining
momentum in the last few years. In addition to the traditional
field of privacy, techniques for data analysis are being designed
or enhanced to take into account moral values such as fairness,
transparency, accountability, and diversity1. We have recentely
developed a novel data-driven technique for addressing segre-
gation of social groups through multi-dimensional data analysis
[4]. The approach is implemented in the SCube system, which
we propose to demonstrate using real case studies.
Social segregation refers to the "separation of socially defined
groups" [11]. People are partitioned into two or more groups
on the grounds of personal or cultural traits that can foster dis-
crimination, such as gender, age, ethnicity, income, skin color,
language, religion, political opinion, membership to a national mi-
nority, etc. Contact, communication, or interaction among groups
are limited by their physical, working or socio-economic distance.
This can be observed when dissecting society in organizational
units (neighborhoods, schools, job types). Due to the ubiquitous
presence and pervasiveness of ICT, segregation is shifting from
ancient forms of well explored spatial segregation2 to novel forms
of digital segregation. For instance, it has been warned that the fil-
ter bubble generated by personalization of online social networks
may foster idelogical segregation [6], opinion polarization [10],
and informational segregation. A data-driven technology that
enables the assessment of the extent, nature, and trends of social
segregation in the offline or online world, is of extreme interest
for a wide audience: social scientistics, public policy makers, reg-
ulation and control authorities, professional associations, civil
rights societies, and investigative journalists. Business decision
1See e.g., the Toronto declaration at www.accessnow.org/toronto-declaration.
2See census stats, e.g., www.census.gov/topics/housing/housing-patterns/data.html
© 2019 Copyright held by the owner/author(s). Published in Proceedings of the
22nd International Conference on Extending Database Technology (EDBT), March
26-29, 2019, ISBN 978-3-89318-081-3. on OpenProceedings.org.
Distribution of this paper is permitted under the terms of the Creative Commons
license CC-by-nc-nd 4.0.
region(CA)
south
north
∗
sex (SA) male
female
0.50
0.71
0.78
young
∗
0.46
0.62
0.83
0.35
0.50
0.43
-
0.88
-
-
0.30
0.30
0.59
0.57
0.22
0.67
0.88
0.86
0.66
0.56
0.76
-
0.75
0.75
-
0.71
0.71
∗
-
0.75
-
0.83
0.63
0.63
middle
elder
age (SA)
Figure 1: A segregation data cube with dissimilarity index.
makers should also care of business practices, particularly auto-
mated decision making, that segregate customers and products
through stereotypes, because this limits diversity and reduces
opportunities of cross-selling. Finally, data scientists and pro-
fessionals should be aware of the unintended consequences of
their models (recommender systems, link suggestion systems,
classifiers) on the cohesion of society at large.
2 SEGREGATION DISCOVERY
From a data analysis perspective, the key problem of assessing
social segregation has been investigated so far by hypothesis
testing, i.e., by formulating one or more possible contexts of seg-
regation against a certain social group, and then in empirically
testing such hypotheses. Such an approach is currently supported
by statistical tools, such as the R packages OasisR3 and seg4 [9],
or by GIS tools such as the Geo-Segregation Analyzer5 [2]. The
formulation of an hypothesis, however, is not straightforward,
and it is potentially biased by the expectations of the data analyst
of finding segregation in a certain context. In addition, explo-
ration of multiple hypothesis can be time consuming, since data
have to be processed multiple times. Finally, this approach is
subject to erroneous conclusions if data is considered at wrong
granularity -- an instance of the Simpson's paradox.
Multi-dimensional segregation data cube. Our approach
consists of providing the analysts with a multi-dimensional data
cube that can be explored in search of candidate contexts of
segregation. An example segregation data cube is shown in Fig.1.
Dimensions of the data cube include two types of attributes:
• segregation attributes (SA), such as sex, age, and ethnicity,
which denote (minority/protected) groups potentially ex-
posed to segregation;
• context attributes (CA), such as region and job type, which
denote contexts where segregation may appear.
Metrics of the data cube are chosen among the social science
indexes proposed for measuring the degree of segregation of
social groups within a society [12]. Here, we recall only one
such index, but the SCube system is parametric to the indexes
3cran.r-project.org/package=OasisR
4cran.r-project.org/package=seg
5geoseganalyzer.ucs.inrs.ca
(cid:12)(cid:12)(cid:12)
− ti − mi
T − M
D = 1
2
(cid:12)(cid:12)(cid:12)mi
M
n
i =1
and it computes 6 of them: dissimilarity, Gini, Information index,
Isolation, Interaction, Atkinson. Also, we restrict to binary groups
(minority/majority). LetT be the size of the total population under
consideration, 0 < M < T be the size of a minority group, T − M
the size of the rest of society (or majority group) and P = M/T
be the overall fraction of the minority group. Assume that there
are n organizational units (or simply, units -- such as schools,
neighboorhoods, job types, etc.), and that for i ∈ [1, n], ti is the
size of the population in unit i, and mi is the size of the minority
group in unit i. The dissimilarity index D measures the absolute
distance between the fractions of minority and majority groups
over the units:
D ranges over [0, 1], with higher values denoting higher segre-
gation. Dissimilarity is minimum when for all i ∈ [1, n], mi/ti =
M/T , namely the distribution of the minority group is uniform
over units. It is maximum when for all i ∈ [1, n], either mi = ti
or mi = 0, namely every unit includes members of only one
group (complete segregation). Dissimilarity and other segrega-
tion indexes can be interpreted as metrics in a cell of a multi-
dimensional cube as follows: set the total population as those
individuals that satisfy the CA coordinates of the cell; and, set
the minority population as those individuals that satisfy the SA
coordinates. For instance, the cube cell in Fig.1 with SA coordi-
nates sex=female, age=young and CA coordinates region=north
contains the dissimilarity index for the population living in the
north region and for the minority group of young women. Notice
that the number n of organizational units here have to be deter-
mined a-priori, while the total population and minority groups in
each unit depend on the values of cell coordinates. As in standard
multi-dimensional modelling [7], the special value "⋆" allows for
considering different granularities of analysis.
Segregation analysis of tabular data. We assume in input
a relational table with a tuple for every individual in the popula-
tion, including SA and CA attributes, and with a further attribute
unitID which denotes the unit an individual belongs to. Unfor-
tunately, segregation indexes are not additive metrics (see [4]).
This gives rise to the problem of efficiently computing a data
cube for segregation analysis. Our approach is more specialized
than generic holistic aggregate computation in datacubes [13].
We resort to frequent closed itemset mining [8]. Data cube co-
ordinates are encoded into itemsets of the form A, B, where A
denotes a minority subgroup and B denotes a context. Recalling
the previous example, A =sex=female, age=young defines the SA
coordinates, and B =region=north defines the CA coordinates.
The SegregationDataCubeBuilder algorithm described in [4] fills
data cube cells with the value of a segregation index by scanning
frequent closed itemsets of the form above. Since relational data
is transformed into transaction database for itemset mining, we
obtain for free that CA or SA attributes can be multi-valued,
e.g., to denote that an individual owns both a house and a car we
admit a relation tuple σ such that σ[owns] = {house, car}.
Segregation analysis of graph data. While transaction data-
bases are able to cover typical analysis from traditional social
science, they are not enough powerful to deal with social net-
work data. We formalize such a case using attributed graphs,
where nodes are assigned values on a specified set of attributes.
However, in this scenario, there is no a-priori defined notion of
organizational unit, i.e., the unitID attribute assumed in input
Figure 2: SCube architecture.
so far. Some forms of community discovery using graph cluster-
ing become necessary in order to determine the organizational
units. Clustering attributed graphs consists of partitioning them
into disjoint communities of nodes that are both well connected
and similar with respect to their attributes [5]. In summary, at-
tributed graph clustering can be used first to partition a social
network into communities. At this stage, every node/individual
in a community is described by its attributes and the community
id, which will be our unitID attribute. We have thus reduced the
problem to the analysis of relational data, for which the Segrega-
tionDataCubeBuilder algorithm can be applied.
Segregation analysis of bipartite graphs. An even more
complex scenario is when individuals are not connected among
them, e.g., because they are friends, but through a connection
with another entity, e.g., because they work in the same company.
Here, a form of projection on unipartite graph is needed to reduce
to the previous case. For instance, in [4], we adopt a bipartite
projection of the bipartite graph of directors and companies to
obtain a graph of companies connected by shared directors. Using
projection, we have reduced the problem to the previous case,
where attributed graph clustering can be adopted to find com-
munities of companies, which then represent the organizational
units for segregation analysis.
3 SCUBE ARCHITECTURE
The architecture of SCube is shown in Fig. 2. The system is
developed in Java, and it relies on a few state-of-the-art libraries6.
Inputs. The user has to provide features for two entities: in-
dividuals and groups. In the reference case studies, individuals
are directors and groups are companies. The input individuals
(a CSV file or a JDBC query) provides for each individual an ID
and a number of attribute values, distinguished into segregation
attributes (e.g., gender, age, birthplace) and context attributes
(e.g., residence). A second input groups provides for each group
an ID and a number of context attributes values (e.g., industrial
sector of a company and its headquarter location). Notice that
individuals are subjects to possible segregation, while groups are
6EWAH for compressed bitmaps (github.com/lemire/javaewah), Apache
POI for OOXML docs (poi.apache.org), Borgelt's FPGrowth for frequent
itemset mining (www.borgelt.net), FastUtil for graph storage (fastu-
til.di.unimi.it).
indexescubeGraph ClusteringTable BuilderGraph BuildernodeUnitgroupsfinalTable edgesSegregationData CubeBuilderdatesINPUTSOUTPUTisolatedVisualizerdatesmembershipindividuals scube.xlsx segregation atts
age
15-38
39-46
55-65
...
birthplace
foreign
south
north
...
gender
M
F
M
...
residence
north
south
south
...
context atts
sector
{electricity, transports}
{education}
{agriculture}
...
unitID
1
2
1
...
Figure 3: The process of segregation discovery supported by SCube (left, top), input to SegregationDataCubeBuilder (left,
bottom), and an output report on dissimilarity segregation index of the Italian provinces (right).
not. For this reasons, groups have no SA feature. A third input
is membership, which includes the edges of the bipartite graph of
individuals and groups, i.e., all pairs (individualID, groupID) for
which the individual is related to the group. In our case studies,
directors are related to companies they sit in the board of. We
also admit that the pairs are labeled with a time interval of valid-
ity, thus allowing for temporal analysis of segregation. We have
such an information for the Estonian dataset. A fourth input is
a list of snapshot dates at which to consider snapshots of the
membership relation.
Modules. SCube consists of five software modules. Graph-
Builder projects the bipartite graph of individuals and groups into
an unipartite attributed graph, where nodes are groups and an
edge connect two groups if they are related by at least one shared
individual. In the case studies, nodes are companies, and edges
connect companies that share at least one director in their boards.
Edges are weighted by the number of shared directors. Graph-
Builder outputs edges of the projection (edges), and nodes that
have zero degree (isolated). The GraphClustering module com-
putes then a clustering of nodes into organizational units (output
file nodeUnit). Methods for clustering available in SCube include:
extraction of connected components (Breadth-First Search), re-
moval of edges from the giant component with weight below
a threshold and then extraction of connected components (de-
signed in [4]), and an attributed graph clustering method for very
large graphs (SToC algorithm [3]). In our case studies, the result
of GraphClustering is a partitioning of companies into clusters
based on connections among companies determined by shared di-
rectors -- which can be readily considered a signal of relationships
(business, personal, or other) between companies. Clusters repre-
sent the organizational units needed for computing segregation
indexes. TableBuilder joins features of individuals with features of
the companies in an organizational unit. This yields a finalTable
with a row per individual and organizational unit she belongs
to. An example is shown in Fig. 3 (left, bottom). This is the input
for the SegregationDataCube builder module, implementing the
algorithm of [4]. Notice that if the data under analysis contains
already the assignment of individuals to units, i.e., it is already
in the form of finalTable, the pre-processing steps of bipartite
projection and graph clustering do not need to be performed. The
Visualizer module transforms the extended datacube in output
of SegregationDataCube into a standard OOXML format that
can be opened by Microsoft Excel, Libre Office, and other office
productivity tools (see Fig. 5). Segregation data cube exploration
can be easily interfaced with visualization tools, as in the map
overlay in Fig. 3 (right).
Process, Wizard, and GUI. The whole process of segrega-
tion discovery supported by SCube is shown in Fig. 3 (left, top).
To facilitate the adoption of SCube by non-technical users, we
have developed two interfaces (see Fig. 4). The first one is a
standalone wizard that guides the user throughout all the steps
of the process, asking for inputs and parameters when appro-
priate, and finish launching Microsoft Excel or Libre Office on
the output file. Using popular desktop tools as GUI's makes
the learning curve of approaching and effectively using SCube
more manageable. The second one is a cloud service offered
by the SoBigDataLab freely accessible research infrastructure
(www.sobigdata.eu/access/virtual), a web front-end comprising
a catalogue of data, services, and virtual research environments
for big data and social mining research.
4 DEMONSTRATION SCENARIO
The demonstration starts with a brief introduction on concepts
and methods of segregation measurement [12] and segregation
discovery [4]. This provides the audience with the basic defi-
nitions for understanding the SCube functionalities. The archi-
tecture of SCube is presented next. For interested participants,
computational efficiency, algorithmic solutions, and source code
internal aspects are discussed. Then, two running case studies in
the context of occupational segregation in the boards of company
directors [1] are introduced. They are based on a 2012 snaphost
of Italian companies (3.6M directors, 2.15M companies), and on a
20-year long dataset of Estonian companies (440K directors, 340K
companies). Such anonymized datasets are the largest ever con-
sidered in the literature of segregation analysis. We summarize
the data pre-processing activities to produce the inputs for SCube.
The demonstration then proceeds by presenting three analysis
scenarios based on input data of increasing complexity. In all
scenarios, gender, age, and birthplace are used as segregation
attributes. The first scenario considers tabular data, where com-
pany sector is used as organizational unitID, and it is intended
Data Pre - processed Data ETL individualGroup . csv group . csv individual . csv finalTable . csv cube . csv Mining SCube cube . xlsx patterns Analysis Knowledge Figure 4: SCube standalone wizard (left) and SCube method at the SoBigData research infrastructure (right).
5 CONCLUSION
This demonstration illustrates the SCube tool for interactive ex-
ploration of social segregation indexes in large and complex data.
The audience is made aware of social exclusion issues that can be
hidden in data and of the indexes that measure segregation. Real
case studies on scenarios of increasing complexity are discussed
and explored. Efficiency issues and algorithmic solutions adopted
for scaling to large datasets and graphs are detailed.
Acknowledgements. This work is partially supported by the
European H2020 Program under the funding scheme "INFRAIA-
1-2014-2015: Research Infrastructures" grant agreement 654024
"SoBigData" (http://www.sobigdata.eu).
REFERENCES
[1] M. Aluchna and G. Aras, editors. Women on Corporate Boards. Routledge,
2018.
[2] P. Apparicio, J. C. Martori, A. L. Pearson, E. Fournier, and D. Apparicio. An
open-source software for calculating indices of urban residential segregation.
Social Science Computer Review, 32(1):117 -- 128, 2014.
[3] A. Baroni, A. Conte, M. Patrignani, and S. Ruggieri. Efficiently clustering very
large attributed graphs. In ASONAM, pages 369 -- 376. ACM, 2017.
[4] A. Baroni and S. Ruggieri. Segregation discovery in a social network of
companies. J. Intell. Inf. Syst., 51(1):71 -- 96, 2018.
[5] C. Bothorel, J. D. Cruz, M. Magnani, and B. Micenková. Clustering attributed
graphs: models, measures and methods. Network Science, 3(03):408 -- 444, 2015.
[6] S. Flaxman, S. Goel, and J. M. Rao. Filter bubbles, echo chambers, and online
news consumption. Public Opinion Quarterly, 80:298 -- 320, 2016. Available at
SSRN: http://ssrn.com/abstract=2363701.
[7] J. Gray, S. Chaudhuri, A. Bosworth, A. Layman, D. Reichart, M. Venkatrao,
F. Pellow, and H. Pirahesh. Data cube: A relational aggregation operator
generalizing group-by, cross-tab, and sub totals. Data Min. Knowl. Discov.,
1(1):29 -- 53, 1997.
[8] J. Han, H. Cheng, D. Xin, and X. Yan. Frequent pattern mining: Current status
and future directions. Data Min. Knowl. Discov., 15(1):55 -- 86, 2007.
[9] S.-Y. Hong, D. O'Sullivan, and Y. Sadahiro. Implementing spatial segregation
measures in R. PLoS ONE, 9(11):e113767, 2014.
Will
[10] M. Maes and L. Bischofberger.
line social networks foster opinion polarization?
http://ssrn.com/abstract=2553436, 2015.
the personalization of on-
Available at SSRN:
[11] D. S. Massey. Segregation and the perpetuation of disadvantage. The Oxford
Handbook of the Social Science of Poverty, pages 369 -- 393, 2016.
[12] D. S. Massey and N. A. Denton. The dimensions of residential segregation.
Social Forces, 67(2):281 -- 315, 1988.
[13] A. Nandi, C. Yu, P. Bohannon, and R. Ramakrishnan. Data cube materialization
and mining over mapreduce. IEEE Trans. Knowl. Data Eng., 24(10):1747 -- 1759,
2012.
Figure 5: Top: sample multidimensional segregation cube.
Bottom: radial plot of segregation indexes for directors in
each of the 20 Italian company sectors.
to answer questions such as: how much are women segregated
in company sectors? The second scenario considers attributed
graph data, where nodes are directors, and edges connect two
directors if they belong to a same company board. Here, the
organizational units are determined through clustering over at-
tributed graphs. This scenario can answer questions such as: how
much are women segregated in communities of connected direc-
tors? Finally, the third scenario considers a bipartite attributed
graph of directors and companies, as presented throughout the
paper. An example of question it can answer is: how much are
women segregated in communities of connected companies? For
each scenario, the output of SCube is interactively explored using
pivot tables and charts. The audience is guided to the discovery
of a few actual cases of a-priori unknown segregation contexts
and to the understanding of which attributes contribute the most
to segregation. Moreover, a cross-comparison of the Italian vs
Estonian segregation findings will be discussed.
|
1710.02475 | 3 | 1710 | 2019-11-22T19:23:01 | Creating Full Individual-level Location Timelines from Sparse Social Media Data | [
"cs.SI",
"cs.CY"
] | In many domain applications, a continuous timeline of human locations is critical; for example for understanding possible locations where a disease may spread, or the flow of traffic. While data sources such as GPS trackers or Call Data Records are temporally-rich, they are expensive, often not publicly available or garnered only in select locations, restricting their wide use. Conversely, geo-located social media data are publicly and freely available, but present challenges especially for full timeline inference due to their sparse nature. We propose a stochastic framework, Intermediate Location Computing (ILC) which uses prior knowledge about human mobility patterns to predict every missing location from an individual's social media timeline. We compare ILC with a state-of-the-art RNN baseline as well as methods that are optimized for next-location prediction only. For three major cities, ILC predicts the top 1 location for all missing locations in a timeline, at 1 and 2-hour resolution, with up to 77.2% accuracy (up to 6% better accuracy than all compared methods). Specifically, ILC also outperforms the RNN in settings of low data; both cases of very small number of users (under 50), as well as settings with more users, but with sparser timelines. In general, the RNN model needs a higher number of users to achieve the same performance as ILC. Overall, this work illustrates the tradeoff between prior knowledge of heuristics and more data, for an important societal problem of filling in entire timelines using freely available, but sparse social media data. | cs.SI | cs | Creating Full Individual-level Location Timelines from Sparse
Social Media Data
Nabeel Abdur Rehman
New York University
[email protected]
Kunal Relia
New York University
[email protected]
Rumi Chunara
New York University
[email protected]
9
1
0
2
v
o
N
2
2
]
I
S
.
s
c
[
3
v
5
7
4
2
0
.
0
1
7
1
:
v
i
X
r
a
ABSTRACT
In many domain applications, a continuous timeline of human lo-
cations is critical; for example for understanding possible locations
where a disease may spread, or the flow of traffic. While data sources
such as GPS trackers or Call Data Records are temporally-rich, they
are expensive, often not publicly available or garnered only in select
locations, restricting their wide use. Conversely, geo-located social
media data are publicly and freely available, but present challenges
especially for full timeline inference due to their sparse nature. We
propose a stochastic framework, Intermediate Location Computing
(ILC) which uses prior knowledge about human mobility patterns
to predict every missing location from an individual's social media
timeline. We compare ILC with a state-of-the-art RNN baseline as
well as methods that are optimized for next-location prediction
only. For three major cities, ILC predicts the top 1 location for all
missing locations in a timeline, at 1 and 2-hour resolution, with
up to 77.2% accuracy (up to 6% better accuracy than all compared
methods). Specifically, ILC also outperforms the RNN in settings
of low data; both cases of very small number of users (under 50),
as well as settings with more users, but with sparser timelines. In
general, the RNN model needs a higher number of users to achieve
the same performance as ILC. Overall, this work illustrates the
tradeoff between prior knowledge of heuristics and more data, for
an important societal problem of filling in entire timelines using
freely available, but sparse social media data.
CCS CONCEPTS
• Social and professional topics → User characteristics; • Ap-
plied computing → Law, social and behavioral sciences;
KEYWORDS
Spatial Information and Society, social media, sparse data
ACM Reference Format:
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara. 2018. Creating
Full Individual-level Location Timelines from Sparse Social Media Data. In
26th ACM SIGSPATIAL International Conference on Advances in Geographic
Information Systems (SIGSPATIAL '18), November 6 -- 9, 2018, Seattle, WA, USA.
ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3274895.3274982
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].
SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
© 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-5889-7/18/11...$15.00
https://doi.org/10.1145/3274895.3274982
1 INTRODUCTION
Using full location timelines (all locations an individual has been
to) is essential to many societal applications including transporta-
tion management [22, 23], urban sensing [18], event detection [1]
and infectious disease dynamics [39]. Combined with additional
information about individuals, location timelines have been used to
predict depressive moods [6], point-of-interest and location recom-
mendation [3, 43], the spread of diseases [25], and contact tracing
for hot spots of infectious diseases [37].
Data sparsity can become a major challenge when predicting full
timelines using publically available data, and can take two forms.
First, the number of users with enough data (full-timelines) to train
the model can be low. Second, increasing the number of users re-
sults in inclusion of users with extremely sparse location timelines.
Therefore, this problem of inference of complete location timelines
is inherently different from mobility prediction, which must priori-
tize accuracy of the prediction for the next location. Accordingly,
mobility prediction models are often built on data sources such as
travel surveys, Call Data Records (CDRs) and Global Positioning
System (GPS) trackers which are high-resolution (provide location
every few seconds or minutes). However, it is not realistic to have
such data in a broad array of contexts; the cost of collecting such
datasets, limited attributes associated with individual records, and
lack of public availability makes them unsuitable for carrying out
large-scale studies for a target population where the impact of lo-
cation over time is to be studied in relation to various secondary
issues. Further, in several emerging real-world modeling applica-
tions such as infectious disease transmission models, knowing a
person's location at such a high temporal frequency is unnecessary.
Instead, locations of where they travel to over the course of a day at
a lower resolution (such as every few hours), provide the relevant
insight [15]. Therefore, we focus on the challenging problem of
constructing the entire mobility timeline of an individual at equal
intervals of time from the geo-location associated with social media
data, which is generated at-will and therefore can be very sparse.
The main challenge here is that the data used is truly sparse in
relation to entire timelines; for example, in six months of social me-
dia sourced from the Twitter Application Programming Interface,
only 5.4% (which is what we use in this study) have a Tweet with
linked-location at each of the daytime hours in a day (independent
of day of the week, over all weeks in the six months). Further, we do
not assume any specific information such as from the text/content
of posts, or network of users is available. To overcome these chal-
lenges, we use several known heuristics about location visitation
patterns of individuals. We also combine patterns both from an
individuals' history, as well as leverage the patterns of similar com-
munity members [19, 32, 42]. Further, we relax criteria about day
and week-specificity of location patterns which enables us to use
SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara
and predict timelines from thousands of social media users with
such sparse data -- our method does not require rich training data
to learn complex patterns of mobility, and works for a realistic
number of users. Finally, our work does not assume any additional
information about the demographics, social networks or content of
tweets of users, allowing for the adoption in situations where such
additional data sources are not available. In accordance with the
sparse nature of the data, these combined approaches enable us to
infer multiple consecutive missing locations from a user's timeline,
and construct a continuous location timeline for individuals using
only sparse geo-tags from their Tweets.
We compare the performance of our model with several models
which, although optimized for next-location prediction, are state-
of-the-art, and show that intermediate location computing (ILC)
has increased accuracy for inferring entire timelines from sparse
data. In particular, while deep learning models have good predic-
tive accuracy, we investigate the tradeoff in performance based on
amount of sparsity, both in terms of number of users or amount
of data per user. By using readily available data to estimate a full
mobility timeline at relevant resolutions, this work opens many
new opportunities to understand and predict human movement for
many domain areas. To the best of our knowledge, this work is the
first to use sparse social media data to infer full individual-level
location timelines. The specific contributions of this work are:
• Developing a framework for filling in entire location time-
lines at reasonable time steps, with personalized forward
and backward timeline prediction.
• Prediction of the timeline from truly sparse, but freely avail-
able and easily accessible data; with smart use of community
data to improve timeline prediction when applicable.
• First use of deep learning for inferring timelines from sparse
data and assessment of amount of data needed for a deep
learning approach to surpass other models.
2 RELATED WORK
Here we summarize related work in two main categories to clarify
differences in data and methodological approaches in other work.
Geo-location data types and sparsity. As the goal of predicting
next location is different than the goal here, the types of data used
in such studies include smartphone data including GPS tracking,
Wi-Fi, Bluetooth and phone usage [11], partial GPS tracks from
automobiles [21], and Foursquare check-in data [30]. These mod-
els are generally designed for temporally rich data-sets and thus
assume that the training data is abundant and collected at frequent
time intervals. Even in the case of Foursquare data, though it can be
sparse, only dense sequences of data (minimum sequence length of
5 locations) have been used in predictive efforts [12]. Hence, studies
have been concerned with data collected at such densities, or small
time intervals (e.g. every 1, 15 or 30 minutes), and individual records
below a threshold number of data points are discarded from the
study entirely [5, 12, 14, 30, 33]. While this restriction increases
confidence in the stay duration of individuals at a location, this
typically (appropriately) limits the problem of prediction to only
a single missing location in the future. Given the inconsistency
of intervals between location tags in an individual's social media
timeline and lack of stay duration information, such models can not
directly be applied in the context of sparse social media data [38].
A method to capture daily habits of individuals using sparse data
has been proposed in [28] (varying the amount of phone GPS data
"seen" by the algorithm). However, the method initially requires
training on users with abundant data histories and hence cannot
be replicated with data sources such as from Twitter, where both
the training and testing datasets are sparse.
Broadly, related social media efforts have been focused on pre-
dicting the location of a given social media post, and not missing
locations from a timeline [19]. Such studies have also included
users with sufficient data and with certain assumptions (e.g. only
on those Twitter users who both themselves and their friends are
extremely active on Twitter, with at least 100 geo-tagged Tweets
in 1 month and assumes that once a user Tweets from a location,
they remain at that location until they Tweet again) [32]. Other
research which use social media in the domain of mobility focus
on Point-of-Interest (POI) and location recommendation, and pro-
vide the insight that similar user behavior can be useful [42]. This
method uses behavior of similar individuals and distances between
pairs of locations to predict the next POI location for an individual.
Thus we incorporate this feature of user similarity into the ILC
approach to address sparsity issues, and also compare our method
to the proposed method for full timeline inference.
Mobility sequence prediction methods. There are many model and
pattern based methods that have been used to infer movement of
individuals. While the focus of these methods has mainly been
to predict the next sequence of locations, and cannot be directly
compared to our goal of filling in an entire timeline, they have still
provided important knowledge about human mobility that can be
used in the timeline problem.
Several variations of Markov models, LZ predictors and pre-
diction by partial matching (PPM), as well as a non-linear spatio-
temporal prediction framework, have been investigated [2, 13, 14,
34, 36]. These methods focus on modeling the probability of visita-
tion to a future location by probability or frequency of past visits
and popular sequences in existing trajectories, each evaluating it's
performance on prediction of a next location. Although that is not
our goal, we can still make use of such probabilities in our data
by incorporating components of the basic Markov model into the
ILC model, though in a manner that promotes filling in all missing
data, not just next location prediction. Besides, we also explicitly
assess performance of each of these these methods on sparse social
media data in comparison to our proposed approach where possible,
including NextPlace non-linear predictor [34], Markov Order-0 and
Markov Order-1 models [36].
More recently, recursive neural networks (RNNs) have been used
to predict individual level mobility timelines [12, 27, 40]. RNN ar-
chitectures have been used to predict where a user will check-in
next [27] and for next location recommendation [40]. Another RNN
architecture to predict next location in the timeline of an individual
has been proposed in [12]. The model is again focused on predict-
ing next location more accurately, and incorporates modules in the
architecture in order to capture more complex multi-scale patterns.
As well, despite the fact that this work aims to predict location value
in a user's timeline when data is sparse, the work only focuses on
predicting the locations in the subset of timelines of users where
Creating Full Individual-level Location Timelines from Sparse Social Media Data SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Figure 1: A: Number of included Tweets by day of the week
(pink: weekdays, blue: weekends). B: Number of Tweets by
hour of day (pink: daytime, blue: nighttime). C: Frequency
distribution of total data points per user before filling in the
timeline. All graphs include all data from all 3 cities.
Figure 2: A&B: Frequency distribution of number of as-
signed locations over all values of k (closest hour of week)
for all posts per user (A: resolution ri = 1 hour, B: ri = 2). C&D:
Frequency distribution of number of assigned locations per
user over all values of h (closest hour of day) (C: ri = 1, D:
ri = 2). All graphs include data from all 3 cities (excluding
users who had no tweet in daytime hours). Users fulfilling
inclusion criteria are highlighted in pink.
richer data is available (described in the previous section), and does
not address the challenge of inferring the complete timeline of a
user. While these new methods provide a fresh approach to address-
ing the problem of mobility prediction by allowing the model to
learn different behaviors on its own as opposed to previous meth-
ods where the behaviors of individuals were manually specified,
they are not specifically tailored for predicting complete timelines.
However, given the potential for high performance of deep learning
models, we do assess what the tradeoff would be, for performance
on our task, in terms of data availability (e.g. with what amount
of data would a standard deep model perform better than a model
incorporating known movement heuristics a priori).
3 METHODS
3.1 Datasets
In order to obtain enough data for training and testing, we used 6
months of publicly available geo-located data from the Twitter API
(1st January -- 30th June 2014) for the cities of New York, Washing-
ton, DC and San Francisco. We collected all the Tweets containing
a 'point' geo-location within defined bounding boxes for all three
cities. The resulting data set consisted of 18,164,503 Tweets by
443,945 users from New York City, 3,385,308 Tweets by 125,873
users from Washington, DC and 1,817,411 Tweets by 111,441 users
from San Francisco.
3.2 Filtering and Preprocessing
Spoofed locations. We identified and excluded any Twitter
3.2.1
accounts that represented impossible movements based on Tweet
locations and times. A threshold speed of 0.5 miles/minute was used
to filter out such Tweets, based on previous work outlining realistic
movement patterns [24], and all accounts with more than 5% of
their Tweets violating the above criteria were excluded. A total of
16,582, 3,342 and 2,750 accounts (from each city, respectively) who
were removed due to having more than 5% of their Tweets marked
as coming from a spoofed location.
3.2.2 Grids. We assessed three grid sizes; 1 × 1, 0.5 × 0.5 and 0.1
× 0.1 miles. For each, we assigned every geo-located Tweet in the
dataset to a grid. These grid sizes are based on previous research
which has identified perception of how large a neighborhood bound-
ary is for temporary movements such as walking (1 mile) [29, 31].
Multiple grid sizes were added to assess the impact of grid size
on the performance of the method. A total of 841, 143 and 736
grids(grid size= 1 × 1 miles), 3,364, 572 and 2,944 grids (grid size=
0.5 × 0.5 miles) and, 84,100, 14,300 and 73,600 grids (grid size= 0.1
× 0.1 miles) were created for NYC, DC and SF respectively.
3.2.3 Temporal Sampling. Tweet timestamps were adjusted for
time zone and daylight savings. Included Tweets were distributed
across all days of the week evenly (Fig. 1A). For each individual
present in the dataset we created separate timelines at resolutions
of ri = 1 and 2 hours. Given the time stamp t of a Tweet, and value
ri, kw := (h, d)w , is computed where k is the closest sampled hour
at an interval of i hours from the start of the week, h is the closest
sampled hour at an interval of ri from the start of the day, d is the
day of the week and w is the week number since the start of the
data. For example, for ri = 1, the time of a Tweet made at t = 19 : 05
on Tuesday would be assigned k = 43 i.e 24+19 and (h, d) = (19, 2)
(assuming week starts on Monday). For ri = 2, for the same Tweet,
k = 22 i.e. 12 + 10 and (h, d) = (10, 2).
Stay Duration. To estimate the stay duration, we interpolated
3.2.4
data points from users who made consecutive tweets from the same
location within a 6-hour or shorter time period. The maximum value
of 6 hours for the interpolation was a conservative estimate chosen
based on research showing how long people generally remain in
their most visited locations, and that an individual generally spends
most of their time in most visited locations [8, 16, 20].
3.2.5 Home Location. Individuals are more likely to stay at their
home location for longer periods and individuals generally don't
change locations at night time [17]. Consistent with previous stud-
ies and Fig. 1B, we consider a location x as the home location of
an individual on a day of week d, if the individual most frequently
tweets from location x between 10 pm of the day of the week d and
051015MonTueWedThuFriSatSunDay of the WeekNumber of Tweets (X 100,000)A024605101520Hour of the DayNumber of Tweets (X 100,000)B012340100020003000Number Of Data PointsLog10 (Number of Users)C0100020003000255075100Number of assigned locations per user over all values of k (ri=1)Number of UsersA01000200030001020304050Number of assigned locations per user over all values of k (ri=2)Number of UsersB0204051015Number of assigned locations over all values of h (ri=1)Number of Users (X 1000)C020402468Number of assigned locations over all values of h (ri=2)Number of Users (X 1000)D SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara
8 am of the day of the week d+1. Given the sparse nature of the data,
for days of the week wherein an individual had no Tweets between
10 pm and 8 am, the home location was assigned where they most
frequently Tweeted from between 10 pm and 8 am, irrespective of
day of the week. We refer to points in an individual's timeline with
location information, either originally from a user or interpolated
from a home or stay duration, as assigned locations.
3.2.6 Personal vs Non-personal accounts. A Twitter account, e.g.
@SearchAmerican, that belongs to an organization, as opposed to
an individual, is likely to be used by multiple individuals in the
organization and hence does not represent the movement patterns
of a single individual. To examine the distribution of personal vs.
non-personal accounts in our dataset, we used Amazon Mechanical
Turk (AMT) labelling on 7,000 randomly selected accounts. Each
account label was manually annotated twice by AMT workers as ei-
ther personal or non-personal accounts. Accounts with conflicting
labels were annotated a third time through AMT and the maxi-
mum vote used. 98% of the 7000 randomly selected accounts were
identified as personal accounts. Cohen's kappa score of the annota-
tors was 93.0% [9]. Given the overwhelming majority of accounts
were identified as personal, it was assumed that most non-personal
accounts must have been removed during the 'spoofed location'
filtering stage. The 2% non-personal accounts from the 7,000 set
were removed from the study but it was deemed unnecessary to
label the remaining accounts. After this stage, no information (e.g.
the Twitter handle) which could link back to an individual Twitter
account holder was retained.
3.2.7 Description of Included Users. We define a relaxed inclusion
criteria to ensure that the performance of all methods is being tested
on users with sparse timelines. From here onward we define two
notations: xkw (u) is the assigned location for a user u, at time kw ,
with kw as above. kw can also be interchangeably written as (h, d)w ,
with h, d and w as above, or as q, which represents the index of kw
in the sampled timeline. Inclusion criteria were defined as follows:
given the timeline of a user, the user must have at least 1 assigned
location for each h during daytime hours (8am-10pm;non-nighttime
hours as defined in the section "Home Location"), irrespective of d
and w. This means that at a resolution of ri = 1 hour, all users were
included in our analysis who, after interpolation of stay duration
had at least 15 assigned location data points (8am-10pm) in the
entire duration of the dataset over all distinct h. For ri = 2, the
number was 8 assigned location data points. This resulted in 29,491,
4,947 and 1,119 users (ri = 1) and 45,710, 8,083 and 2,395 users (ri =2)
from New York, Washington, DC and San Francisco respectively.
Defining a relaxed inclusion criteria based on distinct h instead of
distinct k enabled us to include orders of magnitude more users
(Fig. 2) and allowed us to include up to 45% user (in NYC) who made
a tweet during daytime hours. The above selected users had on
average 82.8% (ri = 1) and 72.0% (ri = 2) of their daytime timelines
with no assigned location.
3.3 Individual Timelines
In this section, we first discuss prediction of a missing location in
a user's timeline at time k, if location information of the user is
available at both k + 1 and k − 1. As described in earlier work, the
movement of individuals is not entirely random and certain features
can be extracted to predict an individual's location based on his
past behavior [35]. Moreover, people often move in groups, and
individuals with similar interests follow similar movement patterns
[26]. Accordingly, here we model the behavior of individuals as
a combination of: i) personal behavior represented as (subscript
I), and ii) community behavior (subscript C). Personal behavior is
further modeled using three behaviors: i) Next Location subscript
(I , a), ii) Previous Location subscript (I , b), and iii) Independent
Location subscript (I , c). Each of these three behaviors are further
treated as either i) day of the week and hour of the day specific
(superscriptW S), ii) workday (weekday) or non-workday (weekend)
and hour of the day specific (superscript RS), and iii) only hour of
the day specific (superscript HS). These three stratifications were
created because of the extremely sparse nature of the dataset in
which we rarely observe users who have at least 1 location value
present for all days of the week and hours of the day. For the
following section we define: any P represents a list of locations
and their corresponding probabilities for a given user at a given
== j) thus represents the probability corresponding to
time. P(xkw
location j at time kw , and P(xkw ) represents the list of all possible
locations and their corresponding probabilities at time kw .
(u) is missing, and
3.3.1 Next Location. Given the location x(k +1)w
given the location xkw (u) = i, we calculate the conditional proba-
bilities of all the possible locations of a user u at time (k + 1)w . This
probability is calculated by taking into account that people often
follow specific patterns of mobility. For example, in the evening at
7pm, given that an individual is at a grocery store, the next loca-
tion of an individual will likely be his home. Given that the same
individual is at home at 7pm, the individual could either choose
to stay home or to go out (e.g. to a restaurant or bar). Given that
the time period is assumed to be 1 week, as contended in previous
work, these conditional probabilities are specific for each sampling
time on a given day and day of the week, irrespective of the date
[5]. Then, for all possible locations j, PW S(I,a)(x(k +1)w′ = j) of a user,
given xkw′ = i, is defined as:
(x(k +1)w
== i)
w
w
== j xkw
(xkw
== i)
For RS we calculate similar proportions, but relax the conditions
by additionally accounting for days which are of the same type,
i.e workday or non-workday, when calculating the proportions.
P RS(I,a)(x(h+1,d′)w′ = j) is thus defined as:
w
w
d =DT(d′)
(x(h+1,d)w
== j x(h,d)w
== i)
(x(h,d)w
d =DT(d′)
== i)
where DT(d′) returns the list of type of days i.e weekdays or
weekends, as d′. P H S(I,a) completely removes the condition of the
proportion being specific to the day of the week (instead of d =
DT(d′), we consider all d).
3.3.2 Previous Location. As a reciprocal of Next Location predic-
tion wherein we used xkw to predict x(k +1)w , here we predict xkw
conditioning over the location value at x(k +1)w . P RS(I,b) and P H S(I,b)
Creating Full Individual-level Location Timelines from Sparse Social Media Data SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
are calculated similarly using relaxed conditions of day of the week,
as defined for P RS(I,a) and P H S(I,a).
Independent Location. Several locations which an individual
3.3.3
visit are specific to the day and time regardless of where the indi-
vidual is coming from or where they plan to go next. For example,
for a weekly meeting or a class at 11am on Tuesday, an individual
will be in the location of the meeting or the class irrespective of
his previous or next location. To incorporate these patterns, we
calculate probabilities for "Independent Location": the probability
of a user being in any location j at time k := (h, d), PW S(I,c). This
is defined as the proportion of times the user was at location j at
time k, in the dataset. P RS(I,c) of a user being in any location j at
time k := (h, d) is defined as the proportion of times the user was
at location j during hour h and days of the week similar to d i.e
(weekday or weekend). And, P H S(I,c) is defined as the proportion of
times the user was at location j during hour h in the dataset.
Combining lists of all probabilities in the individual's (W S) be-
havior gives:
PW S
I
= (λa ∗ PW S(I ,a) + λb ∗ PW S(I ,b) + PW S(I ,c))/3
where λa and λb are information loss factors defined later in the
Intermediate Location Computing section. Probabilities of visit to
each location, from all behaviors, are summed to generate a single
list of locations and their corresponding probabilities:
PI = (PW S
I
+ P RS
I
+ P H S
I
)/3
3.4 Community Behavior
Individuals with similar interests, or those working or living in the
same demographic have a higher chance of visiting similar locations
[5]. Hence, we maximize the use of the data by also including
information about individuals who have shown to follow similar
mobility patterns. For each individual, we identify individuals who
have similar mobility patterns, via a similarity factor. This factor,
s(u1, u2), is defined as the probability that another individual u2
will be in the same location as the individual under consideration
u1 at any given time:
w
k(xkw (u1) == xkw (u2))
k(!N U LL(xkw (u1))&!N U LL(xkw (u2)))
w
Using the similarity factor defined above, we calculated community
behavior (probability list for locations at a time k) using the top m
users in the dataset with the highest similarity factor for a given
individual via:
PC(xkw
= j) =
s(u) ∗ (xkw (u) == j)
m
u =1
Combining individual and community behavior then gives:
= j)
= j) = (1 − βk) ∗ PI(xkw
= j) + βk ∗ PC(xkw
P(xkw
where βk defines the hour and day of week specific effect of com-
munity behavior on an individual. To account for varying behavior
of an individual during a week, we generated separate lists of simi-
lar users for weekdays and weekends. We also examined a range of
values for m (0, 1, 2, 5, 10, 20, 50), to identify the minimum number
of similar users for maximizing prediction accuracy.
3.5 Intermediate Location Computing
Given the sparse nature of social media, in most instances there are
multiple consecutive missing location data points in an individual's
timeline. Thus the issue of predicting location at kw if either or both
(k +1)w and (k −1)w are missing will arise. Hence we introduce the
concept of Intermediate Location Computing. For simplicity, we
will only define the procedure to identify the intermediate location
at sampled time (k − 1)w (which is used to calculate of PW S(I,a)). A
similar approach can be used to identify the location at time (k +1)w
(which is used to calculate PW S(I,b)).
Broadly, our problem is that a location exists at time (k − n)w
such that no location data for an individual is present between kw
and (k − n)w . To address this, we use location data at (k − n)w
to iteratively predict intermediate locations of the individual at
times (k − n + 1)w until we reach (k − 1)w . We define the function
Inter(L1, L2), which for a specific time point, takes in two lists L1
and L2, and returns the location which has the maximum proba-
bility in list L1, and if no location exists, returns the location with
maximum probability in list L2. Here L1 is P(I,a) and L2 is P(I,c).
In simple terms, at each step, we first identify the most probable
location using Next Location. If no location data exists, we resort to
identifying the most probable location using Independent Location.
Given that locations at (k + 1)w and (k − 1)w predicted using
this method are only probable locations and successive predictions
will decrease certainty, we multiply by an information loss factor
λ to account for loss in information in calculating intermediate
locations. This factor λ is defined as: λ = (1 − α)(n−1), where n is
the number of steps required to reach the nearest available point
with an available location, and α is a constant information loss on
each step.
This approach to identify loss in information in sequential predic-
tors has been used in the past, particularly in dynamic belief models
[41]. The basic idea is that at each sequential prediction there is a
probability of α that the prediction will be incorrect. Iterating this
for a data point present n steps away makes the overall probability
of correct prediction (1 − α)(n−1). In the example given in Fig. 3,
when finally calculating the location at kw , given that the value of
n for the left side is 3, PW S(I,a) is multiplied by (1 − α)(3−1). Similarly
given that n for the right side is 2, PW S(I,b) is multiplied by (1−α)(2−1).
The example in Fig. 3 demonstrates the steps performed to compute
the intermediate locations for (W S). We use the same method to
calculate intermediate locations for (RS) and (HS) probabilities.
The complete method to construct complete mobility timeline of
a given user is summarized algorithm 1. In the algorithm, as defined
above, we replace kw with q, to represent the index of each time
step in the timeline. Further given a timeline T of an individual, the
location xq is the q element in T , i.e T[q].
3.6 Training and Testing Data and
Optimization
To select the training data for the entire prediction we, randomly
and uniformly across all distinct values of k, sampled 70% of the
data from each user. It should be noted that the test set contains
the 30% location data of each user which was not used in calcu-
lating the conditional probabilities or training the model. Further,
SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara
if NU LL(T S
end for
a [q − 1]), PS
I,a(xq))
I,a(xqxq−1 = T S
end for
for q in lenдth(T S
a ← T S
T S
for q in 1 : lenдth(T S
b ← T
if NU LL(T S
a ) do
a [q]) then
a [q] ← inter(PS
T S
end if
b ) : 1 do
b [q]) then
b [q] ← inter(PS
T S
end if
Algorithm 1 Constructing complete mobility timeline using ILC
Input: Timeline T of user u, community behavior, PC, of similar
users at each time step, effect of community behavior β and λa and
λb for each q
Output: Complete timeline Tcomplete
1: Tcomplete ← T
2:
3: for each behaviour S in [W S, RS, HS] do
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15: end for
16:
17: for each xq in T do
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
if NU LL(xq−1) then
I,a(xq) ← PS
PS
I,a(xq) ← PS
PS
end if
if NU LL(xq+1) then
I,b(xq) ← PS
PS
I,b(xq) ← PS
PS
end if
I (xq) = (λa ∗ PS(I,a)(xq) + λb ∗ PS(I,b)(xq) +
PS
PS(I,c)(xq))/3
I,b(xqxq+1 = T S
b [q + 1])
I,b(xqxq+1 = T[q + 1])
I,a(xqxq−1 = T S
a [q − 1])
I,a(xqxq−1 = T[q − 1])
for each behaviour S in [W S, RS, HS] do
I,b(xqxq+1 = T S
b [q + 1]), PS
I,a(xq))
if NU LL(xq) then
else
else
end for
PI = (PW S
I
Tcomplete[q] ← arд.max((1−βq)∗PI(xq)+βq∗PC(xq))
(xq) + P H S
I
(xq) + P RS
I
(xq))/3
31:
32:
33:
34:
35: end for
end if
Return Tcomplete
the data spans only the daytime hours wherein an individual is
changing location most frequently. The performance of the model
was calculated only on this test data as not to bias the performance
of the method towards sampled times where an individual is static
(nighttime hours).
Using the training data, we calculated PI and PC lists for every
individual, at each time resolution. These probabilities are then
used to optimize the value of βk and α. For simplicity, we optimize
a fixed value independent of a user or a sampling time for α, but βk
is user, day of the week and hour of the day specific as we would
Figure 3: Intermediate Location Computing algorithm illus-
tration. A: Timeline of an individual for a week w, between
k − 3 and k + 2. Location data for the individual is missing
for k − 2 to k + 1. Shaded area shows the location to be pre-
dicted. B: Intermediate locations (red) calculated after first
iteration. C: Intermediate locations after second iteration. D:
Effect of information loss on PW S(I,a) and PW S(I,b).
expect the contributions of community behavior to vary at different
times and for different people. To select the optimal values of βk,
we vary βk from 0 and 1 ( intervals of 0.05) and select the βk, for a
given k, that maximizes prediction accuracy on the training data. α
was optimized in a similar way, but only using PI (inclusion of PC
would have resulted in concurrent optimization of both α and βk).
The value of α as 0.1 performed well on the training set, and was
used in study.
3.7 Evaluation Versus Baseline Models
For fair comparison and to ensure that the variation in performance
is only due to the inference power of the models and not due to
variation in training data, all baseline models were trained using
the same training data for each user (post processed form of data)
as used for ILC, and the performance of the models was tested on
the same test set.
3.7.1 Home-Work location Model. It has been shown that periodic
behavior accounts for up to 70% of an individual's movement [7].
Given that the periodic behavior Hence, the first baseline model
assumes users follow a simple periodic behavior, switching be-
tween two locations: their inferred home and work locations. Using
the training dataset, we computed and assigned a single home
(nighttime) and a work (daytime) location for each individual by
identifying the most frequent location a user is present in between
10pm and 8am, and between 8am and 10pm.
3.7.2 Markov Models. Markov models have been widely used to
predict individual level mobility patterns [4, 36]. An Order-0 Markov
model identifies the most frequent location a user is in during a
given hour of the day, regardless of where the user came from or
is going[36]. The Order-1 Markov model, given the location x of
Creating Full Individual-level Location Timelines from Sparse Social Media Data SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
an individual at time k, identifies the most frequent location the
individual visits at time k +1 if they were at x during time k. Due to
sparsity of data, multiple missing locations are predicted iteratively.
i.e. each subsequent location at k + n is predicted using the previ-
ously predicted location at k + n − 1. For fair comparison, we use a
fall-back version for both Markov models which first computes the
(W S) likelihoods. If no location data exists, the model falls back to
(RS) likelihood, and then to (HS) likelihood.
3.7.3 Collaborative Point-of-Interest Recommendation Model. The
Point of Interest (POI) recommendation model was initially pre-
sented in [42], to recommend locations of interests of individuals
using data from Location Based Social Networks (LBSNs). The
model, in addition to using geographical distance between loca-
tions, first identifies close users both based on the social network
(friends/followers) of an individual as well as those who follow
similar movement patterns, and uses their location to predict the
individuals location. In line with the conclusion of the original
work, that social ties are not strong predictors, and given that we
are not assuming that the location data for the social network of
individuals is available, we model the movement of an individual
using the geographical distances between locations and location
data of users who follow similar movement patterns. Geographical
influence is modelled based on a power-law distribution between
successive data points, while location of similar users is calculated
similar to the community behavior part of our method.
3.7.4 NextPlace: Spatio-Temporal Non-linear Model. This spatio-
temporal non-linear "NextPlace" prediction model uses a non-linear
framework for predictions and unlike Markov models, which pre-
dict the next location at time k + 1 using historical movement
patterns, or the community based methods, which use location data
of similar users, uses the history of trips to the same location to pre-
dict when an individual will be in the same location the next time.
[34] The method first identifies the start time and stay duration of
each trip, then embeds the timeseries in a multidimensional space
by adding multiple instances of the timeseries with delays to ac-
count for non-linearity. Then, the start times and stay durations of
the user's next visit are averaged to predict when and for how long
the next visit to the location will happen. In our implementation,
we used the delay as the smallest temporal unit in our study (i.e 1
and 2 hours for ri=1 and 2). Given the sparsity of data, we define the
start time when an individual makes a tweet from a location, and
stay duration is either inferred as described in the preprocessing
section of paper, or assumed to be either 1 or 2 hours based the
value of ri.
3.8 Recursive Neural Network
RNNs and specifically LSTMs (Long Short-Term Memory blocks)
have been gaining popularity due to their strength in identifying
and utilizing complex sequences of information to make future
predictions. For the domain of mobility prediction, this provides a
contrast to other work in which the heuristics for mobility mod-
elling are self specified. Hence, here we also study the utility of
an RNN architecture in constructing entire mobility timelines of
individuals in the context of sparse location data. Fig. 4 shows the
architecture of the network. We use a basic architecture, similar to
Figure 4: Architecture of Recursive Neural Network.
those used in previous mobility and sequence prediction work [12],
but adapted for full timeline location inference. Specifically, instead
of using separate inputs for current and historical trajectories of
location, due to the sparse nature of data we input a single trajec-
tory of all locations. Secondly and more importantly, here instead
of only using a historical sequence of locations (left padded input),
[x(k−n)w ,... x(k−2)w , x(k−1)w ] to predict xkw , we also use the fu-
ture sequence of location (right padded input), [x(k +1)w , x(k +2)w ,...
x(k +n)w ], to maximize the utility of sparsely available data and
predict a location value for ever missing time step (not just next
step). Thus, the architecture comprises of left and right padded
input layers which are fed to embedding layers to convert sparse
inputs into dense representations. The outputs from the embedding
layers are then input to recurrent units comprising of an array of
LSTM units. The LSTM outputs are then passed through a fully
connected layer and concatenated before being passed through a
fully connected layer to interpret the output and make prediction.
All fully connected layers use rectified linear unit activation except
for output layer which uses softmax activation. The model uses
categorical cross entropy loss function and uses Adam optimizer
to update weights in the network. The model is trained using the
training dataset. 10% of the training set is set aside for validation.
After each epoch, the performance of the model is tested on the
validation dataset. The training is stopped when no improvement
in prediction accuracy of validation data is observed. Though ar-
chitectures can be further augmented with other types of modules
to model further complexities, the comparison here is meant to
evaluate the pure heuristic versus deep learning approaches.
4 RESULTS
4.1 Comparative Performance of Methods
The ILC, RNN, Home-Work and Markov-0 models predicted a lo-
cation value for every missing data point in the dataset (Table 1).
Amongst the remaining methods, the NextPlace algorithm 'filled-in'
SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara
Table 1: Overall prediction accuracy (%) and average percentage of filled timelines(written in {}) for baseline models and Top 1
and Top 3 locations predicted by the intermediate location computing model.
ri
City
NYC ri=1
ri=2
ri=1
ri=2
ri=1
ri=2
DC
SF
Top 1
72.69{100}
64.78{100}
75.08{100}
68.85{100}
77.20{100}
70.78{100}
Top 3
82.35{100}
77.38{100}
83.61{100}
79.57{100}
86.28{100}
82.06{100}
RNN
73.09{100}
59.33{100}
74.58{100}
63.27{100}
76.26{100}
64.78{100}
Home-Work Markov O(0) Markov O(1)
26.39{32.70}
65.54{100}
32.56{48.69}
59.28{100}
27.75{32.29}
66.91{100}
34.13{48.79}
62.35{100}
67.74{100}
16.78{30.12}
19.52{43.72}
63.66{100}
64.65{100}
57.98{100}
65.76{100}
60.64{100}
67.21{100}
62.91{100}
POI
15.59{56.04}
19.11{76.75}
31.27{70.60}
34.56{82.56}
35.49{60.24}
32.69{67.69}
NextPlace
0.17{18.07}
0.21{28.93}
0.11{17.23}
0.19{28.36}
0.15{17.57}
0.22{28.50}
Table 2: Prediction accuracy (%) for Top 1 (T1) and Top 3 (T3)
locations predicted by the ILC model by grid size. (д) repre-
sents a grid size of д × д miles.
ri
City
NYC ri=1
ri=2
ri=1
ri=2
ri=1
ri=2
DC
SF
T1(0.5) T3(0.5) T1(0.1) T3(0.1)
65.64
59.29
67.32
60.19
70.86
63.37
64.07
57.96
64.10
58.23
67.26
59.81
54.23
46.06
54.27
46.85
57.37
48.07
75.71
71.65
77.65
72.59
80.97
75.47
Figure 5: Cumulative fraction of users vs. prediction accu-
racy for ILC and baseline models, ri = 1, (A) and 2 (B).
Figure 6: Prediction accuracy of ILC (no community data)
and RNN with number of users used to train the model for
NYC ri = 1. Values calculated at # of users=5,10,50,100,200 us-
ing mean of 10 replications. At ~50 users RNN performance
comes close to ILC, and by 200 users the RNN model sur-
passes ILC. Accuracy values for training with all available
users are in Table 1.
the least number of missing data points. ILC and RNN outperformed
all baseline models across all cities (Table 1).
For ri = 1, RNN slightly outperformed ILC in only NYC when
considering the overall performance of methods on test data points
(Table 1). When analyzing prediction accuracy per user in the test
set, RNN slightly outperformed ILC (Fig. 5). For ri = 2 ILC out-
performed RNN across all cities both when considering overall
accuracy on test data and accuracy per user. Additionally, for ri = 2,
despite RNN outperforming Home-Work location model when con-
sidering overall prediction accuracy on test data (Table 1), it per-
formed slightly worse than Home-Work location method when
considering accuracy per user (Fig. 5B).
Amongst the heuristic-based baseline models, simpler models
outperformed more complex models. This was mainly because they
were able to predict a location value for larger number of missing
data points. The Order-0 Markov and Home-Work location model
resulted in similar prediction accuracy and outperformed the re-
maining baseline models. In contrast to previous work, the Order-1
Markov model had a lower prediction accuracy as compared to
Order-0 Markov, largely because it was only able to predict a loca-
tion value for one-third of the data points in timelines at ri=1 and
one-half of the data points in timelines at ri=2. The time-dependent
POI recommendation model outperformed Order-1 Markov model
in SF and DC and underperformed in NYC. This is consistent with
the fact that as shown in Fig. 7B, SF and DC had higher similarity
between locations of individuals as compared to NYC. Additionally,
the POI model was able to predict a much larger portion of users'
timelines as compared to the Markov-1 model, yet accuracy values
for both methods were close. The NextPlace method based on a non-
linear spatio-temporal framework had the least predictive power
given the fact that it relies largely on stay duration information.
Given the lack of this information in social media, the model was
scarcely able to predict missing location values.
While the baseline heuristic based methods have been optimized
for different data types, in general, ILC specifically addresses the
challenge of sparse data by incorporating a wide range of com-
ponents. The simpler components help predict a location value
for each missing point, while the more complex components help
identify complex movement behaviours.
Comparing ILC with RNN shows that RNNs are powerful meth-
ods that can out perform traditional heuristic based methods. How-
ever, we see that in low data settings, heuristics can be used to
outperform the deep learning approach (e.g. when predicting at
less frequent time intervals, or when a lower number of users are
available to train the model). This is evident in Fig. 6, where despite
RNN outperforming ILC in NYC at ri = 1, if trained on a fewer
0.000.250.500.751.000.000.250.500.751.00AccuracyFraction of UsersAMethodNextPlaceMarkov−1POI Recomm.Markov−0Home−WorkRNNILC0.000.250.500.751.000.000.250.500.751.00AccuracyFraction of UsersB020406080050100150200Number of usersAccuracy in %ModelRNNILC w/o community Creating Full Individual-level Location Timelines from Sparse Social Media Data SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Figure 7: A: Prediction accuracy (%) vs. number of similar
users (m). Model based on individual and community behav-
ior (black), and community behavior only i.e β = 1 (red).
All values were computed for ri = 2 and using Washington,
DC data. B: Aggregated similarity factors of closest users
u=1 s(u)) vs. fraction of users by city. X-axis normalized by
dividing by the maximum aggregated similarity factor for a
user in the dataset.
(m
number of users, it under performs. Also in Fig. 5 we see that the
RNN requires data from more users to achieve the same accuracy as
ILC, when considering ri=2. However with increases in the amount
of training data, the RNN outperforms ILC. This is due to the fact
that this implementation of ILC only uses a maximum sequence
length of two time steps, RNNs can learn larger and more complex
sequences of locations. Additionally RNNs can also learn longer
sequences of location data of similar users and help improve pre-
diction. Decision between selecting one over the other is based on
the goal of the study and the availability of data. If sparse data for a
large number of users is available, then an RNN approach should be
preferred. But if the goal of the study is to maximize the number of
users for which complete timelines can be constructed by sampling
their locations at less frequent time intervals, or if the number of
available users is low, then a heuristic based method like ILC should
be preferred given that it does not need data to learn patterns.
4.2 Effect of Community Behavior
We found that the effect of community behavior is consistently
higher on an individual's mobility patterns during weekends as
compared to weekdays (βk higher on weekends across all cities and
ri). The average value for βk during weekdays ranged from 0.449
(NYC, ri =1) to 0.466 (DC, ri=2), while during weekends ranged
from 0.456 (NYC, ri =1) to 0.492 (DC, ri=2).
We observe that inclusion of community data helps the perfor-
mance of the method and the main improvement is seen when
the first similar user is accounted for (Fig. 7A). Moreover, after m
= 20, accuracy improvements begin to plateau with more m (the
inclusion of m closest individuals to compute community behavior
will work best for individuals who have high similarity values with
other individuals and are not outliers in terms of their mobility)
justifying the use of m = 50 in our method.
Figure 8: A: Prediction accuracy (%) for Top 1 vs. number of
distinct locations visited for ri = 1 (A) and 2 (B).
uncertainty . Similar trend is observed for RNN as increase in ri
from 1 to 2 hours decreases the overall training data for the model.
Fig. 8 shows prediction accuracy versus the number of distinct
locations grids by individuals, for 1, and 2-hour resolutions; accu-
racy decreases with an increase in the number of distinct location
grids visited by an individual. The fitted line is generated using a
generalized additive model (GAM).
5 CONCLUSION
In this paper, we present a method for predicting missing locations
from an individual's mobility timeline with good accuracy, using
only sparse location tags from social media data. In order to address
the challenge of sparse data, the model uses several heuristics of
human movement and incorporates similar user data. The proposed
approach consistently outperforms baseline heuristic based meth-
ods across data from three major cities, showing stability of the
approach. We also show how ILC fulfills timeline prediction better
than an RNN in sparse data settings, though use of heuristics should
be incorporated into the RNN architecture design in future work
to further advance the approach.
We recognize limitations of this work. Predictions for an individ-
ual can be biased based on their Tweeting patterns (which can be
specific to the types of people who use Twitter), although the incor-
poration of community behavior helps minimize this bias. Second,
even though our work advances previous work by predicting full
timelines for a large number of users, there are still many users for
whom the location cannot be predicted by our model. Hence, despite
the generalizability of the method and the dataset, the methodology
will not be accurate for every single user. Third, here ILC only uses
one location point in the past i.e k − 1 to predict the location at k
due to the sparse nature of the data and prioritization of filling in
the timeline, but we can expand the approach to use the sequence
of n locations in the past to predict the next location, with more
complex considerations. Overall, this research demonstrates a new
approach for the specific problem of filling in location timelines
from sparse social media data, without assuming any information
besides location data is available. The result can be used in many
real-world applications that require location timelines.
4.3 Performance of ILC in Different Settings
We observe that for ILC performance decreases as the interval (ri)
increases from 1 to 2 hours (Table 1), and as the grid size decreases
(Table 2), which is inline with the findings of [10] that at larger
time intervals and smaller grid sizes there is a higher associated
ACKNOWLEDGMENTS
Support for this project was provided in part by a grant from the
National Science Foundation (1737987). We acknowledge Prof. Ju-
liana Freire of New York University and her group for assistance
with data.
020406001020304050Number of similar usersAccuracy in %A0.000.250.500.751.000.000.250.500.751.00Normalized SimilarityFraction of UsersBcityDCNYCSF0.000.250.500.751.0002040Distinct location grids visitedAccuracy in %A0.000.250.500.751.0002040Distinct location grids visitedAccuracy in %B SIGSPATIAL '18, November 6 -- 9, 2018, Seattle, WA, USA
Nabeel Abdur Rehman, Kunal Relia, and Rumi Chunara
REFERENCES
[1] Shunsuke Aoki, Kaoru Sezaki, Nicholas Jing Yuan, and Xing Xie. 2017. An Early
Event Detection Technique with Bus GPS Data. In Proceedings of the 25th ACM
SIGSPATIAL International Conference on Advances in Geographic Information
Systems. ACM, 49.
[2] Akinori Asahara, Kishiko Maruyama, and Ryosuke Shibasaki. 2012. A mixed
autoregressive hidden-markov-chain model applied to people's movements. In
Proceedings of the 20th international conference on advances in geographic infor-
mation systems. ACM, 414 -- 417.
[3] Jie Bao, Yu Zheng, and Mohamed F Mokbel. 2012. Location-based and preference-
aware recommendation using sparse geo-social networking data. In Proceedings
of the 20th international conference on advances in geographic information systems.
ACM, 199 -- 208.
[4] Mitra Baratchi, Nirvana Meratnia, Paul JM Havinga, Andrew K Skidmore, and
Bert AKG Toxopeus. 2014. A hierarchical hidden semi-Markov model for model-
ing mobility data. In Proceedings of the 2014 ACM International Joint Conference
on Pervasive and Ubiquitous Computing. ACM, 401 -- 412.
[5] Francesco Calabrese, Giusy Di Lorenzo, and Carlo Ratti. 2010. Human mobility
prediction based on individual and collective geographical preferences. In Intelli-
gent Transportation Systems (ITSC), 2010 13th International IEEE Conference on.
IEEE, 312 -- 317.
[6] Luca Canzian and Mirco Musolesi. 2015. Trajectories of depression: unobtrusive
monitoring of depressive states by means of smartphone mobility traces analysis.
In Proceedings of the 2015 ACM international joint conference on pervasive and
ubiquitous computing. ACM, 1293 -- 1304.
[7] Eunjoon Cho, Seth A Myers, and Jure Leskovec. 2011. Friendship and mobility:
user movement in location-based social networks. In Proceedings of the 17th ACM
SIGKDD international conference on Knowledge discovery and data mining. ACM,
1082 -- 1090.
[8] Yohan Chon, Hyojeong Shin, Elmurod Talipov, and Hojung Cha. 2012. Evaluating
mobility models for temporal prediction with high-granularity mobility data.
In Pervasive computing and communications (PerCom), 2012 IEEE international
conference on. IEEE, 206 -- 212.
[9] Jacob Cohen. 1968. Weighted kappa: Nominal scale agreement provision for
scaled disagreement or partial credit. Psychological bulletin 70, 4 (1968), 213.
[10] Andrea Cuttone, Sune Lehmann, and Marta C González. 2018. Understanding
Predictability and Exploration in Human Mobility. EPJ Data Science (2018).
[11] Trinh Minh Tri Do and Daniel Gatica-Perez. 2012. Contextual conditional models
for smartphone-based human mobility prediction. In Proceedings of the 2012 ACM
conference on ubiquitous computing. ACM, 163 -- 172.
[12] Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng
Jin. 2018. DeepMove: Predicting Human Mobility with Attentional Recurrent
Networks. In Proceedings of the 2018 World Wide Web Conference on World Wide
Web. International World Wide Web Conferences Steering Committee, 1459 --
1468.
[13] Sébastien Gambs, Marc-Olivier Killijian, and Miguel Núñez del Prado Cortez.
2010. Show me how you move and I will tell you who you are. In Proceedings of
the 3rd ACM SIGSPATIAL International Workshop on Security and Privacy in GIS
and LBS. ACM, 34 -- 41.
[14] Sébastien Gambs, Marc-Olivier Killijian, and Miguel Núñez del Prado Cortez.
2012. Next place prediction using mobility markov chains. In Proceedings of the
First Workshop on Measurement, Privacy, and Mobility. ACM, 3.
[15] M. C. Gonzalez, C. A. Hidalgo, and A. L. Barabasi. 2008. Understanding individual
human mobility patterns. Nature 453, 7196 (2008), 779 -- 82. https://doi.org/10.
1038/nature06958
[16] Samiul Hasan, Christian M Schneider, Satish V Ukkusuri, and Marta C GonzÃąlez.
2013. Spatiotemporal patterns of urban human mobility. Journal of Statistical
Physics 151, 1-2 (2013), 304 -- 318.
[17] Nabil Hossain, Tianran Hu, Roghayeh Feizi, Ann Marie White, Jiebo Luo, and
Henry A Kautz. 2016. Precise Localization of Homes and Activities: Detecting
Drinking-While-Tweeting Patterns in Communities.. In ICWSM. 587 -- 590.
[18] Shenggong Ji, Yu Zheng, and Tianrui Li. 2016. Urban sensing based on human mo-
bility. In Proceedings of the 2016 ACM International Joint Conference on Pervasive
and Ubiquitous Computing. ACM, 1040 -- 1051.
[19] David Jurgens, Tyler Finethy, James McCorriston, Yi Tian Xu, and Derek Ruths.
2015. Geolocation Prediction in Twitter Using Social Networks: A Critical Anal-
ysis and Review of Current Practice. ICWSM 15 (2015), 188 -- 197.
[20] Daniel Kahneman, Alan B Krueger, David A Schkade, Norbert Schwarz, and
Arthur A Stone. 2004. A survey method for characterizing daily life experience:
The day reconstruction method. Science 306, 5702 (2004), 1776 -- 1780.
[21] John Krumm and Eric Horvitz. 2006. Predestination: Inferring destinations from
partial trajectories. UbiComp 2006: Ubiquitous Computing (2006), 243 -- 260.
[22] Yang Li, Dimitrios Gunopulos, Cewu Lu, and Leonidas Guibas. 2017. Urban Travel
Time Prediction using a Small Number of GPS Floating Cars. In Proceedings of
the 25th ACM SIGSPATIAL International Conference on Advances in Geographic
Information Systems. ACM, 3.
[23] L. Liao, D. J. Patterson, D. Fox, and H. Kautz. 2007. Learning and inferring
transportation routines. Artificial Intelligence 171, 5-6 (2007), 311 -- 331. https:
//doi.org/10.1016/j.artint.2007.01.006
[24] Todd Litman. 2003. Measuring transportation: traffic, mobility and accessibility.
Institute of Transportation Engineers. ITE Journal 73, 10 (2003), 28.
[25] Jiajun Liu, Kun Zhao, Saeed Khan, Mark Cameron, and Raja Jurdak. 2015. Multi-
scale population and mobility estimation with geo-tagged tweets. In Data En-
gineering Workshops (ICDEW), 2015 31st IEEE International Conference on. IEEE,
83 -- 86.
[26] Liang Liu, Anyang Hou, Assaf Biderman, Carlo Ratti, and Jun Chen. 2009. Under-
standing individual and collective mobility patterns from smart card records: A
case study in Shenzhen. In Intelligent Transportation Systems, 2009. ITSC'09. 12th
International IEEE Conference On. IEEE, 1 -- 6.
[27] Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2016. Predicting the Next
Location: A Recurrent Model with Spatial and Temporal Contexts.. In AAAI.
194 -- 200.
[28] James McInerney, Jiangchuan Zheng, Alex Rogers, and Nicholas R Jennings. 2013.
Modelling heterogeneous location habits in human populations for location
prediction under data sparsity. In Proceedings of the 2013 ACM international joint
conference on Pervasive and ubiquitous computing. ACM, 469 -- 478.
[29] Anne Vernez Moudon, Chanam Lee, Allen D Cheadle, Cheza Garvin, Donna
Johnson, Thomas L Schmid, Robert D Weathers, and Lin Lin. 2006. Operational
definitions of walkable neighborhood: theoretical and empirical insights. Journal
of Physical Activity & Health 3 (2006), S99.
[30] Anastasios Noulas, Salvatore Scellato, Neal Lathia, and Cecilia Mascolo. 2012.
Mining user mobility features for next place prediction in location-based services.
In Data mining (ICDM), 2012 IEEE 12th international conference on. IEEE, 1038 --
1043.
[31] Andrew Rundle, Kathryn M Neckerman, Lance Freeman, Gina S Lovasi, Marnie
Purciel, James Quinn, Catherine Richards, Neelanjan Sircar, and Christopher
Weiss. 2009. Neighborhood food environment and walkability predict obesity in
New York City. Environmental health perspectives 117, 3 (2009), 442.
[32] Adam Sadilek, Henry Kautz, and Jeffrey P Bigham. 2012. Finding your friends and
following them to where you are. In Proceedings of the fifth ACM international
conference on Web search and data mining. ACM, 723 -- 732.
[33] Adam Sadilek and John Krumm. 2012. Far Out: Predicting Long-Term Human
Motivations and Recent Changes.. In ICWSM. 250 -- 259.
[39] A. Wesolowski, T. Qureshi, M. F. Boni, P. R. Sundsoy, M. A. Johansson, S. B.
Rasheed, K. Engo-Monsen, and C. O. Buckee. 2015. Impact of human mobility on
the emergence of dengue epidemics in Pakistan. Proc Natl Acad Sci U S A 112, 38
(2015), 11887 -- 92. https://doi.org/10.1073/pnas.1504964112
[40] Cheng Yang, Maosong Sun, Wayne Xin Zhao, Zhiyuan Liu, and Edward Y Chang.
2017. A Neural Network Approach to Jointly Modeling Social Networks and
Mobile Trajectories. ACM Transactions on Information Systems (TOIS) 35, 4 (2017),
36.
[41] A. J. Yu and J. D. Cohen. 2008. Sequential effects: Superstition or rational behavior?
Adv Neural Inf Process Syst 21 (2008), 1873 -- 1880. https://www.ncbi.nlm.nih.gov/
pubmed/26412953
[42] Quan Yuan, Gao Cong, Zongyang Ma, Aixin Sun, and Nadia Magnenat Thalmann.
2013. Time-aware point-of-interest recommendation. In Proceedings of the 36th
international ACM SIGIR conference on Research and development in information
retrieval. ACM, 363 -- 372.
[43] Jia-Dong Zhang, Chi-Yin Chow, and Yanhua Li. 2014. Lore: Exploiting sequential
influence for location recommendations. In Proceedings of the 22nd ACM SIGSPA-
TIAL International Conference on Advances in Geographic Information Systems.
ACM, 103 -- 112.
Mobility.. In AAAI.
[34] S. Scellato, M. Musolesi, C. Mascolo, V. Latora, and A. T. Campbell. 2011.
NextPlace: A Spatio-temporal Prediction Framework for Pervasive Systems. Per-
vasive Computing 6696 (2011), 152 -- 169. <GotoISI>://WOS:000305891700010
[35] C. M. Song, Z. H. Qu, N. Blumm, and A. L. Barabasi. 2010. Limits of Predictability
in Human Mobility. Science 327, 5968 (2010), 1018 -- 1021. https://doi.org/10.1126/
science.1177170
[36] L. B. Song, D. Kotz, R. Jain, and X. N. He. 2006. Evaluating next-cell predictors
with extensive Wi-Fi mobility data. Ieee Transactions on Mobile Computing 5, 12
(2006), 1633 -- 1649. <GotoISI>://WOS:000241195600001
[37] Roberto CSNP Souza, Renato M Assunção, Derick M de Oliveira, Denise EF de
Brito, and Wagner Meira Jr. 2016. Infection Hot Spot Mining from Social Media
Trajectories. In Joint European Conference on Machine Learning and Knowledge
Discovery in Databases. Springer, 739 -- 755.
[38] Dan Tasse, Zichen Liu, Alex Sciuto, and Jason I Hong. 2017. State of the Geotags:
|
1606.08927 | 1 | 1606 | 2016-06-29T01:04:29 | Least Cost Influence Maximization Across Multiple Social Networks | [
"cs.SI",
"physics.soc-ph"
] | Recently in Online Social Networks (OSNs), the Least Cost Influence (LCI) problem has become one of the central research topics. It aims at identifying a minimum number of seed users who can trigger a wide cascade of information propagation. Most of existing literature investigated the LCI problem only based on an individual network. However, nowadays users often join several OSNs such that information could be spread across different networks simultaneously. Therefore, in order to obtain the best set of seed users, it is crucial to consider the role of overlapping users under this circumstances.
In this article, we propose a unified framework to represent and analyze the influence diffusion in multiplex networks. More specifically, we tackle the LCI problem by mapping a set of networks into a single one via lossless and lossy coupling schemes. The lossless coupling scheme preserves all properties of original networks to achieve high quality solutions, while the lossy coupling scheme offers an attractive alternative when the running time and memory consumption are of primary concern. Various experiments conducted on both real and synthesized datasets have validated the effectiveness of the coupling schemes, which also provide some interesting insights into the process of influence propagation in multiplex networks. | cs.SI | cs |
Least Cost Influence Maximization Across Multiple Social Networks
1
Huiyuan Zhang, Dung T. Nguyen, Soham Das, Huiling Zhang and My T. Thai
Department of Computer and Information Science and Engineering
University of Florida, Gainesville, Florida 32611
Email: {huiyuan, dtnguyen, sdas, huiling, mythai}@cise.ufl.edu
Abstract
Recently in Online Social Networks (OSNs), the Least Cost Influence (LCI) problem has become
one of the central research topics. It aims at identifying a minimum number of seed users who can
trigger a wide cascade of information propagation. Most of existing literature investigated the LCI
problem only based on an individual network. However, nowadays users often join several OSNs such
that information could be spread across different networks simultaneously. Therefore, in order to obtain
the best set of seed users, it is crucial to consider the role of overlapping users under this circumstances.
In this article, we propose a unified framework to represent and analyze the influence diffusion
in multiplex networks. More specifically, we tackle the LCI problem by mapping a set of networks
into a single one via lossless and lossy coupling schemes. The lossless coupling scheme preserves
all properties of original networks to achieve high quality solutions, while the lossy coupling scheme
offers an attractive alternative when the running time and memory consumption are of primary concern.
Various experiments conducted on both real and synthesized datasets have validated the effectiveness
of the coupling schemes, which also provide some interesting insights into the process of influence
propagation in multiplex networks.
Coupling, multiple networks, influence propagation, online social networks
Index Terms
I. INTRODUCTION
In the recent decade, the popularity of online social networks, such as Facebook, Google+,
Myspace and Twitter etc., has created a new major communication medium and formed a
promising landscape for information sharing and discovery. On average [1], Facebook users
spend 7 hours and 45 minutes per person per month on interacting with their friends ; 3.2 billion
likes and comments are posted every day on Facebook; 340 million tweets are sent out everyday
on Twitter. Such engagement of online users fertilizes the land for information propagation to a
degree which has never been achieved before in the mass media. More importantly, OSNs also
inherit one of the major properties of real social networks – the word-of-mouth effect, in which
personal opinion or decision can be reshaped or reformed through influence from friends and
colleagues. Recently, motivated by the significant effect of viral marketing, OSNs have been the
most attractive platforms to increase brand awareness of new products as well as strengthen the
relationship between customers and companies. In general, the ultimate goal is to find the least
advertising cost set of users which can trigger a massive influence.
Along with the fast development of all existing OSNs, there have been quite a number of
users who maintain several accounts simultaneously, which allow them to propagate information
This article has been published in IEEE/ACM Transactions on Networking, 24(2), 929-939, March 12, 2015, DOI:
10.1109/TNET.2015.2394793. Copyright (c) 2015 IEEE. Personal use of this material is permitted. Permission from IEEE
must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising
or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted
component of this work in other works. H. Zhang and D. T. Dung are co-first authors of this work.
2
(a) Auto post from Facebook to Twitter (b) Auto post from Twitter to Facebook
Fig. 1.
Information propagation across social networks
(c) The number of shared users between
major OSNs in 2009 [2]
across different networks. For example, Jack, a user of both Twitter and Facebook, knew a
new book from Twitter. After reading it, he found it very interesting and shared this book with
friends in Facebook as well as Twitter. This can be done by configuring both of the accounts
to allow automatically posting across different social networks. As a consequence, the product
information is exposed to his friends and further spreads out on both networks. If we only focus
on an individual network, the spread of the information is estimated inaccurately. As shown
in Fig. 1, the fraction of overlapping users is considerable. Therefore considering the influence
only in one network fails to identify the most influential users, which motivates us to study the
problem in multiplex networks where the influence of users is evaluated based on all OSNs in
which they participate.
Related works. Nearly all the existing works studied different variants of the least cost influence
problem on a single network. Kempe et al. [10] first formulated the influence maximization
problem which asks to find a set of k users who can maximize the influence. The influence is
propagated based on a stochastic process called Independent Cascade Model (IC) in which a user
will influence his friends with probability proportional to the strength of their friendship. The
author proved that the problem is NP-hard and proposed a greedy algorithm with approximation
ratio of (1−1/e). After that, a considerable number of works studied and designed new algorithms
for the problem variants on the same or extended models such as [6], [14], [15], [17]. There are
also works on the linear threshold (LT) model for influence propagation in which a user will
adopt the new product when the total influence of his friends surpass some threshold. Dinh et
al [7] proved the inapproximability as well as proposed efficient algorithms for this problem on
a special case of LT model. In their model, the influence between users is uniform and a user
is influenced if a certain fraction ρ of his friends are active.
Recently, researchers have started to explore multiplex networks with works of Yagan et al.
[16] and Liu et al. [11] which studied the connection between offline and online networks. The
first work investigated the outbreak of information using the SIR model on random networks.
The second one analyzed networks formed by online interactions and offline events. The authors
focused on understanding the flow of information and network clustering but not solving the
least cost influence problem. Additionally, these works did not study any specific optimization
problem of viral marketing. Shen et al. [14] explored the information propagation in multiplex
3
online social networks taking into account the interest and engagement of users. The authors
combined all networks into one network by representing an overlapping user as a super node.
This method cannot preserve the individual networks' properties.
In this article, we studies the LCI problem which aims at finding a set of users with minimum
cardinality to influence a certain fraction of users in multiplex networks. Due to the complex
diffusion process in multiplex networks, it is difficult to develop the solution by directly extending
previous solutions in a single network. Additionally, studying the problem in multiplex networks
introduces several new challenges: (1) how to accurately evaluate the influence of overlapping
users; (2) in which network, a user is easier to be influenced; (3) which network propagates the
influence better. To answer above questions, we first introduce a model representation to illustrate
how information propagate in multiplex networks via coupling schemes. By mapping multiple
networks into one network, different coupling schemes can preserve partial or full properties
of the original networks. After that, we can exploit existing solutions on a single network to
solve the problem in multiplex networks. Moreover, through comprehensive experiments, we
have validated the effectiveness of the coupling schemes, and also provide some interesting
insights into the process of influence propagation in multiplex networks. Our main contributions
are summarized as follows:
• Propose a model representation via various coupling schemes to reduce the problem in
multiplex networks to an equivalent problem on a single network. The proposed coupling
schemes can be applied for most popular diffusion models including: linear threshold model,
stochastic threshold model and independent cascading model.
• Provide a scalable greedy algorithm to solve the LCI problem. Especially, the improvement
factor scales up with the size of the network which allows the algorithm to run on very
large networks with millions of nodes.
• Conduct extensive experiments on both real and synthesized datasets. The results show that
considering multiplex networks instead of a single network can effectively choose the most
influential users.
The rest of the paper is organized as follows. In Section II, we present the influence propagation
model in multiplex networks and define the problem. The lossless and lossy coupling schemes are
introduced in Section IV and Section V. A scalable greedy algorithm is proposed in Section VI.
Section VII shows the experimental results on the performance of different algorithms and
coupling schemes. Finally, Section VIII concludes the paper.
A. Graph notations
II. MODEL AND PROBLEM DEFINITION
We consider k networks G1, G2, . . . , Gk, each of which is modeled as a weighted directed
graph Gi = (V i, Ei, θi, W i). The vertex set V i = {u's} represents the participation of ni = V i
users in the network Gi, and the edge set Ei = {(u, v)'s} contains mi = Ei oriented connections
(e.g., friendships or relationships) among network users. W i = {wi(u, v)'s} is the (normalized)
weight function associated to all edges in the ith network. In our model, weight wi(u, v) can also
interpreted as the strength of influence (or the strength of the relationship) a user u has on another
user v in the ith network. The sets of incoming and outgoing neighbors of vertex u in network Gi
are denoted by N i−
u , respectively. In addition, each user u is associated with a threshold
θi(u) indicating the persistence of his opinions. The higher θi(u) is, the more unlikely that u
will be influenced by the opinions of his friends. Furthermore, the users that actively participate
u and N i+
in multiple networks are referred to as overlapping users and can be identified using methods in
4
[3], [9] (Note that identifying overlapping users is not the focus of this paper). Those users are
considered as bridge users for information propagation across networks. Finally, we denote by
G1...k the system consisting of k networks, and by U the exhaustive set of all users U = ∪k
i=1V i.
B. Influence Propagation Model
We first describe the Linear Threshold (LT) model [7], a popular model for studying informa-
tion and influence diffusion in a single network, and then discuss how LT model can be extended
to cope with multiplex networks. In the classic LT model, each node u can be either active or
inactive: u is in an active state if it is selected into the seed set, or the total influence from the
in-degree neighbors exceeds its threshold θ(u), i.e, Pv∈N (u) w(v, u) ≥ θ(u). Otherwise, u is in
an inactive state.
In multiplex network system, given a number of k networks, the information is propagated
separately in each network and can only flows to other networks via the overlapping users. The
information starts to spread out from a set of seed users S i.e. all users in S are active and the
remaining users are inactive. At time t, a user u becomes active if the total influence from its
active neighbors surpasses its threshold in some network i.e. there exists i such that:
X
v∈N i−
u ,v∈A
wi(v, u) ≥ θi(u)
where A is the set of active users after time (t − 1).
In each time step, some of inactive users become activated and try to influence other users
in the next time step. The process terminates until no more inactive users can be activated. If
we limit the propagation time to d, then the process will stop after t = d time steps. The set
of active users in time d is denoted as Ad(G1...k, S). Note that d is also the number of hops
up to which the influence can be propagated from the seed set, so d is called the number of
propagation hops.
C. Problem definition
In this paper, we address the fundamental problem of viral marketing in multiplex networks:
the Least Cost Influence problem. The problem asks to find a seed set of minimum cardinality
which influences a large fraction of users.
Definition 1. (Least Cost Influence (LCI) Problem) Given a system of k networks G1...k with
the set of users U, a positive integer d, and 0 < β ≤ 1, the LCI problem asks to find a seed set
S ⊂ U of minimum cardinality such that the number of active users after d hops according to
LT model is at least β fraction of users i.e. Ad(G1...k, S) ≥ βU.
When k = 1, we have the variant of the problem on a single network which is NP-hard to
solve [5], Dinh et al. [7] proved the inapproximability and proposed an algorithm for a special
case when the influence between users is uniform and a user is activated if a certain fraction ρ of
his friends are active. In the following sections, we will present different coupling strategies to
reduce the problem in multiplex networks to the problem in a single network in order to utilize
the algorithm design.
III. COUPLING SCHEME
5
A coupling scheme is an approach to project multiple networks to a single network, which can
preserve important network information and reproduce the diffusion process from each individual
network. Such a scheme will facilitate researchers to study various optimization problems that
relate to the diffusion of information on multiple networks. In general, we can mitigate these
problems to the one defined on single network and apply existing solutions to solve them. Next
we specify the requirements for such schemes and the general framework.
A. Coupling scheme general framework
Our goal is to map multiple networks into a single network such that a diffusion process on
multiple networks can be simulated by a process on the projected network. Two most important
points are: (1) which user is active and (2) when a user is activated. Formally, a coupling scheme
that maps a system of networks G1...k with the set of users U to a network G = (V, E) needs
to satisfy following requirements:
(1) There exists a set of nodes U ⊆ V and bijection function that maps users to nodes in the
coupled network: F : U → U.
(2) There exists a time mapping function T : N → N.
(3) User u ∈ U is activated at time t on G1...k iff F (u) is activated at time T (t) on G.
The first constraint reserves the identity of users in the coupled network. The second constraint
allows us the know when a user is activated. The last constraint guarantees that the diffusion
process is preserved, i.e., the diffusion of information on the set of user U is the same on the
set of nodes U. This is the core part of the couple scheme and may be difficult to achieve. Since
the main goal is to construct a solution to the studied problem on multiple networks from the
solution on single network, we can relax the last condition such as u ∈ U is activated at time t
on G1...k if F (u) is activated at time T (t) on G. In this case, the diffusion information is not
totally reserved. The coupling scheme is called lossless coupling scheme if the last condition is
satisfied and lossy coupling scheme otherwise.
Since our main concern is the diffusion of information among users, such coupling scheme
reserve most of the properties of the diffusion process. It helps to answer following questions:
• When a node becomes active?
• How many nodes are activated at a specific time?
• Who are top influencers in the multiple networks?
Another important aspect of the coupling scheme is the activation state of nodes in V \ U.
In some optimization problems, the fraction of active nodes plays an important role. Thus, it is
desirable for the coupling scheme to reserve the fraction of active nodes or the scale-up property.
Definition 2 (Scale-up Property). A coupling scheme is said to have scale-up property if there
exists a constant c = c(G1...k) such that there is cK active nodes on G iff there is K active
users on G1...k.
B. General framework to solve some optimization problems
With the coupling scheme, if we only consider the set of users and its mapped set on the
coupled network, the diffusion process is the same on these two sets. Thus, we can design algo-
rithms to solve various information diffusion optimization problems on multiple networks such
as Influence Maximization problem [10], Limiting the misinformation problem [4], Minimum
Influential Node Selection problem [18], etc, by the following framework: (1) Create a coupled
network following a coupling scheme, (2) Use an algorithm for the studied problem on single
6
networks to identify the set of selected nodes, (3) Use the F function to determine the set of
selected users from the set of selected nodes on the coupled network.
Algorithm 1 General Framework
Input: A set of users U, a system of networks G1...k, and an algorithm A.
Output: A solution S ⊂ U
G ← The coupled network of G1...k
C ← Set of selected users provided by A on G
S ← F (C)
Return I
IV. LOSSLESS COUPLING SCHEMES
In this section, we present the lossless scheme to couple multiple networks into a new single
network with respect to the influence diffusion process on each participant network. A notable
advantage of this newly coupled graph is that we can use any existing algorithm on a single
network to produce the solution in multiplex networks with the same quality.
A. Clique lossless coupling scheme
1, ui
p to ui
2, . . . , ui
In LT-model, the first issue is solved by introducing dummy nodes for each user u in networks
that it does not belong to. These dummy nodes are isolated. Now the vertex set V i of ith network
can be represented by V i = {ui
n} where U = {u1, u2, . . . , un} is the set of all users.
p is called the representative vertex of up in network Gi. In the new representation, there is
ui
an edge from ui
q if up and uq are connected in Gi. Now we can union all k networks
to form a new network G. The approach to overcome the second challenge is to allow nodes
u1, u2, . . . , uk of a user u to influence each other e.g. adding edge (ui, uj) with weight θ(uj).
When ui is influenced, uj is also influenced in the next time step as they are actually a single
overlapping user u, thus the information is transferred from network Gi to Gj. But an emerged
problem is that the information is delayed when it is transferred between two networks. Right
after being activated, ui will influence its neighbors while uj needs one more time step before
it starts to influence its neighbors. It would be better if both ui and uj start to influence their
neighbors in the same time. For this reason, new gateway vertex u0 is added to G such that both
ui and uj can only influence other vertices through u0. In particular, all edges (ui, vi) ((uj, zj))
will be replaced by edges (u0, vi) ((u0, zj)). In addition, more edges are added between u0, ui,
and uj to let them influence each other, since the connection between gateway and representative
vertices of the same user forms a clique, so we call it clique lossless coupling scheme. After
forming the topology of the coupled network, we assign edge weights and vertex thresholds as
following:
Vertex thresholds. All dummy vertices and gateway vertices have the threshold of 1. Any
remaining representative vertex ui
p has the same threshold as up in Gi, i.e., θ(ui
p) = θi(up).
Edge weights. If there is an edge between user u and v in Gi, then the edge (u0, vi) has weight
w(u0, vi) = wi(u, v). The edges between gateway and representative vertices of the same user
u are assigned as w(ui, uj) = θ(uj), ∀ 0 ≤ i, j ≤ k, i 6= j to synchronize their state together.
A simple example of the clique lossless coupling scheme is illustrated in Fig. 2.
7
0.4
0.2
0.5
0.1
0.2
0.7
0.3
0.2
0.6
0.5
0.6
0.5
0.1
0.5
0.1
0.3
0.4
0.2
0.9
0.5
G1
G2
G3
Facebook
Twitter
Foursquare
(a) An instance of multiplex networks with 4 users. Each user is
represented by vertices of the same color with different thresholds in
different networks e.g. green user has thresholds of 0.3 and 0.2 in G2
and G3.
!"#$!%&vertex
'#()#*#+"!",-#&vertex
Dummy ertex
1
1
0.3
0.6
0.5
G2
Twitter
0.6
1
0.4
1
0.7
0.1
0.5
0.2
1
0.2
G1
Facebook
0.5
1
0.1
0.5
0.2
1
1
1
0.9
0.3
0.4
1
G3
0.1
0.2
0.5
Foursquare
0.6
0.7
0.6
0.6
1
0.7
1
0.7
1
0.5
0.7
0.6
0.5
0.5
1
0.5
(b) The influence of gateway vertices on represenative vertices
represent the influence between users in multiplex networks.
(c) The connection between gateway and repre-
sentative vertices of the same user.
Fig. 2. An example of the clique lossless coupling scheme.
Next we will show that the propagation process in the original multiplex networks and the
coupled network is actually the same. Influence is alternatively propagated between gateway and
representative vertices, so the problem with d hops in the multiplex networks is equivalent to
the problem with 2d hops in the coupled network.
Lemma 1. Suppose that the propagation process in the coupled network G starts from the
seed set which contains only gateway vertices S = {s0
p}, then representative vertices are
activated only at even propagation hops.
1, . . . , s0
Proof. Suppose that a gateway vertex u0 is the first gateway vertex that is activated at the odd
hops 2d + 1. u0 must be activated by some vertex ui and ui is the first activated vertex among
vertices u1, u2, . . . , uk. It means that ui is activated in hop 2d. Since all incoming neighbors of
ui are gateway vertices, some gateway vertex becomes active in hop 2d − 1 (contradiction).
Lemma 2. Suppose that the propagation process on G1...k and G starts from the same seed set
S, then following conditions are equivalent:
(1) User u is active after d propagation hops in G1...k.
(2) There exists i such that ui is active after 2d − 1 propagation hops in G.
(3) Vertex u0 is active after 2d propagation hops in G.
Proof. We will prove this lemma by induction. Suppose it is correct for any 1 ≤ d ≤ t, we need
8
to prove it is correct for d = t + 1. Denote A1...k(t) and A(t) as the set of active users and active
vertices after t propagation hops in G1...k and G, respectively.
(1) ⇒ (2): If user u is active at time t + 1 in G1...k, it must be activated in some network Gj.
We have:
X
v∈N j−
u ∩A1...k(t)
wj(v, u) ≥ θj(u)
Due to the induction assumption, for each v ∈ A1...k(t), we also have v0 ∈ A(2t) in G. Thus:
X
uj ∩A(2t)
v0∈N −
w(v0, uj) = X
u ∩A1...k(t)
v∈N j−
= θ(uj)
wj(v, u) ≥ θj(u)
It means that uj is active after (2(t + 1) − 1) propagation hops.
(2) ⇒ (3): If there exists i such that ui is active after 2(t + 1) − 1 propagation hops on G,
then ui will activate u0 in hop 2(t + 1)
(3) ⇒ (1): Suppose that u0 /∈ S is active after 2(t + 1) propagation hops in G, then there
exists uj which activates u0 before. This is equivalent to:
X
uj ,v∈A(2t)
v∈N −
w(v, uj) ≥ θ(uj)
For each v ∈ A(2t), we also have v ∈ A1...k(t). Replace this into the above inequality we
have:
X
v∈N j−
u ∩A1...k(t)
wj(v, u) = X
v0∈N −
uj ∩A(2t)
w(v0, uj)
Thus, u is active in network Gj after t + 1 hops.
Next, we will show that the number of influenced vertices in the coupled network is (k + 1)
≥ θ(uj) = θj(u)
times the number of influenced users in multiplex networks as stated in Theorem 1.
Theorem 1. Given a system of k networks G1...k with the user set U, the coupled network G
produced by the lossless coupling scheme, and a seed set S = {s1, s2, . . . , sp}, if Ad(G1...k, S)
= {a1, a2, . . . , aq} is the set of active users caused by S after d propagation hops in multiplex
networks, then A2d(G, S) = {a0
q } is the set of active vertices caused
by S after 2d propagation hops in the coupled network.
Proof. For each user ai ∈ Ad(G1...k, S) i.e. ai is active after d hops in G1...k, then there exists aj
which is active after 2d − 1 hops in G according to the Lemma 2. As a result, all a0
i , . . . , ak
are active after 2d hops. So B = {a0
1, . . . , ak
Let consider a vertex of A2d(G, S) which is:
Case 1. A gateway vertex u0 which is active after 2d hops in G, so vertex u must be active
q } ⊆ A2d(G, S).
1, . . ., a0
q, a1
1, a1
1, . . . , ak
q, . . . , ak
1, . . ., a0
q, a1
q, . . . , ak
1, a1
i , a1
i
i
after d hops in G1...k. This implies u ∈ Ad(G1...k, S), thus u0 ∈ B.
Case 2. An representative vertex ui. If ui is active after 2d − 1 hops, then u must be active
after d hops due to Lemma 2, thus u ∈ Ad(G1...k, S). Otherwise, ui is activated at hop 2d , it
must be activated by some vertex uj, j > 0 since all gateway vertices only change their state at
even hops. Again, u ∈ Ad(G1...k, S). This results in ui ∈ B.
9
From two above cases, we also have A2d(G, S) ⊆ B. So that A2d(G, S) = B, the proof is
completed.
Theorem 1 provides the basis to derive the solution for LCI in multiplex networks from the
solution on a single network. It implies an important algorithmic property of the lossless coupling
scheme regarding the relationship between the solutions of LCI in G1...k and G. The equivalence
of two solutions is stated below:
Theorem 2. When the lossless scheme is used, the set S = {s1, s2, . . . , sp} influences β fraction
of users in G1...k after d propagation hops if and only if S′ = {s0
p} influences β
fraction of vertices in coupled network G after 2d propagation hops.
2, . . . , s0
1, s0
Size of the coupled network. Each user u has k + 1 corresponding vertices u0, u1, . . . , uk in
the coupled network, thus the number of vertices is V = (k + 1)U = (k + 1)n. The number of
edges equals the total number of edges from all input networks plus the number of new edges
for synchronizing. Thus the total number of edges is E = Pk
B. Star lossless coupling scheme
i=1 Ei + nk(k + 1).
In last subsection, we discussed the clique lossless coupling scheme, however, in this scheme,
the number of edges to synchronize the state of vertices u0, u1, . . . , uk is added up to k(k + 1)
for each user u, which results in nk(k + 1) extra edges in the coupled network. In real networks,
the number of edges is often linear to the number of vertices, while the number of extra edges
greatly increases the size of the coupled network, especially when k is large. Therefore, we
would like to design another synchronization strategy that has less additional edges.
0.7
0.6
0.5
0.7
0.5
0.6
1
1
1
1
1
1
1
Fig. 3. Synchronization of star lossless coupling scheme.
Note that the large volume of extra edges is due to the direct synchronization between each pair
of representative vertices of u in clique lossless coupling scheme, so we can reduce it by using
indirect synchronization. In the new coupling scheme, we create one intermediate vertex uk+1
with threshold θ(uk+1) = 1 and let the active state propagate from any vertex in u1, u2, . . . , uk via
this vertex. Specifically, the synchronization edges are established as follows: w(ui, uk+1) = 1
and w(uk+1, ui) = θ(ui) 1 ≤ i ≤ k; w(uk+1, u0) = w(u0, uk+1) = 1. The synchronization
strategy of star lossless coupling scheme is illustrated in Fig. 3. Now, the number of extra edge
for each user is 2(k +1) and the size of the coupled network is reduced as shown in the following
proposition.
i=1 Ei + 2n(k + 1) edges.
Proposition 1. When star lossless scheme is used, the coupled network has V = (k + 2)U =
10
(k + 2)n vertices and E = Pk
In star lossless coupling scheme, it takes 2 hops to synchronize the states of representative
vertices of each user which leads to delaying the propagation of influence in the coupled network.
Due to the similarity between star lossless scheme and clique lossless scheme, we state the
following property of star lossless scheme without proof.
Theorem 3. When star lossless coupling scheme is used, the set S = {s1, s2, . . . , sp} influences β
fraction of users in G1...k after d propagation hops if and only if S′ = {s0
p} influences
β fraction of vertices in coupled network G after 3d propagation hops.
2, . . . , s0
1, s0
C. Reduced lossless schemes
In all above coupling schemes, we create representative vertices in all networks in G1...k to
guarantee that the number of influenced vertices in the coupled network is scaled up from the
number of influenced users in the original system of networks. This creates an extraordinary
redundant vertices. For example, we have n users and 4 networks with 0.8n, 0.6n, 0.3n, 0.2n
users, then the total number of vertices in all network is only 1.9n while the number of vertices
created in clique lossless scheme and star lossless scheme are 5n and 6n, respectively. The
redundant ratios of these two schemes are 260 % and 315 %. To overcome this redundant, we
use assign weight for vertices in the coupled network and guarantee that the total weight of
active vertices is scaled from the number of active users in the original system. In particular,
we only create representative vertices ui1, ui2, . . . , uip for user u where Gi1, Gi2, . . . , Gip are
networks that u joins in. Each representative vertex is assigned weight 1, and the user vertex is
assigned weight k − p. We have reduced clique lossless scheme or reduced star lossless scheme
corresponding the method to synchronize the state of user and representative vertices is clique
or star type. With this modification, the number of extra vertices is only n and 2n when clique
and star synchronization is used, respectively. The number of extra edges now depends on the
participants of users.
Proposition 2. When reduced clique lossless scheme or reduced star lossless scheme is used,
the coupled network has V = Pk
The relation between the a set of active vertices in coupled network and the set of active users
in original networks is similar to previous schemes. We state this relation without proof below.
Theorem 4. When reduced clique lossless scheme (reduced star lossless scheme) is used, the
set S = {s1, s2, . . . , sp} influences β fraction of users in G1...k after d propagation hops if and
only if the total weight of active vertices caused by S′ = {s0
p} after 2d (3d) hops in
coupled network G is β fraction of the total weight of all vertices.
i=1 V i + n or V = Pk
i=1 V i + 2n vertices, respectively.
2, . . . , s0
1, s0
D. Extensions to other diffusion models
In this section, we show that we can design lossless coupling schemes for some other well-
known diffusion models in each component network. As a result, top influential users can be
identified under these diffusion models. In particular, we investigate two most popular stochastic
diffusion models which are Stochastic Threshold and Independent Cascading models [10].
• Stochastic Threshold model. This model is similar to the Linear Threshold model but the
threshold θi(ui) of each node ui of Gi is a random value in the range [0, Θi(ui)]. Node ui
will be influenced when Pvi∈N −
ui ,v∈A wi(vi, ui) ≥ θi(ui)
• Independent Cascading model. In this model, there are only edge weights representing the
11
influence between users. Once node ui of Gi is influenced, it has a single chance to influence
its neighbor vi ∈ N +(ui) with probability wi(ui, vi).
For both models, we use the same approach of using gateway vertices, representative vertices
and the synchronization edges between gateway vertices and their representative vertices. The
weight of edge (ui, uj), 0 ≤ i 6= j ≤ k will be Θ(uj) for Stochastic Threshold model and 1 for
Independent Cascading model. Once ui is influenced, uj will be influenced with probability 1
in the next time step. The proof for the equivalence of the coupling scheme is similar to ones
for LT-model.
V. LOSSY COUPLING SCHEMES
In the preceding coupling scheme for LT-model, a complicated coupled network is produced
with large numbers of auxiliary vertices and edges. It is ideal to have a coupled network which
only contain users as vertices. This network provides a compact view of the relationship between
users crossing the whole system of networks. The loss of the information is unavoidable when
we try to represent the information of multiplex networks in a compact single network. The
goal is to design a scheme that minimizes the loss as much as possible i.e. the solution for the
problem in the coupled network is very close to one in the original system. Next, we present
these schemes based on the following key observations.
Observation 1. User u will be activated if there exists i such that: Pv∈N i−
u ∩A wi(v, u) ≥ θi(u)
where A is the set of active users. We can relax the condition to activate u with positive
parameters α1(u), α2(u), . . ., αk(u) as follows:
k
X
i=1
(cid:16)αi(u) X
v∈N i−
u ∩A
wi(v, u)(cid:17) ≥
k
X
i=1
αi(u)θi(u)
(1)
u ∩A wi(v, u) − θi(u). In the special case, if Pv∈N i−
Note that sometimes the condition to activate u is met, but the condition (1) still needs more
influence from u's friends to satisfy. The more this need for extra influence is, the looser condition
(1) is. We can reduce this redundancy by increasing the value of αi(u) proportional to the value
of Pv∈N i−
u ∩A wi(v, u) > θi(u) and we choose
αi(u) ≫ αj(u), ∀j 6= i, then there is no redundancy. Unfortunately, we do not know before
hand in which network user u will be activated, so we can only choose parameters heuristically.
Observation 2. When user u participates in multiple networks, it is easier to influence u in
some network than the others. The following simple case illustrate such situation. Suppose that
we have two networks. In network 1, θ1(u) = 0.1 and u has 8 in-neighbors, each neighbor v
influences u with w1(v, u) = 0.1. In network 2, θ2(u) = 0.7 and u has 8 in-neighbors, each
neighbor v influences u with w2(v, u) = 0.1. The number of active neighbors to activate u is 1
and 7 in network 1 and 2, respectively.
Easiness. Intuitively, we can say that u is easier to be influenced in the first network. We
quantify the easiness ǫi(u) that u is influenced in network i as the ratio between the total
influence from friends and the threshold to be influenced: ǫi(u) =
. We can use the
easiness of a user in networks as the parameters of the condition 1.
i−
u
θi(u)
wi(v,u)
P
v∈N
Based on above observations, we couple multiplex networks into one using parameters {αi(u)}.
The vertex set is the set of users V = {u1, u2, . . . , un}. The threshold of vertex u is set to
θ(u) = Pk
i=1 αi(u)θi(u)
12
0.1
1/10
26/25
9/20
5/14
3/5
2.4
2/15
117/50
0.5
3/20
1/2
3/10
0.8
Fig. 4. Lossy coupled network using easiness parameters.
is no edge from v to u in ith network.
The weight of the edge (v, u) is: w(v, u) = Pk
Then the set of edges is E = {(v, u)w(v, u) > 0}. Fig. 4 illustrates the loosy coupled network
i=1 αi(u)wi(v, u) where wi(v, u) = 0 if there
of networks in Fig. 2.
Besides easiness, other metrics can be used for the same purpose. We enumerate here some
other metrics.
Involvement. If a user is surrounded by a group of friends who have high influence on each
other, he tends to be influenced. When a few of his friends are influenced, the whole group
involving him is likely to be influenced. We estimate involvement of a node v in a network Gi
by measuring how strongly the 1-hop neighborhood v is connected and to what extent influence
can propagate from one node to another in the 1-hop neighborhood. Formally we can define
involvement of a node v in network Gi as: σi
is
v∪{v}
the set of all neighbors of v (both in-coming and out-going).
v = Px,y∈N i
Average. All parameters have the same value αi(u) = 1
Next we show the relationship between the solution for LCI in the lossy coupled network
and the original system of networks. As discussed in the above observations, if the propagation
process starts from the same set of users in G1...k and the coupled network G, then the active
state of a user in G implies its active state in G1...k. It means that if the set of users S activates
β fraction of users in G, it also activates at least β fraction of users in G1...k. It implies that if
a seed set is a feasible solution in G, it is also a feasible solution in G1...k. Thus we have the
following result.
Theorem 5. When the lossy coupling scheme is used, if the set of users S activates β fraction
of users in G, then it activates at least β fraction of users in G1...k.
v ∪ N i−
wi(x,y)
θi
y
where N i
v = N i+
v
VI. ALGORITHMS
In this section, we describe a greedy algorithm and its improvement in terms of scalability in
large networks. In the state of the art work, Dinh et al. [7] only solved the problem in a special
case where the threshold is uniform and the required fraction of active nodes is the same for all
nodes.
13
Algorithm 2 Improved Greedy
Input: A system of networks G1...k, fraction β, T , R.
Output: A small seeding set S
G ← The coupled network of G1...k
C ← Set of user vertices
I ← ∅, Counter ← 0
Initialize a heap: H ← ∅
for u ∈ C do
H.push((u, f∅(u)))
end for
while Number of active vertices ≤ βV do
Counter ← Counter + 1
if Counter % R == 0 then
Update key values of all elements in H
else
A ← ∅
for i = 1 to T do
(u, f (u)) ← H.extract−max()
end for
for u ∈ A do
H.push((u, fI(u)))
end for
end if
(u, f (u)) ← H.extract−max()
I ← I ∪ {u}
end while
S ← corresponded users G1...k of nodes in I
Return S
A. Improved greedy algorithm
The bottle neck of the native greedy algorithm is to identify the best node to be selected
in each iteration, thus we focus on reducing the evaluating the computational cost of this step
while maintaining the same quality of selected nodes. We notice that the marginal gain function
fI(·) is recomputed for all unselected nodes and it does not change much after a single iteration.
Since we only select the best one, which indicates that nodes with higher marginal gain in
previous round are necessary to be reevaluated. Therefore, we use a max heap to store the
marginal gain and extra the top one to reevaluate it, if it is not of largest marginal value, it will
be pushed back. Since the time to extract/push an element to the heap is O(log n), the total
computation cost for each iteration is O(T (m + n) + T log n) = O(T (m + n)). Normally, T
is much smaller than n, so the running time is improved significantly. In addition, due to the
property of the Linear Threshold model, the required influence (remained threshold value after
subtracting the influence of activate neighbors) to activate a node is decreasing when the seed
set grows up. When a large number of nodes is selected, there are many nodes are very easy
to be activated. Thus the marginal gain of a node can accumulate to a large value. If we just
apply the proposed strategy, we will never evaluate these nodes again. Therefore, we need to
do exhaustively reevaluation periodically. Combine two strategies together, we present the idea
14
of using heavy and light iterations alternatively. In heavy iteration, we will update fI(·) of all
unselected nodes while only top T nodes are reevaluated in light iteration. Since we do not
want too many heavy iterations, we only use one such iteration after every R iterations. With
this implementation, the running time is reduced much while the quality of the solution only
fluctuates infinitesimal. The improved Greedy is described in Algorithm 2.
This algorithm will terminate when the number of influenced users is larger then the required
fraction of total users. The complexity of this algorithm is O((m + n) · nd) in the worst case
scenario, however, after applying the above discussed update techniques, the running time can
be improved up to 700 times faster than the native greedy from experimental results.
B. Comparing to Optimal Seeding
In this section, we evaluate the performance of proposed algorithm with different coupling
schemes to the optimal solution. We formulate the LCI problem to a 0-1 Integer Linear Pro-
gramming (ILP) problem as follows.
x0
v
v∈V
minimizeX
subject toX
X
v∈V
xd
v ≥ βV
xi−1
u wuv + θv · xi−1
v ≥ θv · xi
v
w∈N (v)
v ≥ xi−1
xi
xi
v ∈ {0, 1}
v
∀v ∈ V, i = 1..d
∀v ∈ V, i = 1..d
∀v ∈ V, i = 0..d
where xi
v = 1 if v is active in round i, otherwise, xi
v = 0.
Since solving IP is NP-hard, we can not run the IP on large networks. Moreover, we have to
run the IP on the coupled network with clique lossless coupling scheme, where there will be
additional nodes and edges created, and the propagation hops need to be doubled (Theorem 2).
Therefore, to evaluate the performance of our algorithm, we compare the result with small size
synthesize networks. For generating networks, 50 nodes are randomly chosen from a 100 users
base, and the probability of connecting each pair of nodes is p = 0.04, which yields a coupled
network of 300 users and with an expected average degree 2. Fig.4(a) shows the obtained seed
size with influence fraction from 0.2 to 0.8 under all coupling schemes. And we also evaluate
the impact of setting different propagation hops on seed size in Fig.4(b) with influence fraction
β = 0.4.
The optimal seeding along with the results of the improved greedy are shown in Fig.4. As
can be seen in Fig. 4(a), the seeding sizes obtained from the proposed algorithm are close to the
optimal solution while varying the influenced fraction β. The same phenomenon is also shown
by varying the number of propagation hops in Fig.4(b). Especially, when the number of hops is
relatively larger, the result is only one or two more than the optimal solution.
(cid:1)(cid:2)(cid:3)
(cid:4)(cid:5)(cid:6)(cid:6)(cid:7)(cid:8)(cid:6)(cid:6)
(cid:9)(cid:10)(cid:8)(cid:11)(cid:12)(cid:13)(cid:8)
(cid:14)(cid:15)(cid:10)(cid:5)(cid:7)(cid:10)(cid:8)(cid:16)(cid:8)(cid:15)(cid:17)
(cid:18)(cid:12)(cid:6)(cid:19)(cid:15)(cid:8)(cid:6)(cid:6)
(cid:2)
(cid:6)
(cid:5)
(cid:1)
(cid:4)
(cid:3)
(cid:2)
(cid:2)
(cid:1)
(cid:1)(cid:6)(cid:2)
(cid:1)(cid:5)(cid:3)
(cid:1)(cid:5)(cid:2)
(cid:1)(cid:4)(cid:3)
(cid:1)(cid:4)(cid:2)
(cid:1)(cid:3)
(cid:1)(cid:2)
(cid:1)(cid:2)(cid:7)(cid:5)
(cid:1)(cid:2)(cid:7)(cid:8)
(cid:1)(cid:2)(cid:7)(cid:9)
(cid:1)(cid:2)(cid:7)(cid:10)
(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:2)(cid:8)(cid:12)(cid:2)(cid:4)(cid:13)(cid:14)(cid:15)(cid:12)(cid:16)(cid:5)(cid:17)(cid:8)
(a)
Fig. 5. Seed size on synthesize network.
(cid:1)(cid:5)(cid:2)
(cid:1)(cid:4)(cid:3)
(cid:1)(cid:4)(cid:2)
(cid:1)(cid:3)
(cid:2)
(cid:6)
(cid:5)
(cid:1)
(cid:4)
(cid:3)
(cid:2)
(cid:2)
(cid:1)
(cid:1)(cid:2)
(cid:1)(cid:4)
(cid:1)(cid:2)(cid:3)
(cid:4)(cid:5)(cid:6)(cid:6)(cid:7)(cid:8)(cid:6)(cid:6)
(cid:9)(cid:10)(cid:8)(cid:11)(cid:12)(cid:13)(cid:8)
(cid:14)(cid:15)(cid:10)(cid:5)(cid:7)(cid:10)(cid:8)(cid:16)(cid:8)(cid:15)(cid:17)
(cid:18)(cid:12)(cid:6)(cid:19)(cid:15)(cid:8)(cid:6)(cid:6)
15
(cid:1)(cid:5)
(cid:1)(cid:6)
(cid:1)(cid:7)
(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:11)(cid:13)(cid:5)(cid:9)(cid:14)(cid:4)(cid:15)(cid:9)(cid:10)
(b)
VII. EXPERIMENTS
In this section, we show the experimental results to compare the proposed coupling schemes
and utilize these coupling schemes to analyze the influence diffusion in multiplex networks.
First, we compare lossless and lossy coupling schemes to measure the trade-off between the
running time and the quality of solutions. In particular, for those different kinds of lossless
coupling methods, all of them can preserve complete information of all networks. As a result,
the quality of seeds are the same, the only difference would be the running time which have
been theoretically proved in Proposition 1 and Proposition 2. Therefore, we only chose the clique
coupling scheme to be evaluated. Second, we investigate the relationship between networks in the
information diffusion to address the following questions: (1) What is the role of overlapping users
in diffusing the information? (2) What do we miss when considering each network separately?
(3) How and to what extent does the diffusion on one network provide a burst of information
in other networks?
A. Datasets
Real networks. We perform experiments on two datasets:
• Foursquare (FSQ) and Twitter networks [14]
• Co-author networks in the area of Condensed Matter(CM) [13], High-Energy Theory(Het)
[13], and Network Science (NetS) [12].
The statistics of those networks are described in Table I. The number of overlapping users
in the first dataset FSQ-Twitter is 4100 [14]. We examine the in and out-degree of overlapping
nodes. For the second dataset, we match overlapping users based on authors' names. The numbers
of overlapping users of the network pairs CM-Het, CM-NetS, and Het-NetS are 2860, 517, and
90, respectively. While the edge weights are provided for co-author networks, only the topology
is available for Twitter and Foursquare networks. Thus to assign the weight of each edge, we
adopt the method in [10], where each edge weight is randomly picked from 0 to 1 and then
normalize it so that the total weight of in-coming edges is sum up to 1 for each node. This is
suitable since the influence of user u on user v tends to be small if v is under the influence of
many friends. Finally, we also adopt the assignment of threshold in [10] where all thresholds
are randomly chosen from 0 to 1.
Synthesized networks. We also use synthesized networks generated by Erdos-Renyi random
network model [8] to test on networks with controlled parameters. There are two networks with
10000 nodes which are formed by randomly connecting each pair of nodes with probability
p1 = 0.0008 and p2 = 0.006. The average degrees, 8 and 60, reflect the diversity of network
TABLE I
DATASETS DESCRIPTION
16
Networks
Twitter
FSQ
CM
Het
NetS
#Nodes
48277
44992
40420
8360
1588
#Edges
16304712
1664402
175692
15751
2742
Avg. Degree
289.7
35.99
8.69
1.88
1.73
densities in reality. Then, we select randomly f fraction of nodes in two networks as overlapping
nodes. We shall refer to 0 ≤ f ≤ 1 as the overlapping fraction. The edge weights and node
thresholds are assigned as Twitter.
Setup. We ran all our experiments on a desktop with an Intel(R) Xeon(R) W350 CPU and 12
GB RAM. The number of hops is d = 4 and the influenced fraction β = 0.8, unless otherwise
mentioned.
B. Comparison of coupling schemes
We evaluate the impact of the coupling schemes on the running time and the solution quality
of the greedy algorithm to solve the LCI problem.
Solution quality. As shown in Figs. 6(a) and 6(b), the greedy algorithm provides larger seed
sets but runs faster in lossy coupled networks than lossless coupled networks. In both Twitter-
FSQ and the co-author networks, the seed size is smallest when the lossless coupling scheme is
used. It is as expected since the lossless coupling scheme reserves all the influence information
which is exploited later to solve LCI. However, the seed sizes are only a bit larger using the lossy
coupling schemes. In the lossy coupling schemes, the information is only lost at overlapping
users which occupy a small fraction the total number of users (roughly 5% in FSQ-Twitter and
7% in co-author networks). Thus, the impact of the lossy coupling schemes on the solution
quality is small especially when seed sets are large to influence a large fraction of users.
A closer examination reveals the relative effectiveness of the coupling methods on the seed
size. That is when the seed size is significantly small, the lossess coupling outperforms all the
lossy methods. For example, when the overlapping fraction f = 0.8, the solution using the
lossless coupling is roughly 55% of that in the solution using the (lossy) Easiness, and the
solution using Easiness is about 15% smaller than the other two lossy methods (Fig. 7(a)).
Running time. The greedy algorithm runs much faster in the lossy coupled networks than
the lossless ones in general. As shown in Figs. 6(c) and 6(d), using the lossy coupling reduces
the running times by a factor of 2 in FSQ-Twitter and a factor 4 in the co-author networks
compared to the lossless coupling. The major disadvantages of the lossless coupling scheme are
the redundant nodes and edges. Therefore, the lossy coupling schemes works better on networks
that are sparse and the number of overlapping users is small.
However, in some other cases, the lossless coupling scheme is more efficient. As shown in Fig.
7(b), the running time in the lossless coupled networks is larger in the beginning, but gradually
reduces down and beats other methods at f = 0.4. The larger f is, the larger the ratio between
seed size in the lossless and lossy coupled networks is. As the running time depends on the seed
size, thus it reduces faster in the lossless coupled network with larger overlapping ratio.
Overall, the lossless coupling scheme returns solution with higher quality, especially when
the seed set is small. However, if the constraint of running time and the memory are of priority,
the lossy Easiness coupling scheme offers an attractive alternative.
Average
Easiness
Involvement
Lossless
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(a) Seed size - co-author networks
Average
Easiness
Involvement
Lossless
4000
3500
3000
2500
2000
1500
1000
500
0
4500
4000
3500
3000
2500
2000
1500
1000
500
0
e
z
i
S
d
e
e
S
s
n
i
i
e
m
T
g
n
n
n
u
R
i
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(c) Running time - co-author networks
Average
Easiness
Involvement
Lossless
17
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(b) Seed size - FSQ-Twitter
Average
Easiness
Involvement
Lossless
1600
1400
1200
1000
800
600
400
200
0
0
40000
35000
30000
25000
20000
15000
10000
5000
0
e
z
i
S
d
e
e
S
s
n
i
i
e
m
T
g
n
n
n
u
R
i
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(d) Running time - FSQ-Twitter
Fig. 6.
Impact of coupling schemes on finding the minimum seed set
C. Advantages of using coupled networks.
To understand the benefit of taking consideration of overlapping users and coupled network,
in this part, we are going to compare the seed size with/out using coupled network. In particular,
we do two comparisons on: 1) influencing a fraction β of the nodes in all networks by selecting
seeds from each network and taking the union to compare with seeds achieved from lossless
coupling scheme; 2) influencing a fraction β of the nodes in a particular network by only
choosing seeds from that network compare to the seeds obtained from lossless coupling scheme.
e
z
i
S
d
e
e
S
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0.1
Average
Easiness
Involvement
Lossless
s
n
i
i
e
m
T
g
n
n
n
u
R
i
Average
Easiness
Involvement
Lossless
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
0.3
0.8
0.2
Overlapping fraction (f)
0.4
0.5
0.6
0.7
0.9
1
0.2
0.4
0.8
0.3
Overlapping fraction (f)
0.5
0.6
0.7
0.9
1
(a) Seed size
(b) Running time
Fig. 7. Comparing coupling schemes in the synthesized networks
The results for the first scenario are shown in Fig. 8. The seed obtained by the lossless coupling
method outperforms other methods. The size of the union set is approximately 30% and 47%
larger than lossless coupling method in co-author and FSQ-Twitter, respectively. This shows that
overlapping users do propagate information through several networks and thus effectively help
18
reduce the overall seed size.
e
z
i
S
d
e
e
S
e
z
i
S
d
e
e
S
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0
3500
3000
2500
2000
1500
1000
500
0
0
Only CM
Only Het
Only NetS
Lossless
Seed Union
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(a) Co-author networks
Separate CM
CM in Lossless
Separate Het
Het in Lossless
Separate NetS
NetS in Lossless
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1600
1400
1200
1000
800
600
400
200
e
z
i
S
d
e
e
S
e
z
i
S
d
e
e
S
Only FSQ
Only Twitter
Lossless
Seed Union
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(b) FSQ and Twitter
Separate FSQ
FSQ in Lossless
Separate Twitter
Twitter in Lossless
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(c) Co-author networks
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Influenced fraction (β)
(d) FSQ and Twitter
Fig. 8. The quality of seed sets with and without using the coupled network
In the second scenario, the lossless coupling scheme achieves the best result in both networks.
When the network is considered as a standalone network and choose seeds individually(labeled
with Only in Fig. 8), the seeds size is relatively larger than choosing from the coupled network.
As shown in Fig. 8), the sizes decrease by 9%, 25%, and 17% in CM, Het, and FSQ, accordingly.
This improvement is also due to the information diffusion across several networks by the
overlapping users. Especially, when the network sizes are unbalanced, like Het with the smaller
size of users seems to get more improvement than CM.
D. Analysis of seed sets
We analyze seed sets with different influenced fraction β to find out: the composition of the
seed set and the influenced set; and the influence contribution of each network. As illustrated in
Fig. 9, a significant fraction of the seed set is overlapping nodes although only 5% (7%) users
of FSQ-Twitter (the co-author networks) are overlapping users. With β = 0.4, the fraction of
overlapping seed vertices is around 24.9% and 25% in the co-author and FSQ-Twitter networks,
respectively. As overlapping users can influence friends in different networks, they are more likely
to be selected in the seed set than ones participating in only one network. Fig. 10 demonstrates
the high influence contribution of the overlapping users, especially when β is small (contribute
more than 50% of the total influence when β = 0.2). However, when β is large, good overlapping
users are already selected, so overlapping users are not favored any more.
Additionally, there is an imbalance between the number of selected vertices and influenced
vertices in each networks. In the co-author dataset, CM contributes a large number of seed
Overlap
Only CM
Only Het
Only NetS
0.2
0.8
Influenced fraction (β)
0.4
0.6
Overlap
Only CM
Only Het
Only NetS
3500
3000
2500
2000
1500
1000
500
0
40000
35000
30000
25000
20000
15000
10000
5000
0
e
z
i
S
d
e
e
S
s
e
d
o
N
d
e
c
n
e
u
l
f
n
I
f
o
.
o
N
1400
1200
1000
800
600
400
200
0
80000
70000
60000
50000
40000
30000
20000
10000
0
e
z
i
S
d
e
e
S
s
e
d
o
N
d
e
c
n
e
u
l
f
n
I
f
o
.
o
N
Overlap
Only FSQ
Only Twitter
19
0.2
0.8
Influenced fraction (β)
0.4
0.6
Overlap
Only FSQ
Only Twitter
0.2
0.8
Influenced fraction (β)
0.4
0.6
(a) Co-author networks
Fig. 9. The bias in selecting seed nodes
0.6
0.4
0.2
0.8
Influenced fraction (β)
(b) FSQ-Twitter
vertices and influenced vertices since the size of CM is significantly larger than other networks.
When β = 0.8, 76.7% of seed vertices and 80.5% of influenced vertices are from CM. In
contrast, the number of seed vertices from FSQ is small but the number of influenced vertices
in FSQ is much higher than Twitter. With β = 0.4, 27% (without overlapping vertices) of seed
vertices belong to FSQ while 70% of influenced vertices are in FSQ. After the major of vertices
in FSQ are influenced, the algorithm starts to select more vertices in Twitter to increase the
influence fraction. This implies that it is easier for the information to propagate in one network
than the other, even when we consider the overlapping between them. Moreover, we can target
the overlapping users in one network (e.g. Twitter) to influence users in another network (e.g.
FSQ).
E. Mutual impact of networks
We evaluate the mutual impact between networks when the number of network k increases. We
use a user base of 10000 users to synthesize networks for the experiment. For each network, we
randomly select 4000 users from the user base and connect each pair of selected users randomly
with probability 0.0025. Thus all networks have the same size and the expected average outgoing
(incoming) degree of 10. The expected overlapping fraction of any network pair is 16%. We
measure the seed size to influence 60% of users (6000 users) with the different number of
networks (Fig. 11(a)). When k increases from 2 to 5, the seed size decreases several times. It
implies that the introduction of a new OSN increases the diffusion of information significantly.
We also compute how much new networks help the existing one to propagate the information.
Using the same seed set found by the greedy algorithm to influence 60% (2400 users) of the
target network (the first created network), we compute the total number of influenced vertices in
Overlap
Only CM
Only Het
Only NetS
y
a
l
e
R
l
a
t
o
T
40000
35000
30000
25000
20000
15000
10000
5000
0
y
a
l
e
R
l
a
t
o
T
80000
70000
60000
50000
40000
30000
20000
10000
0
Overlap
Only FSQ
Only Twitter
20
0.4
0.2
0.8
Influenced fraction (β)
0.6
(a) Co-author networks
Fig. 10. The influence contribution of seed vertices from component networks
0.6
0.4
0.2
0.8
Influenced Fraction (β)
(b) FSQ and Twitter
that network as well as the external influence. Fig. 11(b) shows that the number of influenced
vertices is raised 46% with the support of 3 new networks when k is changed from 2 to 5. In
addition, the fraction of external influence is also increased dramatically from 39% when k = 2
to 67% when k = 5. It means that the majority of influence can be obtained via the support of
other networks. On the hand, these results suggest that the existing networks may benefit from
the newly introduced competitor.
e
z
i
S
d
e
e
S
1600
1400
1200
1000
800
600
400
200
0
Average
Easiness
Involment
Lossless
Internal Influence
External Influence
l
y
a
e
R
l
a
t
o
T
4500
4000
3500
3000
2500
2000
1500
1000
500
0
2
3
4
5
No. of Networks
(a)
2
3
4
5
No. of Networks
(b)
Fig. 11. The impact of additional networks
VIII. CONCLUSIONS
In this paper, we study the least cost influence problem in multiplex networks. To tackle the
problem, we introduced novel coupling schemes to reduce the problem to a version on a single
network. Then we design a new metric to quantify the flow of influence inside and between
networks based on the coupled network. Exhaustive experiments provide new insights to the
information diffusion in multiplex networks.
In the future, we plan to investigate the problem in multiplex networks with heterogeneous
diffusion models in which each network may have its own diffusion model. It is still an ongoing
problem whether they can be represented efficiently, or we have better method to couple them
into one network.
This work is supported in part of NSF CCF-1422116 and DTRA HDTRA1-14-1-0055.
ACKNOWLEDGMENT
[1] 216 social media and internet statistics. http://thesocialskinny.com/216-social-media-and-internet-statistics-september-2012/.
[2] Overlap among major social network services. http://www.tomhcanderson.com/2009/07/09/overlap-among-major-social-network-services/.
[3] Francesco Buccafurri, Gianluca Lax, Antonino Nocera, and Domenico Ursino. Discovering links among social networks.
REFERENCES
21
In Machine Learning and Knowledge Discovery in Databases, pages 467–482. Springer, 2012.
[4] Ceren Budak, Divyakant Agrawal, and Amr El Abbadi. Limiting the spread of misinformation in social networks.
In
Proceedings of the 20th international conference on World wide web, pages 665–674. ACM, 2011.
[5] Ning Chen. On the approximability of influence in social networks. In SODA, pages 1029–1037, 2008.
[6] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for prevalent viral marketing in large-scale social
networks. In SIGKDD, pages 1029–1038, 2010.
[7] T. N. Dinh, H. Zhang, D. T. Nguyen, and M. T. Thai. Cost-effective viral marketing for time-critical campaigns in
large-scale social networks. IEEE/ ACM Transactions on Networking (ToN), 22(6):2001–2011, 2013.
[8] Paul Erdos and A R´enyi. On the evolution of random graphs. Publ. Math. Inst. Hungar. Acad. Sci, 5:17–61, 1960.
[9] Tereza Iofciu, Peter Fankhauser, Fabian Abel, and Kerstin Bischoff. Identifying users across social tagging systems. In
ICWSM, pages 522–525, 2011.
[10] David Kempe, Jon Kleinberg, and ´Eva Tardos. Maximizing the spread of influence through a social network. In KDD,
pages 137–146, 2003.
[11] Xingjie Liu, Qi He, Yuanyuan Tian, Wang-Chien Lee, John McPherson, and Jiawei Han. Event-based social networks:
linking the online and offline social worlds. In KDD, pages 1032–1040, 2012.
[12] M. E. J. Newman. Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E, 74:036104,
2006.
[13] Mark EJ Newman. The structure of scientific collaboration networks. Proceedings of the National Academy of Sciences,
98(2):404–409, 2001.
[14] Y. Shen, T. N. Dinh, H. Zhang, and M. T. Thai.
Interest-matching information propagation in multiple online social
networks. In in Proceedings of ACM Int Conference on Information and Knowledge Management (CIKM), pages 1824–
1828, 2012.
[15] Wen Xu, Zaixin Lu, Weili Wu, and Zhiming Chen. A novel approach to online social influence maximization. Social
Network Analysis and Mining, 4(1):1–13, 2014.
[16] Osman Yagan, Dajun Qian, Junshan Zhang, and Douglas Cochran.
Information diffusion in overlaying social-physical
networks. In CISS, pages 1–6, 2012.
[17] Huiyuan Zhang, Thang N Dinh, and My T Thai. Maximizing the spread of positive influence in online social networks.
In ICDCS, pages 317–326. IEEE, 2013.
[18] Feng Zou, Zhao Zhang, and Weili Wu. Latency-bounded minimum influential node selection in social networks.
In
International Conference on Wireless Algorithms, Systems, and Applications, pages 519–526. Springer, 2009.
|
1801.05042 | 1 | 1801 | 2018-01-15T21:50:28 | Centralized "big science" communities more likely generate non-replicable results | [
"cs.SI",
"nlin.AO",
"physics.data-an",
"q-bio.QM"
] | Growing concern that most published results, including those widely agreed upon, may be false are rarely examined against rapidly expanding research production. Replications have only occurred on small scales due to prohibitive expense and limited professional incentive. We introduce a novel, high-throughput replication strategy aligning 51,292 published claims about drug-gene interactions with high-throughput experiments performed through the NIH LINCS L1000 program. We show (1) that unique claims replicate 19% more frequently than at random, while those widely agreed upon replicate 45% more frequently, manifesting collective correction mechanisms in science; but (2) centralized scientific communities perpetuate claims that are less likely to replicate even if widely agreed upon, demonstrating how centralized, overlapping collaborations weaken collective understanding. Decentralized research communities involve more independent teams and use more diverse methodologies, generating the most robust, replicable results. Our findings highlight the importance of science policies that foster decentralized collaboration to promote robust scientific advance. | cs.SI | cs | Centralized "big science" communities more likely generate
non-replicable results
Valentin Danchev1,2,3*, Andrey Rzhetsky3,4,5, and James A. Evans1,2,3*
1 Department of Sociology, University of Chicago, 1126 E 59th St, Chicago, IL 60637, USA
2 Knowledge Lab, University of Chicago, 5735 South Ellis Avenue, Chicago, IL 60637, USA
3 Computation Institute, University of Chicago, 5735 South Ellis Avenue, Chicago, IL 60637,
USA
4 Departments of Medicine and Human Genetics, University of Chicago, 5841 S. Maryland Ave.,
Chicago, IL 60637, USA
5 Institute for Genomic and Systems Biology, 900 East 57th Street, Chicago, IL 60637, USA
*Please address correspondence to [email protected] and [email protected].
Abstract
Growing concern that most published results, including those widely agreed upon, may
be false are rarely examined against rapidly expanding research production. Replications
have only occurred on small scales due to prohibitive expense and limited professional
incentive. We introduce a novel, high-throughput replication strategy aligning 51,292
published claims about drug-gene interactions with high-throughput experiments
performed through the NIH LINCS L1000 program. We show (1) that unique claims
replicate 19% more frequently than at random, while those widely agreed upon replicate
45% more frequently, manifesting collective correction mechanisms in science; but (2)
centralized scientific communities perpetuate claims that are less likely to replicate even
if widely agreed upon, demonstrating how centralized, overlapping collaborations
weaken collective understanding. Decentralized research communities involve more
independent teams and use more diverse methodologies, generating the most robust,
replicable results. Our findings highlight the importance of science policies that foster
decentralized collaboration to promote robust scientific advance.
1
Introduction
Concern over reliability (1) and reproducibility (2, 3) in science calls into question the
cumulative process of building on prior results published by others. In a publication
environment that rewards novel findings over verifications (4, 5), scientists remain
uncertain about the published claims they assemble into new research designs and
discoveries. In this paper, we demonstrate a claim replication strategy that repurposes
high-throughput experiments to evaluate the replication likelihood for tens of thousands
of scientific claims curated from a wide range of research articles.
Our strategy builds on the synergy of two advances. First, databases of empirical
claims on many topics, ranging from material science to biomedicine (6, 7), are extracted
from literature and linked to digital archives of scientific articles such as MEDLINE and
Web of Science, making possible systematic analysis of features that characterize the
provenance of a scientific claim, such as where, when, who, and how many experiments
provided findings for and against it (8). Second, high-throughput experiments driven by
consistent, programmable robots have become increasingly widespread, initially in
genome sequencing (9) but extending to many domains, including drug-gene interactions.
These experiments provide large-scale biomedical data enabling simultaneous estimation
of the replication likelihood for many prior published claims.
Replication failures are typically attributed to systemic bias in a publication
system that favors positive results (1). This incentivizes questionable research choices
such as p-hacking (10, 11), "flexible" data analysis (12), low statistical power (13),
selective reporting ("the file drawer problem") (14), and confirmation bias (15), which
generate false results unlikely to replicate in future experiments (12).
Here we investigate the social and methodological structure of scientific and
methodological communities that coalesce around scientific claims. What network
dependencies linking scientists and their experimental methods produce the most robust,
replicable claims? We hypothesize that an empirical claim tested by a decentralized
community of non-overlapping teams with diverse training and prior knowledge, using
distinct methods under varying experimental conditions will produce robust knowledge,
likely to replicate in future experiments. By contrast, increased collaboration (16),
growing teams (17), star scientists (18, 19), and expensive, shared equipment trace a
2
pathway to "big science" (20) approaches to knowledge production (Fig. S1), which at
the extreme involve integrated collaboration of an entire community around massive,
singular research efforts like the Human Genome Project and the Laser Interferometer
Gravitational-Wave Observatory (LIGO) (20-22). We hypothesize that a claim tested by
a centralized or densely connected community involving repeated collaborations and a
narrow range of methods, knowledge, and conditions is fragile and less likely to replicate
than those produced by many decentralized or independent labs, while acknowledging
that some claims can only be explored through massive, integrated collaboration.
Prior studies simulated the effect of network structure on scientific outcomes,
suggesting that independent labs should more likely arrive at truth than a connected
network of scientists, which can more easily propagate early, false results (23-25). A
more recent experiment demonstrates that decentralized network ties, rather than
independence, most likely improves the wisdom of crowds (26). Scientific findings could
also be more or less robust as a function of diverse methods (27, 28) used to corroborate
them or distinct theories used to motivate them. Consider Jean Perrin's use of multiple
experimental techniques and theories to precisely determine Avogadro's number (29).
We represent the provenance of a scientific claim with a multilayer network (30,
31), where each network node depicts a research article reporting a finding about that
claim. Articles are linked based on whether they (1) agree and the degree to which they
share (2) scientific collaborators, (3) empirical methodologies, and (4) references to prior
work (see Fig. 2A). This network representation allows us to analyse the replication
likelihood for each claim as a function of the epistemic, social, technical and intellectual
dependencies that underlie it.
High-throughput Claim Replication Strategy
We examined a corpus of 51,292 scientific claims about chemical/drug-gene/mRNA
interactions in human systems. We compiled the corpus by using claims about directed
drug-gene interactions curated from biomedical publications in the most comprehensive
database, the Comparative Toxicogenomics Database (CTD) (6), recording over one
million published claims regarding drug-gene interactions. Each scientific claim is a
triple of drug, gene, and interaction effect. For comparability with high-throughput
3
experiments, we selected interaction effects in which a drug "increases expression" or
"decreases expression" of an mRNA in humans (effect magnitudes were not recorded).
The database identifies source articles in which the finding is claimed (i.e., PubMed ID),
which enabled examination of article content (e.g., methods) and metadata features (e.g.,
authors).
To estimate replication likelihood, we map our corpus of drug-gene claims to
high-throughput experimental data from the NIH LINCS L1000 program. The program
generated 1.3M gene expression profiles from 42,080 chemical and genetic perturbagens
across cell lines, time points, and doses (32). We used profiles induced by chemical
perturbagens, amounting to 19,811 small molecule compounds (including FDA approved
drugs). To measure relative gene expression resulting from chemical perturbations, the
LINCS L1000 team computed robust z-scores that represent differential expression
signatures. We combined
the z-scores across signatures using a bootstrapped
modification of Stouffer's method (33, 34) to generate drug-gene interaction triples of (i)
drug, (ii) gene (mRNA), and (iii) combined z-score indicating experimental effect size
and effect direction. We matched drug-gene interaction triples from the LINCS L1000
experiments to triples in CTD, and found 51,292 drug-gene interaction claims at the
intersection, corresponding to 60,159 supportive and 4,253 opposing findings from the
literature, annotated from 3,363 scientific papers.
We distinguish between robustness to social, methodological and intellectual
dependencies and robustness to experimental conditions (27) like time, dosages, cells, or
tissues, which delimit the potential for a drug-gene interaction to generalize. In all
subsequent models, we control for the variability of drug-gene interactions in LINCS
L1000 to disentangle features of experimental and biological heterogeneity from global
patterns of (in)dependence.
Our high-throughput replication strategy evaluates reliability of the overall
scientific claim rather than of any particular finding. The latter would require precise
reproduction of the original research following the original experimental protocol (35).
Nevertheless, collective agreement across many articles may serve as evidence of a
generalized claim robust to any one technique (28, 36), scientist, research group, or
4
biological setting (e.g., cell line), thereby justifying increased confidence in the claim's
widespread scientific and biomedical relevance.
Results
We observe a long-tailed distribution of published findings in support of a given
scientific claim (Fig. 1A). Most claims are supported by findings in one (89%) or two
(8%) articles, while few appear in many articles. The distribution of experimental effect
sizes over drug-gene interactions in LINCS L1000 decays much more slowly (Fig. 1D).
Published findings widely agree on the direction of drug-gene interactions (93%)
(Fig. 1A). In contrast, the majority of interaction effect-sizes (59%) in LINCS L1000
across experimental conditions are statistically undetermined (Fig. 1D). We find that
consensus in the literature is positively associated with social dependencies. A pair of
findings about a drug-gene interaction reported in papers with shared author(s) are
significantly more likely to agree in the direction of the effect (0.989; 95% CI: 0.984,
0.993; 2,514 pairs of papers) than findings reported in socially independent papers
(0.889; 95% CI: 0.884, 0.893; 20,846 pairs of papers) (Fig. S2).
We also find that generalizable interactions, which converge across experimental
conditions in LINCS L1000, are more likely to replicate published claims (RR = 0.60)
than interactions that vary across conditions (RR = 0.52), which are indistinguishable
from random (RR = 0.501) in our corpus (Fig. 1E). We define the replication success (R
= 1; 0 otherwise) of a claim as the effect size (combined z-score) in LINCS L1000
matching the direction (increase or decrease) of the effect claimed in literature (see SM).
5
A
e
n
e
g
-
g
u
r
d
f
o
r
e
b
m
u
N
s
m
a
c
i
l
Support Oppose
Drug-gene claim
105
104
103
102
101
100
100
Support
Oppose
101
Number of published findings
102
B
2500
2000
y
c
n
e
u
q
e
r
F
Support Oppose
0%
20%
40%
60%
80%
100%
C
Very high
(N = 325)
High
(N = 1,083)
Moderate
(N = 4,743)
Low
(N = 50,807)
Contradictory
µ
>
95%
>
80%
>
68%
~
68%
(N = 77)
<
68%
Probability of support in literature
0
0.5
1
P
P
P
P
P
P
G
10.0
0
0
0
1
L
S
C
N
L
I
n
i
y
t
i
l
i
b
a
i
r
a
V
P = 0.0032
P = 0.0175
P = 0.010
P = 0.0090
n.s.
n.s.
n.s.
7.5
5.0
2.5
0.0
Abs. combined z-scores
D
e
n
e
g
-
g
u
r
d
f
o
r
e
b
m
u
N
s
n
o
i
t
c
a
r
e
t
n
i
105
104
103
102
101
100
100
Effect size in LINCS L1000 experiment
102
101
Significant
40%
20%
Not significant
60%
80%
100%
0%
E
R
R
,
e
t
a
r
n
o
i
t
a
c
i
l
p
e
R
1.00
0.75
0.50
0.25
0.00
Probability of support in literature
Corpus
Contradictory
Low
Moderate
High
Very high
RRobs
RRrand
Not significant
Random
Significant
Random
N = 136
N = 406
N = 21,181
N = 18,688
N = 1,908
N = 43
F
)
%
(
e
s
a
e
r
c
n
d
n
a
r
R
R
i
n
o
i
t
a
c
i
l
d
n
a
r
R
R
p
e
r
e
v
i
t
s
b
o
R
R
Effect size in LINCS L1000 experiment
l
a
e
R
Very high
Contradictory
Moderate
High
Low
All significant
Contradictory
All significant
Moderate
High
Low
Very high
Fig. 1. Aligning drug-gene claims from the literature with interactions from high-
throughput experiments to evaluate replication likelihood. (A) On top, the
distribution of drug-gene claims against the number of supporting and opposing findings
in the literature on a log-log scale. On bottom, the proportion of supporting and opposing
findings on aggregate. (B) Posterior distributions of support in the biomedical literature
for a sample of drug-gene claims. For each claim, we summarize the probability of
support (dashed line) using the lower bound on the 95% posterior credible intervals
(PCI). (C) Typology of drug-gene claims from the literature on the basis of claim's
distribution of support compared to the null value of 0.5. Number of claims included in a
category provided in brackets. (D) Effect sizes (combined z-scores) for drug-gene
interactions derived from LINCS L1000 experiments. On top, the distribution of drug-
gene interactions against absolute experimental effect size in LINCS L1000 (log-log
scale). On bottom, the proportion of significant and insignificant interactions in LINCS
L1000. We established significance at the 0.05 level by bootstrapping (10,000 iterations)
the combined z-score for each of the 51,292 interactions sampling from experiments
across different cell lines, durations, and dosages. (E) Observed, 𝑅𝑅!"#, versus expected,
𝑅𝑅!"#$, replication rate (see SM) for drug-gene claims differentiated by significance in
of claim, indicating the percentage increase of 𝑅𝑅!"# relative to random 𝑅𝑅!"#$ or
100 × RRobs!RRrand
. (G), Variability (coefficient of variation) in LINCS L1000 across cell
lines, durations, and dosages for each type of claim. We established statistical
significance using a nonparametric bootstrap test with 100,000 iterations. Error bars
represent SEM.
LINCS L1000 and support from literature. (F) Relative replication increase for each type
6
RRrand
A central concern is whether the replication problem applies only to novel and
rare claims or if it also afflicts widely supported results, as recently hypothesised (37,
38). To integrate collective support in the literature over multiple claim sources, we
design a binomial Bayesian model with a uniform prior that accommodates skewed
distributions (39) like that of findings per claim we observed. To estimate the probability
of support for each drug-gene claim, we assume that the number of supportive published
findings 𝛾 in 𝑛 findings about that claim is drawn from a binomial distribution,
𝑝(γ θ ~ Bin γ n, θ (see SM). The model allocates higher probability to scientific claims
We find a strong relationship between scientific support in the literature 𝐿!"#$
unanimously supported in a large number of articles and lower probability to infrequent
and disputed claims (Fig. 1B).
and the probability of replication success, ranging from 0.507 (95% CI: 0. 492, 0.522) for
claims with the least support to 0.705 (95% CI: 0.665, 0.743) for claims with the largest,
controlling for biological and experimental variability in LINCS L1000 (Fig. S3A).
When we consider only significant drug-gene interactions (Fig. S3B), which generalize
across experimental conditions, the probability of replication for top-supported claims
reaches 0.798 (95% CI: 0. 751, 0.845).
To examine the hypothesis that convergent drug-gene claims more likely replicate
than those claimed rarely or disputed in literature, we categorized claims on the basis of
their distribution of support (Figs. 1C and S4; see SM). We found that claims with "Very
high" (RRI = 45.4%; SEM ± 13.5%) and "High" (RRI = 34.5%; SEM ± 7.6%) support in
the biomedical literature are significantly more likely to replicate in high-throughput
experiments than claims with "Low" (RRI = 19%; SEM ± 1.1%) and "Moderate" (RRI =
16.2%; SEM = 3.3%) support (Fig. 1F). The replication of "Low" and "Moderate"
support claims is consistent with estimates (~11–25%) from recent biomedical
reproducibility studies (2, 3), but those receiving support from multiple sources is much
higher. "Contradictory" claims are significantly less likely to replicate than random. They
are also associated with greater experimental variability (Fig. 1G), indicating that
collective disagreements among findings truthfully signal unstable drug-gene interactions
with low replication likelihood. These results suggest that findings reported in a single
7
Independent and decentralized sources of evidence should increase claim
scientific article may be fallible, but viewed as a system of converging and diverging
findings, science exhibits collective correction.
robustness. Figure 2A presents a schematic of a multilayer network 𝑀=(𝑉!,𝐸!,𝐿) for
a claim. In each network layer L, nodes 𝑉 are scientific papers and edges 𝐸 between pairs
of papers represent either a binary relationship of agreement (𝐿!) or the amount of
(𝐿!), methodologies (𝐿!), and references to prior literature (𝐿!). Our independence scores
overlap measured with the Jaccard coefficient, between the two papers' sets of authors
for social, methodological, and prior knowledge range between 0 (dependent) and 1
(independent), and can be viewed as the probability that any two randomly chosen
findings about a claim are obtained by disconnected sets of authors, methods, and
references, respectively (see SM). Figure 2B-C present schematics of bipartite networks
of authors connected to the papers they publish. To quantify the centralization of research
communities C, we compute the Gini coefficient of the authors' degree distribution (see
Fig. 2D-E).
in LINCS L1000. Figure 2F shows that the odds ratios (OR) of replication increase
We examine the impact of network dependencies on claim replicability by fitting
logistic regression models that predict replication success against the measures described
above and defined for each claim, using a sub-corpus of 2,493 claims with determined
direction of the drug-gene effect in the literature (𝐿!"#$=68% PCI > µ < 68% PCI) and
substantially with the increase of scientific support in the literature 𝐿!"#$ (OR 23.20; 95%
CI: 9.08, 59.3), social independence 𝑆!"# (OR 6.31; 95% CI: 4.07, 9.79), methodological
independence 𝑀!"# (OR 6.30; 95% CI: 3.44, 11.53), and prior knowledge independence
𝐾!"# (OR 5.53; 95% CI: 2.58, 11.84). By contrast, claim replication decreases with
centralization 𝐶 (OR 0.36; 95% CI: 0.27, 0.48). Our estimates indicate that claim
robustness, defined here as multiple decentralized evidence, increases replication success
(Fig. 2G-M). When all predictors are modelled simultaneously (Fig. 2F), however,
centralization and support in the literature account for the rest, suggesting that centralized
academic training and extensive collaboration mediates the application of research
techniques and attention to prior research.
8
A
Agree-
ment
Authors
overlap
Methods
overlap
References
overlap
6
1
2
1
B
d
e
z
i
l
a
r
t
n
e
C
C
d
e
z
i
l
a
r
t
n
e
e
D
Benzo(a)pyrene SLC22A3
Sulforafan NQO1
Estradiol PRSS23
T
R
O
P
P
U
S
T
R
O
P
P
U
S
L
S
S I N
0
P
U
T = 0.6
D = 0.8 5
D = 0.8 7
M IN
D = 0.6
IIN
T
R
O
P
P
U
S
L
P
U
S
S I N
4
T = 0.8
D = 0.9 9
D = 0.9 1
M IN
D = 0.8 1
IIN
Authors in many papers
Authors in one paper
P
S
U
L
7
OPPOSE
T = 0.4
D = 0.9 6
D = 0.8 8
S I N
M IN
D = 0.8 7
D
P
P
O
Tretinoin F3
O
S
E
S
P
U
L
9
SUPPORT
T = 0.6
D = 0.8 6
D = 0.7 9
M IN
D = 0.9 1
S I N
IIN
Tretinoin F3
C = 28.3%
IIN
T
r
e
t
i
n
o
n
i
F
3
E
s
t
r
a
d
o
i
l
P
R
S
S
2
3
100%
80%
60%
40%
20%
s
e
g
d
e
r
e
p
a
p
-
o
t
-
r
o
h
t
u
a
f
o
e
r
a
h
S
E
s
e
g
d
e
r
e
p
a
p
-
o
t
-
r
o
h
t
u
a
f
o
e
r
a
h
S
0%
0% 20% 40% 60% 80% 100%
Share of authors from low to high degrees
100%
80%
60%
40%
20%
Estradiol PRSS23
C = 12.4%
0%
0% 20% 40% 60% 80% 100%
Share of authors from low to high degrees
Authors
Research papers
G
100%
75%
50%
25%
0%
K
100%
75%
50%
25%
0%
0.00 0.25 0.50 0.75 1.00
Support in literature
H
100%
75%
50%
25%
0%
L
100%
75%
50%
25%
0%
0.00
n
o
i
t
a
c
i
l
p
e
r
i
l
m
a
c
f
o
y
t
i
l
i
b
a
b
o
r
P
1
5 10 25 50
Claim replication (OR, 95% CI)
F
Support
in literature
Journal
prominence
Social
independence
Methodological
independence
Prior knowledge
independence
Variability
in LINCSL1000
Centralization
J
100%
75%
50%
25%
0%
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
o
y
t
i
l
i
b
a
b
o
r
P
0.25
0.75
Social independence
0.50
I
100%
75%
50%
25%
0%
1.00
0.00
M
100%
75%
50%
25%
0%
0.50
0.25
Centralization
0.75
1.00
0.00 0.25 0.50 0.75 1.00
0.25
0.00
1.00
Methodological independence
0.50
0.75
0.25
0.00
1.00
Prior knowledge independence
0.50
0.75
Journal prominence
Fig. 2. Multilayer networks of social, methodological, and prior knowledge
dependences predict replication likelihood. (A) Examples of multilayer claim
networks. Each layer 𝐿 represents a co-paper network: nodes are scientific papers; pairs
of papers are connected by an unweighted edge if they agree on the effect direction (𝐿!),
9
0.00 0.25 0.50 0.75 1.00
Variability in LINCS L1000
two papers' authors (𝐿!), methodologies (𝐿!), and references to prior literature (𝐿!).
and a weighted edge proportional to the overlap (Jaccard coefficient or JC) between the
Dashed red line in the agreement layer indicates opposing edges. For clarity, we only plot
edges above the mean JC value for overlapping methods. (B-C) Example of centralized
and decentralized bipartite networks. Edges connect authors (rectangles) to papers they
published (circles). Centralization C is the Gini coefficient of the authors' degree
distribution. (D-E) Lorenz curves showing different level of centralization, corresponding
to the examples in B and C. (F) Logistic regression models with replication success as the
response variable and predictors modelled
(bottom)
simultaneously. Predictors are rescaled
probabilities (PP) of claim replication PP𝑅=1 = !"# (!!!!!!)
!!!!"# (!)
!"#!!!"# (!) for comparability. (G-M) Predicted
!!!"# (!!!!!!) for logistic models of
(top)
independently and
independent predictors with 95% error bounds.
An alternative explanation of replication success is the biological tendency for
some drug-gene interactions to generalize across conditions and so replicate in future
experiments. Figure 2F shows that experimental variability has a small, insignificant
negative effect, and that support from multiple, decentralized teams are much more
informative predictors of replication success.
Our combined model also controls for journal prominence J, which we measure
with journal eigenfactor (40), a score that credits journals receiving many citations from
highly cited journals. Claim replication increases with journal prominence (Figs. 2F and
2L), but prominent journals are responsible for only a tiny fraction of all claims, which
warrants our evaluation strategy and the practice of extracting and archiving findings
from a wide range of journals (6).
Figure 3 shows that by accounting for scientific support, centralization, and social
independence, we can identify claims with high replication probability. Claims supported
by many publications have about 45 percent higher probability to replicate when
investigated by decentralized versus centralized communities (Figs. 3A and S8). Even if a
claim garners wide support, if studied exclusively by a centralized scientific community,
the claim is indistinguishable in replication probability from a claim reported in a single
paper. This suggests a condition under which collective correction in science is
undermined, when one or several scientists exercise disproportionate influence on team
assembly and research across multiple investigations of a claim. Claims robust to
multiple, socially independent investigations have a 55 percent higher probability to
10
replicate than those studied by small clusters of overlapping collaborators (Figs. 3B and
S9). These effects all account for the experimental and biological variability of drug-gene
interactions.
A
B
Fig. 3. Contour plots of the probability of claim replication based on network
dependencies and support in the literature. (A) The probability of claim replication is
greater in decentralized communities of scientists, after controlling for variability in
LINCS L1000 across cell lines, durations, and dosages. (B) The probability of claim
replication increases with support in the literature and social independence, after
controlling again for experimental variability.
Discussion
Extensive (22) and often overlapping scientific collaboration (16), along with cumulative
advantage processes that create central, star scientists (19), produce centralized "big
science" communities with dense methodological and intellectual dependencies. Such
communities tend to publish fragile findings. Our research points to the importance of
science policies that foster competition and decentralized collaboration to promote robust
and replicable scientific advance. Our analysis demonstrates the utility of large-scale
experiments coupled with enriched article content and metadata to diagnose the
replicability of published research. Our findings suggest a calculus for evaluating the
system-level trade-off between investments in robust, replicable knowledge, which
comes at the price of larger, but weaker, preliminary insight.
11
0.350.350.350.40.40.40.450.450.450.50.50.550.550.60.60.650.650.70.70.750.80.850.9Social independence00.20.40.60.810.20.30.40.50.60.70.80.910.30.350.40.450.450.50.50.550.550.60.60.60.650.650.65Centralization00.20.40.60.81Support in literature00.20.40.60.81Probability of claim replication References
J.
P.
Ioannidis,
Why
most
published
research
findings
are
false.
PLoS
Med
2,
1.
e124
(2005).
F.
Prinz,
T.
Schlange,
K.
Asadullah,
Believe
it
or
not:
how
much
can
we
rely
on
2.
published
data
on
potential
drug
targets?
Nat
Rev
Drug
Discov
10,
712–712
(2011).
C.
G.
Begley,
L.
M.
Ellis,
Drug
development:
Raise
standards
for
preclinical
3.
cancer
research.
Nature
483,
531–533
(2012).
B.
A.
Nosek,
J.
R.
Spies,
M.
Motyl,
Scientific
Utopia:
II.
Restructuring
incentives
4.
and
practices
to
promote
truth
over
publishability.
Perspectives
on
Psychological
Science
7,
615–631
(2012).
B.
Alberts
et
al.,
Self--‐correction
in
science
at
work.
Science
348,
1420
(2015).
5.
A.
P.
Davis
et
al.,
The
Comparative
Toxicogenomics
Database:
update
2017.
6.
Nucleic
Acids
Research
45,
972–978
(2017).
A.
H.
Wagner
et
al.,
DGIdb
2.0:
mining
clinically
relevant
drug–gene
7.
interactions.
Nucleic
acids
research,
gkv1165
(2015).
8.
J.
A.
Evans,
J.
G.
Foster,
Metaknowledge.
Science
331,
721
(2011).
International
Human
Genome
Sequencing.
Consortium,
Initial
sequencing
9.
and
analysis
of
the
human
genome.
Nature
409,
860
(2001).
10.
M.
L.
Head,
L.
Holman,
R.
Lanfear,
A.
T.
Kahn,
M.
D.
Jennions,
The
extent
and
consequences
of
p--‐hacking
in
science.
PLoS
Biol
13,
e1002106
(2015).
11.
U.
Simonsohn,
L.
D.
Nelson,
J.
P.
Simmons,
P--‐curve:
a
key
to
the
file--‐drawer.
Journal
of
Experimental
Psychology:
General
143,
534
(2014).
12.
J.
P.
Simmons,
L.
D.
Nelson,
U.
Simonsohn,
False--‐Positive
Psychology.
Psychological
Science
22,
1359–1366
(2011).
E.
Dumas--‐Mallet,
K.
S.
Button,
T.
Boraud,
F.
Gonon,
M.
R.
Munafò,
Low
13.
statistical
power
in
biomedical
science:
a
review
of
three
human
research
domains.
Royal
Society
Open
Science
4,
160254
(2017).
14.
R.
Rosenthal,
The
file
drawer
problem
and
tolerance
for
null
results.
Psychological
Bulletin
86,
638–641
(1979).
15.
R.
Nuzzo,
Fooling
ourselves.
Nature
526,
182–185
(2015).
R.
Guimerà,
B.
Uzzi,
J.
Spiro,
L.
A.
N.
Amaral,
Team
Assembly
Mechanisms
16.
Determine
Collaboration
Network
Structure
and
Team
Performance.
Science
308,
697
(2005).
17.
S.
Wuchty,
B.
F.
Jones,
B.
Uzzi,
The
Increasing
Dominance
of
Teams
in
Production
of
Knowledge.
Science
316,
1036
(2007).
R.
K.
Merton,
The
Matthew
effect
in
science.
Science
159,
56–63
(1968).
18.
19.
P.
Azoulay,
T.
Stuart,
Y.
Wang,
Matthew:
Effect
or
Fable?
Management
Science
60,
92--‐109
(2013).
D.
J.
de
Solla
Price,
Little
science,
big
science.
(Columbia
University
Press,
New
20.
York,
1963).
21.
A.
M.
Weinberg,
Impact
of
Large--‐Scale
Science
on
the
United
States.
Science
134,
161
(1961).
D.
Hicks,
J.
S.
Katz,
Science
policy
for
a
highly
collaborative
science
system.
22.
Science
and
Public
Policy
23,
39–44
(1996).
12
N.
Payette,
in
Models
of
Science
Dynamics:
Encounters
Between
Complexity
23.
Theory
and
Information
Sciences,
A.
Scharnhorst,
K.
Börner,
P.
van
den
Besselaar,
Eds.
(Springer,
Berlin,
Heidelberg,
2012),
pp.
127–157.
Kevin
J.
S.
Zollman,
The
Communication
Structure
of
Epistemic
Communities.
24.
Philosophy
of
Science
74,
574–587
(2007).
J.
Lorenz,
H.
Rauhut,
F.
Schweitzer,
D.
Helbing,
How
social
influence
can
25.
undermine
the
wisdom
of
crowd
effect.
Proceedings
of
the
National
Academy
of
Sciences
108,
9020–9025
(2011).
J.
Becker,
D.
Brackbill,
D.
Centola,
Network
dynamics
of
social
influence
in
the
26.
wisdom
of
crowds.
Proceedings
of
the
National
Academy
of
Sciences
114,
E5070--‐E5076
(2017).
27.
W.
G.
Kaelin
Jr,
Common
pitfalls
in
preclinical
cancer
target
validation.
Nat
Rev
Cancer
17,
425–440
(2017).
28.
W.
C.
Wimsatt,
in
Characterizing
the
Robustness
of
Science:
After
the
Practice
Turn
in
Philosophy
of
Science,
L.
Soler,
E.
Trizio,
T.
Nickles,
W.
Wimsatt,
Eds.
(Springer
Netherlands,
Dordrecht,
2012),
pp.
61–87.
29.
W.
Salmon,
Scientific
explanation
and
the
causal
structure
of
the
world.
(Princeton
University
Press,
Princeton,
New
Jersey,
1984).
30.
M.
Kivelä
et
al.,
Multilayer
networks.
Journal
of
Complex
Networks
2,
203–271
(2014).
31.
J.
F.
Padgett,
C.
K.
Ansell,
Robust
Action
and
the
Rise
of
the
Medici,
1400–
1434.
American
Journal
of
Sociology
98,
1259–1319
(1993).
A.
Subramanian
et
al.,
A
Next
Generation
Connectivity
Map:
L1000
Platform
32.
and
the
First
1,000,000
Profiles.
Cell
171,
1437–1452.e1417.
33.
M.
C.
Whitlock,
Combining
probability
from
independent
tests:
the
weighted
Z--‐method
is
superior
to
Fisher's
approach.
Journal
of
evolutionary
biology
18,
1368–1373
(2005).
34.
D.
S.
Himmelstein
et
al.,
Systematic
integration
of
biomedical
knowledge
prioritizes
drugs
for
repurposing.
eLife
6,
e26726
(2017).
T.
M.
Errington
et
al.,
An
open
investigation
of
the
reproducibility
of
cancer
35.
biology
research.
Elife
3,
e04333
(2014).
36.
B.
A.
Nosek,
T.
M.
Errington,
Making
sense
of
replications.
eLife
6,
e23383
(2017).
S.
B.
Nissen,
T.
Magidson,
K.
Gross,
C.
T.
Bergstrom,
Publication
bias
and
the
37.
canonization
of
false
facts.
eLife
5,
e21451
(2016).
R.
McElreath,
P.
E.
Smaldino,
Replication,
Communication,
and
the
Population
38.
Dynamics
of
Scientific
Discovery.
PLOS
ONE
10,
e0136088
(2015).
39.
C.
Davidson--‐Pilon,
Bayesian
Methods
for
Hackers:
Probabilistic
Programming
and
Bayesian
Inference.
(Pearson
Education,
2015).
C.
T.
Bergstrom,
J.
D.
West,
M.
A.
Wiseman,
The
Eigenfactor™
Metrics.
The
40.
Journal
of
Neuroscience
28,
11433
(2008).
41.
J.
Kruschke,
Doing
Bayesian
Data
Analysis:
A
Tutorial
with
R,
JAGS,
and
Stan.
(Elsevier
Science,
2014).
42.
A.
Gelman
et
al.,
Bayesian
Data
Analysis.
(Taylor
&
Francis,
ed.
Third
edition,
2014).
13
43.
M.
Kivelä,
Multilayer
Networks
Library
for
Python
(Pymnet),
available
at
http://bitbucket.org/bolozna/multilayer--‐networks--‐library.
2017.
S.
Wasserman,
K.
Faust,
Social
network
analysis:
methods
and
applications.
44.
(Cambridge
University
Press,
Cambridge,
1994),
pp.
825.
A.
Bandrowski
et
al.,
The
Ontology
for
Biomedical
Investigations.
PLOS
ONE
45.
11,
e0154556
(2016).
46.
A.
Gonzalez--‐Beltran,
P.
Rocca--‐Serra,
O.
Burke,
S.--‐A.
Sansone.
(Available
online:
http://bioportal.bioontology.org/ontologies/STATO,
Last
accessed
on
07
November
2017,
2016).
47.
M.
M.
Kessler,
Bibliographic
coupling
between
scientific
papers.
American
Documentation
14,
10–25
(1963).
48.
B.
Peirson,
R.
Erick,
e.
al.
(http://diging.github.io/tethne/,
2017).
A.
Hagberg,
P.
Swart,
D.
S
Chult,
"Exploring
network
structure,
dynamics,
and
49.
function
using
NetworkX,"
(Los
Alamos
National
Laboratory
(LANL),
2008).
J.
Kunegis,
J.
Preusse,
in
Proceedings
of
the
3rd
Annual
ACM
Web
Science
50.
Conference.
(ACM,
2012),
pp.
175–184.
51.
L.
C.
Freeman,
Centrality
in
social
networks
conceptual
clarification.
Soc
Networks
1,
215–239
(1978).
52.
J.
M.
Badham,
Commentary:
Measuring
the
shape
of
degree
distributions.
Network
Science
1,
213–225
(2013).
53.
J.
Fox,
Effect
displays
in
R
for
generalised
linear
models.
Journal
of
statistical
software
8,
1–27
(2003).
D.
Lüdecke,
sjPlot:
data
visualization
for
statistics
in
social
science.
R
package
54.
version
2.3.3
(2017).
Acknowledgments
We thank A. Belikov, D. Centola, E. Duede, J. P.A. Ioannidis, M. Lewis, and T. Stoeger
for helpful discussions and comments, E. Demir and R. Melamed for advice on curated
drug-gene databases and high-throughput experimental data, respectively, W. Catino for
computational help, M. Kivelä for help with his Multilayer Networks Library for Python
(Pymnet), and T. Natoli from Broad Institute for advice on NIH LINCS L1000 data. We
used MEDLINE/PubMed and Web of Science data to build the co-paper networks. We
thank Clarivate Analytics for supplying the Web of Science data and T. Ando for
computing the journal eigenfactor scores. This work was supported by DARPA's Big
Mechanism research program grant 14145043, NSF SciSIP grant 1158803, and AFOSR
grant FA9550-15-1-0162.
14
Supplementary Materials
List of supplementary content:
Materials and Methods
Data Analysis
Figures
Tables
Materials and Methods
DATA
High-throughput drug-gene interactions
We used the Library of Integrated Network-based Cellular Signatures (LINCS) Phase I
L1000 data set to estimate claim replication (32). The data set measures the expression of
978 landmark genes treated with wide range of perturbagens across cell lines, time
points, and doses (concentration), resulting in approximately 1.3M profiles. The
experimental profiles are aggregated to 473,647 signatures, as represented in Level 5 data
we used to perform our analysis. The landmark genes are subsequently used to infer gene
expressions for the remaining genes in the human genome. In addition to the 978
landmark genes, we consider 9,196 for which the LINCS L1000 project estimated to be
well inferred, resulting in 10,174 Best INferred Genes (BING) in total. With respect to
perturbagen types, we used the set of small-molecule compounds (19,811 compounds),
which includes a subset of approximately 1,300 FDA-approved drugs. We accessed the
data file GSE70138_Broad_LINCS_Level5_COMPZ_n118050x12328_2017-03-
06.gctx.gz and metadata from the GEO depository at
https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE70138.
Published drug-gene interactions
We analysed curated data about published interactions between chemicals/drugs and
genes/mRNAs released by the Comparative Toxicogenomics Database (CTD) (6) on June
7 2016. (The current CTD data release is available here: http://ctdbase.org/downloads/.)
To align triples of drug, gene, and interaction effect in CTD to corresponding triples in
the experimental signatures from LINCS L1000, we performed the following procedures.
First, we selected drug-gene interactions about Homo sapiens, comprising approximately
40% of the CTD data. Second, CTD reports the form of the gene (e.g., mRNA, protein)
that is implicated, and we selected only mRNA as LINCS L1000 measures gene
expression at the mRNA level. Third, we mapped chemical names and Entrez gene IDs in
the CTD to perturbagen names and Entrez gene IDs in LINCS L1000. Fourth, to ensure
comparability to the LINCS L1000 signatures, we selected drug-gene interactions with a
single interaction effect, either "decreases expression" or "increases expression", defining
the direction of the effect that chemical/drug manifests on a gene/mRNA. Note that we
do not consider complex interactions with multiple, nested effects in our analysis.
15
Likewise, interactions for which the direction of the effect is not specified, such as
"affects binding", are not considered. The resulting corpus at the intersection of LINCS
L1000 and CTD comprises 51,292 drug-gene claim combinations of 605 unique drugs
and 9,123 unique genes, annotated from 3,363 scientific articles.
METHODS
Combining LINCS L1000 effect sizes
The LINCS L1000 team applied a moderated Z-scoring procedure to estimate weighted
average signatures. For each drug-gene interaction, we computed a combined gene
signature using Stouffer's method (33, 34) Z =
i=1! , where 𝑍! is an experimental z-score
and 𝑘 is the number of signatures. Using this procedure, we aggregated experiments
about drug-gene interactions across cell lines, dosages, and durations into a combined
effect size (z-score). Some drug-gene interactions generalize across conditions, whereas
others are context specific. We define as generalized drug-gene effect sizes that are
significant at the 0.05 level. To measure significance, for each drug-gene interaction, we
performed 10,000 bootstrap iterations on the combined z-score, sampling from
experiments performed under different cell lines, dosages, and durations. Conventionally,
if the 95% confidence intervals do not contain the null value of 0, we consider a drug-
gene interaction as generalized.
Measuring variability in LINCS L1000
To measure how each drug-gene interaction varies across cell lines, dosages, and
durations in LINCS L1000, we computed the coefficient of variation. The coefficient of
variation (CI) is the ratio of the standard deviation to the absolute value of the mean:
Zi
k
measure that allows us to make comparisons across drug-gene interactions.
Modelling claims' support in literature
We apply a binomial Bayesian model (39, 41, 42) to estimate the probability of scientific
𝐶𝐼= !!"#(!). In comparison to scale-dependent measures of variability, CI is a normalized
support 𝜃 for each drug-gene claim given the number of supportive findings 𝛾 in all
prior distribution of 𝜃 is uniform on the interval [0,1]: θ!~Uniform(min=0,max=1).
This uninformative prior considers all possible probability support values 𝜃 as equally
assume that the number of findings in support of a drug-gene claim 𝛾 follows a binomial
distribution 𝑝(𝛾𝑁,𝜃 ~ Binomial𝛾𝑁,𝜃 where Binomial𝛾𝑁,𝜃 ~ 𝑁𝛾 𝜃!(1−
𝜃)!!!. Then, the posterior density for 𝜃 is 𝑝𝜃𝛾 ∝ 𝜃!1−𝜃 !!!. We approximated
the posterior density of 𝜃 for each drug-gene claim by performing 10,000 Markov chain
findings N reported in CTD about a particular drug-gene interaction. We assume that the
possible and is suitable for the skewed distribution of findings over drug-gene claims. We
Monte Carlo (MCMC) sampling iterations (2,500 burn in iterations) for each drug-gene
claim using the Metropolis–Hastings MCMC sampler implemented in the PyMC package
(version 2.3.6) for Python. To improve convergence, we approximate the maximum
posterior (MAP) before running the MCMC sampler (39).
16
1,083 claims, supported and opposed in 3,525 and 42 articles, respectively); Moderate
from multiple papers (~7 papers on average), amounting to 325 claims that are supported
by 2,241 findings, but only opposed by 21; High support claims
Typology of drug-gene claims
To avoid setting arbitrary thresholds to categorize claims by their support in literature, we
used the posterior distributions from our Bayesian model of support. For each claim, we
estimated the overlap between the posterior credible intervals (PCI) and the null value of
µ = 0.5 (Fig. 1C): Very high support claims (𝐿!"#$ 95% PCI exceeds µ) yield agreement
(80% PCI ≥ 𝐿!"#$ < 95% PCI exceeds µ) yield agreement from 3 papers on average (N =
support claims (68% PCI ≥ 𝐿!"#$ < 80% PCI exceeds µ) yield agreement from 2 papers
respectively); Low support claims (68% PCI ≥ 𝐿!"#$ < 80% PCI contains µ) are
claims (𝐿!"#$ 68% PCI is smaller than µ) generate lower support than expected as a
We define replication as a binary variable 𝑅∈{0,1}, where R = 1 indicates that the
random replication rate 𝑅𝑅!"#$ by iteratively randomizing LINCS L1000 drug-gene
greater number of papers reported findings in the opposite direction (N = 77 claims,
supported and opposed in 101 and 484 articles, respectively) (Fig. S4).
Null model for estimating relative replication increase
overwhelmingly supported by a single paper or opposed by virtually the same number of
papers that support them such that the direction of the effect is undetermined (N = 46,064
claims, supported and opposed in 46,735 and 3,668 articles, respectively). Contradictory
direction (increase or decrease) of the effect claimed in literature is matched by the
combined z-score in LINCS L1000, and R = 0 otherwise. To estimate the amount of
relative replication increase (RRI), we first empirically established the expected or
on average (N = 3,743 claims, supported and opposed in 7,557 and 38 articles,
interactions. For each randomization, we computed the replication rate for all significant
drug-gene claims (N = 21,181) as well as for the collection of very high support (N =
136), high support (N = 406), moderate (N = 1,908), low (N = 18,688), and contradictory
(N = 43) claims against the drug-gene interactions from LINCS L1000 that randomly
matched them. We repeated the estimation procedure 100,000 times to generate
randomized distributions, mean randomized replication rates RRrand, and standard errors
of the RRrand (or SEM) for each claim type. This randomization model takes into account
biases in the distribution of published claims, as when disproportionately more claims of
a certain direction, either "increasing" or "decreasing", are present. This allows us to
conservatively estimate the expected replication rate on the basis of empirical data, rather
than assuming a random null value of 0.5. We bootstrapped (100,000 iterations with
replacement) the observed replication vector for each type of claim to generate
distributions of observed replication rates, mean observed replication rates RRobs, and the
standard error of RRobs (or SEM). We used the randomized and bootstrapped distributions
for each type of claim to estimate the percentage relative replication increase: 𝑅𝑅𝐼=
100 × RRobs!RRrand
multilayer network 𝑀= 𝑉!,𝐸!,𝐿 (30). The network consists of four layers, which we
plot using the Multilayer Networks Library for Python (Pymnet) (43). Nodes 𝑉! in the
Claim multilayer networks
To investigate the robustness of drug-gene claims, we represent each claim as a
17
RRrand
.
network are scientific papers. In each layer L, edges 𝐸 between pairs of papers represent
either a binary relationship of agreement and disagreement about effect direction (𝐿!), or
the amount of overlap between the two articles' authors (𝐿!), methods (𝐿!), and
references (𝐿!). For this analysis, we exclude claims with low support as estimated via
our Bayesian model because they are primarily reported by a single paper and therefore,
lack independence and robustness. To examine the impact of claim robustness on the
probability of replication in high-throughput experiments, we interrogated only
interactions that produce a replication signal by posting a statistically significant effect
size. These are claims that generalize across experimental and biological conditions in
LINCS L1000. By exploring only statements having significant agreement within the
literature and within LINCS L1000, we can directly examine the effect of social,
methodological, and epistemic dependencies on shaping agreement between literature
and high-throughput experiment. The sub-corpus of claims both supported in the
literature and significant within LINCS L1000 consists of 2,493 claims, associated with
6,272 supporting and 339 opposing findings from 1,282 papers. For each claim, we create
a multilayer co-paper network and model the predictive power of each layer on the
replicability of a claim in the LINCS L1000 experiments.
Quantifying overlap between articles' authors, methods, and references
To quantify the amount of overlap between papers, we used the Jaccard coefficient (JC).
JC is the size of intersection divided by the size of the union:
For any two sets of paper's attributes-i.e., authors, methods, or references-𝐴! and 𝐴!,
𝐽𝐶𝐴!,𝐴! =
!! ! !! ! !! ∩ !!. The resulting quantity represents the edge weight
!! ∩ !!
between a pair of articles in the respective network layers of shared authors (𝐿!),
methods (𝐿!), and references (𝐿!). Each drug-gene claim constitutes an undirected,
, where 𝑊 is the sum over
in a network layer 𝐸!"#=
multilayer network of papers connected via such weighted edges across layers (see Fig
2A).
Independence of findings
We define an independence score IND as the proportion of maximum possible edges (44)
!(!!!)/! not present, 𝐼𝑁𝐷=!!"#! !!!"#
!
all weighted edges in a claim's respective layer of shared authors, methods, or references.
The independence scores for social, methodological, and prior knowledge approaches 1
when most papers with findings in support of a claim share no common authors, methods,
and references, and 0 when all papers share all of their authors, methods, and references,
respectively.
Social independence
We used the MEDLINE/PubMed database to extract the set of authors for each paper. To
measure the overlap between two sets of authors, we need individual identifiers for
authors. Authors' name disambiguation is a common problem in research on scientific
knowledge production. We used the individual identifiers based on author last name and
initials. We note that because we assessed authors separately for each claim, our
conservative matching procedure is very unlikely to produce false positive author
linkages, and so our author co-paper network should be considered a lower bound for
18
author co-paper density. For the sub-corpus of 2,493 claims, sourced from 1,282 papers,
we estimated a mean of 6.4 authors per paper and a mean of 23.5 authors per scientific
community defined here as the total number of authors that have published papers
reporting a drug-gene claim. For the set of papers supporting a claim, we applied JC to
measure the overlap between any two sets of papers' authors and utilized the output to
compute claim's social independence 𝑆!"# using our independence score.
Methodological independence
We compiled a controlled vocabulary of 3,074 terms (incl. synonyms) concerning
methods, techniques, and experimental apparatus used in biomedical research using
ontologies of biomedical investigations (45) and statistics (46). We then used the
RESTful API Web Service of Europe PMC to query the methods sections from 4.4
million full text articles and extracted, on aggregate, 13,095 terms for 488 articles (38%).
In parallel, for all 1,282 articles that share a drug-gene claim, we applied fuzzy matching
against our vocabulary using the difflib module in Python and extracted 12,135 terms
from abstracts available in MEDLINE/PubMed. We combined the outputs from the two
search procedures. Then, for the set of papers supporting a claim, we again used JC to
measure the overlap between any two sets of papers' methods and then employed our
independence score to measure claim's methodological independence 𝑀!"#.
Prior knowledge independence
To examine whether a pair of publications is exposed to similar or dissimilar prior
information, we use the notion of bibliographic coupling (47), i.e., the number of
citations any two papers share. To compute bibliographic coupling, we used the Web of
Science citation data. Out of 1,282 papers sharing a drug-gene claim with at least one
other paper, we mapped 1,234 PubMed IDs to Web of Science IDs and performed
bibliographic coupling on this subset using Python modules Tethne (48) and NetworkX
(49). 880 of our 1,234 papers were coupled bibliographically by at least one paper.
Consistent with the procedure we applied to measure shared authors and methods, we
used JC to measure the overlap between any two sets of papers' citations and then
employed our independence score to measure claim's prior knowledge independence
𝐾!"#.
Centralization of scientific communities
To quantify the centralization of scientific communities in the bipartite author-article
networks for each claim, we employed the Gini coefficient. The Gini coefficient is used
to measure the heterogeneity of distributions in social and information networks (50).
The coefficient ranges between 0 and 1. In the context of a bipartite author-article
network, the coefficient approaches 0 when all investigators author equal numbers of
articles about a claim and increases to 0.3 and above (depending on the number of
articles), when one investigator authors all articles and all others author only one. The
Gini coefficient can be also represented as a percentage ranging from 0 to 100, as in
Figure 2D-E. While other measures of network centralization are available (e.g.,
Freeman's centralization (51)), the Gini coefficient, and the Lorenz curve on which it is
based, is independent from the underlying degree distribution, making it suitable for
comparisons among networks with different size and mean degree (50, 52).
19
Journal prominence
To measure journal prominence, we employed the journal eigenfactor score (40). The
eigenfactor score acts like a recursively weighted degree index by rating journals highly
that receive citations from journals that are themselves highly cited. Using the Web of
Science database, we computed journal eigenfactor scores for 3,162 papers (94% of all
3,363 papers in our corpus) published in 656 journals between 1995 and 2016. For the
sub-corpus of 1,282 papers with shared drug-gene claims, we recovered 1,212 papers or
95% published in 496 journals. For each claim, we computed mean journal eigenfactor
scores by averaging over the eigenfactor score of all journals that published a paper
reporting findings in support of the claim. The distribution of mean journal eigenfactor
scores (i.e., journal prominence) per claim is highly skewed (Figs. S4 and S5), indicating
that claims receive overwhelmingly support from findings published in low and medium
ranked journals. This highlights the value of archiving findings across a wide range of
journals, as do CTD and other scientific database projects, which makes possible our
large-scale evaluation of scientific output.
Shared authors and scientific agreement
To examine whether papers with shared authors are more likely to agree on the direction
("increasing expression" or "decreasing expression") of the drug-gene interaction, we
identified all pairs of papers with shared or distinct authors that reported findings about
the same drug-gene pair irrespective of the directionality of the effect. We found that
among the 2,514 edges between papers with shared authors, 2,486 (98.9%) agreed on the
direction of the effect. By comparison, among the 20,846 pairs of papers with distinct
authors, 18,527 (88.9%) agreed on the direction of the effect. To examine the differences
in agreement rates among papers with common versus distinct authors, we performed a
two-tailed bootstrap test (100,000 samples) (Fig. S2).
Predictors of replication success
We estimate logistic regression models to predict claim replication R [Replicated = 1,
Non-replicated = 0] as a function of support in the literature 𝐿!"#$, social independence
𝑆!"#, methodological independence 𝑀!"#, prior knowledge independence 𝐾!"#,
centralization 𝐶, journal prominence 𝐽, and experimental variability 𝑉. First, for
exploratory purposes, we model each variable independently (see Table S1):
logit P𝑅=1 =𝛽!+𝛽!× 𝐿!"#$
𝛽!+𝛽!× 𝑆!"#
𝛽!+𝛽!× 𝑀!"#
𝛽!+𝛽!× 𝐾!"#
𝛽!+𝛽!× 𝐶
𝛽!+𝛽!× 𝐽
𝛽!+𝛽!× 𝑉.
Second, we model our variables simultaneously (see Table S2):
logit Ρ𝑅=1 =𝛽!+𝛽!×𝐿!"#$+𝛽!×𝑆!"#+𝛽!×𝑀!"#+𝛽!×𝐾!"#+𝛽!×𝐶+𝛽!×𝐽+𝛽!×𝑉.
20
Third, we estimate two interaction models to examine the effect of support in the
literature on replication success as a function of social independence and centralization,
respectively (see Figs. 3 and S8–9):
logit P𝑅=1 =𝛽!+𝛽!×𝐿!"#$× 𝑆!"#+𝛽!×𝑉
𝛽!+𝛽!×𝐿!"#$× 𝐶+𝛽!×𝑉.
To estimate and visualized the logistic regression models, we used the "glm" function
and the "effects" (53) and "sjPlot" (54) packages for R.
Data Analysis
Robustness analysis
Some of our variables are correlated (e.g., methodological independence and prior
knowledge independence; see Fig. S7), which is to be expected as they capture related
dimensions of scientific knowledge production. We performed a multicollinearity test
using the variance inflation factor (VIF). The variance inflation factors vary from low for
variability in LINCS L1000 (VIF = 1.015), journal prominence (VIF = 1.135), and
support in literature (VIF = 1.534) to moderate for centralization (VIF = 3.162),
methodological independence (VIF = 3.752), prior knowledge independence (VIF =
4.681), and social independence (VIF = 4.900). We observe no predictor with high
variance inflation factor, i.e., VIF ≥ 10. We removed the two variables with the highest
VIF > 4 and refit our logistic regression model. In the refitted model, both support in the
literature and decentralization of scientific communities remain strong and significant
predictors of replication success (Fig. S10). Further, we verified that the effects of
support from the literature and community centralization on claim replication success are
not dominated by outliers. Recall the long-tailed distribution of findings per claim, with
few claims receiving support from many published findings (Fig. 1A). We removed
claims supported by 10 or more findings, amounting to 26 claims supported by 426
findings and found that support from the literature (OR 34.785; 95% CI: 12.518, 96.662,
P = 1.00e-11) and community centralization (OR 0.335; 95% CI: 0.249, 0.451, P =
5.57e-13) remain strong and significant predictors of replication success, after controlling
for biological and experimental variability in LINCS L1000 (OR 0.726; 95% CI: 0.429,
1.232; P = 0.236). Similarly, the distribution of findings per paper and per pair of papers
is heterogeneous (Fig. S11). We removed the largest set of 796 drug-gene claims reported
by a pair of papers and still found that the effect of support from the literature (OR
19.512; 95% CI: 7.193, 52.929, P = 5.37e-09) and centralization (OR 0.432; 95% CI:
0.287, 0.65, P = 5.71e-05) holds and is not explained by variability in LINCS L1000 (OR
0.574; 95% CI: 0.279, 1.181; P = 0.131).
21
A
Small
science
Big
science
B
Small
science
s
e
i
t
i
n
u
m
m
o
c
f
o
r
e
b
m
u
N
104
103
102
101
100
100
N = 51,292
101
102
Size of scientific communities
reporting a claim (# of authors)
s
e
i
t
i
n
u
m
m
o
c
f
o
r
e
b
m
u
N
102
101
100
100
Big
science
N = 2,493
101
102
Size of scientific communities
reporting a claim (# of authors)
Fig. S1. A continuum of "small" to "big" science communities defined as the total
number of authors that have published papers reporting a drug-gene claim. (A-B)
Distribution of scientific communities of different sizes against the number of
communities for the corpus of 51,292 and for the sub-corpus of 2,493 claims,
respectively.
22
A
y
c
n
e
u
q
i
r
F
8000
6000
4000
2000
Papers with
shared authors
Papers with
distinct authors
B
8000
6000
4000
2000
0
0.85
0.9
0.95
1
Agreement rate among pairs of papers
0
Difference between agreement rates
0.05
0.1
0
Fig. S2. Publications that share authors are more likely to agree about the direction
of a drug-gene interaction than publications with distinct authors, computed among
pairs of papers reporting claims in the sub--‐corpus
of
2,493 claims. (A) Bootstrap
samples of agreement (1 = agreement, 0 = disagreement) among pairs of papers that
reported findings about the same drug-gene interaction depending on whether the pairs of
papers shared (2,514 pairs of papers) or did not share author(s) in common (20,846 pairs
of papers). (B) A two-tailed bootstrap test of difference between means indicates that
papers with one or more authors in common have significantly higher agreement rates
than pairs of papers with distinct authors (100,000 samples with replacement). The solid
dark line indicates the sample mean, the dashed line indicates 95% CI, and the red line
indicates the null hypothesis.
23
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
A
n
o
i
t
y
t
i
l
i
a
c
i
l
b
a
b
o
r
P
p
e
r
i
m
a
c
f
l
o
B
n
o
i
t
y
t
i
l
i
a
c
i
l
b
a
b
o
r
P
p
e
r
i
m
a
c
f
l
o
OR = 2.33; 95% CI: 1.82, 2.97
P < 0.0001
N = 51,292
0.0
0.4
0.6
0.2
0.8
Support in literature
N = 21,181
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
y
t
i
l
i
b
a
b
o
r
P
1.0
o
n
o
i
t
OR = 3.50; 95% CI: 2.40, 5.11
P < 0.0001
y
t
i
l
i
b
a
b
o
r
P
a
c
i
l
p
e
r
i
m
a
c
f
l
o
1.0
0.0
0.6
0.4
0.2
0.8
Support in literature
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
OR = 0.04; 95% CI: 0.002, 0.72
P = 0.029
0.2
0.4
0.0
1.0
Variability in LINCS L1000
0.6
0.8
OR = 0.56; 95% CI: 0.44, 0.72
P < 0.0001
0.2
0.4
0.0
1.0
Variability in LINCS L1000
0.6
0.8
Fig. S3. Replication success increases with claim's probability of support in the
literature. (A) Probability of claim replication estimated via a logistic regression model
with replication as the response variable and support in the literature and experimental
variability as predictors, for the whole corpus of 51,292 claims. (B) Same as A but for the
subset of 21,181 significant drug-gene interactions in LINCS L1000. For each variable,
we indicate odds ratios (OR), 95% confidence interval, and significance.
24
B
100
Supporting
Opposing
A
s
g
n
d
n
i
i
f
f
o
r
e
b
m
u
n
n
a
e
M
l
i
m
a
c
r
e
p
10.0
7.5
5.0
2.5
0.0
Low
Moderate
High
Very High
Contradictory
Types of claims
by support
Very High
support
N = 325
)
n
o
%
i
t
(
s
m
a
c
u
b
i
r
t
s
d
i
i
l
e
n
e
g
-
g
u
r
d
e
v
i
t
l
a
u
m
u
C
f
o
High
support
N = 1,083
Moderate
support
N = 4,743
80
60
40
20
Low
support
N = 46,064
Contradictory
N = 77
0
0
0.4
0.2
0.8
Probability of support in literature
0.6
1
Fig. S4. Description of claim types in the whole
corpus
of
51,292
claims. (A) Mean
number of findings supporting or opposing the direction of the effect for a scientific
claim across our typology. (B) Cumulative distribution of the probability of support in the
literature, including a schematic of our claim typology.
25
y
t
i
s
n
e
D
30
20
10
0
0
0.3
0.2
0.1
y
t
i
s
n
e
D
N = 51,292
0.5
Support in literature
1
y
t
i
s
n
e
D
30
20
10
0
0
Abs. combined z-scores
N = 51,292
y
t
i
s
n
e
D
0
0
Effect sizes in LINCS L1000
150
100
50
N = 50,346
12
Journal prominence
× 10-5
N = 51,292
2
1.5
1
0.5
0
Variability in LINCS L1000
46
2
× 104
Fig. S5. Estimates of probability density functions for variables of interest in our
corpus using a normal kernel function. We used the normal kernel function to estimate
probability density functions.
26
20
10
0
0
10
5
0
-0.5
10
5
0
-0.2
y
t
i
s
n
e
D
y
t
i
s
n
e
D
y
t
i
s
n
e
D
y
t
i
s
n
e
D
N = 2,493
20
10
N = 2,493
0.5
Support in literature
1
N = 2,493
0
0.5
1
Methodological
independence
N = 2,493
0
-0.5
0
1
Social independence
0.5
40
20
0
-0.5
100
50
N = 2,493
0
0.5
1
Prior knowledge
independence
N = 2,491
0
0.2
Centralization
0.4
0
0
12
Journal prominence
0.1
0.05
N = 2,493
0.4
0.2
N = 2,493
0
-200
Effect size in LINCS L1000
-100
100
0
0
20
Variability in LINCS L1000
10
0
Fig. S6. Estimates of probability density functions for our variables for the sub-
corpus of claims with determined direction of the drug-gene effect in CTD and
LINCS L1000. As in Fig. S5, we used the normal kernel function to estimate probability
density functions.
27
Prior knowledge independence
Methodological independence
0.79
Social independence
0.74
0.66
Centralization
−0.62
−0.33
−0.03
Corr
1.0
0.5
0.0
−0.5
−1.0
Centralization
Social independence
Prior knowledge independence
Methodological independence
Fig. S7. Pearson correlation coefficients between network indices. Social,
methodological, and prior knowledge (in)dependencies are positively correlated with
each other and negatively correlated with the network centralization of scientific
communities. N = 2,493 claims.
28
A
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
o
y
t
i
l
i
b
a
b
o
r
P
0
0.2
Support in literature
1
0.5
0.8
1.0
0.8
0.6
0.4
0.2
0.0
C
0.0
0.2
0.6
0.4
Centralization
0.8
1.0
Intercept
Support in
literature*Centralization
Variability in LINCS L1000
B
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
o
y
t
i
l
i
b
a
b
o
r
P
1.0
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Variability in LINCS L1000
B
SE Exp(B)
0.823*** 0.079 0.553
–1.96*** 0.35 28.786
95% CI
[1.951, 2.657]
[0.071, 0.28]
–0.47 0.265 0.704
[0.372, 1.052]
Fig. S8. Claims reported by centralized communities less likely replicate. (A-C),
Logistic interaction model with claim replication as the response variable regressed on
support in the literature and centralization as interacting predictors, controlling for
variability in LINCS L1000 across cell lines, durations, and dosages. Predictors are
rescaled
!"#!!!"# (!) for comparability. N = 2,493 claims. *** P < 0.001; ** P < 0.01;
!!!!"# (!)
* P < 0.05.
29
A
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
o
y
t
i
l
i
b
a
b
o
r
P
0
0.2
Support in literature
1
0.5
0.8
1.0
0.8
0.6
0.4
0.2
0.0
C
0.0
0.4
0.2
0.8
Social independence
0.6
Intercept
Support in literature*Social
independence
Variability in LINCS L1000
B
n
o
i
t
a
c
i
l
p
e
r
i
m
a
c
f
l
o
y
t
i
l
i
b
a
b
o
r
P
1.0
1.0
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Variability in LINCS L1000
B
SE Exp(B)
–0.59*** 0.148 0.553
3.36*** 0.384 28.786
95% CI
[0.414, 0.738]
[13.56, 61.1]
–0.351 0.268 0.704
[0.417, 1.189]
Fig. S9. Claims reported by multiple socially independent teams more likely
replicate. (A-C) Logistic interaction model with claim replication as the response
variable regressed on support from the literature and social independence as interacting
predictors, controlling for variability in LINCS L1000 across cell lines, durations, and
dosages. Predictors are rescaled
!!!!"# (!)
*** P < 0.001; ** P < 0.01; * P < 0.05.
!"#!!!"# (!) for comparability. N = 2,493 claims.
30
Support in literature
20.72 ***
Journal prominence
2.69
Methodological independence
Variability in LINCS L1000
1.23
0.66
Centralization
0.39 ***
0.2 0.5 1 2
Claim replication (OR, 95% CI)
5 10 20 50
Fig. S10. Support in the literature and decentralization of scientific communities
remain strong and significant predictors of claim replication success after we
account for multicollinearity. Logistic model with replication as the response variable,
including only predictors with relatively low variance inflation factor (VIF < 4).
Predictors are rescaled
!!!!"# (!)
!"#!!!"# (!) for comparability. N = 2,491 claims. *** P < 0.001;
** P < 0.01; * P < 0.05.
31
A
s
r
e
p
a
p
f
o
r
e
b
m
u
N
102
101
100
100
101
102
Number of findings
reported by a paper
B
s
r
e
p
a
p
f
o
s
r
i
a
p
f
o
r
e
b
m
u
N
105
104
103
102
101
100
100
Number of findings reported by
101
102
a pair of papers
Fig. S11. Papers and pairs of papers are differentiated by the number of findings
they report (in the sub-corpus of 2,493 claims). (A) Distribution of papers against the
number of supporting findings they report. (B) Distribution of pairs of papers against the
number of supporting findings they report.
32
Variable
B
SE Exp(B)
95% CI
Intercept
Support in literature
Intercept
Journal prominence
Intercept
Social independence
Intercept
Methodological independence
Intercept
Prior knowledge independence
Intercept
Variability in LINCS L1000
Intercept
Centralization
–0.762*** 0.195
3.144*** 0.479
0.47
23.2
[0.318, 0.684]
[9.078, 59.287]
0.414***
1.513
2.379*** 0.564 10.795
0.046
[1.384, 1.654]
[3.571, 32.638]
–1.053***
0.193
1.843*** 0.224
0.349
6.312
[0.239, 0.509]
[4.069, 9.793]
–1.004***
0.255
1.841*** 0.308
–1.14**
1.71***
0.585***
-0.444
0.376
0.388
0.065
0.264
0.8***
0.062
–1.018*** 0.149
0.366
6.301
0.32
5.53
1.794
0.641
2.224
0.361
[0.222, 0.604]
[3.444, 11.527]
[0.153, 0.668]
[2.584, 11.835]
[1.579, 2.039]
[0.382, 1.077]
[1.972, 2.51]
[0.27, 0.484]
33
Table S1. Logistic regression models with claim replication R [Replicated = 1, Non-
replicated = 0] as response variable and predictors modelled independently.
Predictors are rescaled
!!!!"# (!)
!"#!!!"# (!) for comparability. N = 2,493 claims (N = 2,491 in
the Journal eigenfactor model); *** P < 0.001; ** P < 0.01; * P < 0.05.
Variable
B
SE Exp(B)
95% CI
Intercept
Support in literature
Journal prominence
Social independence
Methodological independence
Prior knowledge independence
Variability in LINCS L1000
Centralization
0.5
–0.693* 0.352
2.729*** 0.595 15.316
2.769
1.881
0.833
0.971
0.664
0.48
0.563
0.486
0.585
0.769
0.271
–0.734** 0.268
1.019
0.632
–0.183
–0.029
–0.41
[0.251, 0.996]
[4.772, 49.153]
[0.918, 8.355]
[0.725, 4.877]
[0.265, 2.621]
[0.215, 4.383]
[0.391, 1.128]
[0.284, 0.812]
Table S2. Logistic regression models with claim replication R [Replicated = 1, Non-
replicated = 0] as response variable and predictors modelled simultaneously.
Predictors are rescaled
!!!!"# (!)
!"#!!!"# (!) for comparability. N = 2,491 claims; *** P < 0.001;
** P < 0.01; * P < 0.05.
34
|
1805.03348 | 1 | 1805 | 2018-05-09T02:01:19 | Wisdom in Sum of Parts: Multi-Platform Activity Prediction in Social Collaborative Sites | [
"cs.SI",
"cs.DL",
"cs.SE"
] | In this paper, we proposed a novel framework which uses user interests inferred from activities (a.k.a., activity interests) in multiple social collaborative platforms to predict users' platform activities. Included in the framework are two prediction approaches: (i) direct platform activity prediction, which predicts a user's activities in a platform using his or her activity interests from the same platform (e.g., predict if a user answers a given Stack Overflow question using the user's interests inferred from his or her prior answer and favorite activities in Stack Overflow), and (ii) cross-platform activity prediction, which predicts a user's activities in a platform using his or her activity interests from another platform (e.g., predict if a user answers a given Stack Overflow question using the user's interests inferred from his or her fork and watch activities in GitHub). To evaluate our proposed method, we conduct prediction experiments on two widely used social collaborative platforms in the software development community: GitHub and Stack Overflow. Our experiments show that combining both direct and cross-platform activity prediction approaches yield the best accuracies for predicting user activities in GitHub (AUC=0.75) and Stack Overflow (AUC=0.89). | cs.SI | cs | Wisdom in Sum of Parts: Multi-Platform Activity
Prediction in Social Collaborative Sites
Roy Ka-Wei Lee
David Lo
Living Analytics Research Centre
Singapore Management University
[email protected]
Singapore Management University
[email protected]
8
1
0
2
y
a
M
9
]
I
S
.
s
c
[
1
v
8
4
3
3
0
.
5
0
8
1
:
v
i
X
r
a
ABSTRACT
In this paper, we proposed a novel framework which uses user in-
terests inferred from activities (a.k.a., activity interests) in multiple
social collaborative platforms to predict users' platform activities.
Included in the framework are two prediction approaches: (i) direct
platform activity prediction, which predicts a user's activities in a
platform using his or her activity interests from the same platform
(e.g., predict if a user answers a given Stack Overflow question
using the user's interests inferred from his or her prior answer and
favorite activities in Stack Overflow), and (ii) cross-platform activity
prediction, which predicts a user's activities in a platform using his
or her activity interests from another platform (e.g., predict if a user
answers a given Stack Overflow question using the user's interests
inferred from his or her fork and watch activities in GitHub). To
evaluate our proposed method, we conduct prediction experiments
on two widely used social collaborative platforms in the software
development community: GitHub and Stack Overflow. Our experi-
ments show that combining both direct and cross platform activity
prediction approaches yield the best accuracies for predicting user
activities in GitHub (AUC=0.75) and Stack Overflow (AUC=0.89).
KEYWORDS
Social Collaborative Platforms; Prediction; Stack Overflow; GitHub
ACM Reference Format:
Roy Ka-Wei Lee and David Lo. 2018. Wisdom in Sum of Parts: Multi-Platform
Activity Prediction in Social Collaborative Sites. In WebSci'18: 10th ACM
Conference on Web Science, May 27–30, 2018, Amsterdam, Netherlands. ACM,
New York, NY, USA, 10 pages. https://doi.org/10.1145/3201064.3201067
1 INTRODUCTION
Software developers are increasingly adopting social collaborative
platforms for software development. GitHub and Stack Overflow are
two of such popular platforms. GitHub is a collaborative software
development platform that allows code sharing and version control.
Users can participate in various activities in GitHub, for example,
users may fork (i.e., create a copy of) repositories of other users or
watch the activities of repositories of interest. Stack Overflow is a
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].
WebSci'18, May 27–30, 2018, Amsterdam, Netherlands
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-5563-6/18/05...$15.00
https://doi.org/10.1145/3201064.3201067
technical question-and-answer community-based website where
users post and answer questions relating to software development.
As these social collaborative platforms gain popularity, many
research studies have proposed recommender systems to improve
the usability of these platforms. For example, there are work which
predict and recommend relevant Stack Overflow questions and
answers to aid users in software development [7, 32, 34]. While
for GitHub, researchers have proposed methods to predict which
software repositories are more relevant to a target user [9, 11, 43].
Nevertheless, many of these studies only consider the users' be-
haviours and interests in a single platform when predicting and
recommending user platform activities.
There have been few existing inter-platform studies on GitHub
and Stack Overflow. Vasilescu et al. [29] studied how users' involve-
ment in Stack Overflow impacted their productivity in GitHub.
Badashian et al. [2] did an empirical study on the correlation be-
tween different types of user activities in GitHub and Stack Over-
flow. In a more recent study by Lee and Lo [15], the researchers
found that users who have accounts on both GitHub and Stack
Overflow do share similar interests across the two platforms. For
example, a user who commits to Java-related repositories in GitHub,
is likely to also answer Java-related questions in Stack Overflow.
In this paper, we aim to extend the study in [15], and propose a
multi-platform activity prediction method, which predicts a user's
activities in a platform using the user's interests inferred from his
or her activities in multiple platforms.
Figure 1 illustrates an example for activity prediction in a multi-
platform setting. Consider user u, who has accounts on both GitHub
and Stack Overflow. If we adopt a direct platform activity prediction
approach, i.e., predicts a user's activities in a platform using his or
her activity interests from the same platform, we could predict that
u is likely to answer or favorite1 question X in Stack Overflow as u
has previously answered a LSTM related question. However, if we
adopt a cross-platform activity prediction approach, i.e., predicts a
user's activities in a platform using his or her activity interests from
another platform, we could predict that u is also likely to answer
or favorite a SVM related question Y as u has previously watched2
a SVM related repository B in GitHub.
The social collaborative nature of these platforms could also
be exploited for activity prediction. According to Lee and Lo [15],
users who participated in the same GitHub repositories and Stack
Overflow questions tend to share common interests. Therefore, we
also explore the possibility to expand a user's interests to include
the interests of users whom he or she had co-participated activities
with. Referencing to the same example in Figure 1, user v co-forked
1Bookmark a question in Stack Overflow
2Subscribe and receive updates on a repository in GitHub
Figure 1: Example of Activity Prediction in Multi-Platform Setting
the same scikit-learn related repository as u and forked another
XGBoost related repository. Although u did not participate in any
XGBoost related repositories and questions, we can "expand" u's
interests to include XGBoost as it is an interest of user v. Finally, we
could predict that u is likely to also answer or favorite a XGBoost
related question. This expansion of interests could be particularly
useful when a user has participated in very few activities on either
platform. Note that the example also work for predicting GitHub
activities using user's activities in Stack Overflow.
There are a number of benefits for using user interests from
multi-platforms for activity prediction. Firstly, it enables predic-
tion and recommendation of user activities in social collaborative
platforms even when past activity history of a user is minimal or
unavailable, i.e, cold-start problem [23]. For example, if we learn
from a user's activities in GitHub that she is interested in Python
and text mining techniques, we would predict that she will likely
participate in Python and text mining related Stack Overflow ques-
tions even when she has just newly joined Stack Overflow and has
not participated in any questions. Second, it could cover the blind
spots of activity recommender systems which use only data from a
single platform. For example, if a user has forked Android related
repositories in GitHub, recommender systems which are built on
user's past activity in GitHub will likely to recommend the user
more Android related repositories. However, the same user may
have also participated in some iOS related questions in Stack Over-
flow, and such observations can be used to make relevant GitHub
activity recommendations to the user.
Contributions. This work improves the state-of-the-art of inter-
platform studies on multiple social collaborative platforms. Key
contributions of this work include: Firstly, we proposed a novel
framework which enables predicting users' activities using inter-
ests inferred from their activities in multiple social collaborative
platforms. Secondly, we evaluate our method using large real-world
datasets from Stack Overflow and GitHub. The results from our
prediction experiments show that our proposed method is able to
predict users' activities in GitHub and Stack Overflow with good
accuracy, achieving an AUC score of up to 0.75 and 0.89 respectively.
Paper outline. The rest of the paper is organized as follows.
Section 2 introduces the social collaborative platform activity pre-
diction problem and describe our proposed the multi-platform pre-
diction framework. Section 3 describe the data extraction process
and the two real-world datasets, Stack Overflow and GitHub, that
we used in our prediction experiments. Section 4 presents our ex-
periments to predict user activities in the two social collaborative
platforms using our proposed framework. Threat to validity of our
study are discussed in Section 5. Section 6 reviews the literature
related to our study. Finally, we summarize and conclude our work
in Section 7.
2 PROPOSED METHOD
In this section, we first present our proposed multi-platform activity
prediction framework. We then define the prediction problem and
describe the features used in our proposed prediction method.
2.1 Multi-Platform Prediction Framework
Figure 2 shows the framework that we adopt for multi-platform
activity prediction. We begin with data extraction from two social
collaborative platforms: Stack Overflow and GitHub. There are
three sub-processes in data extraction: (i) matching of users Stack
Overflow and GitHub accounts, (ii) extracting the users' platform
activities, and (iii) inferring users' interests from their activities.
The details of these sub-processes will be covered in Section 3. Next,
we construct the Stack Overflow and GitHub user features which
we will use in our prediction.
Our framework also incorporates two approaches to predict
users' platform activities, namely: direct and cross platform activity
prediction. We define direct platform activity prediction as predict-
ing a user's platform activity using features from the same platform.
For example, we predict if a given user will answer a given Stack
Overflow question using the user's Stack Overflow features. Con-
versely, we define cross-platform activity prediction as predicting a
platform activity to a user using features from a different platform.
For example, we predict if a given user will answer a given Stack
Overflow question using the user's GitHub features. The perfor-
mance of both prediction approaches will be evaluated on four
prediction tasks, which will be described in Section 4.
Repository A<LSTM>Repository B<SVM>Repository C<scikit-learn>Repository D<XGBoost>Question Z<XGBoost>Question Y<SVM>Question X<LSTM>Question W<LSTM>user uuser vforkwatchforkforkforkansweranswer/favorite?favoriteanswer/favorite?answer/favorite? Figure 2: Cross-Platform Activity Prediction Framework
2.2 Problem Statement
Given a pair of query user and item (i.e., question or repository),
(u, k), we aim to predict if u will perform an activity (e.g. answer,
favorite, fork or watch) on k. There are various ways to measure
the likelihood of u performing an activity on k. For example, we
could consider the similarity between k's description and u's inter-
ests inferred from different activities, or the similarity between k's
description and the inferred interests of the user who co-participate
activities with u. In our proposed framework, we propose two types
of user features, namely:user activity interest similarity features
and user co-activity interest similarity features. The notations used
throughout this paper are summarized in Table 1.
We denote the estimated interests of a user given a repository
r that he or she forked and watched in GitHub as I(r). Similarly,
we denote the estimated interests of a user given a question q that
he or she answered and favorited in Stack Overflow as I(q). Since
the estimated interests given a repository or a question is the same
for all users participated in it, we also refer to I(r) and I(q) as the
interests in r and q. For simplicity, we also refer to them as r's
interests and q's interests respectively.
2.3 User Activity Interest Similarity Features
This set of features measures the similarity between a query item
k and a query user u's fork, watch, answer and favorite activity
interests in GitHub and Stack Overflow. The intuition behind this
set of features comes from the empirical study from Lee and Lo
[15], where they found that users in GitHub and Stack Overflow
shared similarities between their interests in different types of
activities and across the two platforms. Suppose that we want to
predict if a user would fork a given repository in GitHub, we would
measure the similarity between the given repository's interests and
the developer's interests for the different activity types. Intuitively,
Table 1: List of notations used
Symbol
u
k
v
r
q
I(r)
I(q)
I(k)
u.RF
u.RW
u.QA
u.QF
CoF ork(u)
CoW atch(u)
CoAns(u)
CoF av(u)
Description
Query user
Query item
User who co-participated activities with user u
Repository
Question
Interests of repository r
Interests of question q
Interests of query item k
Set of repositories forked by user u
Set of repositories watched by user u
Set of questions answered by user u
Set of questions favorited by user u
Set of users who co-forked at least one reposi-
tory with user u
Set of users who co-watched at least one repos-
itory with user u
Set of users who co-answered at least one ques-
tion with user u
Set of users who co-favorited at least one ques-
tion with user u
the higher the similarity scores, the more likely the user would fork
the given repositories. Equation 1 captures the above intuition and
measures similarity between k and u's fork activity interests (i.e.,
SimF ork(u, k)), by dividing {r ∈ u.RFI(r) ∈ I(k))}, which is the
number of u's forked repositories that shared common interests
with the item interests of k, by the total number of repositories
forked by u (i.e.,u.RF).
Answer FavoriteQuestionsGitHubUsersForkWatchRepositoriesStack Overflow UsersMatch Users with SO and GH AccountsInfer User Interests in Stack OverflowInfer User Interests in GitHubAnswer Interests SimilarityFavorite Interests SimilarityCo-Answer Interests SimilarityCo-Favorite Interests SimilarityStack Overflow FeaturesFork Interests SimilarityWatch Interests SimilarityCo-Fork Interests SimilarityCo-Watch Interests SimilarityGitHub FeaturesData ExtractionFeature ConstructionEvaluationStack OverflowTask S1: Predict if user answer a given questionTask S2: Predict if user favorite a given questionGitHubTask G1: Predict if user fork a given repositoryTask G2: Predict if user watch a given repositoryDirect PlatformDirect Platform Cross PlatformPredictionCross PlatformStack OverflowGitHub Repository C), while the total number of repositories forked by v
1
is 2 (i.e., Repository C and D). Finally, SimCoF ork(u, k) =
1 = 0.5.
2
SimCoF ork(u, k) =
(1)
(2)
(3)
SimCoW atch(u, k) =
SimCoAns(u, k) =
v ∈CoF or k(u)
{r ∈v .RF I(r)∈I(k)}
v .RF
CoF ork(u)
v ∈CoW atch(u)
{r ∈v .RW I(r)∈I(k)}
v .RW
CoW atch(u)
v ∈CoAns(u)
{q∈v .QAI(q)∈I(k)}
v .QA
CoAns(u)
(cid:21)
(cid:21)
(cid:21)
(cid:21)
(5)
(6)
(7)
Example. Referencing to the earlier example in Figure 1, we
could predict if user u will answer question X by computing the
similarity between question X and u's fork activity interests. In this
example, the common interests between u and question X will be
LSTM. The number of u's forked repositories that shared common
interests with question X (i.e.,{r ∈ u.RFI(r) ∈ I(k))}) will then be
1 (i.e., Repository A), while the total number of repositories forked
by u is 2 (i.e., Repository A and B). Thus, SimF ork(u, k) = 1
2 = 0.5.
SimF ork(u, k) =
SimW atch(u, k) =
SimAns(u, k) =
{r ∈ u.RFI(r) ∈ I(k))}
{r ∈ u.RW I(r) ∈ I(k)}
{q ∈ u.QAI(q) ∈ I(k)}
u.RF
u.RW
u.QA
{q ∈ u.QFI(q) ∈ I(k)}
SimF av(u, k) =
u.QF
(4)
We compute the similarities between k and u's watch, answer
and favorite activities interests in similar ways as shown in Equation
2, 3 and 4 respectively.
2.4 User Co-Activity Interest Similarity
Features
This set of features measures the similarity between a query item k
and the activity interests of other users v who have co-participated
in an activity with a query user u. The intuition behind this set
of features also comes from the empirical study from Lee and Lo
[15], where they found that users share similar interests with other
users who they co-participated an activity (even minimally) in a
social collaborative platform. Suppose that we want to predict if a
user would fork a given repository in GitHub, we would measure
the similarity between the given repository's interests and the
interests of other users who had co-forked repositories with the
user in GitHub. Intuitively, we would also expect that the higher the
similarity score, the more likely the user would answer the given
question. Equation 5 captures the above intuition and measures the
average similarity between k and fork activity interests of all users
v, who had co-forked at least one question with u (i.e., CoF ork(u)).
As users also share common interests across different activities
and platforms, we would expect that considering other users who
had co-participated in other types of platform activities with the
target user can also potentially help to predict if the target user
would participate in a given platform activity. For instance, we are
potentially able to predict if a user would fork a given repository
by measuring the similarity between the given repository's interest
and the interests of other users who have co-participated with the
user in watch, answer and favorite activities.
Example. Referencing to the example in Figure 1, we could
predict if user u will favorite question Z by computing the similarity
between question Z and the fork activity interests of other users
who have co-fork a repository with user u. Assuming that user u
only has 1 other user, v, who co-fork repositories with him or her,
the common interests between v and question Z will be XGBoost.
The number of v's forked repositories that shared common interests
with question Z (i.e.,{r ∈ v.RFI(r) ∈ I(k))}) will then be 1 (i.e.,
(cid:20)
(cid:20)
(cid:20)
(cid:20)
v ∈CoF av(u)
{q∈v .Q F I(q)∈I(k)}
v .Q F
SimCoF av(u, k) =
We compute the similarities between k and activity interests of
other users v who have co-watched, co-answered and co-favorited
with a target user u in similar ways as shown in Equation 6, 7 and
8 respectively.
CoF av(u)
(8)
3 DATA EXTRACTION & EXAMINATION
In this section, we first introduce the two large real-world datasets
that we use in our activity prediction experiments. Next, we dis-
cuss the user accounts linkage process to retrieve users who are
active in multiple social collaborative sites, and a summary of the
users' activities retrieved. We then discuss the heuristic used to
infer user interests from their participated activities. Finally, we
empirically examine the similarity between the GitHub repositories
and Stack Overflow questions participated by users on both social
collaborative sites.
3.1 Datasets
There are two main datasets used in our study. For the GitHub
dataset, we use the MongoDB database dump released on March
2015 [8]. The dataset contains GitHub activities from October 2013
to March 2015 of about 2.5 million users. Specifically, we are inter-
ested in the fork and watch repositories activities of the GitHub
users. For Stack Overflow, we use the XML dataset released on
March 20153. This dataset contains information of estimated 1 mil-
lion Stack Overflow users and their activities from October 2013
to March 2015. We are particularly interested in the answer and
favorite activities of the Stack Overflow users.
3.2 User Account Linkage
As this study intends to investigate user interests across GitHub
and Stack Overflow, we need to identify users who were using both
platforms. For this work, we used the dataset provided by Badashian
et al. [2], where they utilized GitHub users' email addresses and
Stack Overflow users' email MD5 hashes to find the intersection
between the two datasets. We also filter out users who do not have
3https://archive.org/details/stackexchange
at least 1 activity on both platforms between October 2013 and
March 2015. In total, we identify 92,427 users, which forms our base
users set. After the base users have been identified, we extract their
GitHub and Stack Overflow activities from the datasets. In total, we
have extracted 416,171 fork, 2,168,871 watch, 766,315 answer and
427,093 favorite activities from the base users.
Figure 3: Base users' Stack Overflow and GitHub Activity
Distributions
Figure 3 shows the distributions of base users' activities in GitHub
and Stack Overflow. Most of the base users forked 1-10 repositories
(64% of the base users), answered 1-10 questions (54% of the base
users). There are also quite a number of developers who watched
11-100 repositories (26% of the base users). We also observe that
more than half of base users have at least answered 1 questions
(71%) and a substantial number of base users also answered 11-100
questions (12%).Interestingly, the high contribution of answers to
questions in Stack Overflow could also suggest that the many of
these active developers in our study were experts in their domain or
areas of interest. Lastly, we also notice that there are developers (al-
beit very few in number) who were extremely active in GitHub and
Stack Overflow; they forked, watched, committed, pull-requested
more than 1000 repositories, or asked, answered and favorited more
than 1000 questions.
3.3 Inferring User Interests
Next, we infer user interests by observing repositories and questions
that users participated in GitHub and Stack Overflow. We use the
following heuristics to infer user interests:
(1) To infer user interests in Stack Overflow, we use the descrip-
tive tags of the questions that they answered and favorited.
For example, consider a question related to mobile program-
ming for Android smartphones which contain the following
set of descriptive tags: {Java, Android}. If a user answered,
or favorited that question, we infer that his interests include
Java and Android.
(2) In the time period covered in our dataset, GitHub does not
allow users to tag repositories but it allows users to describe
their repositories. These descriptions often contain impor-
tant keywords that can shed light on user interests. To infer
user interests from the repositories that a user had partic-
ipated, we first collect all descriptive tags that appear in
our Stack Overflow dataset. In total, 39,837 unique descrip-
tive tags are collected. Next, we perform keyword matching
between the collected Stack Overflow tags and a GitHub
repository description. We consider the matched keywords
as the inferred interests. We choose to use Stack Overflow
tags to ensure that developer interests across the two plat-
forms can be mapped to the same vocabulary.
3.4 Similarity between GitHub repositories and
Stack Overflow Questions
In [15], Lee and Lo had empirically studied the similarity between
user activity interests in GitHub and Stack Overflow. We extend
their study by examining the descriptive tags of GitHub repositories
and Stack Overflow questions participated by the base users. The
objective is to investigate what are the popular descriptive tags
used by the users on two sites and if there are overlaps among the
popular descriptive tags.
Table 2: Top 10 most used descriptive tags in Stack Overflow
and GitHub
Stack Overflow
% Questions
GitHub
% Repositories
Rank
1
2
3
4
5
6
7
8
9
10
Tag
javascript
java
python
c
php
android
jquery
ios
ruby
css
3.758
3.104
2.760
1.940
1.884
1.787
1.577
1.459
1.031
0.963
Tag
javascript
ruby
python
c
java
objective-c
php
go
css
shell
8.163
2.698
2.604
2.163
1.928
1.731
1.330
1.275
1.236
0.884
Table 2 shows the top 10 most used descriptive tags and the
percentages of Stack Overflow questions and GitHub repositories
containing these tags. We observe quite a significant number of
overlap in the top 10 descriptive tags between the two social col-
laborative sites (known as overlapped tags); i.e., javascript, java,
python, c, php, ruby and css. This suggests that generally, the base
users participated in questions and repositories of similar domain
and nature in Stack Overflow and GitHub. Another interesting
observation is the proportions of questions and repositories with
the top 10 descriptive tags; the overlapped tags are used in similar
proportions of Stack Overflow questions and GitHub repositories,
with the exception of javascript, which seems to be more popular
in GitHub (i.e., ∼8% of repositories) than Stack Overflow (i.e., ∼3%
of questions).
We further investigate the usage of these descriptive tags in the
different Stack Overflow and GitHub user activities. Figure 4 shows
the activity bar charts for the most used descriptive tags. For exam-
ple, ∼7% of the repositories watched by the base users contained the
descriptive tag javascript, while only ∼3% of the questions answered
by the base users contained the same tag. We observe that for a
Figure 4: Bar chats of percentages of activities that use the different descriptive tags. Note that "A", "FA", "FO" and "W" denote
answer, favorite, fork and watch activities respectively.
given descriptive tag, the proportion of activities involving the tag
is not uniform even within the same social collaborative site. For
example, the base users answered more Stack Overflow questions
on php than favoriting them. Conversely, for android related ques-
tions, the base users favorited these question more than answering
them. This unevenness in activity proportions is also observed to
be greater for activities involving non-overlaps descriptive tags.
For example, the base users fork and watch significantly more shell
related repositories than answering and favoriting tshell related
questions. This suggests that although the users do participate in
questions and repositories of similar domains, the activity prefer-
ences involving the similar domains varies. Thus, it would be more
natural to learn the user interests at the activity level instead of
aggregating the interests at the platform level.
AFAFOW0.02.55.07.5% activitiesjavascriptAFAFOW0123% activitiesjavaAFAFOW0123% activitiespythonAFAFOW012% activitiescAFAFOW012% activitiesphpAFAFOW012% activitiesandroidAFAFOW0.00.51.01.5% activitiesjqueryAFAFOW0.00.51.01.5% activitiesiosAFAFOW0123% activitiesrubyAFAFOW0.00.51.01.5% activitiescssAFAFOW0.00.51.01.5% activitiesobjective-cAFAFOW0.00.51.0% activitiesgoAFAFOW0.000.250.500.75% activitiesshell 4 EXPERIMENTS
In this section, we describe the supervised prediction experiments
conducted to evaluate our proposed method. Specifically, we con-
sider the following activity prediction tasks:
predict if the user will answer the question
• Answer Prediction. Given a Stack Overflow user-question pair,
• Favorite Prediction. Given a Stack Overflow user-question,
• Fork Prediction. Given a GitHub user-repository, predict if the
• Watch Prediction. Given a GitHub user-repository, predict if
predict if the user will favorite the question
user will fork the repository
the user will watch the repository
4.1 Experiment Setup
Data Selection. For answer prediction task, we retrieve all the Stack
Overflow questions that the base users have answered and define
a positive instance as a user-question pair where a base user had
answered the particular question in Stack Overflow. For negative
instances, we randomly assign a Stack Overflow question to the
base users and check that the randomly assigned pair does not exist
in the positive instance set. For the training datasets used in answer
prediction task, we randomly generated 5,000 negative instances
and randomly selected 5,000 positive instances from the questions
answered by users between October 2013 and June 2014 (9 months).
The same approach was used to generate the positive and negative
instances for test sets using the questions answered by the users
between July 2014 and March 2015 (9 months). Similar approach
was used to generate the user-question and user-repository pairs
for positive and negative instances used in favorite, fork and watch
prediction tasks.
Note that we have repeated the prediction experiments for five
runs, and the random selection of train and tests set are repeated for
each of the runs. Also, although we know the true labels of the user-
question and user-repository pairs, we do not take the labels into
consideration when deriving the values of our proposed features,
i.e., we assume that we do not know the labels of the pairs.
Feature Configuration. To compare the performance of direct
and cross platform activity prediction approaches, we use Support
Vector Machine (SVM) with linear kernel and apply the following
feature sets on all prediction tasks:
• SO_Act: This set of features includes the Answer (Eqn. 3)
and Favorite (Eqn. 4) Interests Similarity scores for a given
user-question or user-repository pair.
• SO_CoAct: This set of features includes the Co-Answer (Eqn.
7) and Co-Favorite (Eqn. 8) Interests Similarity scores for a
given user-question or user-repository pair.
• GH_Act: This set of features includes the Fork (Eqn. 1) and
Watch (Eqn. 2) Interests Similarity scores for a given user-
question or user-repository pair.
• GH_CoAct: This set of features includes the Co-Fork (Eqn. 5)
and Co-Watch (Eqn. 6) Interests Similarity scores for a given
user-question or user-repository pair.
• ALL: This set of features is the union of all features.
4.2 Prediction Results
We measure the prediction accuracy for each feature configuration
by computing the average area under the ROC curve (AUC) over
a set of positive and negative examples drawn from the test set
for each of the five runs. The results for the four prediction tasks
are shown in Figure 5. We observe that feature configuration ALL
performed the best in all prediction tasks, achieving an AUC of 0.89,
0.77, 0.75 and 0.67 for answer, favorite, fork and watch prediction
tasks respectively.
Performance of cross-platform prediction approach. Al-
though the cross-platform prediction approach did not outperform
the direct platform prediction approach in user activity prediction,
they still yield good accuracy. For example, when predicting user's
answer and favorite activities in Stack Overflow, the GitHub user
activity interests similarity features (i.e., GH_Act) has AUC of 0.71
and 0.64 respectively, and when predicting user's fork and watch
activities in GitHub, the Stack Overflow user activity interests simi-
larity features (i.e., SO_Act) has AUC of 0.65 and 0.58 respectively.
The AUC for predicting user's answer activities in Stack Overflow
using user activity interests similarity features (i.e., GH_Act) is ob-
served to be slightly higher than the prediction for other activities.
A possible explanation for this could be the difference between the
nature of user activities; answering a question in Stack Overflow
would require that a user possesses a particular domain expertise,
whereas other activities such as watching a GitHub repository or
favoriting a Stack Overflow question depend on the user's interests.
As such, we observe higher AUC score for predicting answer activity
task as the users' expertise are usually more specialized and less
diverse than their interests.
More interestingly, using cross-platform prediction approach with
user co-activity interests similarity features (i.e., GH_CoAct and
SO_CoAct), have also yielded reasonable prediction accuracies.
For example, when predicting user's answer activities in Stack
Overflow, GH_CoAct has yielded an AUC of 0.62. This suggests
that even with no information about a user's past activities in the
Stack Overflow and only minimal information such as the user's
co-activities in GitHub, we are still able to reasonably predict user's
activity in Stack Overflow. Similar observations are made when
predicting user activities in GitHub using user's co-activities in
Stack Overflow.
4.3 Discussion
The results of the four prediction tasks offer us some insights in
performing recommendations in social collaborative platforms.
Solving cold-starts. The reasonably good accuracies of cross-
platform prediction approach also demonstrate its potential to solve
the cold-start problem; i.e., predicting and recommending a user's
activities without knowing the users' past activity history on the
platform. For example, when predicting user's answer activities in
Stack Overflow, we are able to achieve AUC as high as 0.71 with-
out using any Stack Overflow features (i.e, using GitHub features
GH_Act only). Similar observations were made for fork, watch and
favorite activities.
We further conduct a small case study to retrieve and review
fork predictions of users who did not have any past fork activities.
For example, we successfully predicted that user U420338 would
Figure 5: ROCs for Four Prediction Tasks
forked repository R12172473 in GitHub even when this was the
first repository forked by the user (i.e., no past user fork activity).
Examining into details, we found that R12172473 has description
tags ⟨svд, javascript⟩, and among the 95 questions U420338 had
answered in Stack Overflow, 83 contain the tags ⟨javascript⟩ or
⟨svд⟩ or both. By analyzing U420338's Stack Overflow activities,
our approach can identify his interests, which ultimately help in
predicting the user's GitHub activities.
Heterogeneous behaviors in cross-platform. There are ex-
isting research work on recommendations in cross-platform set-
ting. For example, Yan et al.[38] addressed the cold-start friend
recommendation problem by combining random walk with Flickr
platform features to recommend friends on Twitter. Lee and Lim
[13] performed similar cross-platform friendship recommendation
in Instagram and Twitter using friendship maintenance features
derived from users' friendship behaviours in multiple social plat-
forms. In a more recent study, Lee et al [14] proposed a probabilistic
model to predict which social media platform would a user publish
a given post.
However, most of these cross-platform recommendation studies
only focus on recommending homogeneous behaviours (e.g. con-
necting users, publishing post, etc) in online social platforms. In our
study, we handled recommendation of heterogeneous behaviors
(e.g. user fork and watch repositories in GitHub) in a cross-platform
setting. Although our prediction experiments are conducted on so-
cial collaborative platforms, the cross-platform activity prediction
framework can also be applied to other online social platforms. For
example, we can predict if a user will like, comment or retweet a post
in Twitter by learning the same user's heterogeneous behaviors in
Facebook.
5 THREATS TO VALIDITY
Threats to Internal Validity. Threats to internal validity often
refer to experimenter biases. In this study, most of our processes
are automated. The positive and negative instances in our training
and testing datasets were also randomly selected and generated.
when estimating the developer interests on GitHub, we matched
the keywords in repositories' descriptions to the tags collected from
Stack Overflow questions. There could be cases where the words
from a repository's descriptions did not match any of the collected
tags, and thus we are not able to infer the interest of the developer
using that repository. In the future, we plan to mitigate this by
labelling the repositories using a tag recommendation approach
[27, 28, 33, 35, 36].
Threats to External Validity. Threats to external validity refer
to the generalizability of our findings. We have used large datasets
from GitHub and Stack Overflow for our analysis and experiments.
Our study findings are based on 92,427 developers, which is close to
10% of the Stack Overflow users who are active during the studied
period, the interests of these users were derived using the data from
the whole GitHub and Stack Overflow datasets (2.5 million and 1
million active users respectively). The current studied users were
also obtained from the dataset provided in previous research by
Badashian et al. [2]. Both Vasilescu et al. [29] and Badashian et al.
0.00.20.40.60.81.0False Positive Rate0.00.20.40.60.81.0True Positive RateROC for Answer PredictionSO_ACT (AUC = 0.86)SO_CoACT (AUC = 0.82)GH_ACT (AUC = 0.71)GH_CoACT (AUC = 0.62)ALL (AUC = 0.89)0.00.20.40.60.81.0False Positive Rate0.00.20.40.60.81.0True Positive RateROC for Favorite PredictionSO_ACT (AUC = 0.75)SO_CoACT (AUC = 0.66)GH_ACT (AUC = 0.64)GH_CoACT (AUC = 0.56)ALL (AUC = 0.77)0.00.20.40.60.81.0False Positive Rate0.00.20.40.60.81.0True Positive RateROC for Fork PredictionSO_ACT (AUC = 0.65)SO_CoACT (AUC = 0.57)GH_ACT (AUC = 0.69)GH_CoACT (AUC = 0.68)ALL (AUC = 0.75)0.00.20.40.60.81.0False Positive Rate0.00.20.40.60.81.0True Positive RateROC for Watch PredictionSO_ACT (AUC = 0.58)SO_CoACT (AUC = 0.55)GH_ACT (AUC = 0.63)GH_CoACT (AUC = 0.57)ALL (AUC = 0.67) [2] had utilized GitHub users' email address and Stack Overflow
users' email MD5 hashes to find the intersection between the two
platform datasets. However, this existing method is no longer valid
as Stack Overflow no longer make the email hash of their user
available [26], thus motivating new methods to match users in two
social collaborative platforms. As part of our future work, we will
explore new user matching methods such as matching users by
their username to generate a large dataset of cross-platform users.
Threats to Construct Validity. Threats to construct refer to
the appropriateness of metrics used. In this work, we use precision,
recall and F1-measure to evaluate the results of our prediction ex-
periments. These metrics were commonly used for other prediction
experiments, e.g., [5, 12, 16, 17].
6 RELATED WORK
In this section, we review three groups of existing research work
related to our research. The first group discuss the studies on user
interests in social collaborative platforms, in particularly, GitHub
and Stack Overflow. The second group focuses on studies on pre-
diction and recommendation in GitHub and Stack Overflow. The
last group reviews the inter-platform studies on users in the two
platforms.
The user interests in social collaborative platforms have been
a widely studied research area. There were research work that
focused on analyzing topics asked by the user in Stack Overflow [3,
4, 22, 40, 44]. Barua et al. [4] conducted an extensive empirical
study to mine the topics discussed by users in Stack Overflow. Bajaj
et al. [3] performed a similar study specific to web development
while Rosen and Shihab [22] performed a similar study specific to
mobile application development. Similarly, there were also work
on analyzing programming languages used by users in GitHub and
their relationships to GitHub contributions [18, 20, 24, 30]. Our
work extends this group of research by comparing user interests in
the two social collaborative platforms.
Prediction and recommendation in social collaborative platforms
have been widely studied. These work can be further categorized
into two groups: (i) finding experts to perform a certain platform
tasks or activities [1, 6, 10, 19, 21, 31, 37, 39, 41, 42] and (ii) recom-
mending content or activities to users in the platforms [7, 9, 11, 32,
34, 43]. For work in group (i), there were work which proposed
methods to find experts to answer questions in Stack Overflow
[6, 21, 31, 37, 39], while for GitHub, experts are predicted if they
will review pull-requests and code for repositories [19, 41, 42]. For
work in group (ii), Wang et al.[34] conducted a study in Stack Over-
flow to recommend questions and answers concerning API issues
to users. de Souza et al. [7] conducted an experiment to recommend
Stack Overflow question-answer pairs relevant to selected software
programming problems. Zhang et al. [43] predict and recommend
relevant repositories to users based on the users' past activities (e.g.
fork, watch, etc) in the platform. In a more recent work, Jiang et
al. [11] proposed to use user programming language preferences
and one-class collaborative filtering to improve prediction of which
GitHub repositories are relevant to a user. Our study adds on to the
state-of-the-art in group (ii) by proposing a novel method that use
multiple platform data to predict platform activities
There have been few existing inter-platform studies on GitHub
and Stack Overflow. Vasilescu et al. performed a study on de-
velopers' involvement and productivity in Stack Overflow and
GitHub [29]. They found that users who are more active on GitHub
(in terms of GitHub commits), tend to ask and answer more ques-
tions on Stack Overflow. Badashian et al. [2] did an empirical study
on the correlation between different types of activities in the two
platforms. Silvestri et at. [25] proposed a user linkage model to link
users' Stack Overflow, GitHub and Twitter accounts. More recently,
Lee and Lo [15] did an extensive study on users' interests across
GitHub and Stack Overflow. In that study, the researchers found
that users who have accounts with GitHub and Stack Overflow
do exhibit similar interests observed from their activities in the
two social collaborative platforms. Our work builds upon insights
reported in [15] by proposing a prediction method based on user
interests across and within social collaborative platforms.
7 CONCLUSION AND FUTURE WORK
In this paper, we propose a novel framework which predicts users
activities in multiple social collaborative platforms. We conducted
experiments on large real-world datasets which contain activities
of 92,427 users who are active in GitHub and Stack Overflow. Our
proposed methods achieved good accuracy in predicting various
user activities (up to an AUC score of 0.89).
Our experiments have shown that user activities in Stack Over-
flow can be predicted with reasonable accuracy using the same
user's interests inferred from his or her activities in GitHub. The
same observation was made when predicting a user's activities in
GitHub using his or her interests inferred from his or her activi-
ties in Stack Overflow. The reasonable accuracies yield by cross-
platform prediction approach demonstrates its potential in solving
the cold-start problem in user activity prediction and recommenda-
tion in social collaborative platforms.
For future work, we intend to consider more advanced techniques
(e.g., topic models or deep learning models) to derive and measure
user interests similarity across multiple platforms. We will also
consider platforms aside from Stack Overflow and GitHub (e.g.
Quora).
ACKNOWLEDGMENTS
This research is supported by the Living Analytics Research Centre
(LARC), a research centre set up in Singapore Management Univer-
sity (SMU) with Carnegie Mellon University (CMU) that focuses
on behavioural and social network analytics in the fields of Ur-
ban and Community Liveability, Personalised Urban Mobility, Jobs
and Skills Intelligence, as well as Smart Consumption and Healthy
Lifestyle.
REFERENCES
[1] Mohammad Y. Allaho and Wang-Chien Lee. 2014. Increasing the Responsiveness
of Recommended Expert Collaborators for Online Open Projects. In CIKM.
[2] Ali Sajedi Badashian, Afsaneh Esteki, Ameneh Gholipour, Abram Hindle, and
Eleni Stroulia. 2014. Involvement, contribution and influence in GitHub and
stack overflow. In CSSE.
[3] Kartik Bajaj, Karthik Pattabiraman, and Ali Mesbah. 2014. Mining Questions
Asked by Web Developers. In MSR.
[4] Anton Barua, Stephen W. Thomas, and Ahmed E. Hassan. 2014. What Are
Developers Talking About? An Analysis of Topics and Trends in Stack Overflow.
Empirical Software Engineering 19, 3 (2014).
[5] Gerardo Canfora, Andrea De Lucia, Massimiliano Di Penta, Rocco Oliveto, Anni-
bale Panichella, and Sebastiano Panichella. 2013. Multi-objective cross-project
defect prediction. In ICST.
[6] Morakot Choetkiertikul, Daniel Avery, Hoa Khanh Dam, Truyen Tran, and Aditya
Ghose. 2015. Who Will Answer My Question on Stack Overflow?. In ASWEC.
[7] Lucas B. L. de Souza, Eduardo C. Campos, and Marcelo de A. Maia. 2014. Ranking
Crowd Knowledge to Assist Software Development. In ICPC.
[8] Georgios Gousios. 2013. The GHTorrent dataset and tool suite. In MSR.
[9] Mohamed Guendouz, Abdelmalek Amine, and Reda Mohamed Hamou. 2015.
Recommending relevant GitHub repositories: a collaborative-filtering approach.
on Networking and Advanced Systems (2015).
[10] Chaoran Huang, Lina Yao, Xianzhi Wang, Boualem Benatallah, and Quan Z Sheng.
2017. Expert as a Service: Software Expert Recommendation via Knowledge
Domain Embeddings in Stack Overflow. In ICWS.
[11] Jyun-Yu Jiang, Pu-Jen Cheng, and Wei Wang. 2017. Open Source Repository
Recommendation in Social Coding. In SIGIR.
[12] Sunghun Kim, E James Whitehead, and Yi Zhang. 2008. Classifying software
changes: Clean or buggy? Software Engineering, IEEE Transactions on 34, 2 (2008),
181–196.
[13] Ka-Wei Roy Lee and Ee-Peng Lim. 2016. Friendship maintenance and prediction in
multiple social networks. In Proceedings of the 27th ACM Conference on Hypertext
and Social Media. ACM, 83–92.
[14] Roy Ka-Wei Lee, Tuan-Anh Hoang, and Ee-Peng Lim. 2017. On Analyzing
User Topic-Specific Platform Preferences Across Multiple Social Media Sites. In
Proceedings of the 26th International Conference on World Wide Web. International
World Wide Web Conferences Steering Committee, 1351–1359.
[15] Roy Ka-Wei Lee and David Lo. 2017. GitHub and Stack Overflow: Analyzing
Developer Interests Across Multiple Social Collaborative Platforms. In SOCINFO.
[16] Jaechang Nam, Sinno Jialin Pan, and Sunghun Kim. 2013. Transfer defect learning.
[17] Foyzur Rahman, Daryl Posnett, and Premkumar Devanbu. 2012. Recalling the
imprecision of cross-project defect prediction. In FSE.
[18] Mohammad Masudur Rahman and Chanchal K. Roy. 2014. An Insight into the
Pull Requests of GitHub. In MSR.
[19] Mohammad Masudur Rahman, Chanchal K Roy, and Jason A Collins. 2016. CoR-
ReCT: code reviewer recommendation in GitHub based on cross-project and
technology experience. In ICSE.
[20] Baishakhi Ray, Daryl Posnett, Vladimir Filkov, and Premkumar Devanbu. 2014.
A Large Scale Study of Programming Languages and Code Quality in Github. In
FSE.
[21] Fatemeh Riahi, Zainab Zolaktaf, Mahdi Shafiei, and Evangelos Milios. 2012. Find-
ing Expert Users in Community Question Answering. In WWW.
[22] Christoffer Rosen and Emad Shihab. 2015. What are mobile developers asking
about? A large scale study using stack overflow. Empirical Software Engineering
(2015).
[23] Andrew I Schein, Alexandrin Popescul, Lyle H Ungar, and David M Pennock.
2002. Methods and metrics for cold-start recommendations. In SIGIR.
[24] Jyoti Sheoran, Kelly Blincoe, Eirini Kalliamvakou, Daniela Damian, and Jordan
Ell. 2014. Understanding "Watchers" on GitHub. In MSR.
[25] Giuseppe Silvestri, Jie Yang, Alessandro Bozzon, and Andrea Tagarelli. 2015.
Linking Accounts across Social Networks: the Case of StackOverflow, Github
In ICSE.
and Twitter.. In KDWeb.
[26] StackExchange. 2014. Where did EmailHash go?
(2014).
stackexchange.com/questions/221027/where-did-emailhash-gom
[27] Clayton Stanley and Michael D Byrne. 2013. Predicting tags for stackoverflow
http://meta.
posts. In ICCM.
[28] Santiago Vargas-Baldrich, Mario Linares-Vasquez, and Denys Poshyvanyk. 2015.
Automated Tagging of Software Projects Using Bytecode and Dependencies (N).
In ASE.
[29] Bogdan Vasilescu, Vladimir Filkov, and Alexander Serebrenik. 2013. StackOver-
flow and GitHub: associations between software development and crowdsourced
knowledge. In SocialCom.
[30] Bogdan Vasilescu, Yue Yu, Huaimin Wang, Premkumar Devanbu, and Vladimir
Filkov. 2015. Quality and Productivity Outcomes Relating to Continuous Integra-
tion in GitHub. In FSE.
[31] Jian Wang, Jiqing Sun, Hongfei Lin, Hualei Dong, and Shaowu Zhang. 2017. Con-
volutional neural networks for expert recommendation in community question
answering. Science China Information Sciences 60, 11 (2017).
[32] Lin Wang, Bin Wu, Juan Yang, and Shuang Peng. 2016. Personalized recommen-
dation for new questions in community question answering. In ASONAM.
[33] S. Wang, D. Lo, B. Vasilescu, and A. Serebrenik. 2014. EnTagRec: An Enhanced
Tag Recommendation System for Software Information Sites. In ICSME.
[34] Wei Wang, Haroon Malik, and Michael W. Godfrey. 2015. Recommending Posts
Concerning API Issues in Developer Q&sA Sites. In MSR.
[35] Xin Xia, David Lo, Xinyu Wang, and Bo Zhou. 2013. Tag recommendation in
software information sites. In MSR.
[36] Bowen Xu, Deheng Ye, Zhenchang Xing, Xin Xia, Guibin Chen, and Shanping Li.
2016. Predicting Semantically Linkable Knowledge in Developer Online Forums
via Convolutional Neural Network. In ASE.
[37] Congfu Xu, Xin Wang, and Yunhui Guo. 2016. Collaborative Expert Recommen-
dation for Community-Based Question Answering. In ECML-KDD.
[38] Ming Yan, Jitao Sang, Tao Mei, and Changsheng Xu. 2013. Friend transfer:
Cold-start friend recommendation with cross-platform transfer learning of social
knowledge. In 2013 IEEE International Conference on Multimedia and Expo (ICME).
IEEE, 1–6.
[39] Liu Yang, Minghui Qiu, Swapna Gottipati, Feida Zhu, Jing Jiang, Huiping Sun, and
Zhong Chen. 2013. CQArank: Jointly Model Topics and Expertise in Community
Question Answering. In CIKM.
[40] Xin-Li Yang, David Lo, Xin Xia, Zhi-Yuan Wan, and Jian-Ling Sun. 2016. What
Security Questions Do Developers Ask? A Large-Scale Study of Stack Overflow
Posts. Journal of Computer Science and Technology 31, 5 (2016).
[41] Yue Yu, Huaimin Wang, Gang Yin, and Charles X Ling. 2014. Reviewer recom-
mender of pull-requests in GitHub. In ICSME.
[42] Yue Yu, Huaimin Wang, Gang Yin, and Tao Wang. 2016. Reviewer recommenda-
tion for pull-requests in GitHub: What can we learn from code review and bug
assignment? Information and Software Technology 74 (2016).
[43] Lingxiao Zhang, Yanzhen Zou, Bing Xie, and Zixiao Zhu. 2014. Recommend-
ing relevant projects via user behaviour: an exploratory study on github. In
CrowdSoft.
[44] Jie Zou, Ling Xu, Weikang Guo, Meng Yan, Dan Yang, and Xiaohong Zhang. 2015.
Which Non-functional Requirements Do Developers Focus On? An Empirical
Study on Stack Overflow Using Topic Analysis. In MSR.
|
1802.06189 | 1 | 1802 | 2018-02-17T04:11:58 | Contrast Subgraph Mining from Coherent Cores | [
"cs.SI"
] | Graph pattern mining methods can extract informative and useful patterns from large-scale graphs and capture underlying principles through the overwhelmed information. Contrast analysis serves as a keystone in various fields and has demonstrated its effectiveness in mining valuable information. However, it has been long overlooked in graph pattern mining. Therefore, in this paper, we introduce the concept of contrast subgraph, that is, a subset of nodes that have significantly different edges or edge weights in two given graphs of the same node set. The major challenge comes from the gap between the contrast and the informativeness. Because of the widely existing noise edges in real-world graphs, the contrast may lead to subgraphs of pure noise. To avoid such meaningless subgraphs, we leverage the similarity as the cornerstone of the contrast. Specifically, we first identify a coherent core, which is a small subset of nodes with similar edge structures in the two graphs, and then induce contrast subgraphs from the coherent cores. Moreover, we design a general family of coherence and contrast metrics and derive a polynomial-time algorithm to efficiently extract contrast subgraphs. Extensive experiments verify the necessity of introducing coherent cores as well as the effectiveness and efficiency of our algorithm. Real-world applications demonstrate the tremendous potentials of contrast subgraph mining. | cs.SI | cs | Contrast Subgraph Mining from Coherent Cores
Jingbo Shang1, Xiyao Shi1, Meng Jiang2, Liyuan Liu1, Timothy Hanra(cid:138)y3, Jiawei Han1
1Department of Computer Science, University of Illinois Urbana-Champaign, IL, USA
2Department of Computer Science and Engineering, University of Notre Dame, IN, USA
3US Army Research Laboratory, MD, USA
1{shang7, xshi27, ll2, hanj}@illinois.edu
[email protected]
3timothy.p.hanra(cid:138)[email protected]
8
1
0
2
b
e
F
7
1
]
I
S
.
s
c
[
1
v
9
8
1
6
0
.
2
0
8
1
:
v
i
X
r
a
ABSTRACT
Graph pa(cid:138)ern mining methods can extract informative and useful
pa(cid:138)erns from large-scale graphs and capture underlying principles
through the overwhelmed information. Contrast analysis serves as
a keystone in various (cid:128)elds and has demonstrated its e(cid:130)ectiveness in
mining valuable information. However, it has been long overlooked
in graph pa(cid:138)ern mining. (cid:140)erefore, in this paper, we introduce the
concept of contrast subgraph, that is, a subset of nodes that have
signi(cid:128)cantly di(cid:130)erent edges or edge weights in two given graphs of
the same node set. (cid:140)e major challenge comes from the gap between
the contrast and the informativeness. Because of the widely existing
noise edges in real-world graphs, the contrast may lead to subgraphs
of pure noise. To avoid such meaningless subgraphs, we leverage
the similarity as the cornerstone of the contrast. Speci(cid:128)cally, we
(cid:128)rst identify a coherent core, which is a small subset of nodes
with similar edge structures in the two graphs, and then induce
contrast subgraphs from the coherent cores. Moreover, we design
a general family of coherence and contrast metrics and derive a
polynomial-time algorithm to e(cid:129)ciently extract contrast subgraphs.
Extensive experiments verify the necessity of introducing coherent
cores as well as the e(cid:130)ectiveness and e(cid:129)ciency of our algorithm.
Real-world applications demonstrate the tremendous potentials of
contrast subgraph mining.
ACM Reference format:
Jingbo Shang1, Xiyao Shi1, Meng Jiang2, Liyuan Liu1, Timothy Hanra(cid:138)y3,
Jiawei Han1. 2018. Contrast Subgraph Mining from Coherent Cores. In
Proceedings of ACM SIGKDD Conference on Knowledge Discovery and Data
Mining, London, UK, August 2018 (KDD'18), 9 pages.
DOI: 10.475/123 4
1 INTRODUCTION
Graph pa(cid:138)ern mining methods can extract useful information from
large-scale graphs and capture underlying principles through the
overwhelmed information. Such information has bene(cid:128)ted many
(cid:128)elds, such as community analysis, document retrieval, and human
mobility analysis.
Recently, graph pa(cid:138)erns across multi-layer/multi-view graphs,
or graphs having di(cid:130)erent edge structures of the same node set,
have drawn lots of a(cid:138)entions. For example, frequent subgraph
mining aims to discover the similar pa(cid:138)erns in di(cid:130)erent graphs [33,
34]. Many researchers tried to identify the subsets of nodes that
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 pro(cid:128)t or commercial advantage and that copies bear this notice and the full citation
on the (cid:128)rst page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
KDD'18, London, UK
© 2018 Copyright held by the owner/author(s). 123-4567-24-567/08/06...$15.00
DOI: 10.475/123 4
Figure 1: Work(cid:131)ow visualization using hypothetical graphs.
form dense subgraphs in most of the graphs [11, 12, 24] or identify
discriminative (sometimes also named as "contrast" [29]) structures
from frequent subgraphs using node labels [13, 28, 29].
Contrast analysis serves as a keystone in many data mining
problems, while the related study in graph pa(cid:138)ern mining has
been long overlooked. For example, in the (cid:128)elds of text mining
and document retrieval, given a major corpus and a background
corpus of the same vocabulary, much e(cid:130)ort has been paid to identify
informative phrases whose frequencies in the major corpus are
signi(cid:128)cantly higher than those in the background corpus [2, 6, 18,
19, 21, 27]. Meanwhile, very few a(cid:138)empt has been made to contrast
subgraph mining. Analogous to document comparative analysis,
if we compare two graphs, GA and GB, of the same set of nodes V ,
we should be able to (cid:128)nd some informative subgraph g whose edge
weights in GA are signi(cid:128)cantly di(cid:130)erent from those in GB. In this
paper, we name such subgraph g as a contrast subgraph.
Contrast subgraphs have great potentials to facilitate many
downstream applications such as temporal role identi(cid:128)cation [9,
20, 25, 31], online-o(cid:132)ine community detection [5, 15, 32, 32, 37],
and spatio-tempral event detection [10, 36]. For example, contrast
subgraphs can provide useful signals for advisor-advisee identi(cid:128)-
cation tasks [31]. Moreover, when both online and o(cid:132)ine social
networks are provided, mining contrast subgraphs can detect some
unusual relationships [5, 32]. Also, by contrasting the incoming
and outgoing real-time tra(cid:129)cs, one can detect speci(cid:128)c types of
spatio-temporal events [10].
Graph AGraph BContrastSubgraphSeedsWorkflow: Seeds àCoherent Core àContrast Subgraph(similar parts) (different parts)Coherent Core KDD'18, August 2018, London, UK
J. Shang et al.
Figure 2: Spatio-Temporal Event Detection. Heat maps visualize the adjacency matrices within the corresponding regions. (cid:135)e
darker, the more tra(cid:129)c. Contrasting two road networks of the real-time tra(cid:129)c and historical averages, the identi(cid:128)ed coherent
core demonstrates the usual, busy tra(cid:129)c on the East 2nd Ring in Beijing, while the extracted contrast subgraph reveals an
unusually large volume of tra(cid:129)c, which indicates the ending of a concert event at the Beijing Workers' Sports Complex.
Besides extracting contrast subgraphs, de(cid:128)ning the contrast itself
remains an open problem. Moreover, trying to induce informative
subgraphs from the contrast makes the task even more challenging.
In real-world graphs, there usually exist noise edges. For the algo-
rithms directly optimizing the contrast, it is likely to end up with a
subset of nodes which have heavy noise in one of the graphs and
no edge in the other graph. To avoid such numerous but meaning-
less subgraphs, we propose to induce contrast subgraphs from the
coherent cores, i.e., a subset of nodes with similar edge structures
in GA and GB. We visualize the work(cid:131)ow using two hypothetical
graphs in Figure 1. Given the seed node 5, we (cid:128)rst identify the co-
herent core of nodes {3, 4, 5, 6}, and then add nodes 1, 7, and 8 into
the contrast graph. More experiments in Section 5.2 demonstrate
that inducing from the coherent cores is necessary and it acts as an
anchor to guarantee the informativeness of the induced contrast
subgraphs.
Beyond the problem formulation, it is also challenging to develop
an e(cid:129)cient and scalable algorithm for mining contrast subgraphs
from large-scale graphs. A straightforward solution is to apply co-
dense graph mining algorithms on (GA, GB) and then on (GA, GB).
However, the dense complementary graph (i.e., GA or GB) makes
most of, if not all, pruning techniques initially designed for sparse
graphs ine(cid:130)ective.
In this paper, we derive a polynomial-time
algorithm to e(cid:129)ciently identify coherent subgraph cores and then
extract contrast subgraphs. More speci(cid:128)cally, we apply a binary
search on the coherent/contrast score and construct a network
such that whether the current score is achievable is equivalent
to whether the min S − T cut in the network is above a certain
threshold. (cid:140)anks to the duality between the min cut and the max
(cid:131)ow, we can solve the reduced problem in a polynomial time [22].
Our experimental results based on real-world datasets demon-
strate the identi(cid:128)ed coherent core and the extracted contrast sub-
graph are quite insightful and encouraging in a variety of tasks.
For example, based on the tra(cid:129)c data and road network in Beijing,
contrast subgraphs between the real-time tra(cid:129)c and the historical
tra(cid:129)c indicate the events happening in the city. Figure 2 shows
the coherent core and the contrast subgraph discovered by our
proposed method through a comparison between the tra(cid:129)c data on
Nov 24, 2012, and the historical averages. (cid:140)e usual, busy tra(cid:129)c on
the major roads (i.e., the East 2nd Ring in Beijing) form the coher-
ent core. (cid:140)e contrast subgraph reveals the unusual tra(cid:129)c around
the Beijing Workers' Sports Complex between 10:30 – 11:00 PM.
We (cid:128)nd that the contrast subgraph indicates the ending of Dang
Ding's concert event, which further re(cid:131)ects the precious value of
the proposed contrast subgraph mining.
Reproducibility: We release our code at the GitHub1.
1h(cid:138)ps://github.com/shangjingbo1226/ContrastSubgraphMining
To our best knowledge, this is the (cid:128)rst work that aims to mine
informative contrast subgraphs between two graphs. Our contribu-
tions are highlighted as follows.
• We formulate the contrast subgraph mining problem and avoid
meaningless contrast subgraphs by inducing from coherent cores.
(cid:140)e problem is formulated for multiple applications using a gen-
eral family of metrics.
• We derive a polynomial-time algorithm to e(cid:129)ciently extract
• Extensive experiments verify the necessity of introducing co-
herent cores as well as the e(cid:130)ectiveness and e(cid:129)ciency of our
algorithm. Real-world applications demonstrate the tremendous
potentials of contrast subgraph mining.
contrast subgraphs from large-scale graphs.
DangDing'sconcertonNov24,2012atBeijingWorkers'SportsComplex.Theconcertendedaround10:30PM,whichledtoasignificanttrafficjam.Given SeedsOur Identified Coherent CoreEast 2ndRing in Beijing,one of the major roadsNov 24, 2012Historical AvgNov 24, 2012Historical AvgBeijing, China10:30 PM -11:00 PMOur Extracted Contrast Subgraph Contrast Subgraph Mining from Coherent Cores
KDD'18, August 2018, London, UK
(cid:140)e remainder of this paper is organized as follows. We (cid:128)rst
formulate the contrast subgraph mining problem in Section 2. (cid:140)e
related work is discussed in Section 3. Section 4 covers the technical
details of our derived polynomial-time algorithm. In Section 5, we
use a real-world task to verify the necessity of inducing contrast
subgraph from coherent cores and evaluate the e(cid:129)ciency of our al-
gorithm. Section 6 uses two real-world applications to demonstrate
the importance of contrast subgraph mining. Section 7 concludes
this paper and outlines future directions.
2 CONTRAST SUBGRAPH MINING
In this section, we (cid:128)rst formulate the problem and propose a general
family of metrics. And then, we discuss our choices of speci(cid:128)c
metrics used in our experiments.
2.1 Problem Formulation
(cid:140)e input of the contrast subgraph mining problem consists of two
weighted, undirected graphs GA = (V , EA) and GB = (V , EB). (cid:140)e
node set V is the same in both GA and GB. But the edge weights in
EA and EB are di(cid:130)erent. As summarized in Table 1, we use EA(u, v)
and EB(u, v) to denote the edge weights of the edge between nodes
u and v in these two graphs, respectively. Because GA and GB
are undirected graphs, we have ∀u, v, EA(u, v) = EA(v, u) and
EB(u, v) = EB(v, u).
In this paper, without loss of generality, we make two assump-
tions for the input graphs. First, we assume EA(u, v), EB(u, v) ≥ 0,
where the edge weight 0 means there is no edge. In most of the
applications, the edge weights are mainly about the connection
strengths between nodes. And we only care about the weight di(cid:130)er-
ence between the corresponding edges. (cid:140)erefore, it is reasonable
to assume the weights are non-negative. Note that this assumption
includes boolean edges as a special case. Second, we assume the
preprocessing has been conducted on EA and EB, thus the ranges
or distributions of EA(u, v) and EB(u, v) would be similar.
(cid:140)e user can provide seed nodes, if necessary, to be(cid:138)er facili-
tate her interests. To make sure the nodes are really around the
given seeds or the coherent core, we de(cid:128)ne the neighbor nodes by a
parameter r,
Nr(s) = {udA(s, u) ≤ r ∨ dB(s, u) ≤ r}
(1)
where s is a subset of nodes, dA(s, u) and dB(s, u) are the minimum
number of non-zero edges to traverse from the node u to any node in
s. In practice, the value of r can be chosen based on the diameter of
the two input graphs. For example, the collaborator graph induced
from the DBLP publication dataset has a small diameter, so r = 1
and r = 2 are good choices. (cid:140)e road network has a relatively large
diameter, so we choose r from 10 to 20. Reasonable values of r can
always lead to meaningful results.
In summary, the inputs are two undirected, weighted graphs GA
and GB, a parameter r, and seed nodes seeds (could be empty). As
we discussed, to avoid meaningless cases, we propose to induce
contrast subgraphs from the coherent cores, where a coherent core
is a subset of nodes with similar edges in the two graphs. (cid:140)e goal
is to (cid:128)rst identify the coherent core c around the seed nodes seeds
(seeds ⊂ c ⊂ Nr(seeds)), and then induce the contrast subgraph g
from this coherent core (c ⊂ g ⊂ Nr(c)).
GA, GB
GA, GB
V
EA(u, v), EB(u, v)
seeds
c
g
Nr(s)
coherence(c)
coherence(u, v)
contrast(g)
contrast(u, v)
penalty(u)
Table 1: Notation Table.
Two undirected, weighted input graphs.
Complementary graphs considering whether the edge exists.
(cid:140)e node set for both GA and GB.
Edge weights between nodes u and v in GA and GB.
A set of seed nodes. It could be empty.
Coherent core. A subset of V . seeds ⊂ c.
Contrast subgraph. A subset of V . c ⊂ g
Neighbor node set of the node set s. r is a parameter.
Coherence metric for a subgraph.
Coherence metric for an edge between u and v.
Contrast metric for a subgraph.
Contrast metric for an edge between u and v.
Node penalty metric.
2.2 A General Family of Metrics
With a speci(cid:128)c coherence metric, i.e., coherence(c), and a speci(cid:128)c
contrast metric, i.e., contrast(g), the contrast subgraph mining prob-
lem becomes the following two maximization problems.
(1) Identify the coherent core around the seed nodes:
c =
argmax
seeds⊂c⊂Nr (seeds)
coherence(c).
(2) Induce the contrast subgraph from the identi(cid:128)ed coherent core:
g = argmax
c⊂g⊂Nr (c)
contrast(g).
It is natural to start from a single edge, since an edge is also a sub-
graph of two nodes. We denote the edge coherence as coherence(u, v)
and the edge contrast as contrast(u, v). Inspired by the formula-
tion of the maximum dense subgraphs [8], we design the coher-
ence/contrast metric based on the edge coherence/contrast as fol-
lows:
coherence(c) =
(2)
u,v ∈c∧u <v coherence(u, v)
u,v ∈❕∧u <v contrast(u, v)
u∈c penalty(u)
u∈g penalty(u)
,
,
contrast(g) =
(3)
where penalty(u) is a function to control the subgraph size. (cid:140)e
numerators in Equation 2 and Equation 3 become bigger when the
subgraph grows. (cid:140)e idea of introducing node penalty is to penalize
these large subgraphs. In fact, the form of these two de(cid:128)nitions is
a generalization of the density function in [8].
Remark. (1) When there is no user-provided seed (i.e., seeds = ∅),
we will identify the most coherent subgraph among all non-empty
subgraphs, which is the most "similar" part between GA and GB.
(2) We can (cid:128)nd a series of non-overlapping contrast subgraphs by
iteratively removing the newly added nodes in the most contrasting
subgraph, i.e., g \ c, until g \ c = ∅. Here g is the most contrasting
subgraph in the current iteration, while c is the same coherent core
in all iterations.
2.3 Speci(cid:128)c Metrics
Our algorithm works with any (1) non-negative edge coherence
coherence(u, v); (2) non-negative edge contrast contrast(u, v); and
(3) positive node penalty penalty(u). (cid:140)erefore, there are countless
ways to de(cid:128)ne these metrics. We discuss the speci(cid:128)c metrics utilized
in both Section 5 and Section 6 here.
KDD'18, August 2018, London, UK
J. Shang et al.
Edge Coherence. Considering the real-world graphs are usually
sparse, the non-zero edges are more telling than the zero-weighted
edges. For example, in co-authorship graphs, the observed co-
authorship node pairs are more important compared to any two
unrelated nodes. Additionally, as discussed before, the edge weights
are mainly about the connection strengths, and their distributions
are similar. So in our experiments, for a given edge, we adopt
its smaller weight in two graphs to describe its edge coherence.
Formally,
coherence(u, v) = min{EA(u, v), EB(u, v)}.
(4)
Edge Contrast. A good contrast metric must meet the following
requirements.
(1) Symmetric. (cid:140)e order of GA and GB should have no e(cid:130)ect on
the contrast metric. Formally, given two nodes, u and v, if we
swap EA and EB, contrast(u, v) should be the same.
(2) Zero. It is natural to require contrast(u, v) = 0 when the edge
weights in two graphs are the same. (cid:140)at is, when EA(u, v) =
EB(u, v), contrast(u, v) must be 0.
(3) Monotonicity. Suppose EA(u, v) ≤ EB(u, v). If we increase
EB(u, v) and keep the other same, the score contrast(u, v) should
increase. If we decrease EA(u, v) and keep the other same, the
score contrast(u, v) should also increase.
Starting from these three requirements, we (cid:128)nd a neat de(cid:128)nition
as follows.
contrast(u, v) = EA(u, v) − EB(u, v)
(5)
It meets all the requirements and is therefore adopted in our exper-
iments.
Node Penalty. If there is no prior knowledge about the two graphs,
a uniform node penalty is always a safe choice. (cid:140)erefore, we use
penalty(u) = 1 in all our experiments.
3 RELATED WORK
In this section, we introduce the related work on dense graph
pa(cid:138)erns within a single graph, cross-graph pa(cid:138)erns, and interesting
phrase pa(cid:138)erns of contrast in the text mining (cid:128)eld.
3.1 Dense Subgraphs within a Single Graph
Our coherence and contrast metrics are inspired by the dense sub-
graph problem, i.e., (cid:128)nding a subset of nodes that maximizes the
ratio of the number of edges between these nodes over the number
of selected nodes. A polynomial time algorithm using a network
(cid:131)ow model is (cid:128)rst proposed to solve this problem [8]. [1] tries to
solve it approximately using a greedy algorithm. Beyond traditional
dense subgraphs, it has been proved that if one wants to specify
the subgraph size, the problem becomes NP-hard [16]. Gibson et
al. [7] design an e(cid:129)cient algorithm speci(cid:128)cally for giant dense sub-
graphs. Rossi et al. [26] develop a fast, parallel maximum clique
algorithm for sparse graphs; and Leeuwen et al. [30] explore the
most interesting dense subgraphs based on the user's prior belief.
3.2 Cross-Graph Patterns
Frequent Subgraph. Given a large collection of graphs, frequent
subgraph mining aims to (cid:128)nd frequent structures across di(cid:130)erent
graphs [33, 34]. It is an unsupervised task but focuses on shared
pa(cid:138)erns instead of contrast pa(cid:138)erns. Moreover, it usually requires
more than two, usually tens of graphs as input.
Discriminative Subgraph. In discriminative subgraph mining,
given multiple graphs with their class labels, the goal is to (cid:128)gure
out the subgraphs that discriminate between classes [13, 28, 29]. Dis-
criminative subgraph is sometimes named as contrast subgraph [29],
however, it's completely di(cid:130)erent from our contrast subgraph min-
ing, due to its supervised learning nature.
Co-Dense Subgraph. Density is an important measurement for
graphs. Among various models and de(cid:128)nitions of dense subgraphs,
quasi-clique is one of the most prominent ones [17]. As a natural
extension, researchers have explored how to e(cid:129)ciently (cid:128)nd cross-
graph quasi-cliques [24] and coherent subgraphs [3, 4]. Jiang et al.
further extend cross-graph quasi-cliques to frequent cross-graph
quasi-cliques [12]. As an application, coherent dense subgraph
mining has shown its usefulness in biological networks [11]. We
can also generate some contrast subgraph in terms of density by
straightforwardly applying co-dense subgraph mining methods
on GA, GB and then on GA, GB. However, as shown in our ex-
periments, these algorithms become extremely slow because the
complementary graphs are so dense that pruning techniques are
no longer e(cid:130)ective.
3.3 Informative Phrases from the Contrast
Given two corpora (i.e., the main corpus and the background cor-
pus), informative phrases are the phrases that occur more frequently
in the main corpus than the background corpus [2, 6, 18, 19, 21, 27].
With the concept of contrast, informative phrases are more salient
than routine phrases like those consisted of stopwords, although
their frequencies in the main corpus are not super high. (cid:140)e def-
inition of informative phrases is based on individual phrase fre-
quencies without considering a set of phrases as a whole. Such
de(cid:128)nition is not applicable for subgraphs.
4 METHODOLOGY
In this section, we derive the polynomial-time algorithm for the
contrast subgraph mining. Since the forms of coherence and con-
trast metrics are similar, we focus on extracting contrast subgraphs
in the derivation. For identifying coherent cores, the same deriva-
tion applies if one replaces the edge contrast contrast(u, v) with
the edge coherence coherence(u, v).
4.1 Binary Search
First of all, we reduce the original maximization problem to a veri-
(cid:128)cation problem by applying the binary search technique on the
contrast score. In order to conduct the binary search, we need to
(cid:128)gure out the lower/upper bounds of the subgraph contrast as well
as the minimum contrast di(cid:130)erence between two subgraphs.
Lower Bound. In the worst case, there is only one di(cid:130)erent edge
between GA and GB (assuming they are not exactly same). Let
c > 0 be the minimum non-zero edge contrast. We have
contrast(g) ≥
c
u∈V penalty(u)
Upper Bound. As an ideal case, all edges are self-loops of the same
node in one graph, while that node has no self-loops in the other
Contrast Subgraph Mining from Coherent Cores
KDD'18, August 2018, London, UK
We (cid:128)rst construct a network with a source node S and a sink
node T as shown in Figure 3, then show a one-to-one mapping
between its S − T (cid:128)nite-value cuts and the contrast subgraphs g
meeting the constraints. Finally, we prove that the min cut gives
us the most contrasting subgraph.
Constructed Network. (cid:140)e network is constructed as follows.
• Creates a source node S and a sink node T .
• ∀u ∈ c, adds an edge between S and u of a weight +∞.
• Chooses a large enough constant U to make sure that all follow-
ing edge weights are positive. In our implementation, U is set as
u,v contrast(u, v), which is no smaller than any d(u).
• ∀u ∈ Nr(c) \ c, adds an edge between S and u of a weight U .
• ∀u ∈ Nr(c), adds an edge between u and T of a weight U + 2 ·
δ · penalty(u) − d(u).
• ∀contrast(u, v) (cid:44) 0, adds an edge between u and v of a weight
contrast(u, v).
One-to-One Mapping. Here, we show the one-to-one mapping
between its S − T (cid:128)nite-value cuts and the contrast subgraphs g
meeting the constraints. AnyS−T cut can be uniquely identi(cid:128)ed by
the set of nodes connected to the source node S, which is denoted
as s(S − T). Note that, we have a hard constraint that c ⊂ g.
(cid:140)erefore, any valid g should contain all nodes in c, which means
there are only 2Nr (c)−c valid g's. Discarding those cuts of +∞
cost (i.e., including those edges from S to c), we will also have
2Nr (c)−c di(cid:130)erent cuts le(cid:137). So mapping the remaining S − T
cuts to s(S − T) forms a one-to-one mapping between the cuts and
contrast subgraphs.
Moreover, the value of the S − T cut can be transformed to the
hδ(g) function. Suppose we have a S − T cut. According to the
one-to-one mapping, we know that g = s(S − T). (cid:140)e cut edges
can be grouped into three types: (1) (cid:140)e edges between g and T ;
(2) (cid:140)e edges between S and Nr(c) \ g; And (3) the edges between
g and Nr(c) \ g. (cid:140)e sum of these edge weights are
u∈g
cutδ(S − T) =
+
= Nr(c) · U +
+
v ∈Nr (c)\g
u∈g
u∈g,v ∈Nr (c)\g
(U + 2 · δ · penalty(u) − d(u))
U +
contrast(u, v)
u∈g,v ∈Nr (c)\g
(2 · δ · penalty(u) − d(u))
contrast(u, v)
(cid:140)erefore, if we follow the one-to-one mapping, we have
hδ(g) = cutδ(S − T) = cutδ(g) − Nr(c) · U
Final Solution. As a result, minimizing hδ(g) is equivalent to (cid:128)nd
the min cut in the constructed network. (cid:140)anks to the duality be-
tween the min cut and the max (cid:131)ow, we can locate the min-cut
among the exponential number of candidates within a polynomial
time [23]. More speci(cid:128)cally, treating the edge weights in the con-
structed network as capacities, we calculate the max (cid:131)ow from S
to T . In the residual network, the nodes connected to S form g. To
obtain the result node set, we do a depth-(cid:128)rst-search in the min-cut
residual graph from S and return all reachable nodes. Algorithm 1
presents the work(cid:131)ow.
Figure 3: An Illustration for the Min-Cut Model.
graph. Let p > 0 be the minimum node penalty. We have
contrast(g) ≤
u,v ∈V ∧u <v contrast(u, v)
p
Minimum Contrast Di(cid:130)erence. Given any two subgraphs g1, g2,
because the edge contrast di(cid:130)erence is at least c and the denomi-
u∈V penalty(u), we
have
nator is no more than the total node penalty
(
contrast(g1) − contrast(g2) >
u∈V penalty(u))2
(cid:140)en, the key problem becomes that given a binary searched
contrast score δ, whether we can (cid:128)nd a subgraph g such that c ⊂
g ⊂ Nr(c) and contrast(g) ≥ δ. Speci(cid:128)cally, we need an algorithm
that can e(cid:129)ciently check whether contrast(g) ≥ δ for any δ ≥ 0. It
is equivalent to checking the following inequalities.
c
u,v∈g∧u <v
contrast(u, v) */
u∈g
u∈❕
v∈Nr (c)
u,v∈g∧u <v
u∈g penalty(u)
u,v∈g∧u <v contrast(u, v)
≥ δ
/* because penalties are positive */
contrast(u, v) −
δ · penalty(u) −
/* let d(u) =
(cid:16)
d(u) −
u∈g
− 1
(cid:16)2 · δ · penalty(u) − d(u)(cid:17)
2
+
(cid:16)2 · δ · penalty(u) − d(u)(cid:17)
+
contrast(u, v) ≤ 0
δ · penalty(u)
u∈g,v∈Nr (c)\g
u∈g,v∈Nr (c)\g
u∈g
u∈g
maxc⊂g⊂Nr (c)
⇔ maxc⊂g⊂Nr (c)
⇔ minc⊂g⊂Nr (c)
⇔ minc⊂g⊂Nr (c)
⇔ minc⊂g⊂Nr (c)
hδ(g) =
u∈g
δ · penalty(u) ≥ 0
contrast(u, v) ≤ 0
contrast(u, v)(cid:17) ≤ 0
(cid:140)erefore, the problem is transformed to minimize the following
function hδ(g) with the constraints c ⊂ g ⊂ Nr(c).
contrast(u, v)
u∈❕,v(cid:60)❕
4.2 Min-Cut Derivation
Fortunately, we (cid:128)nd the aforementioned minimization problem for
the function hδ(g) is solvable in a polynomial time.
STabU+2⋅%⋅penalty(.)−1(.)2∈4567Never cut here!ccontrast(a, b)g =; ; is connected to A} ={2,.,…}contrast(b, c)Uan F−G4HIConclusion: ℎKL=M;NK(L) – P⋅RS(M)Minimize ℎKLóMinimize cutK(L)…… KDD'18, August 2018, London, UK
J. Shang et al.
Algorithm 1: Contrast Subgraph Mining from Coherent Core
Require: GA, GB, the coherent core c, and the parameter r.
Return: (cid:140)e most contrasting subgraph g (c ⊂ g ⊂ Nr(c)).
Compute Nr(c) using a breadth-(cid:128)rst search (BFS).
c ← minu,v ∈V ∧u <v contrast(u, v)
p ← minu,v ∈V ∧u <v penalty(u, v)
l ←
h ←
←
(cid:16)
u∈V penalty(u)(cid:17)2
u∈V penalty(u)
u,v ∈V ∧u <v contrast(u, v)
p
c
c
g ← c
while h − l > do
m ← l + h
2
Construct the network as described in Section 4.2.
(cid:131)ow ← the max (cid:131)ow of the constructed network
if (cid:131)ow - Nr(c) · U ≤ 0 then
l ← m
Update g by the current S − T cut.
h ← m
else
return g.
About the constraints c ⊂ g ⊂ Nr(c), it is clear that the min-cut
solution will only involve nodes in Nr(c) because they are the only
nodes in the constructed network. Moreover, because the edge
weights of all edges from S to ∀u ∈ c are set to +∞, they will
never be included in the min cut. (cid:140)erefore, the g induced from the
min-cut will never miss any nodes in the coherent core c.
4.3 Time Complexity Analysis
We analyze the time complexity step by step. Given the parameter
r and seed nodes s, we can obtain Nr(s) in an O(V + EA + EB)
time through a breadth-(cid:128)rst-search. Computing c and p costs
O(V + EA + EB) time. (cid:140)e binary search will have log2 h−l
iterations before it stops. (cid:140)e bo(cid:138)leneck inside the binary search
is the max (cid:131)ow part. (cid:140)e state-of-the-art max (cid:131)ow algorithm [22]
is O(nm) by combining James B Orlin's algorithm (for the sparse
network) and the KRT algorithm (for the dense network), where
n and m are the numbers of nodes and edges in the constructed
network. (cid:140)erefore, the time complexity of the max (cid:131)ow part is
O(cid:0)Nr(c)(Nr(c) + EA + EB)(cid:1). So the overall time complexity
becomes O(cid:0)V +EA+EB+Nr(c)(Nr(c)+EA+EB) log2 h−l
(cid:1).
the worst case time complexity is O(cid:0)V (V +EA +EB) log2 h−l
(cid:1)
In the worst case, when r is big enough, Nr(c) becomes V . So
Moreover, as both log2(h − l) and log2 are no more than the input
length, this time complexity is polynomial.
5 EXPERIMENTS
In this section, we utilize the collaboration change detection task
as a sanity check. We (cid:128)rst present the useful results from contrast
subgraph mining. (cid:140)en, we discuss the importance of the coher-
ent core and neighborhood constraints. In the end, we compare
our algorithm with an adapted co-dense subgraph mining method,
regarding both e(cid:130)ectiveness and e(cid:129)ciency. We implement both
our algorithm and the compared method in C++. (cid:140)e following
execution time experiments were conducted on a machine with
Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz using a single thread.
5.1 Collaboration Change Detection
(cid:135)e Two Graphs. Two co-authorship graphs are constructed by
spli(cid:138)ing the DBLP publication network dataset2 into two time
periods (2001–2008 and 2009–2016). (cid:140)e nodes are authors who
published at least one paper in KDD, ICDM, WWW, ICML, and
NIPS. (cid:140)e co-authorship from 2001 to 2008 forms the edges in GA,
while GB consists of co-authorships from 2009 to 2016. It ends up
with 6, 999 nodes and 17, 806 edges in total. (cid:140)e edge weight is the
log value (i.e., log(x) + 1) of the actual collaboration time x. We
choose r = 1.
Visualization. We use two aligned heat maps (a black vertical
line in between) to visualize adjacency matrices of the subgraphs.
(cid:140)e darker color, the larger weight. (cid:140)e coherent core is always
positioned in the bo(cid:138)om-right corner and enclosed by red squares,
while the nodes are ranked by their total edge weights of EA within
this subgraph. Moreover, we append a table of the node names in
the top-down order of the rows in the heat map.
Expectations. In this task, coherent cores are expected to be some
long-term collaborators of the seed author spanned over 2001 –
2016, and contrast subgraphs will be likely the researchers who
have more collaborations with the seed author in only one of the
two periods.
Results. Since the time periods start from 2001, we choose a senior
professor, Prof. Jiawei Han from the University of Illinois at Urbana-
Champaign as the seed. As shown in Figure 4, we (cid:128)nd that all
authors in the coherent core are long-term collaborators of Prof.
Jiawei Han. (cid:140)is contrast subgraph achieves a contrast score of
8.99. Note that Jing Gao was Prof. Jiawei Han's Ph.D. student
from 2006 to 2011, and 2008/2009 happens to be her midpoint.
Moreover, they keep collaborating a(cid:137)er Jing Gao's graduation. (cid:140)e
(cid:128)rst 11 researchers in g \ c have much more collaborations with
the researchers in the coherent core in 2001 – 2008 than those in
2009 – 2016. (cid:140)e later 15 researchers have more collaborations
from 2009 to 2016. In summary, the semantic meanings of these
results match our observation and reality. (cid:140)erefore, the contrast
subgraph mining is helpful to the collaboration change detection
task, and our algorithm is e(cid:130)ective.
5.2 Coherent Core and Neighbor Constraints
If NOT inducing from coherent cores. If we directly extract
the most contrasting subgraph without the coherent core step and
the neighborhood step, we will (cid:128)nd, of course, a subgraph with a
potentially higher contrast score. We run the collaboration change
detection task using Prof. Jiawei Han as the seed again but skip
these two steps. As shown in Figure 5, we obtain a subgraph of a
contrast score 11.08, which is higher than the previous score (i.e.,
8.99). Despite the high contrast, as highlighted by the light blue box
in the (cid:128)gure, the seed node Prof. Jiawei Han never collaborated with
anyone in the contrast subgraph. (cid:140)erefore, although this subgraph
2h(cid:138)ps://aminer.org/citation
Contrast Subgraph Mining from Coherent Cores
KDD'18, August 2018, London, UK
seeds
c
g \ c
Jiawei Han
Jiawei Han, Philip S. Yu, Jing Gao, Wei Fan, Charu C. Aggarwal,
Latifur Khan, Bhavani M. (cid:140)uraisingham, Mohammad M. Masud
Hong Cheng−, Chen Chen−, Xifeng Yan−, Haixun Wang−, Dong
Xin−, Chao Liu−, Feida Zhu−, Qiaozhu Mei−, Chengxiang Zhai−,
Jian Pei−, Bo Zhao−, Hanghang Tong, Dan Roth+, (cid:140)omas S.
Huang+, Cindy Xide Lin+, David Lo+, Chi Wang+, Yizhou Sun+,
Fangbo Tao+, Xiao Yu+, Brandon Norick+, Marina Danilevsky+,
Feng Liang+, Jialu Liu+, Ahmed El-Kishky+, Jie Tang+
8.99
contrast(g)
Figure 4: Collaboration Change Detection. (cid:135)e darker, the
more collaborations. In g \ c, − means they have more col-
laborations with Prof. Jiawei Han in the (cid:128)rst time period,
while + means they have more collaborations with Prof. Ji-
awei Han in the second time period.
has a high contrast score, we believe its semantic is unrelated to
the expected collaboration change detection. (cid:140)is example shows
the necessity of inducing from coherent cores and enforcing the
neighbor constraint, which acts as anchors to make the induced
contrast subgraph semantically stable and smooth.
If NOT considering the Nr(c) constraint. Suppose we only uti-
lize the coherent core, but we drop the Nr(c) constraint. Again, we
can (cid:128)nd a subgraph of a higher contrast score, as shown in Figure 6.
Its contrast score is 9.37, which is higher than the previous result,
8.99. However, the results are noisy because of involving many
researchers who never worked with any researcher in the coherent
core. We cannot locate the researchers that are speci(cid:128)cally related
to the researchers in the core because there is no guarantee for the
distance from the researchers in our coherent core. For example, as
highlighted by two blue rectangles in the (cid:128)gure, those researchers
have absolutely no connection to our coherent core. For sure, this
result looks be(cid:138)er than the one without any constraint, but it is
still not as good as Figure 4, in terms of semantic meanings. (cid:140)ere-
fore, the neighbor constraint is necessary to bridge the contrast
subgraph to the coherent core and make them more semantically
connected.
5.3 Running Time and Contrast Comparisons
To our best knowledge, we are the (cid:128)rst to mine informative contrast
subgraphs between two graphs. (cid:140)ere is few a(cid:138)empt before to
a(cid:138)ack this problem directly. In order to make comparisons, we
adapt the co-dense subgraph mining algorithm [3, 4] by using
(GA, GB) and (GB , GA) as inputs. We name it as adapted co-dense
algorithm. (cid:140)is algorithm (cid:128)rst (cid:128)nds several subgraphs which are
dense in one graph but sparse in the other graph. From these results,
seeds
c
g \ c
Jiwei Han
(no coherent core is computed) Jiawei Han
Chang-Tien Lu, Feng Chen, Achla Marathe, Naren Ramakrish-
nan, Nathan Self, Patrick Butler, Rupinder Paul Khandpur, Parang
Saraf, Sathappan Muthiah, Wei Wang, Jose Cadena, Liang Zhao,
Andy Doyle, Anil Vullikanti, Chris J. Kuhlman, Bert Huang, Jaime
Arredondo, Graham Katz, David Mares, Lise Getoor, Kristen Maria
Summers, Fang Jin
11.08
contrast(g)
Figure 5: [Without c and Nr(c)] Collaboration Change Detec-
tion. In the g\c part, none of them has collaborated with the
seed. (cid:135)e results become meaningless.
seeds
c
g \ c
Jiawei Han
Jiawei Han, Philip S. Yu, Wei Fan, Jing Gao, Charu C. Aggarwal,
Latifur Khan, Bhavani M. (cid:140)uraisingham, Mohammad M. Masud
Haixun Wang, Jian Pei, Wei Wang, Chang-Tien Lu, Feng Chen,
Achla Marathe, Naren Ramakrishnan, Nathan Self, Patrick Butler,
Rupinder Paul Khandpur, Parang Saraf, Sathappan Muthiah, Jose
Cadena, Liang Zhao, Andy Doyle, Anil Vullikanti, Chris J. Kuhlman,
Bert Huang, Jaime Arredondo, Graham Katz, David Mares, Lise
Getoor, Kristen Maria Summers, Fang Jin
9.37
contrast(g)
Figure 6: [Without Nr(c)] Collaboration Change Detection.
Most researchers in g \ c have no collaboration with the re-
searchers in c. (cid:135)e results are noisy and less meaningful.
we then pick the subgraph of the highest contrast(g) score as the
contrast subgraph.
We randomly pick a node as the seed and only pass the edges
within Nr(seeds) to the adapted co-dense algorithm. Figure 7(a)
plots the running times of both methods a(cid:137)er the same seeds is
given. Note that the Y-axis is log-scale. From the (cid:128)gure, one can
observe that the adapted co-dense algorithm requires an exponen-
tial running time w.r.t. the number of neighbor nodes. (cid:140)e primary
reason is that the complementary graphs are so dense that most of,
KDD'18, August 2018, London, UK
J. Shang et al.
(a) Running Time
(b) Contrast Scores
Figure 7: Comparison between the adapted co-dense algo-
rithm and our contrast subgraph mining algorithm.
if not all, pruning techniques are no longer e(cid:130)ective. Our algorithm
demonstrates a polynomial growth as same as our previous time
complexity analysis.
We also compare the contrast scores achieved by the two al-
gorithms, as shown in Figure 7(b). In this (cid:128)gure, each data point
corresponds to two runs of the same seed. For a given seed, we
record the contrast score of the adapted co-dense algorithm as X
and the contrast score of our contrast subgraph mining algorithm
as Y. (cid:140)erefore, the data points above y = x indicates be(cid:138)er perfor-
mance than the adapted co-dense algorithm, vice versa. We plot
the y = x line for reference. It is not surprising that the contrast
subgraph mining algorithm will always have a higher score than
the adapted co-dense algorithm because they have slightly di(cid:130)er-
ent objectives. (cid:140)e "contrast" obtained from the adapted co-dense
algorithm emphasizes more on the density vs. the sparsity. (cid:140)is
is from a macro view, while our contrast(g) is de(cid:128)ned in a micro
view by considering every edge.
In summary, our contrast subgraph mining algorithm signi(cid:128)-
cantly outperforms the adapted co-dense subgraph mining algo-
rithm regarding both e(cid:129)ciency and e(cid:130)ectiveness.
6 REAL-WORLD APPLICATIONS
In this section, we present two real-world applications of contrast
subgraph mining: (1) spatio-temporal event detection based on
a taxi trajectory dataset and (2) trend detection on e-commerce
platforms based on an Amazon rating dataset.
6.1 Spatio-Temporal Event Detection
(cid:135)e Two Graphs. We construct two graphs using the tra(cid:129)c data
and road network in Beijing. (cid:140)e same taxi trajectory dataset in [10]
is utilized. We use the road network data of Beijing, which contains
148,110 road joints as nodes and 96,307 road segments as edges.
(cid:140)e GPS trajectories were generated by 33,000 taxis during a period
of 30 days in November 2012. Taxi trajectories are mapped onto
the road network using map matching algorithm proposed in [35].
For each road segment, its weight in GA is set as the number of
taxis traveled along this road segment between 10:30 PM and 11:00
PM on November 24th, 2012, while its weight in GB is set as the
average number of taxis traveled along this road segment between
10:30 PM and 11:00 PM from November 1st to November 23rd. (cid:140)e
neighborhood parameter r is set as 20 considering the roads are
segmented in a (cid:128)ne-grained way.
Visualization. A(cid:137)er identi(cid:128)ed the coherent core and extracted the
contrast subgraph, we project them on a map and utilize a red dot,
seeds
c
g \ c
Games & Strategy Guides
Computers & Internet, Games & Strategy Guides, Strategy Guide,
Programming, Video Games
Tomb Raider−, Half-Life−, Doom+
1.44
contrast(g)
Figure 8: Trend Detection on E-commerce Platform. (cid:135)e
darker, the more popular. In g \ c, − means more popular
in 2002, while + means more popular in 2004.
a blue circle, and a black dashed polygon to highlight the seed, the
coherent core, and the contrast subgraph, respectively.
Expectations. In this task, coherent cores are expected to contain
those road segments that are busy in both the real-time graph and
the historical graph, and contrast subgraphs will be likely those
road segments that are only busy in either the real-time graph of
the historical graph.
Results. We choose some seeds on the East 2nd Ring, a major road
in Beijing. As shown in Figure 2, our algorithm (cid:128)rst identi(cid:128)es a
coherent core about the usual, busy tra(cid:129)c on the East 2nd Ring,
and later extracts a contrast subgraph including many nodes and
road segments around the Beijing Workers' Sports Complex. (cid:140)e
contrast score is 23.87. (cid:140)is contrast subgraph demonstrates a
signi(cid:128)cantly busier real-time tra(cid:129)c than usual, therefore, it may
re(cid:131)ect some unusual event. A(cid:137)er some search on the Internet,
interestingly, on November 24, 2012, there was a concert hosted at
the Beijing Workers' Sports Complex, whose ending time was just
around 10:30 PM. (cid:140)is (cid:128)nding further consolidates the usefulness
of contrast subgraph mining and the e(cid:130)ectiveness of our algorithm
in real world applications.
6.2 Trend Detection on E-commerce Platform
(cid:135)e Two Graphs. We construct two trees based on Amazon's
product hierarchy. (cid:140)e Amazon rating dataset [14] is adopted to ex-
tract the tree and assign weights to each tree edge. (cid:140)ere are 14,222
nodes in total. (cid:140)e root node is "Subjects". A node is represented
as its pre(cid:128)x path from the root, such as "Subjects→Computers &
Internet→Games & Strategy Guides". (cid:140)e weight of an edge linking
from node u to node v equals to the log value (i.e., log(x) + 1) of
the total number of reviews x received by the products under the
node v. In GA, the total number of received reviews are calculated
based on the year 2002, while those in GB come from the year 2004.
We set r = 1.
Visualization. Considering these two graphs are trees, we directly
visualize the tree structures. Only nodes within the contrast sub-
graphs are colored. (cid:140)e darker, the bigger weights of the edge from
v's parent node to v. (cid:140)e blue polygon shows the coherent core
and the dashed black polygon encloses the contrast subgraph.
Contrast Subgraph Mining from Coherent Cores
KDD'18, August 2018, London, UK
Expectations. In this task, we expect the nodes in the coherent
core represent popular products types in both years, while the
nodes in the contrast subgraph demonstrate the speci(cid:128)c trends in
one of the two years.
Results. We choose "Subjects→Computers & Internet→Games &
Strategy Guides" as the seed. As shown in Figure 8, because the
popularity of each sub-category under "Games & Strategy Guides" is
similar in both years, they form the coherent core. Moreover, "Tomb
Raider" and "Half-Life" received many reviews in 2002 but not in
2004, while "Doom" became popular in term of the review amount
in 2004. (cid:140)erefore, these three nodes become the contrast subgraph
of a contrast score 1.44. A(cid:137)er some brief research in Wikipedia,
we found out that "Tomb Raider" released "(cid:140)e Prophecy" episode
in 2002 but nothing in 2004, while "Doom" released "Doom 3" in
2004 and its predecessor (i.e., "Doom 64") was long time ago in
1997. (cid:140)ese results are constructive for the sales trend analysis on
e-commerce platforms and further prove the e(cid:130)ectiveness of our
algorithm on graphs of abstract nodes like hierarchical trees.
7 CONCLUSION & FUTURE WORK
In this paper, we formulated the important contrast subgraph min-
ing problem. To avoid meaningless contrast subgraphs, we pro-
posed to (cid:128)rst identify coherent cores as cornerstones. Our frame-
work can admit a general family of coherence, contrast, and node
penalty metrics. A(cid:137)er rigorous derivations, we developed an el-
egant polynomial-time algorithm to (cid:128)nd the global optimum for
this problem. Extensive experiments veri(cid:128)ed the necessity of intro-
ducing coherent cores as well as the e(cid:129)ciency and e(cid:130)ectiveness of
our proposed algorithm. Real-world applications demonstrated the
tremendous potentials of the contrast subgraph mining.
In future, we will (1) extend the problem se(cid:138)ing to more than
two graphs; (2) (cid:128)nd the top-k contrast subgraphs with overlaps; and
(3) apply our contrast subgraph mining algorithm to other tasks,
like the abnormality detection in social networks.
REFERENCES
[1] Y. Asahiro, K. Iwama, H. Tamaki, and T. Tokuyama. Greedily (cid:128)nding a dense
subgraph. Journal of Algorithms, 34(2):203–221, 2000.
[2] S. Bedathur, K. Berberich, J. Di(cid:138)rich, N. Mamoulis, and G. Weikum. Interesting-
phrase mining for ad-hoc text analytics. Proceedings of the VLDB Endowment,
3(1-2):1348–1357, 2010.
[3] B. Boden, S. Gunnemann, H. Ho(cid:130)mann, and T. Seidl. Mining coherent subgraphs
in multi-layer graphs with edge labels. In Proceedings of the 18th ACM SIGKDD
international conference on Knowledge discovery and data mining, pages 1258–
1266. ACM, 2012.
[4] B. Boden, S. Gunnemann, H. Ho(cid:130)mann, and T. Seidl. Mimag: mining coherent
subgraphs in multi-layer graphs with edge labels. Knowledge and Information
Systems, 50(2):417–446, 2017.
[5] V. M. Buote, E. Wood, and M. Pra(cid:138). Exploring similarities and di(cid:130)erences between
online and o(cid:132)ine friendships: (cid:140)e role of a(cid:138)achment style. Computers in Human
Behavior, 25(2):560–567, 2009.
[6] C. Gao and S. Michel. Top-k interesting phrase mining in ad-hoc collections using
sequence pa(cid:138)ern indexing. In Proceedings of the 15th International Conference on
Extending Database Technology, pages 264–275. ACM, 2012.
[7] D. Gibson, R. Kumar, and A. Tomkins. Discovering large dense subgraphs in
massive graphs. In Proceedings of the 31st international conference on Very large
data bases, pages 721–732. VLDB Endowment, 2005.
[8] A. V. Goldberg. Finding a maximum density subgraph. University of California
Berkeley, CA, 1984.
[9] K. Henderson, B. Gallagher, T. Eliassi-Rad, H. Tong, S. Basu, L. Akoglu, D. Koutra,
C. Faloutsos, and L. Li. Rolx: structural role extraction & mining in large graphs.
In Proceedings of the 18th ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 1231–1239. ACM, 2012.
[10] L. Hong, Y. Zheng, D. Yung, J. Shang, and L. Zou. Detecting urban black holes
based on human mobility data. In Proceedings of the 23rd SIGSPATIAL International
Conference on Advances in Geographic Information Systems, page 35. ACM, 2015.
[11] H. Hu, X. Yan, Y. Huang, J. Han, and X. J. Zhou. Mining coherent dense subgraphs
across massive biological networks for functional discovery. Bioinformatics,
21(suppl 1):i213–i221, 2005.
[12] D. Jiang and J. Pei. Mining frequent cross-graph quasi-cliques. ACM Transactions
on Knowledge Discovery from Data (TKDD), 2(4):16, 2009.
[13] N. Jin and W. Wang. Lts: Discriminative subgraph mining by learning from search
history. In Data Engineering (ICDE), 2011 IEEE 27th International Conference on,
pages 207–218. IEEE, 2011.
[14] N. Jindal and B. Liu. Opinion spam and analysis. In Proceedings of the 2008
International Conference on Web Search and Data Mining, pages 219–230. ACM,
2008.
[15] A. Kavanaugh, J. M. Carroll, M. B. Rosson, T. T. Zin, and D. D. Reese. Community
networks: Where o(cid:132)ine communities meet online. Journal of Computer-Mediated
Communication, 10(4):JCMC10417, 2005.
[16] S. Khot. Ruling out ptas for graph min-bisection, dense k-subgraph, and bipartite
clique. SIAM Journal on Computing, 36(4):1025–1071, 2006.
Machine Learning and Knowledge Discovery in Databases, pages 33–49, 2008.
J. Liu, J. Shang, C. Wang, X. Ren, and J. Han. Mining quality phrases from massive
text corpora. In Proceedings of the 2015 ACM SIGMOD International Conference
on Management of Data, pages 1729–1744. ACM, 2015.
[19] D. P. A. D. D. Majumdar. Fast mining of interesting phrases from subsets of text
corpora. 2014.
[17] G. Liu and L. Wong. E(cid:130)ective pruning techniques for mining quasi-cliques.
[20] A. McCallum, A. Corrada-Emmanuel, and X. Wang. Topic and role discovery in
[18]
[22]
social networks. In Ijcai, volume 5, pages 786–791. Citeseer, 2005.
[21] B. L. Monroe, M. P. Colaresi, and K. M. (cid:139)inn. Fightin'words: Lexical feature
selection and evaluation for identifying the content of political con(cid:131)ict. Political
Analysis, 16(4):372–403, 2008.
J. B. Orlin. Max (cid:131)ows in o (nm) time, or be(cid:138)er. In Proceedings of the forty-(cid:128)(cid:135)h
annual ACM symposium on (cid:138)eory of computing, pages 765–774. ACM, 2013.
[23] C. H. Papadimitriou and K. Steiglitz. Combinatorial optimization: algorithms and
complexity. Courier Corporation, 1998.
J. Pei, D. Jiang, and A. Zhang. On mining cross-graph quasi-cliques. In Proceedings
[24]
of the eleventh ACM SIGKDD international conference on Knowledge discovery in
data mining, pages 228–238. ACM, 2005.
[25] R. A. Rossi and N. K. Ahmed. Role discovery in networks. IEEE Transactions on
Knowledge and Data Engineering, 27(4):1112–1131, 2015.
[26] R. A. Rossi, D. F. Gleich, A. H. Gebremedhin, and M. M. A. Patwary. Fast maximum
clique algorithms for large graphs.
In Proceedings of the 23rd International
Conference on World Wide Web, pages 365–366. ACM, 2014.
J. Shang, J. Liu, M. Jiang, X. Ren, C. R. Voss, and J. Han. Automated phrase
mining from massive text corpora. arXiv preprint arXiv:1702.04457, 2017.
[28] M. (cid:140)oma, H. Cheng, A. Gre(cid:138)on, J. Han, H.-P. Kriegel, A. Smola, L. Song, P. S.
Yu, X. Yan, and K. M. Borgwardt. Discriminative frequent subgraph mining
with optimality guarantees. Statistical Analysis and Data Mining: (cid:138)e ASA Data
Science Journal, 3(5):302–318, 2010.
[29] R. M. H. Ting and J. Bailey. Mining minimal contrast subgraph pa(cid:138)erns. In
Proceedings of the 2006 SIAM International Conference on Data Mining, pages
639–643. SIAM, 2006.
[30] M. van Leeuwen, T. De Bie, E. Spyropoulou, and C. Mesnage. Subjective inter-
[34] X. Yan and J. Han. Closegraph: mining closed frequent graph pa(cid:138)erns.
estingness of subgraph pa(cid:138)erns. Machine Learning, 105(1):41–75, 2016.
[31] C. Wang, J. Han, Y. Jia, J. Tang, D. Zhang, Y. Yu, and J. Guo. Mining advisor-
advisee relationships from research publication networks. In Proceedings of the
16th ACM SIGKDD international conference on Knowledge discovery and data
mining, pages 203–212. ACM, 2010.
[32] B. Wellman, J. Boase, and W. Chen. (cid:140)e networked nature of community: Online
and o(cid:132)ine. It & Society, 1(1):151–165, 2002.
[33] X. Yan and J. Han. gspan: Graph-based substructure pa(cid:138)ern mining. In Data
Mining, 2002. ICDM 2003. Proceedings. 2002 IEEE International Conference on,
pages 721–724. IEEE, 2002.
In
Proceedings of the ninth ACM SIGKDD international conference on Knowledge
discovery and data mining, pages 286–295. ACM, 2003.
J. Yuan, Y. Zheng, C. Zhang, X. Xie, and G.-Z. Sun. An interactive-voting based
map matching algorithm. In Mobile Data Management (MDM), 2010 Eleventh
International Conference on, pages 43–52. IEEE, 2010.
[36] C. Zhang, G. Zhou, Q. Yuan, H. Zhuang, Y. Zheng, L. Kaplan, S. Wang, and
J. Han. Geoburst: Real-time local event detection in geo-tagged tweet streams.
In Proceedings of the 39th International ACM SIGIR conference on Research and
Development in Information Retrieval, pages 513–522. ACM, 2016.
[37] S. Zhang, H. Jiang, and J. M. Carroll. Integrating online and o(cid:132)ine commu-
nity through facebook. In Collaboration Technologies and Systems (CTS), 2011
International Conference on, pages 569–578. IEEE, 2011.
[27]
[35]
|
1912.07551 | 1 | 1912 | 2019-12-16T17:56:56 | Following the footsteps of giants: Modeling the mobility of historically notable individuals using Wikipedia | [
"cs.SI",
"physics.soc-ph"
] | The steady growth of digitized historical information is continuously stimulating new different approaches to the fields of Digital Humanities and Computational Social Science. In this work, we use Natural Language Processing techniques to retrieve large amounts of historical information from Wikipedia. In particular, the pages of a set of historically notable individuals are processed to catch the locations and the date of people's movements. This information is then structured in a geographical network of mobility patterns. We analyze the mobility of historically notable individuals from different perspectives to better understand the role of migrations and international collaborations in the context of innovation and cultural development. In this work, we first present some general characteristics of the dataset from a social and geographical perspective. Then, we build a spatial network of cities, and we model and quantify the tendency to explore by a set of people that can be considered historically and culturally notable. In this framework, we show that by using a multilevel radiation model for human mobility, we are able to catch important features of migration's behavior. Results show that the choice of the target migration place for historically and culturally relevant people is limited to a small number of locations and that it depends on the discipline a notable is interested in and on the number of opportunities she/he can find there. | cs.SI | cs |
Lucchini et al.
RESEARCH
Following the footsteps of giants: Modeling the
mobility of historically notable individuals using
Wikipedia
Lorenzo Lucchini1,2*, Sara Tonelli1 and Bruno Lepri1
*Correspondence: [email protected]
1Fondazione Bruno Kessler, via
Sommarive, 18, 38122 Trento,
Italy
Full list of author information is
available at the end of the article
Abstract
The steady growth of digitized historical information is continuously stimulating
new different approaches to the fields of Digital Humanities and Computational
Social Science. In this work we use Natural Language Processing techniques to
retrieve large amounts of historical information from Wikipedia. In particular, the
pages of a set of historically notable individuals are processed to catch the
locations and the date of people's movements. This information is then
structured in a geographical network of mobility patterns.
We analyze the mobility of historically notable individuals from different
perspectives to better understand the role of migrations and international
collaborations in the context of innovation and cultural development. In this
work, we first present some general characteristics of the dataset from a social
and geographical perspective. Then, we build a spatial network of cities, and we
model and quantify the tendency to explore of a set of people that can be
considered as historically and culturally notable. In this framework, we show that
by using a multilevel radiation model for human mobility, we are able to catch
important features of migration's behavior. Results show that the choice of the
target migration place for historically and culturally relevant people is limited to a
small number of locations and that it depends on the discipline a notable is
interested in and on the number of opportunities she/he can find there.
Keywords: Human behavior understanding; Natural language processing;
Computational social science; Network theory; Human mobility
1 Introduction
Ever since the first villages were built by primitive people, humankind has moved
from one community to another, in search of better life conditions or new oppor-
tunities [1]. These conditions can be represented by different factors, such as the
opportunity of a job, better living standards, or the distance from the home country
[2, 3]. The set of all these factors is difficult to define a-priori.
For example, according to the dominant neo-classical theory, people tend to make
choices in order to maximize their income or level of well-being [4, 5]. Thus, the
search for better economic conditions is one of the most important factors in the
decision of moving from one location to a more attractive one. However, there are
other factors that could play an important role in the decision-making process of a
specific group of people. The attractiveness of an opportunity can also depend on
cultural and linguistic barriers, and on the presence of particular communities at
Lucchini et al.
Page 2 of 17
the destination [4, 6, 7]. Hence, in order to define a complete migration framework
it is important to consider all the linked aspects that intervene in modifying the
attractiveness of a location (e.g. economical, environmental, cultural, and political
aspects) [4, 7, 8]. As a consequence, we believe that, in the same way as the economic
conditions play a central role for those seeking employment, it is very relevant to
investigate if there are other factors playing a role for specific kinds of migration, for
example the migration of notable people and intellectuals in the course of history
[9].
For the general problem of human migration different mathematical models have
been built both in a descriptive [10, 11, 12, 13] and predictive [14] framework. How-
ever, these models have never been applied to the specific scenario of modeling the
historical movements and migrations of intellectual figures. Instead, mobility and
migrations historically played, and still play, an important role in the process of cul-
tural evolution introducing seeds of change in different places around the world [15].
Thus, understanding the patterns that historically notable individuals followed dur-
ing their lives and how these affected the cultural evolution and the human history
is an intriguing and still open research question [16]. This approach introduces new
challenges because of the specificity of the problem and the relatively small num-
ber of people that had an impact on cultural evolution and human history. New
perspectives were opened by a recent approach proposed by Schich et al. [17]. By
analyzing birth and death locations historically notable individuals they captured,
from a network perspective, the key characteristics of their exploratory behavior.
In our paper, we propose a way of using Natural Language Processing (NLP)
and Network Science techniques to further characterize and model the mobility of
historically and culturally notable individuals and to investigate the factors playing
a role in their migration patterns.
In the last decade, the NLP community has developed technologies for extract-
ing information from unstructured texts, thus enabling their application also to
interdisciplinary research areas. Understanding and modeling historical migration
phenomena require specific historical data. To this end, we propose to use NLP
techniques to process the digital biographies contained in Wikipedia and to extract
migratory events from its encyclopedic information. In particular, a subset of the
biographies available in the English version of Wikipedia is used as raw data source.
From this source, for each notable person we search for her/his footsteps hidden in
her/his Wikipedia page and collect the following information: (i) the place and date
of birth, (ii) the place and date of death, and (iii) the place and date of the various
in-life migrations (e.g. moving from one city to a different one). This results in a
more complete set of data with respect to [17], enriching the global picture of the
mobility of notable individuals with a finer temporal granularity, while enabling, to
the best of our knowledge, to model this process from a historical point of view for
the first time.
As proposed in [18], we use the term culture to focus our attention on the set of
notable contributions to the development of human history in its broadest sense:
from poetry to sports, from music to physics and mathematics. In particular, we
model the mobility dynamics of notable people, namely those people whose cultural
production is known at a global level. To this end, we introduce a modification to
Lucchini et al.
Page 3 of 17
the radiation model for human mobility [13]. The assumption behind the radiation
model resides in using the city size (i.e. the city population) as a proxy of the
number of job opportunities. In our paper, we modify the radiation model to take
into account, in addition to the role played by the city size, also the attractive role
played by the different disciplines and by the number of notable people as proxies
of cultural opportunities.
Then, we compare the predictive performances of our cultural-based radiation
model and of the state-of-the art population-based radiation model on three main
aspects of the migration processes of notable people: (i) the radius of gyration of
each notable person, (ii) the number of different cities the notable people lived in
during their lives, and (iii) the distances between the source of a migration and its
destination (jump lengths). Interestingly, our results show that the radius of gyra-
tion and the jump lengths are best modeled considering three different factors: (i)
the population of the city, as a proxy for the economical wealth and job opportuni-
ties, (ii) the number of notable individuals that spent some time of their lives in a
given city, as a proxy for the role played by the city as a cultural attractor, and (iii)
the specific discipline an historically and culturally notable person is working in,
as a proxy both for the interests a city has in investing on a specific cultural area
and for the tendency that people, interested or working on that discipline, have to
follow notable figures from the same domain.
Our results pave the way for further investigations on the historical role played
by places and cities (e.g. ancient Athens, Renaissance Florence, Song Dynasty
Hangzhou, Vienna of 1900, Silicon Valley, etc.) in becoming cultural attractors
for historically and culturally notable individuals and, thus, making them grow into
flourishing places for novel artistic and literary movements, for new scientific and
philosophical theories, for social, technological and political innovations [19].
2 Materials and Methods
2.1 Data
In this section, we present and discuss how we build a dataset containing biographi-
cal information on thousands of historically relevant people in the context of cultural
production and innovation. Together with information about their field of influence,
we also extract information about their mobility patterns.
We start from the set of notable people identified by the Pantheon project [18].
This project collected the biographies of 11, 341 historically and culturally relevant
people that lived from 3500 B.C. to 2010 A.D. More precisely, the dataset was built
by extracting information from the Wikipedia biography and info-boxes. Here, a
person is defined as notable if the corresponding Wikipedia page is translated in
25 or more different languages, since the focus is on global historical and cultural
contributions.
For each notable person, several features were annotated and manually verified,
among which we consider the following ones:
• birth place (geo-localized), state, and date;
• occupation, work area, and discipline.
While the process of selection of notable people in Yu et al. [18] resulted in 11, 341
biographies selected from more than one million biographies available in Wikipedia,
Lucchini et al.
Page 4 of 17
Table 1 Composition of the selected subset of notable people, organized by the different work
areas and disciplines present in the Pantheon dataset [18]. In particular, 15 work areas are
presented, sorted by discipline.
Work Area
Discipline
Percentage (%)
Film and Theatre
Music
Fine Arts
Design
Natural Sciences
Social Sciences
Medicine
Math
Language
Philosophy
Government
Military
Activism
Arts
Arts
Arts
Arts
Science and Technology
Science and Technology
Science and Technology
Science and Technology
Humanities
Humanities
Institutions
Institutions
Public Figure
Individual Sports
Sports
Business
Business and Law
10.35
8.81
5.61
1.68
13.74
3.92
2.25
2.04
17.39
3.51
15.90
3.81
1.23
1.28
1.07
we further restrict our analyses on the notable people that were actively working and
producing cultural outcomes in a specific time-window, the first fifty years of the
20th century. The focus on this specific time-window is justified by two combined
needs. The first is to have a sufficiently short time span so that the properties
of the global mobility do not change. The second is to have the highest possible
number of notable people with complete and precise migratory information, i.e.
with birth, in-life, and death locations. The first requirement reduces the width of
the time-window, while the second one requires to consider relatively recent years.
We therefore consider active the notable people that were at least 20 years old by
the end of the considered time-window to ensure that their historical and cultural
contribution was made during this time span, and that no bias in the type of
migratory event (birth, in-life, and death) is introduced.
By applying this filtering procedure based on the time-window of interest, we
reduce the number of Wikipedia biographies to be processed to 2, 407. We report
in Table 1 the distribution of work areas and disciplines present in this subset of
Wikipedia biographies, which will be further processed as described in the next
Section.
2.1.1 Extracting migration footsteps
For the purposes of our study, we are interested in identifying the different locations
that were visited by the selected set of notable people during their lives and the
year of their visits, in order to build a trajectory (made of multiple footsteps)
for each notable person. This kind of information is not present in the Pantheon
dataset and we therefore need to extract it automatically. The approach we adopt
follows the one recently proposed with Ramble-On [20], a text processing pipeline
dealing with two main tasks: (i) the identification of predicates of migration and
their arguments (i.e. the subject of the migration frame) in Wikipedia biography
pages, and (ii) the recognition and classification of dates, places, and mentions. We
Lucchini et al.
Page 5 of 17
Table 2 Example of identification and classification of a sentence using Semaphor and FrameNet.
Snippet
Predicate
Frame
Place
Time
"Paul moved to Chicago in 1934,
where he continued to
perform on radio."
moved
Motion
Chicago
in 1936
focus our attention on migration processes because they are more likely to describe
a motion action that resulted in a long time permanence in a new location. As a
consequence, if the permanence in a specific location is long, it is more likely that the
notable person had the time to provide his/her cultural contribution there. While
there have been recent attempts to automatically extract people's trajectories and
an associated time period from Wikipedia biographical pages [21], these rely on
shallow NLP approaches based on the presence of keywords and geo-links in the
pages. Furthermore, trajectories are coupled with time spans and not time points
like in our study, leading to a lower granularity of the extracted information. In our
case, the use of semantic parsing associated with a selection of predicates describing
possible trajectories enables a very precise analysis of the resulting data.
To identify the predicates related to migration events, the Ramble-On applica-
tion[1] calls PIKES [22], a suite for NLP that extracts information from English
plain texts, which in turn automatically assigns to each predicate a semantic frame
based on the FrameNet classification [23]. Also the arguments attached to each pred-
icate are automatically labeled with semantic roles, relying on the frame-semantic
parser Semafor [24] . To better distinguish migration predicates, we again follow
the approach proposed in [20], thus removing 16 motion frames (e.g. Escaping,
Getting underway, Touring) out of 45 because of the high number of false posi-
tives found during the identification. Hence, 29 motion frames (e.g. Arriving, Be-
ing employed, Transfer, Travel ) were used for the identification of notable people's
migration actions described in their biographies.
Once a migration frame was identified in a sentence extracted from a Wikipedia
biography page, three elements are required to be present in order to extract a
migration trajectory: (i) the time/date of the motion, (ii) the traveler, and (iii)
the destination. Again, the Ramble-On application selects only sentences satisfying
these constraints, where the date, the notable person (or a reference to him/her)
and the destination have been identified. With this approach, precision is favoured
over recall, requiring that these three elements are explicitly mentioned in the same
sentence.
In Table 2 we show as an example the snippet of a sentence identified as a move-
ment. In this case the predicate "moved" is assigned to the frame "Motion" and it
is selected as a migration frame. Then, the different arguments of the sentence are
identified and labelled depending on their role in the sentence (e.g. the time, the
traveller and the destination/place).
Once the movements have been identified and the information about the date
and the location have been extracted, Ramble-On geo-locates each word related
to the identified destinations using OpenStreetMap Nominatim [2], a search engine
for geo-referenced OpenStreetMap locations. Destinations that lack coordinates are
[1]Code available here: https://github.com/dhfbk/rambleon
[2]https://nominatim.openstreetmap.org/
Lucchini et al.
Page 6 of 17
discarded from the movements' list since they may be erroneously annotated as
destinations. Besides, place and date of death are retrieved for each biography by
Ramble-On using DBpedia [3], where structured data about each notable person are
stored.
In Table 3 we present the information retrieved by processing the same biogra-
phy as in Table 2. Together with the birth information retrieved from Pantheon,
other important information such as date and place of death are extracted from
DBpedia as discussed above. The Migration 1 element, M1, is the first additional
trajectory retrieved by running the Ramble-On tool on the biography page. Merg-
ing the information retrieved about the inventor and musician Les Paul (see Table
2), we identify two jumps, i.e. two migratory events: (i) the first one from the birth
location to Chicago, and (ii) the second one from Chicago to the death location.
Table 3 Results obtained by processing Les Paul's biography using the Ramble-On pipeline.
Birth
Movement 1 Death
date
place
latitude
longitude
predicate
resource
place frame
resource frame
19150000
Waukesha
43.0117
-88.2317
null
dbpedia
null
Birth
19340000
Chicago
41.8369
-87.6847
moved
FrameNet
@Goal
Motion
20091231
White Plains
41.0400
-73.7786
null
dbpedia
null
Death
We refer to [20] for a detailed discussion on the performance of the extraction pro-
cess. In brief, Ramble-On has a precision of 0.86 in correctly identifying migration
frames. As mentioned before, however, the strategy adopted to identify trajectories
may penalize recall, failing to extract movements whose date or destination are
mentioned implicitly or in two different sentences. In order to estimate the amount
of locations visited (birth and death locations included) that we are not able to
capture in our study, we manually annotate the trajectories in the biographies of 50
notable people randomly sampled stratifying over the number of locations found.
In this way, we have estimated that the recall of Ramble-on approach is equal to
0.59.
2.1.2 Dataset composition
The processing of the 2, 407 biographies results in a set of 7, 240 locations connected
with notable persons' trajectories. Among these, we consider the 4, 028 movements
taking place in the 1900-1950 time-window. Each movement with the associated
date and destination was then manually checked by comparing the extracted in-
formation with the source Wikipedia sentence, and corrected if necessary. Also the
coordinates associated with the extracted locations by Nominatim were manually
checked, since the geographical information associated with trajectories is at the
core of our migration model and possible errors must be minimized. These are then
collapsed to the nearest great city, where we adopted as a definition of great cities
the list proposed in [25]. In their work, Reba et al. [25] collected also precious his-
torical demographic data for most of these cities, that we used to test our baseline
for the migration model. More specifically, geo-localized locations are merged based
[3]http://wiki.dbpedia.org/
Lucchini et al.
Page 7 of 17
Figure 1 Two examples of the trajectories obtained after the extraction. Panel A on the left
shows the travels made by Albert Einstein while panel B those made by Maria Montessori.
on a Voronoi tessellation of the Earth. In this framework each cell is built from a list
of cities for which historical population data was available [25]. The space is built
using the great-circle approximation to associate each identified location with the
corresponding Voronoi cell. The distribution of the clustering process is reported in
Fig. SM7 of the Supplementary Materials (SM). In Fig. 1 we present two examples
of trajectories for Albert Einstein, the famous physicist, and Maria Montessori, the
renowned physician and educator. The arcs connects different locations where these
two notable figures spent a part of their lives. The blue coloured side of an arc
indicates the origin of the migration while the red one its destination. For example,
the extraction well captures Einstein's movements from Zurich to Berlin and from
Berlin to US. His first movement from Ulm, his home town, is missing since it hap-
pened before the beginning of 20th century. We also notice that his short period as
visiting professor to Caltech is detected by Ramble-On. Similarly, Maria Montes-
sori's experiences around Europe (i.e. Barcelona, Amsterdam, Vienna, Rome) are
correctly identified. In contrast, we stress that, due to the lack of population data,
her trips to Sri-Lanka are collapsed to cities in India. This shows how the collapsing
process might impact the actual migration distribution.
The merging step results in a set of 629 different cities visited by our notable
people during their lifetime. Figure 2.A shows the members of our set of notable
individuals listed by discipline (as labelled by [18]); while Figure 2.B shows the
distribution of different visited cities for the top two discipline communities, namely
"Arts" and "Science and Technology". The colored dots represent the data while
the lines the geometrical fit to the data. Both the distributions can be described
using a geometrical distribution with parameter p, representing the probability of
successfully settling in a city, p ∼ 1
2 .
2.2 Cultural network and migration modeling
In our framework, we assume that a culturally notable person living in a place
for a certain period of time contributed in some way to make such place a cultural
attractor for other people interested in cultural innovation and development. At the
same time, when a culturally notable person was moving from one place to another,
s/he linked the cultural relations s/he had and the work s/he did in the first place
with the relations and work in the second place. In this way, each movement creates
a cultural connection between two different places around the world. Depending
on the number of notable people moving from one place to another, we can add
Lucchini et al.
Page 8 of 17
Figure 2 A: Size of disciplinary communities present in the dataset. From the bar-plot it is evident
the difference in the number of notable individuals between arts, science&technology, humanities,
institutions and public figures, sports, business&law, exploration. B: Distribution of the number of
trajectories per person in the set of notable people. The lines represent the geometric fit to the
distributions of the two most numerous disciplines, i.e. Arts and Science and Technology. Both
distributions can be described in terms of a geometric distribution with parameter p ∼ 1
2 .
a weight to the links of this cultural network. Thus, the nodes correspond to the
different cities visited by our set of notable people, while the edges are cultural links,
built by interpersonal relations together with the cultural contamination a person
brings with herself/himself while migrating from one place to another, weighted by
the number of occurrences.
In Figure 3 we show a representation of this weighted directed network, where
the weight is the percentage of notable people migrating from a location to their
destination. In the left panel and right panel we show a sector of the network
considering different time frames. On the left, we restrict our focus on a 3-year time-
window centering the map to highlight the connections between Eastern Europe
and Asia (Union of Soviet Socialist Republics in particular) during the October
revolution in 1917. It is interesting to notice that the structure of the network
nicely catches the known phenomenon of Siberian exile of aristocratic families and
important political personalities such as representatives of the previous tsarists'
power and relevant persons not aligned with the current regime. In this example,
we are capturing the movements of people that were forced to migrate to specific
locations. Hence, our notion of cultural attractor both includes locations to which
people moved voluntarily and locations to which culturally notable people were
forced to move. In particular, the number of forced movements (i.e. people sent to
concentration camps and imprisoned people) is 36 out of 3474 identified movements.
Top right panel shows the strong migratory flux of intellectuals from Europe to the
US during the Second World War.
As previously said, we are interested in modeling the mobility of culturally relevant
figures and in investigating the factors playing a role in their migration patterns.
To this end, we modify the radiation model to take into account the role played
by the city size (i.e. proxy for job opportunities) as well as the ones played by the
different disciplines and by the number of notable people (i.e. proxies for cultural
opportunities). The radiation model [13] describes the mobility of people seeking
job opportunities in terms of job openings per number of inhabitants. The model
is developed in the framework of network theory since it treats cities as nodes of a
246810TRIPS102101100B: Number of tripsSCIENCE & TECHNOLOGYp = 0.491rmse = 0.014ARTSp = 0.506rmse = 0.010102103MEMBERSARTSSCIENCE & TECHNOLOGYHUMANITIESINSTITUTIONSPUBLIC FIGURESPORTSBUSINESS & LAWEXPLORATIONA: Discipline communitiesMembers Lucchini et al.
Page 9 of 17
Figure 3 A representation of the spatial directed network encoding the mobility information of the
chosen set of notable people during the first half of the 20th century. We use the blue hexagon to
indicate the source point and the red to indicate the destination of the migratory jump. In this
figure, we present the network structure considering the migratory events during two different
time snapshots. Figure 3.a on the left shows the escaping and exiling from Saint Petersburg of the
aristocratic families during the years of the red revolution. Similarly, Figure 3.b on the left shows
the migration flows from Europe to North America during World War 2, due to the fascist regimes
and persecutions.
completely connected weighted network. Specifically, the radiation model describes
the human mobility behavior at long distances, e.g. at the country or global scale,
better than other often used models (e.g. gravity model) [26]. It is also important to
notice that it undershoots the real flows. Moreover, its performances are dependent
on the structure of the system even though it directly accounts for variations of
the population between the source and the destination of a migration, i.e. the less
population you have between two cities the more probable is to migrate from one
to the other.
Simini et al. [13] show that, using this formulation of the problem, the flow of
people between cities only depends on the population of the two cities (namely mi
and ni), and the population living in the circle of radius rij is equal to the distance
between the two cities (namely sij). The relation can be summarized in a simple and
parameter-free equation. We report here the formula for the probability to move
from city i to city j:
Pij = Pi
(mi + sij)(mi + nj + sij)
nj
,
(1)
j∈N
nj
ability of moving from i to every city (i included): Pi =(cid:80)
where Pi is the normalization coefficient for city i that ensures that Pij is the prob-
(mi+sij )(mi+nj +sij )
,
where N is the set of all nodes present in the network. In our work we make use of
these concepts to model in a similar way the mobility of culturally relevant people.
In particular, inspired by multidimensional network theory and it recent applica-
tions in modeling human mobility [27, 28, 29], we propose a multilevel approach to
cultural mobility. In this framework, every cultural discipline works as a separate
system described by a cultural radiation model. Formally, a level is a fully connected
weighted and directed network in which the nodes are the cities visited by all the
notables of a specific discipline and the links represent the probability of migrating
from a city to a different one. Each node has also a link pointing to itself, represent-
ing the probability of remaining in the same city instead of moving to a different
Lucchini et al.
Page 10 of 17
place. The different levels, l ∈ L where L is the set of all disciplines, do not interact
with each other but their contribution to the overall migratory exploration sums
up. Each level contributes to the global migration model with a factor proportional
to the share of notable people the discipline has, N S. Thus, the probability of this
multilevel migratory network can be described by
Pij =
N SlPil
njl
(mil + sijl )(mil + njl + sijl )
,
(2)
where N Sl is the notable share of discipline l ∈ L, mil and njl are the population
of locations i and j respectively in the discipline level l, and sijl is equivalent to
sij for the specific level l. In a similar way, the generalized Pi normalizes Pij to a
probability following the idea of equation (1):
(cid:88)
∀l∈L
(cid:88)
(mil + sijl )(mil + njl + sijl )
Pil =
j∈N
njl
(3)
Starting from these two equations, we propose different implementations of the
radiation model. In particular, we stress that the radiation model introduces the
concept of attractiveness of a city based on the number of job opportunities that a
city can provide. The assumption that this number is directly proportional to the
number of people living in a city [13] directly connects the concept of attractive-
ness with the population size of a city. Here, with similar assumptions we propose
different formulations of the model based on different possible ways of modeling
cultural attractiveness. In particular, we use, as possible alternatives to the stan-
dard formulation of the model, the number of notable people that visited a city
and its combination with the population of the city. In the case of notable people
that visited a city, we count this quantity considering all the visits during the whole
time-window. As a consequence, we model cultural dynamics of individuals within
this period of observation considering the effects of the notable people distribu-
tion as a constant feature of our model, as we do for general population, which is
not updated after every step of the dynamics. This is equivalent to assume that
a single step of the dynamics has a latency larger than the size of the considered
time-window in affecting the importance of cultural attractors. Moreover, counting
notable people in this way also relies on the simplifying assumption that they all
equally contributed to the importance of a city from a cultural perspective. More
realistic modeling will require a relevance score based on the historical relevance of
the notable people.
We test each of these three possible definitions, namely the standard one based on
population size, the one based only on the number of notable people, and the com-
bination of the two, using both a single level formulation of the radiation model and
a multilevel formulation. Using this probabilistic model, our aim is to understand if
the radiation model abstraction can be used to describe the level of exploration of
the historically and culturally relevant figures (namely, the radius of gyration of each
notable people, the number of different cities visited, and the distance distribution
of the migration jumps) and which formulation better captures these properties. In
the next Section, we discuss first the general information that can be obtained by
Lucchini et al.
Page 11 of 17
analyzing the system in terms of network theory metrics and then the comparison
between the different formulations proposed.
3 Results
3.1 Properties of the migration network
One of the most interesting characteristics of cultural migration patterns is the
tendency of notable figures to explore different cities. To study this property, we
can define S(t) as the number of different cities and N (t) as the number of notable
people's birth locations, the number of their death locations, and the number of
their jumps during the selected time-window, as displayed in Figure 4.A. by the
curves for Birth, Death and In-life respectively. The growth of S(t) is modeled as a
function of N (t) using a Heap's law S(t) = N (t)α. Our result is consistent with the
estimate of the parameter α for the Birth curve obtained by Schich et al. in [17].
A similar result is obtained also for the Death and In-life curves representing the
growth of the location for which the exponent, α = 0.85, suggests a tendency to
migrate to a smaller number of cities with respect to the number of different cities
where notable individuals were born. This finding may be interpreted as a general
and global tendency of notable figures to migrate to a more culturally renowned
subset of cities with respect to all the possible available locations.
Focusing on the migration jumps that notable people made during their life, we
can study the most central cities both from a global and discipline-based perspective.
Here, we use the Page-Rank centrality [30] to measure the importance in terms of
the number of incoming links that point to a city and the relative importance of
the cities from which these links are coming. In Figure 4.B, we measure Page-Rank
centrality for two different time-windows, namely 1900-1925 and 1926-1950, to show
the structural changes of the network during the first half of the 20th century. It
is interesting to notice how the development of the film industry in Los Angeles
attracted several figures to the city. It is also worth noticing how, due to the Second
World War (WW2), Berlin loses positions in the ranking of the more central cities.
In Figure SM4, we highlight the specific effect of WW2, evaluating Page-Rank
centrality before and after the rise of the Nazism regime in Germany, showing the
overall loss in cultural centrality for most of the European cities.
We also evaluated Page-Rank centrality for the sub-network built only by consid-
ering the migration jumps of the four top disciplines in those years. Figure 5 shows
how notable people from different disciplines migrated to different cities, suggesting
that the cultural centrality of a city depends on its cultural characteristics, e.g. Los
Angeles and the film industry. This, indeed, results in Los Angeles being a central
node for Arts (in particular for actors) and Sports, but a more peripheral one for
Institutions, having their more central nodes in capital cities such as London, Paris
and Moscow. In Fig. SM3 we performed a discipline Page-Rank analysis for two
time windows as in Fig. 4, showing that an important change is present also at the
discipline level. An example is given by the dramatic change (i.e. a decrease) in the
centrality of Berlin for the scientific community before and after 1933.
3.2 Cultural attractiveness in a multilevel radiation model
With the results obtained so far we aim at modeling notable figures' migratory
patterns to better understand what is the process driving the choice of the location
Lucchini et al.
Page 12 of 17
Figure 4 A: Number of different visited cities as a function of the notable birth/death/in-life
visited locations. In the case of in-life migrations the number of visited locations does not
necessarily corresponds to the number of notable figures in the dataset, but it represents the
different cities visited in time as a function of the number of trips made by the notable figures. B:
Page-rank centrality for the global mobility network evaluated in two different time-windows. In
blue we report the centrality values for the 1900 -- 1925 window, while in red the centrality in the
window 1926 -- 1950.
where to migrate. The radiation model proposed in [13] finds the motivating factor
of mobility for those seeking a job in the number of opportunities a city can provide.
Following this idea and using the results obtained by Simini et al. [13], we propose
a similar approach to understand if such a model can catch the main factors of
cultural mobility. We assume that the number of opportunities that are available in
the selected time-window is directly proportional to the number of notable figures
that lived in a city during the same time interval. In particular, we explore the
following different configurations:
• cultural opportunities are uniformly distributed among cities;
• cultural opportunities are proportional to the population of a city;
• cultural opportunities of a city are directly proportional to the number of
notable figures that lived in that city;
• cultural opportunities are directly proportional both to the population of a
city and to the number of notable figures that lived in that city.
In addition, we also want to check whether cultural opportunities depend on the
discipline a notable individual is part of. We study all these possibilities using the
formulation proposed in Section 2.2.
To find which model better describes the historical mobility of the first half of
20th century, we simulate mobility using a set of walkers that can move following a
radiation model over the cultural network, based on different equations depending
on the model we are simulating. Thus, depending on the selected configuration, we
are using respectively (i) a random walker dynamics' model over the network, (ii) a
standard (population-related) radiation model, and (iii) an implementation of the
radiation model that specifically considers the cultural opportunities as a subset of
job opportunities not necessarily driven by the same factors. The starting location
of each simulated notable individual is chosen based on a random choice weighted
by the population size of each city.
To compare the models we analyze their impact on predicting (i) the number of
different cities visited, as a proxy of the availability to explore different and new
Lucchini et al.
Page 13 of 17
Figure 5 Page rank centrality per discipline for the five most visited cities. This figures shows the
different importance of cities in terms of discipline specific attractiveness.
Table 4 Models' performances. Results obtained for the five models on predicting the number of
destinations, the radius of gyration, and the distribution of the length of the migration jumps. The
metrics used are the adjusted-R2, the Pearson correlation coefficient, ρ, between models and data,
the Kullback-Leibler distance (K-L dist), and the first Wasserstein distance (Wasserstein dist).
Model
adj − R2
Pearson ρ
K-L dist
Wass. dist
Radius of gyration
pop-notable-multilevel
0.2414 ± 0.0027
pop-multilevel −0.2004 ± 0.0034
notable-multilevel −0.6849 ± 0.0041
notable-singlelevel −1.0249 ± 0.0048
random-singlelevel −2.2673 ± 0.0054
Different destinations
pop-notable-multilevel
pop-multilevel
notable-multilevel
notable-singlelevel
random-singlelevel
0.9547 ± 0.0004
0.9612 ± 0.0004
0.9619 ± 0.0003
0.9624 ± 0.0003
0.9606 ± 0.0004
Length of migration jumps
pop-notable-multilevel
pop-multilevel
0.5104 ± 0.0019
0.2249 ± 0.0023
notable-multilevel −0.0640 ± 0.0029
notable-singlelevel −0.2192 ± 0.0029
random-singlelevel −0.8313 ± 0.0034
0.962∗∗∗
0.953∗∗∗
0.947∗∗∗
0.923∗∗∗
0.886∗∗∗
0.978∗∗∗
0.981∗∗∗
0.982∗∗∗
0.982∗∗∗
0.982∗∗∗
0.982∗∗∗
0.974∗∗∗
0.967∗∗∗
0.962∗∗∗
0.947∗∗∗
0.00554 ± 0.00004
0.00655 ± 0.00005
0.00836 ± 0.00005
0.01006 ± 0.00006
0.01559 ± 0.00009
0.000100 ± 2e − 7
0.000125 ± 1.e − 7
0.000139 ± 1e − 7
0.000143 ± 1e − 7
0.000173 ± 1e − 7
0.0649 ± 0.002
0.0561 ± 0.001
0.0570 ± 0.001
0.0623 ± 0.002
0.0724 ± 0.002
0.0150 ± 0.0001
0.0154 ± 0.0001
0.0155 ± 0.0001
0.0159 ± 0.0001
0.0163 ± 0.0001
0.00533 ± 0.00005
0.00686 ± 0.00005
0.00795 ± 0.00005
0.00790 ± 0.00006
0.01265 ± 0.00006
0.000080 ± 1e − 7
0.000099 ± 1e − 7
0.000109 ± 1e − 7
0.000112 ± 1e − 7
0.000131 ± 1e − 7
destinations, (ii) the radius of gyration of each notable figure simulated, and (iii)
the distribution of the length of the migration jumps. We report in Table 4 the
results obtained for five representative models: (i) the random walker model, (ii)
the notable-based jump probability on a single-level structure, (iii) the notable-
based jump probability on a multilevel structure, (iv) the population-based jump
probability on a multilevel structure, and (v) the mixed population-notable-based
jump probability on a multilevel structure. Results for the other models are reported
in the SM, specifically in Tables SM1, Table SM2, and Table SM3.
Among the different possibilities tested, we found that a modeling approach con-
sidering cultural attractors as a product of both job-opportunities and cultural
SCI & TECHARTSINSTITUTIONSHUMANITIESSPORTSBUSINESS & LAWPUBLIC FIGUREEXPLORATION0.000.020.040.060.080.100.12Page rank centralityMoscowLondonLos AngelesParisNew York Lucchini et al.
Page 14 of 17
Figure 6 A: The distribution of the radius of gyration per notable person; B: The distributions of
the data and the five different models for the number of different cities visited per notable person;
C: The distribution of the lengths of the all the jumps simulated for the different model (lines)
and for the data (stepped density histogram).
interests, by means of the population number and the effective number of notable
people migrated in the time-window under investigation, better captures key fea-
tures of notables' mobility. Moreover, we stress that the model treating different
disciplines as different dynamics outperforms the single-level models in terms of
Kullback-Leibler divergence [31] and first Wasserstein distance [32].
These quantities are measured after simulating the mobility for 2, 000 notable
people, whose number of migration jumps was randomly sampled from a geometric
distribution with parameter p ∼ 0.5. We repeated the simulation for 500 times
to estimate the stability and the standard error of the different metrics. Figure
6.A shows the distribution of the mean radius of gyration of the 500 simulations
against the data (blue-stepped line). Figure 6.B shows the distribution of different
cities visited by a notable person during her/his lifetime, while the distribution
presented in Figure 6.C shows the probability of jumping to a destination that is
at a specific distance from the origin. These distributions are highly dependent on
the geographical distance. Short distance migrations are largely privileged while at
∼ 2500km and ∼ 5500km respectively a second smaller peak appears capturing
the overseas migrations (across the Atlantic Ocean) mainly from Europe to US.
2468Different cities103102101100Pr (density)B: Different visited cities per notablepop-not_multipop_multinot_multinot_singlerand_singledata0200040006000Km106105104103102Pr (density)C: Jump lengthspop-not_multipop_multinot_multinot_singlerand_singledata0100020003000400050006000Km106105104103102101Pr (density)A: Radius of gyration distributionpop-not_multipop_multinot_multinot_singlerand_singledata Lucchini et al.
Page 15 of 17
We stress that the effect of slightly underestimating the log-distance trips, which
also affects the radius-of-gyration distribution, has been proven to be a structural
feature of the radiation model under irregular geographical configurations such as
those imposed by oceans [26].
4 Discussion and Conclusions
A complex question has been posed in [17], i.e. whether it is possible to describe
the dynamical properties of the cultural migration phenomenon. Starting from their
idea of using network theory to tackle the problem, we make some further steps in
understanding this issue. First of all, we use NLP tools to capture a more detailed
representation of the lives of historically notable people that can be considered
as cultural developers or important actors in the evolutionary process of culture.
Our approach gathers information not only from the birth and death events but
also from in-life migratory events, enabling us to study in a more detailed way
the cultural migration processes and to include in our model the years in which a
person is professionally more active. Indeed, there is a difference between the birth
location of a person and the locations s/he migrates to or where s/he decide to
spend her/his last years. The first birth location is not determined by a decision of
the born individual, while the in-life migrations and the death places are more likely
to be chosen following some precise interests and motivations. Using our data we
are able to capture this difference and quantify the contribution of the exploration
level of notable people during these phases of life.
Moreover, we focus our attention on understanding the main features that drive
this kind of mobility. Our results provide evidence that the mobility of historically
and culturally notable individuals is best described by simultaneously considering
three different factors: (i) the population of a city, as a proxy of economic wealth
and generic job opportunities; (ii) the number of culturally notable people that
spent some time of their lives there, as a proxy of the attractive role played by this
city as a cultural hub and of the proneness of this city to invest in culture; and
(iii) the discipline a culturally notable person is working in, as a proxy both of the
interest a city has in investing on a specific cultural area and the tendency of people,
interested or working on a given discipline, to follow notable ones from the same
discipline. The solution proposed in our work represent a functional integration, in
a quantitative theoretical model, of these components.
It is also worth highlighting some limitations of our work. First of all, we rely on
Wikipedia as data source, which shows a clear bias towards the Western culture
and male figures. This limitation is even more relevant since we focus only on pages
written in English. Then, we consider a specific time-window in the cultural his-
tory, i.e. the first half of the 20th century. So, our results may be dependent on the
time chosen (for example, we may observe different behaviors during wider time-
windows) and on the small available dataset of historically and culturally notable
personalities. Besides, while our data on trajectories were extracted automatically
but manually revised, we estimate a recall of our information extraction pipeline of
0.59, as pointed out in Section 2.1.1. This implies that some migration destinations
that are mentioned in the Wikipedia biographies should be added (e.g. by improving
the extraction performances of Ramble-On) in order to make our set of data even
Lucchini et al.
Page 16 of 17
richer. A richer dataset will also help in stabilizing, constructing and precisely char-
acterizing the structure of the network. However, we also stress that the extension
to the radiation model proposed here only uses the visited locations and not the
migration timelines of notables. Thus, the present recall level in extracting complete
timelines does not directly affect the dynamical structure of our model. Thanks to
these considerations, it is also interesting to discuss the specific limitations of the
mobility model proposed. In particular, while the fit of the number of trips might
depend on the recall limitations discussed above and thus affect the number of dif-
ferent visited locations, this does not explain the systematical underestimation of
the probability of high-distance jumps and high-radius of gyration. We expect the
geographical constraints (as discussed in Section 2.2) and the specific time-window
we selected (e.g. WW2 forced to migrate many notable people whose choice was
biased towards US) to be two determinants of these discrepancies.
Overall, our results open interesting possibilities on further investigating the his-
torical role played by places and cities in attracting culturally relevant figures as
well as on better analyzing the level of contribution of each of the factors identified
by our approach (i.e. city's population, number of intellectuals living in the city, and
strength of a specific cultural discipline in the city).Similarly, changing the perspec-
tive, it will become possible to quantify the impact of cultural communities on local
well-being, helping our understanding on how individuals from similar or different
disciplines combine and collaborate to seed the vital growth of cities' economies.
Availability of data and material
The data used in this work are available at: https://doi.org/10.7910/DVN/PJS21L or
https://figshare.com/articles/Following_the_footsteps_of_giants/7352987.
The code used to extract the destinations from the Wikipedia biographies is publicly released at
https://github.com/dhfbk/rambleon.
Competing interests
The authors declare no competing financial or non-financial interests.
Author's contributions
All authors conceptualized the project. L.L acquired and cleaned the data, performed the investigation, the
statistical analyses and drafted the original manuscript. All authors contributed revising the manuscript and gave
final approval for publication.
Funding
No funding supported our research.
Abbreviations
NLP: Natural Language Processing; WW2: World War 2; US: United States; SM: Supplementary Material.
Author details
1Fondazione Bruno Kessler, via Sommarive, 18, 38122 Trento, Italy. 2Department of Information Engineering and
Computer Science, University of Trento, via Sommarive, 9, 38122 Trento, Italy.
References
1. Manning, P.: Migration in World History. London: Routledge, ??? (2005)
2. Lee, E.S.: A theory of migration. Demography 3, 47 -- 57 (1966)
3. Arango, J.: Explaining migration: A critical view. International Social Science Journal 52, 283 -- 296 (2000)
4. Castles, S.: International migration at the beginning of the twenty-first century: Global trends and issues.
International Social Science Journal 52, 269 -- 281 (2000). doi:10.1111/1468-2451.00258
5. Todaro, M.P., Smith, S.: Economic Development. Boston: Addison Wesley, ??? (2000)
6. Guiso, L., Sapienza, P., Zingales, L.: Cultural biases in economic exchange? The Quarterly Journal of
Economics 124, 1095 -- 1131 (2009)
7. Belot, M., Ederveen, S.: Cultural barriers in migration between oecd countries. Journal of Population
Economics 25, 1077 -- 1105 (2012)
8. Beine, M., Parsons, C.: Climatic factors as determinants of international migration. The Scandinavian Journal
of Economics 117, 723 -- 767 (2015)
9. Khoshkish, A.: Intellectual migration: A sociological approach to 'brain drain'. Journal Of World History 10,
1095 -- 1131 (1966)
Lucchini et al.
Page 17 of 17
10. Zipf, G.K.: The p1 p2/d hypothesis: On the intercity movement of persons. American Sociological Review
11(6), 677 -- 686 (1946)
11. Stouffer, S.A.: Intervening opportunities: A theory relating mobility and distance. American Sociological Review
5(6), 845 -- 867 (1940)
12. Block, H.D.: Random Orderings and Stochastic Theories of Responses (1960), pp. 172 -- 217. Springer,
Dordrecht (1974). doi:10.1007/978-94-010-9276-0 8. http://dx.doi.org/10.1007/978-94-010-9276-0 8
13. Simini, F., Gonz´alez, M.C., Maritan, A., Barab´asi, A.-L.: A universal model for mobility and migration patterns.
Nature 484, 96 -- 100 (2012). doi:10.1038/nature10856. 1111.0586
14. Ahmed, M., Barlacchi, G., Braghin, S., Calabrese, F., Ferretti, M., Lonij, V., Nair, R., Novack, R., Paraszczak,
J., Toor, A.: A multi-scale approach to data-driven mass migration analysis. CEUR Workshop Proceedings 1831
(2016)
15. Cavalli-Sforza, L.L., Feldman, M.W.: Cultural Transmission and Evolution: A Quantitative Approach. Princeton
University Press, ??? (1981). http://groups.lis.illinois.edu/amag/langev/paper/cavallisforza81book.html
16. Brewer, J., Gelfand, M., Jackson, J.C., MacDonald, I.F., Peregrine, P.N., Richerson, P.J., Turchin, P.,
Whitehouse, H., Wilson, D.S.: Grand challenges for the study of cultural evolution. Nature Ecology and
Evolution 1, 0070 (2017). Comment
17. Schich, M., Song, C., Ahn, Y.-Y., Mirsky, A., Martino, M., Barab´asi, A.-L., Helbing, D.: A network framework
of cultural history. Science 345(6196), 558 -- 562 (2014). doi:10.1126/science.1240064.
http://science.sciencemag.org/content/345/6196/558.full.pdf
18. Yu, A.Z., Ronen, S., Hu, K., Lu, T., Hidalgo, C.A.: Pantheon 1.0: A manually verified dataset of globally
famous biographies. Scientific Data 3, 150075 (2016). doi:10.1038/sdata.2015.75
19. Weiner, E.: The Geography of the Genius: Lessons from the World's Most Creative Places. Simon & Schuster
Paperbacks, ??? (2016)
20. Menini, S., Sprugnoli, R., Moretti, G., Bignotti, E., Tonelli, S., Lepri, B.: Ramble on: Tracing movements of
popular historical figures. In: Proceedings of EACL 2017 (2017)
21. Gergaud, O., Laou´enan, M., Wasmer, E., et al.: A brief history of human time: Exploring a database of'notable
people'. Technical Report 63, Sciences Po Departement of Economics (2016)
22. Corcoglioniti, F., Rospocher, M., Aprosio, A.P.: Extracting knowledge from text with pikes. In: ISWC 2015
Posters & Demonstrations Track, Within the 14th International Semantic Web Conference (ISWC 2015),
Bethlehem, USA, October 11-15, 2015 (2015). http://ceur-ws.org/Vol-1486/paper 66.pdf
23. Baker, C.F., Fillmore, C.J., Lowe, J.B.: The berkeley framenet project. In: Proceedings of the 17th International
Conference on Computational Linguistics - Volume 1. COLING '98, pp. 86 -- 90. Association for Computational
Linguistics, Stroudsburg, PA, USA (1998). doi:10.3115/980451.980860.
https://doi.org/10.3115/980451.980860
24. Kshirsagar, M., Thomson, S., Schneider, N., Carbonell, J., Smith, N., Dyer, C.: Frame-semantic role labeling
with heterogeneous annotations. In: Proceedings of the 53rd Annual Meeting of the Association for
Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (2015).
doi:10.3115/v1/P15-2036
25. Reba, M., Reitsma, F., Seto, K.C.: Spatializing 6,000 years of global urbanization from 3700 BC to AD 2000.
Nature Scientific Data, Volume 3, id. 160034 (2016). 3, 160034 (2016). doi:10.1038/sdata.2016.34
26. Masucci, A.P., Serras, J., Johansson, A., Batty, M.: Gravity versus radiation models: On the importance of
scale and heterogeneity in commuting flows. Physical Review E 88(2), 022812 (2013).
doi:10.1103/PhysRevE.88.022812. 1206.5735
27. De Domenico, M., Sol´e-Ribalta, A., Cozzo, E., Kivela, M., Moreno, Y., Porter, M.A., G´omez, S., Arenas, A.:
Mathematical formulation of multilayer networks. Phys. Rev. X 3, 041022 (2013).
doi:10.1103/PhysRevX.3.041022
28. Belyi, A., Bojic, I., Sobolevsky, S., Sitko, I., Hawelka, B., Rudikova, L., Kurbatski, A., Ratti, C.: Global
multi-layer network of human mobility. CoRR abs/1601.05532 (2016). 1601.05532
29. Ferretti, M., Barlacchi, G., Pappalardo, L., Lucchini, L., Lepri, B.: Weak nodes detection in urban transport
systems: Planning for resilience in singapore. In: 2018 IEEE 5th International Conference on Data Science and
Advanced Analytics (DSAA), pp. 472 -- 480 (2018). doi:10.1109/DSAA.2018.00061
30. Brin, S., Page, L.: The anatomy of a large-scale hypertextual web search engine. In: Proceedings of WWW
1998 (1998)
31. Kullback, S., Leibler, R.A.: On information and sufficiency. Annals of Mathematical Statistics 22(1), 257 -- 263
(1951)
32. Olkin, I., Pukelsheim, F.: The distance between two random vectors with given dispersion matrices. Linear
Algebra Applied 48, 257 -- 263 (1982)
|
1802.08961 | 1 | 1802 | 2018-02-25T07:32:58 | Optimal Containment of Epidemics over Temporal Activity-Driven Networks | [
"cs.SI",
"math.OC",
"physics.soc-ph"
] | In this paper, we study the dynamics of epidemic processes taking place in temporal and adaptive networks. Building on the activity-driven network model, we propose an adaptive model of epidemic processes, where the network topology dynamically changes due to both exogenous factors independent of the epidemic dynamics as well as endogenous preventive measures adopted by individuals in response to the state of the infection. A direct analysis of the model using Markov processes involves the spectral analysis of a transition probability matrix whose size grows exponentially with the number of nodes. To overcome this limitation, we derive an upper-bound on the decay rate of the number of infected nodes in terms of the eigenvalues of a $2 \times 2$ matrix. Using this upper bound, we propose an efficient algorithm to tune the parameters describing the endogenous preventive measures in order to contain epidemics over time. We confirm our theoretical results via numerical simulations. | cs.SI | cs | OPTIMAL CONTAINMENT OF EPIDEMICS OVER TEMPORAL
ACTIVITY-DRIVEN NETWORKS∗
MASAKI OGURA† , VICTOR M. PRECIADO‡ , AND NAOKI MASUDA§
Abstract. In this paper, we study the dynamics of epidemic processes taking place in tempo-
ral and adaptive networks. Building on the activity-driven network model, we propose an adaptive
model of epidemic processes, where the network topology dynamically changes due to both exogenous
factors independent of the epidemic dynamics as well as endogenous preventive measures adopted
by individuals in response to the state of the infection. A direct analysis of the model using Markov
processes involves the spectral analysis of a transition probability matrix whose size grows expo-
nentially with the number of nodes. To overcome this limitation, we derive an upper-bound on the
decay rate of the number of infected nodes in terms of the eigenvalues of a 2 × 2 matrix. Using this
upper bound, we propose an efficient algorithm to tune the parameters describing the endogenous
preventive measures in order to contain epidemics over time. We confirm our theoretical results via
numerical simulations.
Key words. Temporal networks, adaptive networks, epidemics, stochastic processes, convex
optimization.
AMS subject classifications. 39A50, 60J10, 90C25, 91D10, 91D30
1. Introduction. Accurate prediction and cost-effective containment of epi-
demics in human and animal populations are fundamental problems in mathematical
epidemiology [7, 16, 28]. In order to achieve these goals, it is indispensable to develop
effective mathematical models describing the spread of disease in human and animal
contact networks [8, 9].
In this direction, we find a broad literature on modeling,
analysis, and containment of epidemic processes in static contact networks. However,
these works neglect an important factor: the temporality of the interactions [15,39,40],
which arises either independently of or dependent on epidemic propagations. A frame-
work for modeling temporal interactions in human and animal populations is temporal
networks (i.e., time-varying networks), where individuals and interactions are modeled
as nodes and edges, respectively, which can appear and disappear over time [13,14,24].
Under this framework, the effect of temporal interactions on epidemic propagations
has been investigated numerically and theoretically [22, 23]. For containing epidemic
processes on temporal networks, we find a plethora of heuristic approaches [18, 31]
and analytical methods [19, 26].
Adaptive networks refer to the case in which changes in nodes or edges occur
in response to the state of the dynamics taking place on the network [9, 11, 24, 36].
A common temporality of agent-agent interaction in epidemic dynamics arises from
social distancing behavior [1–3], which let the structure of contact networks change
over time as a result of adaptation to the state of the epidemics. Several models of
such adaptive networks have been proposed. For example, Gross et al. proposed a
rewiring mechanism where a healthy node actively avoids to be adjacent to infected
nodes [10]. Extensions of this model are found in [9, 17, 21, 41, 42]. Guo et al. pro-
posed an alternative model in which links connecting an infected node and a healthy
8
1
0
2
b
e
F
5
2
]
I
S
.
s
c
[
1
v
1
6
9
8
0
.
2
0
8
1
:
v
i
X
r
a
([email protected]).
∗Submitted to the editors DATE.
†Graduate School of Information Science, Nara Institute of Science and Technology, Ikoma, Nara
‡Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia,
§Department of Engineering Mathematics, University of Bristol, Clifton, Bristol BS8 1UB
PA USA ([email protected]).
([email protected]).
1
2
M. OGURA, V. M. PRECIADO, AND N. MASUDA
node are deactivated [12]. As for the containment of epidemic processes on adap-
tive networks, various heuristic [5, 20] and analytical [25, 27] approaches have been
proposed. However, in many studies, the effects of exogenous temporal factors and
endogenous adaptive measures on epidemic processes have been separately examined,
leaving unclear how their combination affects the dynamics of the spread.
In this paper, we study epidemic processes and containment strategies in a tempo-
ral network model where the effect of exogenous factors and that of adaptive measures
are simultaneously present. Our model is based on the activity-driven temporal net-
work model [30]. In this model, a node is stochastically activated and connects to
other nodes independently of the dynamics taking place in the network. In order to
analyze the joint effect of exogenous factors and endogenous adaptations, we add a
mechanism of social distancing to the standard activity-driven model. In other words,
we allow an infected node to endogenously adapt to the state of the epidemics by 1) de-
creasing its activation probability and 2) refusing interactions with other activated
nodes. On top of this temporal network, we adopt the standard susceptible-infected-
susceptible (SIS) model of epidemic dynamics (see, e.g., [28]) and derive an analytical
upper bound on the decay rate of the number of infected nodes over time. Based on
this result, we then propose an efficient strategy for tuning the social distancing rates
in order to suppress the number of infected nodes.
Our work is related to [35], in which an infected individual is allowed to decrease
its activation probability. However, in [35], the durations of temporal interactions are
assumed to be sufficiently short compared with the time scale of the epidemic dynam-
ics, leaving out the interesting case where the time scale of the network dynamics and
that of the epidemic process are comparable. In addition, our results hold true for
networks of any size, while the results in [35] require the networks to be sufficiently
large.
This paper is organized as follows. In Section 2, we introduce a model of epidemic
processes on temporal and adaptive networks. In Section 3, we derive an upper bound
on the decay rate of the infection size. Based on this bound, in Section 4 we formulate
and solve optimization problems for containing the spread of epidemics. The obtained
theoretical results are numerically illustrated in Section 5.
2. Problem setting. In this section, we first describe the activity-driven net-
work proposed in [30]. We then introduce an adaptive SIS (A-SIS) model on activity-
driven networks, which allows nodes to adapt to the state of the nodes (i.e., susceptible
or infected) in their neighborhoods.
2.1. Activity-driven networks. Throughout this paper, we let the set of nodes
in a network be given by V = {v1, . . . , vn}. The activity-driven network is a temporal
network in discrete time and is defined as follows.
Definition 2.1 ([30]). For each i = 1, . . . , n, let ai be a positive constant less
than or equal to 1. We call ai the activity rate of node vi. Let m be a positive
integer less than or equal to n − 1. The activity-driven network is defined as an
independent and identically distributed sequence of undirected graphs created by the
following procedure (see Figure 1 for a schematic illustration):
1. At each time t = 0, 1, 2, . . . , each node vi becomes "activated" with probabil-
ity ai independently of other nodes.
2. Each activated node, say, vi, randomly and uniformly chooses m other nodes
independently of other activated nodes. For each chosen node, say, vj, an
edge {vi, vj} is created. These edges are discarded at time t + 1 (i.e., do not
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
3
Fig. 1: Schematic on an activity-driven network. We set n = 10 and m = 2. Filled
circles represent active nodes. Open circles represent inactive nodes. The time is
denoted by t.
exist at time t + 1).
3. Steps 2 and 3 are repeated for each time t ≥ 0, independently of past realiza-
tions.
Remark 2.2. We do not allow multiple edges between a pair of nodes. In other
words, even when a pair of activated nodes choose each other as their neighbors at a
specific time, we assume that one and only one edge is spanned between those nodes.
Although the activity-driven network is relatively simple, the model can reproduce
an arbitrary degree distribution [30]. Several properties of activity-driven networks
have been investigated, including structural properties [30,38], steady-state properties
of random walks [29,34], and spreading dynamics [30,35,37]. However, the model does
not allow nodes to adapt to the state of the epidemics and, therefore, is not suitable
for discussing how social distancing affects the dynamics of the spread. In the next
subsection, we extend the activity-driven network by incorporating social distancing
behaviors of nodes.
2.2. Activity-driven A-SIS model. Building upon the activity-driven net-
work described above, we consider the scenario where nodes change their neighbor-
hoods in response to the state of the epidemics over the network [25]. Specifically,
we propose the activity-driven adaptive-SIS model (activity-driven A-SIS model for
short) as follows:
Definition 2.3 (Activity-driven A-SIS model). For each i, let ai, χi, πi ∈ (0, 1]
be constants. We call ai, χi, and πi the activity rate, adaptation factor, and accep-
tance rate of node vi, respectively. Also, let m ≤ n − 1 be a positive integer, and
β, δ ∈ (0, 1] be constants. We call β and δ the infection rate and recovery rate, re-
spectively. The activity-driven A-SIS model is defined by the following procedures (see
Figure 2 for an illustration):
1. At the initial time t = 0, each node is either susceptible or infected.
2. At each time t = 0, 1, 2, . . . , each node vi randomly becomes activated inde-
pendently of other nodes with the following probability:
(cid:40)
(1)
Pr(node vi becomes activated) =
ai,
χiai,
if vi is susceptible,
if vi is infected.
3. Each activated node, say, vi, randomly and uniformly chooses m other nodes
independently of other activated nodes. For each chosen node, say, vj, an
𝑡=0𝑡=1𝑡=2 4
M. OGURA, V. M. PRECIADO, AND N. MASUDA
Fig. 2: Adaptation of nodes in the activity-driven A-SIS model. Filled and empty
circles represent active and inactive nodes, respectively. (a) A susceptible node is
activated with probability ai. (b) An infected node is activated with probability χiai.
(c) An infected node (vj) accepts an edge spanned from an activated node with
probability πj.
edge {vi, vj} is created with the following probability:
(2)
Pr({vi, vj} is created) =
1,
πj,
if vj is susceptible,
if vj is infected.
(cid:40)
These edges are discarded at time t + 1 (i.e., do not exist at time t + 1). As
in Remark 2.2, we do not allow multiple edges between a pair of nodes.
4. The states of nodes are updated according to the SIS model. In other words,
if a node vi is infected, it transits to the susceptible state with probability δ.
If vi is susceptible, its infected neighbors infect node vi with probability β
independently of the other infected neighbors.
5. Steps 2–4 are repeated for each time t ≥ 0.
Steps 2 and 3 in Definition 2.3 model social distancing behavior by infected nodes.
In Step 2, an infected node decreases its activity rate to avoid infecting other nodes.
Step 2 can also be regarded as modeling reduction of social activity by infected nodes
due to sickness. In Step 3, an infected node, say, vj, establishes a connection with an
activated node only with probability πj to avoid infecting other nodes (when πj < 1).
A susceptible node behaves in the same way as in the standard SIS model in the
original activity-driven network.
3. Decay rate. In order to quantify the persistence of epidemic infections in
the activity-driven A-SIS model, in this section, we introduce the concept of decay
rate of the epidemics. A direct computation of the decay rate requires computing the
eigenvalues of a matrix whose size grows exponentially with the number of the nodes.
To overcome this difficulty, we present an upper bound on the decay rate in terms of
the eigenvalues of a 2 × 2 matrix.
3.1. Definition. For each time t and node vi, define the random variable
(cid:40)
(3)
xi(t) =
0,
1,
if vi is susceptible at time t,
if vi is infected at time t.
𝜋𝑗SSS𝑎𝑖𝑣𝑖1−𝑎𝑖III𝜒𝑖𝑎𝑖𝑣𝑖1−𝜒𝑖𝑎𝑖I𝑣𝑗I𝑣𝑗(a)(b)(c)𝑣𝑖𝑣𝑖 CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
5
Define the vector p(t) = [p1(t) ··· pn(t)](cid:62) of the infection probabilities by
(4)
pi(t) = Pr(vi is infected at time t).
In this paper, we measure the persistence of infection by the rate of convergence of
infection probabilities to the origin.
Definition 3.1. We define the decay rate of the activity-driven A-SIS model by
(5)
α = sup
x(0)
lim sup
t≥0
log(cid:107)p(t)(cid:107)
t
,
where (cid:107)·(cid:107) denotes the (cid:96)1 norm.
The infection-free equilibrium, x1 = ··· = xn = 0, is the unique absorbing state
of the Markov process {x1(t), . . . , xn(t)}t≥0. Moreover, the infection-free equilibrium
is reachable from any other states by our assumption δ > 0. This implies α < 1.
In fact, α is difficult to compute for large networks for the following reason. The
Markov process {x1(t), . . . , xn(t)}t≥0 has 2n states. Let Q denote its 2n×2n transition
probability matrix. Since the disease-free state is the unique absorbing state, it follows
that
(6)
α = max{λ : λ is an eigenvalue of Q, λ < 1}.
Because the size of the matrix Q grows exponentially fast with respect to the number
of the nodes, a direct computation of the decay rate is difficult for large networks.
3.2. An upper bound. We start with the following proposition, which allows
us to upper-bound the infection probabilities using a linear dynamics:
Proposition 3.2. Let
(7)
¯m = m/(n − 1)
and, for all i, define the constants
(8)
Then,
(9)
φi = ¯mχiai, ψi = ¯mπiai.
pi(t + 1) ≤ (1 − δ)pi(t) + β
n(cid:88)
j=1
[1 − (1 − ψi)(1 − φj)]pj(t)
for all nodes vi and t ≥ 0.
Proof. By the definition of the A-SIS dynamics on the activity-driven network,
the nodal states x1, . . . , xn obey the following stochastic difference equation
(10) xi(t + 1) = xi(t)− xi(t)N (i)
δ (t) + (1− xi(t))
1 − aij(t)xj(t)N (ij)
β
(t)
where
(11)
aij(t) =
(cid:40)
if an edge {vi, vj} exists at time t,
1,
0, otherwise,
1 −(cid:89)
j(cid:54)=i
(cid:16)
(cid:17) ,
M. OGURA, V. M. PRECIADO, AND N. MASUDA
t=0 are independent and identically distributed ran-
6
and {N (i)
(t)}∞
dom Bernoulli variables satisfying
t=0 and {N (ij)
δ (t)}∞
β
(12)
and
(13)
N (i)
δ (t) =
N (ij)
β
(t) =
(cid:40)
(cid:40)
1, with probability δ,
0, with probability 1 − δ,
1, with probability β,
0, with probability 1 − β.
On the right-hand side of equation (10), the second and third terms represent recovery
and transmission events, respectively (a similar equation for the case of static networks
can be found in [6]).
(10) is upper-bounded by (1 − xi(t))(cid:80)n
By the Weierstrass product inequality, the third term on the right-hand side of
(t). Since the expectation
j=1 aij(t)xj(t)N (ij)
β
of xi(t) equals pi(t), taking the expectation in (10) gives
(14)
pi(t + 1) ≤ pi(t) − δpi(t) + β
E[(1 − xi(t))aij(t)xj(t)],
(cid:88)
j(cid:54)=i
where E[·] denotes the expectation of a random variable.
Now, assume i (cid:54)= j. By the definition of the variables xi and aij, it follows that
(15)
E[(1 − xi(t))aij(t)xj(t)]
= Pr(vi and vj are adjacent, vi is susceptible, and vj is infected at time t)
= Pr(vi and vj are adjacent at time t Ξt
i,j) Pr(Ξt
i,j),
where the event Ξt
i,j is defined by
(16)
Ξt
i,j = "vi is susceptible and vj is infected at time t".
If we further define the event
(17)
Γt
i→j = "vi is activated and chooses vj as its neighbor at time t",
then, we obtain
(18)
Pr(vi and vj are adjacent at time t Ξt
i,j)
i,j) − Pr(Γt
i→j Ξt
j→i Ξt
i,j)(cid:3)(cid:2)1 − Pr(Γt
=1 −(cid:2)1 − Pr(Γt
i,j) + Pr(Γt
i→j Ξt
i,j)(cid:3).
i→j Ξt
j→i Ξt
= Pr(Γt
i,j) Pr(Γt
j→i Ξt
i,j)
The event Γt
i→j occurs when and only when vi is activated, chooses vj as a potential
neighbor, and actually connects to vj (according to the probability given by equa-
tion (2)). Therefore, equation (8) implies
(19)
Pr(Γt
i→jΞt
i,j) = ψi.
Similarly, the event Γt
j→i occurs when and only when vj is activated (with probabil-
ity χjaj if vj is infected at time t) and chooses vi as one of its m neighbors. Therefore,
we have
(20)
Pr(Γt
j→iΞt
i,j) = φj.
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
7
Hence, for i (cid:54)= j, combination of equations (15) and (18)–(20) yields
E[(1 − xi(t))aij(t)xj(t)] = [1 − (1 − ψi)(1 − φj)] Pr(Ξt
≤ [1 − (1 − ψi)(1 − φj)]pj(t),
(21)
i,j)
i,j) ≤ pj(t). Moreover, inequality (21)
where we have used the trivial inequality Pr(Ξt
trivially holds true also when i = j. Inequalities (14) and (21) prove (9), as desired.
Using Proposition 3.2, we obtain the following theorem that gives an explicit
upper bound on the decay rate of the activity-driven A-SIS model. For a vector ξ ∈
Rn, introduce the notations
(22)
(cid:104)ξ(cid:105)a =
1
n
Theorem 3.3. Define
n(cid:88)
i=1
aiξi,
(cid:104)ξ(cid:105)a2 =
1
n
a2
i ξi.
n(cid:88)
i=1
(23)
where
(24)
κ =
(cid:104)χ(cid:105)a + (cid:104)π(cid:105)a − ¯m(cid:104)χπ(cid:105)a2 +
Then, the decay rate α satisfies
(25)
αu = 1 − δ + κ ¯mnβ,
(cid:113)
((cid:104)χ(cid:105)a + (cid:104)π(cid:105)a − ¯m(cid:104)χπ(cid:105)a2)2 + 4((cid:104)χπ(cid:105)a2 − (cid:104)χ(cid:105)a(cid:104)π(cid:105)a)
2
α ≤ αu.
.
Proof. Inequality (9) implies that there exists a nonnegative variable i(t) such
that
(26)
pi(t + 1) = (1 − δ)pi(t) + β
n(cid:88)
(1 − (1 − ψi)(1 − φj)) pj(t) − i(t)
for all nodes vi and t ≥ 0. Let us define the vectors (t) = [(t) ··· n(t)](cid:62), φ =
[φ1 ··· φn](cid:62), and ψ = [ψ1 ··· ψn](cid:62). Equation (26) is rewritten as
j=1
(27)
where
p(t + 1) = Fp(t) − (t),
F = (1 − δ)I + β(cid:2)11(cid:62) − (1 − ψ)(1 − φ)(cid:62)(cid:3) ,
(28)
1 denotes the n-dimensional vector whose entries are all one, and I denotes the n× n
identity matrix. Since F and (t) are nonnegative entrywise, equation (27) leads to
p(t) = F tp(0) −(cid:80)t
(cid:96)=0 F k−(cid:96)((cid:96)) ≤ F tp(0). This inequality shows
α ≤ ρ(F),
(29)
where ρ(·) denotes the spectral radius of a matrix.
Now, we evaluate ρ(F). Equation (28) is rewritten as F = (1 − δ)I + βA, where
A = 11(cid:62) − (1 − ψ)(1 − φ)(cid:62). Since A is nonnegative entrywise and 1 − δ ≥ 0, we
obtain
(30)
ρ(F) = 1 − δ + βρ(A).
8
M. OGURA, V. M. PRECIADO, AND N. MASUDA
Fig. 3: Characteristic equation (34)
Furthermore, as we prove in Appendix A, it holds that
ρ(A) = ρ(nB),
(cid:104)φ(cid:105) + (cid:104)ψ(cid:105) − (cid:104)φψ(cid:105) ±(cid:113)
having the roots
(35)
λ =
((cid:104)φ(cid:105) + (cid:104)ψ(cid:105) − (cid:104)φψ(cid:105))2 + 4((cid:104)φψ(cid:105) − (cid:104)φ(cid:105)(cid:104)ψ(cid:105))
2
.
The roots are real because
((cid:104)φ(cid:105) + (cid:104)ψ(cid:105) − (cid:104)φψ(cid:105))2 + 4((cid:104)φψ(cid:105) − (cid:104)φ(cid:105)(cid:104)ψ(cid:105)) ≥ ((cid:104)ψ(cid:105) − (cid:104)φ(cid:105))2 + (cid:104)φψ(cid:105)2 > 0,
(36)
which follows from the trivial inequality 4(cid:104)φψ(cid:105) ≥ 2(cid:104)φ(cid:105)(cid:104)φψ(cid:105) + 2(cid:104)ψ(cid:105)(cid:104)φψ(cid:105). Therefore,
by substituting equation (8) into equation (35), we obtain ρ(B) = κ ¯m. This equation
and (29)–(31) complete the proof of the theorem.
The following corollary shows that an epidemic will become extinct more quickly
when the adaptation factor and acceptance rate are less correlated in a weighted sense.
Corollary 3.4. Let (χ, π) and (χ(cid:48), π(cid:48)) be pairs of adaptation factors and accep-
tance rates of nodes, and denote the corresponding upper-bounds on the decay rates
u, respectively. If (cid:104)χ(cid:105)a = (cid:104)χ(cid:48)(cid:105)a, (cid:104)π(cid:105)a = (cid:104)π(cid:48)(cid:105)a, and (cid:104)χπ(cid:105)a2 < (cid:104)χ(cid:48)π(cid:48)(cid:105)a2, then
by αu and α(cid:48)
(37)
αu < α(cid:48)
u.
(31)
where
(32)
(33)
(34)
(cid:20)
n(cid:88)
i=1
B =
1
1 − (cid:104)ψ(cid:105)
−1 + (cid:104)φ(cid:105) −1 + (cid:104)φ(cid:105) + (cid:104)ψ(cid:105) − (cid:104)φψ(cid:105)
1
n
(cid:104)φψ(cid:105) =
(cid:104)ψ(cid:105) =
n(cid:88)
1
n
φi,
ψi,
i=1
(cid:104)φ(cid:105) =
1
n
(cid:21)
n(cid:88)
,
i=1
φiψi.
As shown in Figure 3, matrix B has the characteristic equation
(1 − λ)(cid:104)φψ(cid:105) = (λ − (cid:104)φ(cid:105))(λ − (cid:104)ψ(cid:105))
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
9
Proof. By the proof of Theorem 3.3, we have αu = 1 − δ + ρ(B)nβ. Figure 3
implies that ρ(B) increases with (cid:104)φψ(cid:105) when (cid:104)φ(cid:105) and (cid:104)ψ(cid:105) are fixed. This proves the
claim of the corollary because (cid:104)χπ(cid:105)a2 = (cid:104)φψ(cid:105)/ ¯m2, (cid:104)χ(cid:105)a = (cid:104)φ(cid:105)/ ¯m, and (cid:104)π(cid:105)a = (cid:104)ψ(cid:105)/ ¯m.
As another corollary of Theorem 3.3, we also present an upper bound on the
decay rate when nodes do not adapt to the states of the nodes.
Corollary 3.5. Assume χi = πi = 1 for all i. Let
2(cid:104)a(cid:105) − ¯m(cid:104)a2(cid:105) +(cid:112)4(cid:104)a2(cid:105) − 4 ¯m(cid:104)a(cid:105)(cid:104)a2(cid:105) + ¯m2(cid:104)a2(cid:105)2
(38)
κ0 =
2
.
Then, the decay rate of the activity-driven SIS model is at most 1 − δ + κ0 ¯mnβ.
Remark 3.6. If m is sufficiently small compared with n and, furthermore, n is
sufficiently large (as implicitly assumed in [30]), the upper-bound in Corollary 3.5
reduces to 1 − δ + ((cid:104)a(cid:105) +(cid:112)(cid:104)a2(cid:105))mβ, which coincides with the result in [30].
4. Cost-optimal adaptations. In this section, we study the problem of eradi-
cating an epidemic outbreak by distributing resources to nodes in the activity-driven
network. We consider the situation in which there is a budget that can be invested
on strengthening the preventative behaviors of each node. We show that the optimal
budget allocation is found using geometric programs, which can be efficiently solved
in polynomial time.
4.1. Problem statement. We consider an optimal resource allocation problem
in which we can tune the adaptation factors and acceptance rates of nodes. Assume
that, to set the adaptation factor of node vi to χi, we need to pay a cost fi(χi).
Similarly we need to pay a cost gi(πi) to set the acceptance rate of node vi to πi. The
total cost for tuning the parameters to the values χ1, . . . , χn, π1, . . . , πn equals
n(cid:88)
(39)
C =
(fi(χi) + gi(πi)).
i=1
Throughout this section, we assume the following box constraints:
(40)
0 <
χi ≤ χi ≤ ¯χi,
¯
0 <
πi ≤ πi ≤ ¯πi.
¯
In this paper, we consider the following two types of optimal resource allocation
problems.
Problem 4.1 (Cost-constrained optimal resource allocation). Given a total bud-
get ¯C, find the adaptation rates and acceptance rates that minimize αu while satisfying
the budget constraint
(41)
C ≤ ¯C.
Problem 4.2 (Performance-constrained optimal resource allocation). Given a
largest tolerable decay rate ¯α, find the adaptation rates and acceptance rates that
minimize the total cost C while satisfying the performance constraint
(42)
αu ≤ ¯α.
10
M. OGURA, V. M. PRECIADO, AND N. MASUDA
4.2. Cost-constrained optimal resource allocation. In this subsection, we
show that Problem 4.1 can be transformed to a geometric program [4], which can be
efficiently solved. Before stating our main results, we give a brief review of geometric
programs. Let x1, . . . , xn denote positive variables and define x = (x1, . . . , xn). We
say that a real function q(x) is a monomial if there exist c ≥ 0 and a1, . . . , an ∈ R
such that q(x) = cxa1
n . Also, we say that a function r(x) is a posynomial if
it is a sum of monomials of x (we point the readers to [4] for more details). Given
a collection of posynomials r0(x), . . . , rk(x) and monomials q1(x), . . . , q(cid:96)(x), the
optimization problem
1 ··· xan
(43)
r0(x)
minimize
subject to ri(x) ≤ 1,
qj(x) = 1,
i = 1, . . . , k,
j = 1, . . . , (cid:96),
is called a geometric program. A constraint of the form r(x) ≤ 1 with r(x) being
a posynomial is called a posynomial constraint. Although geometric programs are
not convex, they can be efficiently converted into equivalent convex optimization
problems [4].
We assume that the cost functions fi and gi decrease with the adaptation fac-
tor χi and acceptance rate πi, respectively. This assumption implies a natural sit-
uation in which it is more costly to suppress χi and πi to a larger extent. We also
expect diminishing returns with increasing investments [33]. For a fixed > 0, let
∆fi(χi) = fi(χi − ) − fi(χi) denote the cost for improving the adaptation factor
from χi to χi − . Then, diminishing returns imply that ∆fi decreases with χi, which
implies the convexity of fi. Therefore, we place the following assumption on the cost
functions.
Assumption 4.3. For all i ∈ {1, . . . , n}, decompose fi and gi into the differences
of their positive and negative parts as follows:
(44)
(45)
fi = f +
gi = g+
i − f−
i ,
i − g−
i ,
where f +
Then, f +
i = max(f, 0), f−
i and g+
i are posynomials, and f−
i = max(−f, 0), g+
i and g−
i = max(g, 0), and g−
i = max(−g, 0).
i are constants.
Assumption 4.3 allows us to use any cost functions that are convex on the log-
log scale because any function convex on the log-log scale can be approximated by a
posynomial with an arbitrary accuracy [4, Section 8]. We now state our first main
result in this section, which allows us to efficiently solve Problem 4.1 via geometric
programming:
Theorem 4.4. Let χ(cid:63)
i and π(cid:63)
i be the solutions of the following optimization prob-
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
11
minimize
λ, χi, πi, ζ, η>0
subject to
lem:
(46a)
(46b)
(46c)
(46d)
(46e)
(46f)
1/λ
n(cid:88)
i=1
(40),
¯m2λ(cid:104)χπ(cid:105)a2ζη ≤ 1,
ζ−1 + λ + ¯m(cid:104)χ(cid:105)a ≤ 1,
η−1 + λ + ¯m(cid:104)π(cid:105)a ≤ 1,
(f +
i (χi) + g+
i (πi)) ≤ ¯C +
n(cid:88)
i=1
(f−
i + g−
i ).
Then, the adaptation factor χi = χ(cid:63)
i solve Prob-
lem 4.1. Moreover, under Assumption 4.3, the optimization problem (46) is a geo-
metric program.
i and the acceptance rate πi = π(cid:63)
To prove this theorem, we show an alternative characterization of the decay rate
in terms of inequalities.
Lemma 4.5. Let λ > 0. The upper bound αu satisfies
(47)
if and only if
(48)
(49)
(50)
αu ≤ 1 − δ + λnβ
(1 − λ)(cid:104)φψ(cid:105) ≤ (λ − (cid:104)φ(cid:105))(λ − (cid:104)ψ(cid:105)),
(cid:104)φ(cid:105) < λ,
(cid:104)ψ(cid:105) < λ.
Proof. By the proof of Theorem 3.3, inequality (47) holds true if and only if
λ ≥ ρ(B). Figure 3 indicates that λ ≥ ρ(B) is equivalent to conditions (48)–(50).
We can now prove Theorem 4.4:
Proof of Theorem 4.4. By Lemma 4.5, the solutions of Problem 4.1 are given by
those of the following optimization problem:
(51a)
1 − δ + λnβ
minimize
λ, χi, πi>0
subject to (40), (41), and (48)–(50).
(51b)
Define the auxiliary variables ζ = 1/(λ − (cid:104)φ(cid:105)) and η = 1/(λ − (cid:104)ψ(cid:105)). Then, con-
ditions (48)–(50) hold true if and only if (1 − λ)(cid:104)φψ(cid:105)ζη ≤ 1, ζ > 0, and η > 0.
Therefore, the optimization problem (51) is equivalent to the following optimization
problem:
(52a)
(52b)
(52c)
(52d)
(52e)
minimize
λ, χi, πi, ζ, η>0
subject to
λ
(40) and (41),
(1 − λ)(cid:104)φψ(cid:105)ζη ≤ 1,
ζ−1 − λ + (cid:104)φ(cid:105) = 0,
η−1 − λ + (cid:104)ψ(cid:105) = 0,
M. OGURA, V. M. PRECIADO, AND N. MASUDA
12
where we minimize λ instead of 1 − δ + λnβ. We claim that the optimal value of the
objective function is equal to the one in the following optimization problem:
(53a)
(53b)
(53c)
(53d)
(53e)
minimize
λ, χi, πi, ζ, η>0
subject to
λ
(40) and (41),
(1 − λ)(cid:104)φψ(cid:105)ζη ≤ 1,
ζ−1 − λ + (cid:104)φ(cid:105) ≤ 0,
η−1 − λ + (cid:104)ψ(cid:105) ≤ 0.
i , π(cid:63)
i , π(cid:63)
Let λ(cid:63)
1 and λ(cid:63)
1 ≥ λ(cid:63)
1 ≤ λ(cid:63)
2. Assume that the optimal value λ(cid:63)
2 be the optimal values of the objective functions in problems (52) and
(53), respectively. We have λ(cid:63)
2 because the constraints in problem (52) are more
strict than those in (53). Let us show λ(cid:63)
2 in
problem (53) is attained by the parameters (λ, χi, πi, ζ, η) = (λ(cid:63), χ(cid:63)
i , ζ (cid:63), η(cid:63)). Since
the left-hand sides of constraints (53d) and (53e) decease with ζ and η, there exist
nonnegative constants ∆ζ and ∆η such that ζ = ζ (cid:63) − ∆ζ and η = η(cid:63) − ∆η satisfy
the equality constraints (52d) and (52e). Moreover, since the left-hand side of the
constraint (53c) increases with ζ and η, the new set of parameters (λ, χi, πi, ζ, η) =
i , ζ (cid:63) − ∆ζ, η(cid:63) − ∆η) still satisfies (53c). Furthermore, these changes of pa-
(λ(cid:63), χ(cid:63)
rameters do not affect the feasibility of the box constraints (40) and the budget
constraint (41) because the constraints are independent of the values of ζ and η.
Therefore, we have shown the existence of parameters achieving λ = λ(cid:63)
2 but still sat-
isfying the constraints in the optimization problem (52). This shows λ(cid:63)
2, as
desired.
Now, by rewriting the optimization problem (53) in terms of the variables λ = 1−λ
and substituting (8) in (53), we obtain the optimization problem (46). Notice that
minimizing λ is equivalent to maximizing 1 − λ, which is equivalent to minimizing
1/λ.
1 ≤ λ(cid:63)
Let us finally show that (46) is a geometric program. The objective function, 1/λ,
is a posynomial in λ. The constraints (40) and (46c)–(46e) are posynomial constraints.
Finally, Assumption 4.3 guarantees that constraint (46f) is a posynomial constraint
as well. This completes the proof of the theorem.
4.3. Performance-constrained optimal resource allocation. In the same
way as in the previous section, we can efficiently solve Problem 4.2 via geometric
programming:
Theorem 4.6. Let χ(cid:63)
i and π(cid:63)
i be the solution of the following optimization prob-
n(cid:88)
minimize
λ, χi, πi, ζ, η>0
subject to
lem:
(54a)
(54b)
(54c)
(f +
i (χi) + g+
i (πi))
i=1
(40) and (46c)–(46e),
βn + 1 − δ − ¯α
λ−1 ≤ 1.
βn
Then, the adaptation factor χi = χ(cid:63)
i solve Prob-
lem 4.2. Moreover, under Assumption 4.3, the optimization problem is a geometric
program.
i and the acceptance rate πi = π(cid:63)
Proof. Constraint (54c) is equivalent to the performance constraint (42). The
rest of the proof is almost the same as the proof of Theorem 4.4 and is omitted.
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
13
Fig. 4: Comparison between the numerically obtained decay rates α and their upper
bounds αu in Case 1 when β = 0.8. (a) m = 2, (b) m = 10, and (c) m = 50.
Fig. 5: Discrepancy between the true decay rates and their upper bounds in Case 1.
(a) m = 2, (b) m = 10, and (c) m = 50.
5. Numerical simulations. In this section, we illustrate the theoretical results
obtained in previous sections by numerical simulations.
5.1. Accuracy of the upper bound. We first illustrate the accuracy of the
upper bound (23) on the decay rate. We use an activity-driven network with n = 250
nodes and study the following two cases:
Case 1. Activity rates following a uniform distribution over [0, 10−2];
Case 2. Activity rates following a probability distribution F (a) that is propor-
tional to a−2.8 in the interval [10−3, 1] and equal to zero elsewhere [30].
We assume that both the adaptation factors χi and acceptance rates πi follow a
uniform distribution over [0, 1]. For various values of the infection rate β, the recovery
rate δ, and m, we compute the decay rate α based on numerical simulations of the
model and its upper bound αu. To compute the decay rate α, we use Monte Carlo
simulations. For each triple (β, δ, m), we run 10,000 simulations of the activity-driven
A-SIS model with all nodes being infected at time t = 0. In each numerical simulation,
we compute the probability vector p(t) for each t = 0, 1, 2, . . . until the norm (cid:107)p(t)(cid:107)
falls below 0.1. We then estimate the decay rate by α = maxt t−1 log(cid:107)p(t)(cid:107).
In Figure 4, we let β = 0.8 and compare the decay rates and their upper bounds
in Case 1 for various values of δ and m. We confirm that αu bounds the numerically
obtained decay rates. The discrepancy αu − α increases with m. To further examine
how the discrepancy depends on the parameters, we present the discrepancy for vari-
ous values of β, δ, and m in Figure 5. Besides the aforementioned dependence of the
discrepancy on m, we also see that the discrepancy tends to be large when β is large.
00.050.10.1500.050.10.15 14
M. OGURA, V. M. PRECIADO, AND N. MASUDA
Fig. 6: Comparison between the numerically obtained decay rates α and their upper
bounds αu in Case 2 when β = 0.8. (a) m = 2, (b) m = 10, and (c) m = 50.
Fig. 7: Discrepancy between the true decay rates and their upper bounds in Case 2.
(a) m = 2, (b) m = 10, and (c) m = 50.
Fig. 8: Cost function fi(χi) for p = 0.01, 1, 10, and 100 when χi = 0.5.
We observe the same trend in the case of the power-law distribution of the activity
rate (Case 2; shown in Figures 6 and 7).
5.2. Optimal resource distribution. We numerically illustrate our framework
to solve the optimal resource allocation problems developed in Section 4. We assume
¯χi = 1 and ¯πi = 1 in the box constraints (40). We use the following cost functions
00.020.040.060.0800.020.040.060.08 CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
15
Fig. 9: Optimal investments on adaptation factors (circles) and acceptance rates
(squares) in Case 1. Left column: m = 2, middle column: m = 10, right column:
m = 50. Top row: πi = 0.2, middle row: πi = 0.7, bottom row: πi = 0.9.
(similar to the ones in [32]):
(55)
fi(χi) = (1 −
−p
i − 1
−p
i − 1
,
χi)
¯
χ
χ
¯
gi(πi) = (1 −
−q
i − 1
−q
i − 1
.
π
π
¯
πi)
¯
These cost functions satisfy Assumption 4.3. Parameters p, q > 0 tune the shape of
the cost functions as illustrated in Figure 8. Because the cost functions are normalized
as fi(
πi, and gi(¯πi) = gi(1) = 0, the
πi) (1 ≤ i ≤ n) in the network is achieved with
¯
maximum adaptation (χi, πi) = (
¯
the budget
πi) = 1 −
χi, fi( ¯χi) = fi(1) = 0, gi(
¯
¯
χi) = 1 −
¯
(56)
χi,
¯
Cmax = 2n − n(cid:88)
(
χi +
¯
πi).
¯
i=1
As in our previous simulations, we consider the activity-driven A-SIS model over
χi = 0.8.
a network with n = 250 nodes. We let m be either 2, 10, or 50. Let
¯
πi be either 0.2, 0.7, or 0.9, and use p = q = 0.01 for the
We let the value of
¯
cost functions (55). We use the fixed budget ¯C = Cmax/4. For each pair (m,
πi),
¯
we determinate the adaptation factors and acceptance rates for the cost-constrained
16
M. OGURA, V. M. PRECIADO, AND N. MASUDA
Fig. 10: Optimal investments on adaptation factors (red circles) and acceptance rates
(blue squares) in Case 2. Left column: m = 2, middle column: m = 10, right column:
m = 50. Top row: πi = 0.2, middle row: πi = 0.7, bottom row: πi = 0.9.
optimal resource allocation problem (Problem 4.1) by solving the geometric program
shown in Theorem 4.4.
The optimal investments on the adaptation factors and acceptance rates (i.e.,
fi(χi) and gi(πi)) are shown in Figure 9 for Case 1. We see that the smaller the
lower limit of the acceptance rate
πi, the more we should invest on decreasing the
¯
acceptance rates. We also find that, in the case of m = 2 and 10, the resulting
investment is almost independent of the activity rates of nodes. This trend disappears
for larger values of m. In the case of m = 50, the optimal solution disproportionately
invests on the nodes having high activity rates. This result reflects the structure of the
optimization problem (46) for the following reason. If m (cid:28) n, then ¯m (cid:28) 1. Therefore,
the set of constraints in the optimization problem (46) approximately reduces to the
set of constraints (40),
(57)
(58)
ζ−1 + λ ≤ 1,
η−1 + λ ≤ 1,
and (46f). These four constraints do not involve activity rates, which leads to optimal
adaptation factors and acceptance rates that are independent of the activity rates.
On the other hand, for a large m, constraints (46c)–(46e) involving the weighted
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
17
sums (cid:104)χ(cid:105)a, (cid:104)π(cid:105)a, and (cid:104)χπ(cid:105)a2 become tighter, rendering investments on high-activity
nodes more effective.
The optimal investments for Case 2 are shown in Figure 10. As in Case 1, the
smaller
πi, the optimal solution invests more on decreasing the acceptance rates.
¯
However, the dependence of the optimal solution on the value of m is not as strong
as in Case 1.
6. Conclusion. In this paper, we have studied epidemic processes taking place
in temporal and adaptive networks. Based on the activity-driven network model, we
have proposed the activity-driven A-SIS model, where infected individuals adaptively
decrease their activity and reduce connectivity with other nodes to prevent the spread
of the infection.
In order to avoid the computational complexity arising from the
model, we have first derived a linear dynamics able to upper-bound the infection
probabilities of the nodes. We have then derived an upper-bound on the decay rate of
the expected number of infected nodes in the network in terms of the eigenvalues of a
2× 2 matrix. Then, we have shown that a small correlation between the two different
adaptation mechanisms is desirable for suppressing epidemic infection. Furthermore,
we have proposed an efficient algorithm to optimally tune the adaptation rates in
order to suppress the number of infected nodes in networks. We have illustrated our
results by numerical simulations.
Appendix A. Proof of (31). The image of A is spanned by 1 and 1 − ψ. All
but two eigenvalues of A are zero. Therefore, an eigenvector of A corresponding to
the spectral radius of A is a linear combination of vectors 1 and 1 − ψ. Assume that
v = s11 + s2(1 − ψ) is such an eigenvector of A with eigenvalue λ. By comparing the
coefficients of the vectors 1 and 1 − ψ in the eigenvalue equation Av = λv, we obtain
s1n + s2
ψi
= λs1,
(cid:32)
n − n(cid:88)
(cid:33)
(cid:32)
n − n(cid:88)
i=1
φi
(cid:33)
− s2
(59)
(60)
− s1
i=1
We rewrite (59) and (60) as
(61)
(cid:33)
ψiφi
= λs2.
n(cid:88)
i=1
φi +
(cid:32)
n − n(cid:88)
(cid:20)s1
(cid:21)
i=1
nB
s2
ψi − n(cid:88)
(cid:20)s1
(cid:21)
i=1
,
= λ
s2
where we have used the notations in (32) and (33). Therefore, λ is an eigenvalue of
the matrix nB, as desired.
Acknowledgments. N.M. acknowledges the support provided through JST
CREST Grant Number JPMJCR1304 and the JST ERATO Grant Number JPM-
JER1201, Japan. V.M.P. acknowledges the support of the US National Science Foun-
dation under grant CAREER-ECCS-1651433.
REFERENCES
[1] J. E. Aledort, N. Lurie, J. Wasserman, and S. A. Bozzette, Non-pharmaceutical pub-
lic health interventions for pandemic influenza: an evaluation of the evidence base,
BMC Public Health, 7 (2007), p. 208, https://doi.org/10.1186/1471-2458-7-208, http:
//bmcpublichealth.biomedcentral.com/articles/10.1186/1471-2458-7-208.
18
M. OGURA, V. M. PRECIADO, AND N. MASUDA
[2] D. Bell, A. Nicoll, K. Fukuda, P. Horby, A. Monto, F. Hayden, C. Wylks, L. Sanders,
and J. Van Tam, Nonpharmaceutical interventions for pandemic influenza, national and
community measures, Emerging Infectious Diseases, 12 (2006), pp. 88–94, https://doi.org/
10.3201/eid1201.051371.
[3] M. C. J. Bootsma and N. M. Ferguson, The effect of public health measures on the 1918
influenza pandemic in U.S. cities, Proceedings of the National Academy of Sciences, 104
(2007), pp. 7588–7593, https://doi.org/10.1073/pnas.0611071104, http://www.pnas.org/
cgi/doi/10.1073/pnas.0611071104.
[4] S. Boyd, S.-J. Kim, L. Vandenberghe, and A. Hassibi, A tutorial on geometric pro-
gramming, Optimization and Engineering, 8 (2007), pp. 67–127, https://doi.org/10.1007/
s11081-007-9001-7.
[5] Y. Bu, S. Gregory, and H. L. Mills, Efficient local behavioral-change strategies to reduce
the spread of epidemics in networks, Physical Review E, 88 (2013), p. 042801, https:
//doi.org/10.1103/PhysRevE.88.042801.
[6] D. Chakrabarti, Y. Wang, C. Wang, J. Leskovec, and C. Faloutsos, Epidemic thresholds
in real networks, ACM Transactions on Information and System Security, 10 (2008).
[7] O. Diekmann and J. A. P. Heesterbeek, Mathematical Epidemiology of Infectious Diseases:
Model Building, Analysis and Interpretation, Wiley, 2000.
[8] S. Funk, M. Salath´e, and V. A. A. Jansen, Modelling the influence of human behaviour on
the spread of infectious diseases: a review., Journal of the Royal Society, Interface / the
Royal Society, 7 (2010), pp. 1247–1256, https://doi.org/10.1098/rsif.2010.0142.
[9] T. Gross and B. Blasius, Adaptive coevolutionary networks: a review., Journal of the Royal
Society, Interface / the Royal Society, 5 (2008), pp. 259–271, https://doi.org/10.1098/rsif.
2007.1229.
[10] T. Gross, C. J. D. D'Lima, and B. Blasius, Epidemic dynamics on an adaptive network,
Physical Review Letters, 96 (2006), p. 208701, https://doi.org/10.1103/PhysRevLett.96.
208701.
[11] T. Gross and H. Sayama, Adaptive Networks, Springer-Verlag Berlin Heidelberg, 2009.
[12] D. Guo, S. Trajanovski, R. van de Bovenkamp, H. Wang, and P. Van Mieghem, Epi-
demic threshold and topological structure of susceptible-infectious-susceptible epidemics
in adaptive networks, Physical Review E, 88 (2013), p. 042802, https://doi.org/10.1103/
PhysRevE.88.042802.
[13] P. Holme, Modern temporal network theory: a colloquium, The European Physical Journal B,
88 (2015), p. 234, https://arxiv.org/abs/1508.01303.
[14] P. Holme and J. Saramaki, Temporal networks, Physics Reports, 519 (2012), pp. 97–125.
[15] L. Isella, J. Stehl´e, A. Barrat, C. Cattuto, J. F. Pinton, and W. Van den Broeck,
What's in a crowd? Analysis of face-to-face behavioral networks, Journal of Theoretical
Biology, 271 (2011), pp. 166–180, https://doi.org/10.1016/j.jtbi.2010.11.033, https://arxiv.
org/abs/1006.1260.
[16] I. Z. Kiss, J. C. Miller, and P. L. Simon, Mathematics of Epidemics on Networks, Springer,
2017.
[17] C. Lagorio, M. Dickison, F. Vazquez, L. A. Braunstein, P. A. Macri, M. V. Migue-
les, S. Havlin, and H. E. Stanley, Quarantine-generated phase transition in epidemic
spreading, Physical Review E, 83 (2011), p. 026102, https://doi.org/10.1103/PhysRevE.
83.026102.
[18] S. Lee, L. E. C. Rocha, F. Liljeros, and P. Holme, Exploiting temporal network structures
of human interaction to effectively immunize populations, PloS One, 7 (2012), p. e36439,
https://doi.org/10.1371/journal.pone.0036439.
[19] S. Liu, N. Perra, M. Karsai, and A. Vespignani, Controlling contagion processes in activity
driven networks, Physical Review Letters, 112 (2014), p. 118702, https://doi.org/10.1103/
PhysRevLett.112.118702.
[20] S. Maharaj and A. Kleczkowski, Controlling epidemic spread by social distancing: Do
it well or not at all, BMC Public Health, 12 (2012), p. 679, https://doi.org/10.1186/
1471-2458-12-679.
[21] V. Marceau, P. A. Noel, L. H´ebert-Dufresne, A. Allard, and L. J. Dub´e, Adaptive
networks: Coevolution of disease and topology, Physical Review E, 82 (2010), p. 036116,
https://doi.org/10.1103/PhysRevE.82.036116.
[22] N. Masuda and P. Holme, Predicting and controlling infectious disease epidemics using tem-
poral networks, F1000prime reports, 5 (2013), p. 6, https://doi.org/10.12703/P5-6.
[23] N. Masuda and P. Holme, Temporal Network Epidemiology, Springer, 2017, https://doi.org/
10.1007/978-981-10-5287-3.
[24] N. Masuda and R. Lambiotte, A Guide to Temporal Networks, World Scientific Publishing,
CONTAINING EPIDEMICS ON ACTIVITY-DRIVEN NETWORKS
19
2016.
[25] M. Ogura and V. M. Preciado, Epidemic processes over adaptive state-dependent net-
works, Physical Review E, 93 (2016), p. 062316, https://doi.org/10.1103/PhysRevE.
93.062316, http://journals.aps.org/pre/abstract/10.1103/PhysRevE.93.062316, https://
arxiv.org/abs/1602.08456.
[26] M. Ogura and V. M. Preciado, Stability of spreading processes over time-varying large-
scale networks, IEEE Transactions on Network Science and Engineering, 3 (2016), pp. 44–
57, https://doi.org/10.1109/TNSE.2016.2516346, http://ieeexplore.ieee.org/document/
7377119/, https://arxiv.org/abs/1507.07017.
[27] M. Ogura and V. M. Preciado, Optimal Containment of Epidemics in Temporal and
Adaptive Networks, in Temporal Networks Epidemiology, Springer, 2017, pp. 241–266,
https://doi.org/10.1007/978-981-10-5287-3, https://goo.gl/vRFGcA, https://arxiv.org/
abs/1612.06832.
[28] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, and A. Vespignani, Epidemic
processes in complex networks, Reviews of Modern Physics, 87 (2015), pp. 925–979, https:
//arxiv.org/abs/1408.2701.
[29] N. Perra, A. Baronchelli, D. Mocanu, B. Gonc¸alves, R. Pastor-Satorras, and
A. Vespignani, Random walks and search in time-varying networks, Physical Review
Letters, 109 (2012), p. 238701, https://doi.org/10.1103/PhysRevLett.109.238701, https:
//arxiv.org/abs/1206.2858.
[30] N. Perra, B. Gonc¸alves, R. Pastor-Satorras, and A. Vespignani, Activity driven mod-
eling of time varying networks, Scientific Reports, 2 (2012), https://doi.org/10.1038/
srep00469.
[31] B. A. Prakash, H. Tong, N. Valler, M. Faloutsos, and C. Faloutsos, Virus propaga-
tion on time-varying networks: theory and immunization algorithms, in Joint European
Conference on Machine Learning and Knowledge Discovery in Databases, 2010, pp. 99–114.
[32] V. M. Preciado, M. Zargham, C. Enyioha, A. Jadbabaie, and G. J. Pappas, Optimal
resource allocation for network protection against spreading processes, IEEE Transactions
on Control of Network Systems, 1 (2014), pp. 99–108, https://doi.org/10.1109/TCNS.2014.
2310911.
[33] T. C. Reluga, Game theory of social distancing in response to an epidemic, PLoS Computa-
tional Biology, 6 (2010), pp. 1–9, https://doi.org/10.1371/journal.pcbi.1000793.
[34] B. Ribeiro, N. Perra, and A. Baronchelli, Quantifying the effect of temporal resolution
on time-varying networks, Scientific Reports, 3 (2013), p. 3006, https://doi.org/10.1038/
srep03006, https://arxiv.org/abs/1211.7052.
[35] A. Rizzo, M. Frasca, and M. Porfiri, Effect of individual behavior on epidemic spreading
in activity-driven networks, Physical Review E, 90 (2014), p. 042801, https://doi.org/10.
1103/PhysRevE.90.042801.
[36] H. Sayama, I. Pestov, J. Schmidt, B. J. Bush, C. Wong, J. Yamanoi, and T. Gross,
Modeling complex systems with adaptive networks, Computers & Mathematics with Ap-
plications, 65 (2013), pp. 1645–1664, https://doi.org/10.1016/j.camwa.2012.12.005, https:
//arxiv.org/abs/1301.2561.
[37] L. Speidel, K. Klemm, V. M. Egu´ıluz, and N. Masuda, Temporal interactions facilitate
endemicity in the susceptible-infected-susceptible epidemic model, New Journal of Physics,
18 (2016), p. 073013, https://doi.org/10.1088/1367-2630/18/7/073013.
[38] M. Starnini and R. Pastor-Satorras, Topological properties of a time-integrated activity-
driven network, Physical Review E, 87 (2013), p. 062807, https://doi.org/10.1103/
PhysRevE.87.062807, https://arxiv.org/abs/arXiv:1304.3370v1.
[39] J. Stehl´e, N. Voirin, A. Barrat, C. Cattuto, L. Isella, J. F. Pinton, M. Quaggiotto,
W. van den Broeck, C. R´egis, B. Lina, and P. Vanhems, High-resolution measurements
of face-to-face contact patterns in a primary school, PLoS ONE, 6 (2011), p. e23176,
https://arxiv.org/abs/1109.1015.
[40] L. Sun, K. W. Axhausen, D.-H. Lee, and X. Huang, Understanding metropolitan pat-
terns of daily encounters, Proceedings of the National Academy of Sciences, 110
(2013), pp. 13774–13779, https://doi.org/10.1073/pnas.1306440110, http://www.pnas.
org/lookup/doi/10.1073/pnas.1306440110, https://arxiv.org/abs/1301.5979.
[41] I. Tunc and L. B. Shaw, Effects of community structure on epidemic spread in an adaptive
network, Physical Review E, 90 (2014), p. 022801, https://doi.org/10.1103/PhysRevE.90.
022801.
[42] D. H. Zanette and S. Risau-Gusm´an, Infection spreading in a population with evolving
contacts, Journal of Biological Physics, 34 (2008), pp. 135–148, https://doi.org/10.1007/
s10867-008-9060-9.
|
1810.05101 | 2 | 1810 | 2018-10-12T21:23:33 | Centrality in Modular Networks | [
"cs.SI",
"physics.soc-ph"
] | Identifying influential nodes in a network is a fundamental issue due to its wide applications, such as accelerating information diffusion or halting virus spreading. Many measures based on the network topology have emerged over the years to identify influential nodes such as Betweenness, Closeness, and Eigenvalue centrality. However, although most real-world networks are modular, few measures exploit this property. Recent works have shown that it has a significant effect on the dynamics on networks. In a modular network, a node has two types of influence: a local influence (on the nodes of its community) through its intra-community links and a global influence (on the nodes in other communities) through its inter-community links. Depending of the strength of the community structure, these two components are more or less influential. Based on this idea, we propose to extend all the standard centrality measures defined for networks with no community structure to modular networks. The so-called "Modular centrality" is a two dimensional vector. Its first component quantifies the local influence of a node in its community while the second component quantifies its global influence on the other communities of the network. In order to illustrate the effectiveness of the Modular centrality extensions, comparison with their scalar counterpart are performed in an epidemic process setting. Simulation results using the Susceptible-Infected-Recovered (SIR) model on synthetic networks with controlled community structure allows getting a clear idea about the relation between the strength of the community structure and the major type of influence (global/local). Furthermore, experiments on real-world networks demonstrate the merit of this approach. | cs.SI | cs | Noname manuscript No.
(will be inserted by the editor)
Centrality in Modular Networks
Zakariya Ghalmane · Mohammed El
Hassouni · Chantal Cherifi · Hocine
Cherifi
8
1
0
2
t
c
O
2
1
]
I
S
.
s
c
[
2
v
1
0
1
5
0
.
0
1
8
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract Identifying influential nodes in a network is a fundamental issue due
to its wide applications, such as accelerating information diffusion or halting virus
spreading. Many measures based on the network topology have emerged over the
years to identify influential nodes such as Betweenness, Closeness, and Eigenvalue
centrality. However, although most real-world networks are modular, few measures
exploit this property. Recent works have shown that it has a significant effect on
the dynamics on networks. In a modular network, a node has two types of influ-
ence: a local influence (on the nodes of its community) through its intra-community
links and a global influence (on the nodes in other communities) through its inter-
community links. Depending of the strength of the community structure, these
two components are more or less influential. Based on this idea, we propose to
extend all the standard centrality measures defined for networks with no commu-
nity structure to modular networks. The so-called "Modular centrality" is a two
dimensional vector. Its first component quantifies the local influence of a node
in its community while the second component quantifies its global influence on
the other communities of the network. In order to illustrate the effectiveness of
the Modular centrality extensions, comparison with their scalar counterpart are
performed in an epidemic process setting. Simulation results using the Susceptible-
Infected-Recovered (SIR) model on synthetic networks with controlled community
Zakariya Ghalmane
LRIT URAC No 29, Faculty of Science, Rabat IT center, Mohammed V University, Rabat,
Morocco
E-mail: [email protected]
Mohammed El Hassouni
LRIT URAC No 29, Faculty of Science, Rabat IT center, Mohammed V University, Rabat,
Morocco
E-mail: [email protected]
Chantal Cherifi
DISP Lab, University of Lyon 2
E-mail: [email protected]
Hocine Cherifi
LE2I UMR 6306 CNRS, University of Burgundy, Dijon, France
E-mail: [email protected]
2
Zakariya Ghalmane et al.
structure allows getting a clear idea about the relation between the strength of the
community structure and the major type of influence (global/local). Furthermore,
experiments on real-world networks demonstrate the merit of this approach.
Keywords Influential nodes · Centrality measures · Community Structure · SIR
model
1 Introduction
Identifying the most influential nodes in a network has gained much attention
among researchers in the recent years due to its many applications. Indeed, they
play a major role in controlling the epidemic outbreak [1], increasing the publicity
on a new product [2], controlling the rumor spreading [3]. The most popular ap-
proach to uncover these central nodes is to quantify their influence using centrality
measures. Various centrality measures have been proposed to quantify the influ-
ence of nodes based on their topological properties. Degree centrality, betweenness
centrality, closeness centrality are one of the most basic centrality measures, yet
they are widely used.
In real-world networks, community structure is a very common and important
property [4 -- 7]. Although there is no formal definition of a community, it is often
apprehended as a group of nodes tightly connected that are loosely connected with
nodes in other communities. Previous works have shown that community struc-
ture has important effect on the spreading process in networks [8 -- 11]. However,
classical centrality measures [12] do not take into account the influence of this
major topological property on the spreading dynamics. In a modular network, we
can distinguish two types of links that support the diffusion process: the links
that connect nodes belonging to the same community (intra-community links or
strong ties) and the links that bridge the communities (inter-community links or
weak ties). The former exercise a local influence on the diffusion process (i.e, at
the community level), while the latter have a global influence (at the network
level). Therefore, we believe that this two types of links should be treated differ-
ently. Indeed, the intra-community links contribute to the diffusion in localized
densely connected areas of the networks, while the inter-community links allows
the propagation to remote areas of the network. Suppose that an epidemic starts
in a community, as it is highly connected, the intra-community links will tend to
confine the epidemic inside the community, while the inter-community links will
tend to propagate it to the other communities. As their role is quite different, we
propose to represent the centrality of modular networks by a two dimensional vec-
tor where the first component quantifies the intra-community (or local) influence
and the second component quantifies the inter-community (or global) influence of
each individual node in the network. To compute these components, we need to
split the original network into a local and a global network. The local network
is obtained by removing all the inter-community links from the original network.
The global network is obtained by removing all the intra-community links from the
original network. Note that if the original network is made of a single connected
component the global and local networks split into many connected components.
Therefore, care must be taken to adapt the centrality definition to networks with
multiple components. In the following, we restrict our attention to non-overlapping
Centrality in Modular Networks
3
community structure (i.e a node belongs to a single community). Furthermore, we
consider undirected and unweighted networks for the sake of simplicity, but results
can be easily extended to more general situations.
The proposed approach can be summarized as follows:
-- Choose a standard centrality measure.
-- Compute the local network by removing all the inter-community links from
the modular network.
-- Compute the Local component of the Modular centrality using the standard
centrality.
-- Compute the global network by removing all the intra-community links from
the modular network.
-- Compute the Global component of the Modular centrality using the standard
centrality.
As nodes need to be ranked according to their centrality values, it is necessary
to adopt a strategy based on a combination of the two components of the Modu-
lar centrality. Various strategies, that include different levels of information about
the community structure, may be used. As our main concern, in this paper, is
to highlight the multidimensional nature of centrality in modular networks rather
than devising optimal ranking methods, elementary strategies are evaluated. Two
straightforward combination strategies (the modulus and the tangent of the argu-
ment of the Modular centrality) and a weighted linear combination of the compo-
nents of the Modular centrality are investigated.
Experiments are conducted on modular synthetic networks in order to better
understand the relative influence of the local and Global component of the Modu-
lar centrality in the propagation process. Extensive comparisons with the standard
centrality measures show that Modular centrality measures provide more accurate
rankings. Simulations on real-world networks of diverse nature have also been
performed. As their community structure is unknown, a community detection al-
gorithm has been used. Results confirm that node rankings based on the Modular
centrality are more accurate than those made by the standard centrality measures
which have been designed for networks with no community structure.
The rest of the paper is organized as follows. Related modular-based measures
are discussed in the next section. In Section 3, a general definition of the Mod-
ular centrality is given. In this framework, we present the extensions to modular
networks of the most influential centrality measures (closeness, betweenness and
eigenvector centrality). The experimental setting is described in Section 4. We re-
port and analyze the results of the experiments performed on both synthetic and
real-world networks in Section 5. Finally, the main conclusions are presented in
Section 6.
2 Related works
Ranking the nodes according to their centrality constitutes the standard deter-
ministic approach to uncover the most influential nodes in a network. These
measures rely usually on various network topological properties. However, the
community structure of the network is rarely taken into consideration. Few re-
searchers have payed attention to this property encountered in many real-world
4
Zakariya Ghalmane et al.
networks [7 -- 11, 13 -- 18]. In this section, we give a brief overview of the main deter-
ministic methods that motivates our proposition.
a. Comm centrality. N. Gupta et al. [19] proposed a degree-based centrality mea-
sure for networks with non-overlapping community structure. It is based on a non-
linear combination of the number of intra-community links and inter-community
links. The goal is to select nodes that are both hubs in their community and
bridges between the communities. This measure gives more importance to commu-
nity bridges. Indeed, the number of inter-community links is raised to the power of
two. Comparison have been performed with deterministic and random immuniza-
tion strategies using the S.I.R epidemic model and both synthetic and real-world
networks. Nodes are immunized sequentially from each community in the decreas-
ing order of their centrality value in their respective community. Number of nodes
to be removed from a community are kept proportional to the community size.
Results show that the Comm strategy is more effective or at least works as well as
Degree and Betweenness centrality while using only information at the community
level.
b. Number of neighboring communities centrality. In a previous work [20],
we proposed to rank the nodes according to the number of neighboring commu-
nities that they reach in one hop. The reason for selecting these nodes is that
they are more likely to have a big influence on nodes belonging to various com-
munities. Simulation results on different synthetic and real-world networks show
that it outperforms Degree, Betweenness and Comm centrality in networks with a
community structure of medium strength (i.e when the average number of intra-
community links is of the same order than the number of inter-community links).
c. Community Hub-Bridge centrality. We also proposed the Community Hub-
Bridge centrality measure in [20]. It is based on the combination of the number
of intra-community links weighted by the size of the community and the inter-
community links weighted by the number of neighboring communities. This mea-
sure tends to select preferentially nodes that can be considered as hubs inside
large communities and bridges having many connections with various neighboring
communities. According to experimental results, on both synthetic and real-world
networks, this centrality measure is particularly suited to networks with strong
community structure (i.e, when there is few inter-community links as compared
to the number intra-community links). In this situation, it can identify effectively
the most influential spreaders as compared to Degree, Betweenness, Comm and
the Number of Neighboring Communities centrality measures. A variation of this
centrality measure called the Weighted Community Hub-Bridge centrality has also
been introduced. It is weighted such that, in networks with well-defined community
structure, more importance is given to bridges (inter-community links), while in
networks with weak community structure the hubs in the communities dominate.
The goal is to target the bridges or the hubs according to the community structure
strength. This measure has proved its efficiency as compared to the alternatives
particularly in networks with weak community structure.
d. K-sell with community centrality. Luo et al. proposed a variation of the
K-shell decomposition for modular networks [21]. They suggest that the intra-
Centrality in Modular Networks
5
community and the inter-community links should be considered separately in the
K-core decomposition process. Their method works as follows:
(i) After the removal of nodes with intra-community links, the K-shell decompo-
sition of the remaining nodes is computed. It is associated with an index of kW
core.
(ii) After the removal of nodes with inter-community links, the K-shell of the re-
maining nodes is computed. It is associated with an index of kS
(iii) A new measure is then calculated and assigned to each node based on the
linear combination of both kW
core in order to find nodes that are at the
same time bridges and hubs located in the core of the network.
Experiments have been performed using SIR simulations on Facebook friendship
networks at US Universities. Results show that this strategy is more efficient than
the classical K-shell decomposition, the Degree and the Betweenness centrality
measures.
core and kS
core.
e. Global centrality. In [22], M. Kitromilidis et al. propose to redefine the stan-
dard centrality measures in order to characterize the influence of Western artist.
Based on the idea that influential artists have connections beyond their artistic
movement, they propose to define the centrality of modular networks by consid-
ering only the inter-community links. In other words, an influential artist must
be related to multiple communities, rather than being strongly embedded in its
own community. Considering a painter collaboration network where edges between
nodes represent biographical connections between artists, they compared Between-
ness and Closeness centrality measures with their classical version. Results show
that the correlation values between the standard and modified centrality measures
are quite high. However, the modified centrality measures allows to highlight in-
fluential nodes who might have been missed as they do not necessary rank high in
the standard measures.
All these works suggest that it is of prime interest to disentangle the local influ-
ence from the global influence in order to characterize a node centrality in modular
networks. Indeed, these complementary types of influence may carry very different
meanings and be more or less important in different situations. This is the rea-
son why we propose to exploit the mesoscopic granularity level in order to extend
the definition of the centrality measures that are agnostic about the community
structure to modular networks. We propose to represent the centrality measures in
modular networks as a two dimensional vector made of its local and Global com-
ponent. If needed, these two components can be merged in a scalar value, but the
combination can be made in multiple ways according to complementary available
information about the network nature and topological properties.
3 Modular centrality
Our main objective is to take into account the community structure in order to
identify influential nodes. Indeed, in modular networks, a node has two types
of influence: a local influence which is linked to its community features and a
global influence related to its interactions with the other communities. Under this
assumption, we provide a general definition of centrality in modular networks. We
design a generic algorithm for computing the centrality of a node under this general
6
Zakariya Ghalmane et al.
definition. The Modular centrality extension can be naturally inferred from the
various existing definitions of centrality designed for networks without community
structure. To illustrate this process, we give the modular extensions of the most
influential centrality measures (Betweenness, Closeness, and Eigenvector).
3.1 Definitions
of the network Gl =(cid:83)m
3.1.1 Local component of the Modular centrality
Let's consider a network denoted as G(V, E), where V = {v1, v2, ...., vn} and
E = {(vi, vj) \ vi, vj ∈ V } denotes respectively the set of vertices and edges. Its
non-overlapping community structure C is a partition into a set of communities
C = {C1, ..Ck., Cm} where Ck is the kth community and m is the number of com-
munities. The local network Gl is formed by the union of all the disjoint modules
k=1 Ck. These components are obtained by removing all the
inter-community links between modules from the original network G. Each mod-
ule represents a community Ck denoted as Ck(Vk, Ek). Where Vk = {vk
i \ vi ∈ V }
and Ek = {(vk1
i refers to any node vi
belonging to the community Ck.
i ) as the local centrality of
the node vi ∈ Vk. It is computed separately in each module Ck of the local graph
Gl .
j ) \ vi, vj ∈ V and k1 = k2}, while vk
, vk2
For a selected centrality measure β, we define βL(vk
i
3.1.2 Global component of the Modular centrality
i , vq2
j )\ vi, vj ∈ V and q1 = q2}, while vq
p = C(cid:48) is the size of the set S, the global network is defined by Gg =(cid:83)p
Let's consider the network G(V, E), the global network Gg is formed by the union
of all the connected components of the graph that are obtained after removing
all the intra-community links from the original network G(V, E). Let's suppose
that S = {S1, ..Sq.., Sp} is the set of the revealed connected components and
q=1 Sq.
Each component Sq is denoted as Sq(Vq, Eq). Where Vq = {vq
i \ vi ∈ V } and
Eq = {(vq1
i refers to any node vi belonging
to the component Sq. In this network, there may be some isolated nodes (i.e, nodes
that are not linked directly to another community). These nodes are removed from
Gg in order to obtain a trimmed network formed only by nodes linked to different
communities by one hop. Consequently, the set of nodes of Gg is defined then by
Vg = {vi ∈ V \ N 1
vi is the neighborhood set of nodes reachable
in n hops. It is defined by N n
vi = {vj ∈ V \ vi (cid:54)= vj and dG(vi, vj) <= n}, dG is
the geodesic distance.
i ) as the global centrality
of the node vi ∈ Vq. It is computed over each connected component Sq included in
the global graph Gg. Remember that the global centrality measure of the removed
isolated nodes is set to 0.
For a selected centrality measure β, we define βG(vq
vi (cid:54)= 0}. Where N n
Centrality in Modular Networks
7
Fig. 1 A toy example representing the Local network (Gl) and the Global network
(Gg)associated to a a modular network (G) made of four non-overlapping communities.
3.1.3 Modular centrality
It is a vector with two components. The first component quantifies the local influ-
ence of the nodes in their own community through the local graph Gl, while the
second component measures the global influence of the nodes on the other com-
munities trough the connected components of the global graph Gg. The Modular
centrality of a node vi is given by:
BM (vi) = (βL(vk
i ), βG(vq
i ))
k ∈ {1, ..., m} and q ∈ {1, ..., p}
(1)
Where βL and βG represent respectively the local and global centrality of the
node vi.
3.2 Algorithm
The Modular centrality is computed as follows:
Step 1. Choose a standard centrality measure β.
Step 2. Remove all the inter-community edges from the original network G to ob-
tain the set of communities C forming the local network Gl.
Step 3. Compute the local measure βL for each node in its own community.
Step 4. Remove all the intra-community edges from the original network to reveal
the set of connected components S formed by the inter-community links.
Step 5. Form the global network Gg based on the union of all the connected com-
ponent. Isolated nodes are removed from this network and their global centrality
value is set to 0.
Step 6. Compute the global measure βG of the nodes linking the communities
based on each component of the global network.
Step 7. Add βL and βG to the Modular centrality vector BM .
The pseudo-code of the algorithm to compute the Modular centrality is given in
algorithm 1 .
C1C2C34236951871011121415161713C41819202122C1C2C34236951871011121415161713C41819202122S2S1426914161822=+GGlGg7 8
Zakariya Ghalmane et al.
Algorithm 1: Generic computation of the Modular centrality
Input : Graph G(V, E), Centrality measure β
Output: A map M (node : centrality vector)
1 Remove all the inter-community links from G to form the local network Gl
2 Remove all the intra-community links from G to form the global network Gg
3 Create and initialize an empty map M (node : BM )
4 BM (vi) = (βL(vk
i )) represent the centrality vector, where each node vi of the
network should be associated with its local and global value according to the selected
centrality measure β
i ), βG(vq
5 for each Ck ⊂ Gl, where k ∈ {1, ..., m} do
6
7
8
9
12
13
14
for each vk
i ∈ Vk do
Calculate βL(vk
i )
M.add(vi, BM (vi))
end
for each vq
i ∈ Vq do
Calculate βG(vi)
BM (vi).add(βG(vi))
10 end
11 for each Sq ⊂ Gg, where q ∈ {1, ..., p} do
end
15
16 end
17 for each vi ∈ V do
M.add(vi, BM (vi))
18
19 end
20 Return the map M
3.3 Modular extensions of standard centrality measures
In order to illustrate the process allowing to extend a given centrality defined for a
network without community structure to a modular network, we give as examples
the modular definitions of the Betweenness, Closeness and Eigenvector centrality.
3.3.1 Modular Betweeness centrality
The modular Betweeness centrality takes into account separately paths that start
and finish in the same community and those which starts and finish in different
communities. For a given node vi, it is represented by the following vector:
k ∈ {1, ..., m} and q ∈ {1, ..., p}
BM (vi) = (βL(vk
i ), βG(vq
(2)
i ))
Where:
βL(vk
i ) =
βG(vq
i ) =
(cid:88)
(cid:88)
vs,vt∈Ck
vs,vt∈Sq
σst(vi)
σst
σst(vi)
σst
(3)
(4)
βL measures the Betweenness centrality of nodes in their own community and βG
measures the Betweenness centrality of nodes linking the communities. σst is the
Centrality in Modular Networks
9
number of shortest paths connecting nodes vs and vt, while σst(vi) is the number
of shortest paths connecting nodes vs and vt and passing through vi.
3.3.2 Modular Closeness centrality
Modular Closeness centrality considers separately the shortest distances of nodes
originating from the same or from another community than the starting node vi.
It is defined as follows:
BM (vi) = (βL(vk
i ), βG(vq
i ))
k ∈ {1, ..., m} and q ∈ {1, ..., p}
Where:
βL(vk
i ) =
βG(vq
i ) =
vj∈Ck
1(cid:80)
1(cid:80)
vj∈Sq
dij
dij
(5)
(6)
(7)
βL and βG measure respectively the local and Global component of the Modular
Closeness centrality. dij is the length of the geodesic from node vi to node vj.
3.3.3 Modular Eigenvector centrality
Modular Eigenvector takes into account separately both the number and the im-
portance of the neighbors belonging to the same community and those belonging to
different communities to measure its centrality. The community based Eigenvector
of a node vi is defined by the following vector:
BM (vi) = (βL(vk
i ), βG(vq
i ))
k ∈ {1, ..., m} and q ∈ {1, ..., p}
Where:
(cid:88)
(cid:88)
vj∈Ck
vj∈Sq
1
λ
1
λ
aijβL(vk
j )
aijβL(vq
j )
βL(vk
i ) =
βG(vq
i ) =
(8)
(9)
(10)
βL and βG measure respectively the local and Global component of the Modular
Eigenvector centrality. A = (ai,j) is the network adjacency matrix, i.e. ai,j = 1 if
vertex vi is linked to vertex vj, 0 otherwise, and λ is a constant.
10
Zakariya Ghalmane et al.
Table 1 Standard Degree centrality, Global and Local Component of the Modular Degree
Centrality of the nodes in the toy example
Node ID 4
β
7
3
βL
βG
4
Node ID 9
3
β
2
βL
βG
1
11
7
7
0
10
3
3
0
18
5
3
2
14
3
2
1
22
5
3
2
15
3
3
0
1
4
4
0
17
3
3
0
2
4
2
2
19
3
3
0
3
4
4
0
21
3
3
0
16
4
2
2
6
2
1
1
5
3
3
0
12
2
2
0
7
3
2
1
20
2
2
0
8
3
3
0
13
1
1
0
3.4 Toy example
The toy example reported in Figure 1 allows to illustrate the two types of influence
that can occur in a modular network. For the sake of simplicity, we consider the
Degree centrality measure. In this case, v4 and v11 are the most influential nodes as
they have the highest degree value (β(v11) = β(v4) = 7). Even though they share
the same degree value the influence they have on the other nodes of the network
is not comparable. Indeed, their position in the network are quite different: v11 is
embedded in its community, while v4 is at the border of its community. Inspecting
the local and global networks give us a clear picture of their differences. As shown
in the local network Gl, node v11 is the most influential at the community level
since it is linked to all the of nodes of its community (βL(v2
11) = 7), while v4 is
only linked to 3 nodes of its community (βL(v1
4) = 3). Actually, both nodes v1
and v3 are more influential than node v4 in the community C1 with their higher
local Degree values. Looking at the Global network Gl, it appears clearly than
node v4 is the most influential node at the network level since it is connected to
4 nodes inside its component (βG(v1
4) = 4).These nodes belong to all the other
communities of the network (C1, C2 and C3). Therefore, v4 is more influential
than node v11 in the global network Gl because of its ability to reach the different
modules of the network as compared to v11 which is influential only locally (in the
community C2).
To sum up, it can be noticed from this example that when we consider the
Degree centrality, the community hubs are the most influential spreaders locally
due to their ability to reach a high number of nodes in their own communities. The
bridges which are linked to various communities are the most influential spreaders
globally as they allow to reach a high number of communities all over the network.
3.5 Modular centrality ranking strategies
In order to rank the nodes according to their centrality, it is necessary to derive a
scalar value from the Modular centrality vector. To do so, we can proceed in many
different ways. In order to highlight the essential features of centrality in modular
networks, we choose to consider three strategies. The first two are straightforward.
Indeed, a simple way to combine the components of the Modular centrality is to
use the modulus and the argument of this vector. The third strategy uses more
information about the community structure in order to see if this can be beneficial.
Centrality in Modular Networks
11
The modulus r of the modular vector BM of a node vi is defined by:
r(vi) = BM (vi) =
i ))2 + (βG(vq
(βL(vk
i ))2
k ∈ {1, ..., m} and q ∈ {1, ..., p}
(cid:113)
(11)
The argument ϕ of the modular vector BM of a node vi is defined as follows:
ϕ(vi) = arctan
k ∈ {1, ..., m} and q ∈ {1, ..., p}
(12)
(cid:19)
(cid:18) βG(vq
i )
βL(vk
i )
Rather than using the argument, we propose to use the tangent of the argument
because it as a higher range. It is defined by:
tan(ϕ(vi)) =
βG(vq
i )
βL(vk
i )
k ∈ {1, ..., m} and q ∈ {1, ..., p}
(13)
Note that in these ranking strategies the information used about the community
structure is very limited. As we expect that integrating more knowledge about
the community structure in the combination strategy of the Modular centrality
components may improve the efficiency of the ranking method, we also investigate
the so-called "Weighted Modular measure". It is based on a linear combination
of the components of the Modular centrality vector weighted by a measure of the
strength of the communities.
The Weighted Modular measure αW of a node vi is given by:
i ) + µCk ∗ βG(vq
i )
αW (vi) = (1 − µCk ) ∗ βL(vk
(14)
Where k ∈ {1, ..., m}, q ∈ {1, ..., p} and:
(cid:80)
(cid:80)
µCk =
vi∈Ck
kinter(vk
i )
k(vk
i )
vi∈Ck
(15)
i ) is the number of inter-community links of node vk
Where µCk is the fraction of inter-community links of the community Ck.
kinter(vk
degree of node vk
i .
The Weighted Modular measure works as follows:
i and k(vi) is the
-- A community Ck, where the intra-community links predominate is densely
connected and therefore it has a very well-defined community structure. If an
epidemic starts in such a cohesive community, it has more chance to stay con-
fined than to propagate trough the few links that allows to reach the other
communities of the network. In this case, priority must be given to local im-
munization. Consequently, more weight is given to the Local component of the
Modular centrality βL to target the most influential nodes in the community
since it is well separated from the other communities of the network.
-- A community Ck where the inter-community links predominate has a non-
cohesive community structure. It is more likely that an epidemic starting in
this community diffuses to the other communities trough the many links that
it shares with the other communities. Consequently, more weight is given to
the Global component of the Modular centrality measure βG in order to target
nodes that can propagate the epidemic more easily all over the network due to
the loose community structure of Ck.
12
Zakariya Ghalmane et al.
Table 2 LFR network parameters
Number of nodes
Average degree
Maximum degree
Exponent for the degree distribution
Exponent for the community size distribution
Mixing parameter
Community size range
4000
7
80
2.8
2
0.1, 0.4, 0.7
[15 200]
4 Experimental setting
In this section, we give some information about the synthetic and real-world
dataset used in the empirical evaluation of the centrality measures. The SIR sim-
ulation process is recalled, together with the measure of performance used in the
experiments.
4.1 Dataset
4.1.1 Synthetic networks
In order to generate artificial modular networks with controlled topological prop-
erties, the LFR benchmark is used [23]. It allows to generate small-world networks
with power-law distributed degree and community size. The input parameters of
the model are the number of nodes, the desired average and maximum degree, the
exponents for the degree and the community size distributions, and the mixing
coefficient. The mixing coefficient parameter µ value ranges from 0 to 1. It repre-
sents the average proportion of links between a node and the ones located outside
of its community. This parameter allows to control the strength of the community
structure. If its value is low, there is few links between the communities and they
are well separated from each other. A high value of µ indicates a very loose com-
munity structure. Indeed, in this case, a node share more links with nodes outside
its community than with nodes inside its community. Experimental studies have
shown that typical value of the degree distribution exponent in real-world net-
works vary in the range 2 ≤ γ ≤ 3. Networks can have different size going from
tens to millions of nodes. In addition, it is also difficult to characterize the average
and the maximal degree since they are very variable. Consequently, we choose for
these parameters some consensual values while considering also the computational
aspect of the simulations. They are reported in Table 2.
4.1.2 Real-world networks
Although the LFR model produces pretty realistic networks, uncontrolled prop-
erties such as transitivity and degree correlation can deviate significantly from
those observed in real-world networks [24]. Therefore, it is necessary to use real-
world networks in the evaluation process. In order to cover a wide range of situa-
tions, we selected networks from various origin: online social networks, collabora-
tion networks, technological networks, communication networks. All networks are
Centrality in Modular Networks
13
Table 3 Description of the structural properties of the real-world networks. N is the total
numbers of nodes, E is the number of edges. < k >, kmax are respectively the average and
the max degree. C is the average clustering coefficient. αth is the epidemic threshold of the
network.
Network
ego-Facebook
Caltech
Princeton
Georgetown
Email-Eu-core
Power grid
CR-QC
N
4039
620
5112
7651
986
4941
4158
E
88234
7255
28684
163225
25552
6594
13428
< k > kmax
1045
43.69
248
43.31
628
88.93
1235
90.42
33.24
347
19
2.66
5.53
81
C
0.605
0.443
0.298
0.268
0.399
0,107
0.529
αth
0.009
0.012
0.006
0.006
0.013
0.092
0.059
undirected and unweighted. Experiments are performed on their largest connected
component. The Louvain Algorithm is used to unveil the community structure of
these networks. We choose this greedy optimization method for its simplicity. Fur-
thermore, this popular algorithm has proved to be a good compromise between
efficiency and complexity when used in many different type of networks [25] [26].
- Social networks: Four Samples of the Facebook Network are used. The ego-
Facebook network collected from survey participants using the Facebook app. 1 [27]
and the Facebook friendship network at 3 US universities (Caltech, Princeton,
Georgetown) collected by Traud et al [28]. Nodes represent individuals (survey
participant or members of the University), and edges represent online friendship
links between two individuals. In the University network, in order to obtain data
that are relevant for the spread of epidemic infections, only the relationship of
individuals who live in the same dormitory or study the same major are considered.
- Communication network: The Email-Eu-core1 network has been gener-
ated using email data from a large European research institution. The dataset con-
tains only communication between institution members. Each node corresponds
to an email address and an edge is established between two nodes u and v, if at
least one email has been exchanged between address u and address v.
- Technological network: Power-Grid2 is a network containing information
about the topology of the Western States Power Grid of the United States. An
edge represents a power supply line. A node is either a generator, a transformator
or a substation.
- Collaboration network: GR-QC1 (General Relativity and Quantum Cos-
mology) collaboration network has been collected from the e-print arXiv and covers
scientific collaborations between authors of papers submitted to the General Rela-
tivity and Quantum Cosmology category. If an author i co-authored a paper with
author j, the graph contains an edge from i to j. If the paper is co-authored by k
authors this generates a completely connected (sub)graph on k nodes.
The basic topological properties of these networks are given in Table 3.
1http://snap.stanford.edu/data
2http://www-personal.umich.edu/~mejn/netdata/
14
Zakariya Ghalmane et al.
4.2 SIR simulations
To evaluate the efficiency of the centrality measures, we consider an epidemic
spreading scenario using the Susceptible-Infected-Recovered (SIR) model [29]. In
this setting, nodes can be classified into three classes: S(Susceptible), I(Infected)
and R(Recovered). Initially, all nodes are set as susceptible nodes. Then, a given
fraction f0 of the top-ranked nodes according to the centrality measure under
test are set to the state infected. After this initial setup, at each iteration, each
infected node affects one of its susceptible neighbors with probability α. Besides,
the infected nodes turn into recovered nodes with the recover probability σ. To
better characterize the spreading capability, the value of the transmission rate α
is chosen to be greater than the network epidemic threshold αth given by [30]:
αth =
< k >
< k2 > − < k >
(16)
Where < k > and < k2 > are respectively the first and second moments of the
degree distribution. The epidemic threshold values αth for the networks used in
this study are reported in Table 3. In all the experiments, we use the same value of
the transmission rate (α = 0.1). Naturally, it is much larger than the values of the
epidemic threshold αth of all the dataset. The value of the recover probability is
also constant (σ = 0.1). We choose this small value so that each infected node may
have many chances to infect its neighbors with the probability α before changing
to the recovered status. The process continues until there is no more infected node
in the network. Finally, when the spreading process stops, the number of nodes in
the state "Recovered" R is used to measure the spreading efficiency of the fraction
of the initially infected nodes. The larger the value of R, the more influential
the initially selected nodes. To ensure the effectiveness of the evaluation process,
results of the SIR simulations are averaged over 200 independent realizations. A
more detailed description of the SIR simulation process is given in algorithm 2.
4.3 Evaluation criteria
In order to compare the ranking efficiency of a centrality measure with the one
obtained by a reference centrality measure, we compute the relative difference of
the outbreak size. It is given by:
∆r =
(17)
Rm − Rs
Rs
Where Rm is the the final number of recovered nodes of the ranking method under
test, and Rs is the final number of recovered nodes for the reference method. Thus,
a positive value of ∆r indicates a higher efficiency of the method under test as
compared to the reference.
5 Experimental Results
Extensive experiments have been performed in order to evaluate the effectiveness
of the most popular Modular centrality extensions (Degree, Betweenness, Close-
ness and Eigenvector centrality) as compared to their standard definition. First,
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Select one infected node v from the infected list L Inf ected
for each node v(cid:48) neighbor of v do
if v(cid:48) is susceptible then
With a probability α set the node v(cid:48) as infected
L Inf ected.add(v(cid:48))
end
else
With a probability σ set the node v(cid:48) as recovered
L Inf ected.remove(v(cid:48))
nR ← nR + 1
end
end
Centrality in Modular Networks
15
Algorithm 2: Decription of the SIR simulation process
Input : Graph G(V, E),
Centrality measure: β,
Fraction of the initial spreaders: f0,
Transmission rate: α,
Recovery rate: σ,
The number of simulations: n
Output: The average number of the recovered nodes after the SIR simulations: rav,
The standard deviation: rdev
1 Rank the nodes according to the centrality measure β
2 Set all nodes as susceptible nodes
3 Compute nI the number of initially infected nodes: nI ← card(V ) ∗ f0
4 Select nI of the top ranked nodes and change their state to the infected state
5 Add the initially infected nodes to the infected list L Inf ected
6 Initialize the list of the numbers of recovered nodes nR obtained after each simulation:
L nbrR ← EmptyList()
7 for counter from 0 to n do
8
nR ← 0
while L Inf ected (cid:54)= N ull do
end
nR ← nR − nI
L nbrR.add(nR)
24
25 end
26 Compute the average number rav and the standard deviation rdev of the recovered
nodes over the n simulations based on the list L nbrR
27 Return rav, rdev
the local and Global component of the various Modular centrality measures are
compared to their standard counterpart. Next, the three ranking methods based on
the combination of the components of the Modular centrality are also evaluated.
These experiments are conducted on both synthetic and real-world networks.
5.1 Synthetic networks
Networks with different mixing parameter values have been generated in order to
better understand the effect of the community structure strength on the perfor-
mance of the various centrality measures. Figure 2 represents the relative difference
of the outbreak size as a function of the fraction of immunized nodes with the stan-
dard measure used as a reference. The mixing parameter values µ cover all the
16
Zakariya Ghalmane et al.
Fig. 2 The relative difference of the outbreak size ∆r as a function of the fraction of the
initial spreaders f0 for Synthetic networks. The Degree (a), Betweenness (b), Closeness (c)
and Eigenvector (d) centrality measures derived from the Modular centrality are compared
to the standard counterpart designed for networks with no community structure. Networks
are generated using the LFR algorithm with various µ values. For each µ value, five sample
networks are generated. The final epidemic sizes are obtained by averaging 200 SIR model
simulations per network for each initial spreading coverage value. Positive ∆r value means
higher efficiency of the measure under test as compared to the standard centrality.
0246810121416f0 (%)−20020406080100Ɗr (%)=0.10246810121416f0 (%)−1001020Δ04050Ɗr (%)=0.4(a) Degree centrality0246810121416f0 (%)−20−1001020Δ04050Ɗr (%)=0.7Local meas reGlobal meas retan(Arg)Mod l sWeighted mod lar meas re0246810121416f0 (%)−20020406080100Ɗr (%)=0.10246810121416f0 (%)−1001020Δ04050Ɗr (%)=0.4(b) Betweeness centrality0246810121416f0 (%)−20−1001020Δ04050Ɗr (%)=0.7Local meas reGlobal meas retan(Arg)Mod l sWeighted mod lar meas re0246810121416f0 (%)−20020406080100Ɗr (%)=0.10246810121416f0 (%)−1001020Δ04050Ɗr (%)=0.4(c) Closeness centrality0246810121416f0 (%)−20−1001020Δ04050Ɗr (%)=0.7Local meas reGlobal meas retan(Arg)Mod l sWeighted mod lar meas re0246810121416f0 (%)−20020406080100Ɗr (%)=0.10246810121416f0 (%)−1001020Δ04050Ɗr (%)=0.4(d) Eigenvector centrality0246810121416f0 (%)−20−1001020Δ04050Ɗr (%)=0.7Local meas reGlobal meas retan(Arg)Mod l sWeighted mod lar meas re Centrality in Modular Networks
17
range of community structure strength.
5.1.1 Evaluation of the local and the Global component of the Modular centrality
a. Strong community structure strength
In networks with well-defined community structure, the Local component of the
Modular centrality always outperforms the standard measures for all the centrality
measures as it is shown in the left panels of the Figure 2 (when µ = 0.1). The
gain is around 20% as compared to the standard measure for Closeness, Degree
and Eigenvector centrality. The smallest gain is for Betweenness centrality with
an average value of 10%. On the contrary, the Global component of the Modu-
lar centrality is always less performing than the standard measures. These results
clearly demonstrates that it is more efficient to immunize the influential nodes
inside the communities when there is few inter-community links in the networks.
Indeed, as there is few inter-community edges, the infection may die out before
reaching other communities. So, the local influence of nodes is more important
than global influence in networks with strong community structure.
b. Medium community structure strength
The middle panels of Figure 2 show the performance of the various ranking meth-
ods in networks with community structure of medium strength (µ = 0.4). In this
case, both the Global and Local components of the Modular centrality are always
more efficient than the standard centrality. The gain in performances of the Global
component of the Modular centrality is always greater than for the Local com-
ponent. Indeed, the Global component outperforms the standard measure with
a Gain around 12% for Betweenness, Closeness and Eigenvector centrality. The
largest gain is for the Degree centrality with an average value of 17%. The Local
component of the Modular centrality performs better than the standard measure
with a gain around 5% for Betweenness, Closeness and Eigenvector centrality and
around 12% for Degree centrality. These results send a clear message: In networks
with medium community structure strength, the global influence is more impor-
tant than the local influence. Indeed, with a greater number of inter-community
links, there is more options to spread the epidemics to the other communities of
the network.
c. Weak community structure strength
The right panel of Figure 2 reports the comparison between the Modular centrality
and the traditional centrality measures in networks with non-cohesive community
structure (µ = 0.7). It appears that the relative difference of the outbreak size
between the Global component of the Modular centrality and the standard cen-
trality is always positive while it is always negative for the Local component of the
Modular centrality. And this is true for all the centrality measures under test. In
fact, there is a gain around 5% using the Global component of the Modular cen-
trality, while the Local component performs worse than the traditional measure
with an average of 5% for the Degree, Betweenness, Closeness and Eigenvector
centrality measures. Consequently, we can conclude that in networks with a loose
community structure the global influence is dominant, even if the difference with
18
Zakariya Ghalmane et al.
the standard measure is not as important than for networks with a medium com-
munity structure. Indeed, in this situation (µ = 0.7), the inter-community edges
constitute the majority of edges in the network (around 70% of links lie between
the communities). In fact, as the community structure is not well defined, minor
differences are observed with a a network that has no community structure.
5.1.2 Evaluation of the ranking methods of the Modular centrality
Figure 2 reports also the relative difference of the outbreak size ∆r as a function of
the fraction of the initial spreaders f0 for the three ranking methods (Modulus and
Tangent of the argument of the Modular centrality, Weighted Modular measure)
and for the various centrality measure and community structure strength under
study. The first observation that can be made from these results is that combin-
ing the components of the Modular centrality is always more efficient than using
either a single component or the conventional centrality. This remark holds for all
the centrality measures studied and whatever the community structure strength.
Additionally, the ranking of the three combination methods in terms of efficiency
is always the same. The Weighted Modular measure ranks first. It is followed by
the Modulus and then the Tangent of the argument of the Modular centrality. We
believe that it is due to the fact that the Weighted Modular measure uses more
information about the topology of the community structure than its alternatives.
Indeed, the weights introduced in this measure allow tuning locally the relative
importance of the local and Global component for each community in the net-
work. Thus, the Weighted Modular measure can adapt to the structure of each
community in the network. As a result, it is more efficient than the other proposed
ranking methods. One of the main benefits of this result is to highlight the fact
that significant gains can result from improving the way the local and Global com-
ponent are combined and that there may be still room for improvement in this
direction. In other words, even more effective measures can be obtained if rele-
vant additional information about the community structure is used. Furthermore,
it is noticed that the ranking strategies show their best performance in networks
with well defined community structure. For instance, The Modulus of the Modu-
lar centrality outperforms the standard measure with a gain, on average, of 40%
in networks with strong community structure, 25% in networks with community
structure of medium strength and 20% in networks with unclear community struc-
ture for all the centrality measures. For the Weighted Modular measure, the gain
is around 42% in networks with strong community structure, 29% in networks
with community structure of medium strength and 25% in networks with unclear
community structure for all the centrality measures. The gain therefore decreases
as the community structure becomes more and more loose. The reason behind
that behavior is that the local centrality is typical of networks with a community
structure while the global centrality is also a feature of networks with no commu-
nity structure. As the mixing proportion increases, the differences with networks
without community structure become less and less important. Indeed, the global
network size increases until it tends to represent the major part of the original
network. In the limiting case, it is a network with no community structure and
the Modular centrality reduces to its Global component which is identical to the
classical centrality measures.
Centrality in Modular Networks
19
Table 4 The estimated mixing parameter µ and modularity Q of the real-world networks.
Network
µ
Q
Network
µ
Q
ego-Facebook
Power-grid
ca-GrQc
Princeton
0.03
0.834
0.034
0.934
0.095
0.86
0.354
0.753
Email-Eu-core
Caltech
Georgetown
0.42
0.569
0.448
0.788
0.522
0.662
5.2 Real-world networks
In this section, we report the results of the set of experiments on real-world net-
works. Experiments performed with synthetic networks have shown that the com-
munity structure strength plays a major role in determining the performance of
the various centrality measures. Therefore, we adopt the same presentation for
real-world networks in order to link the results of this set of experiments with
those obtained using synthetic networks. Once the community structure has been
uncovered using the Louvain algorithm, the mixing proportion parameter is com-
puted for each network. Estimated values are reported in Table 4. According to
these results we can classify the ego-Facebook network, Power Grid and the ca-
GrQc as networks with strong community structure. Princeton, Email-Eu-core and
Caltech have a community structure of medium strength while Princeton has a
weak community structure.
5.2.1 Evaluation of the local and Global component of the Modular centrality
a. Strong community structure strength
The relative difference of the outbreak size between the community based measures
and the standard measure is reported in Figure 3. To evaluate their performances
in networks with strong community structure, ego-Facebook, power grid and the
ca-GrQc networks are used. For these networks, the estimated mixing parameter
value ranges from to 0.03 to 0.095. In this figure, we notice that for all the central-
ity measures under test the standard measure outperforms the Global component
of the Modular centrality while it is less performing that its Local component.
Let's consider for example the Betweenness centrality. With a fraction of the ini-
tial spreaders equal to 8%, the gain in terms of the outbreak size for the Local
component of the Modular centrality as compared to the standard Betweenness is
19% for the ego-Facebook network, 14% for Power-Grid and and 9% for ca-QrGc.
Conversely, in the same situation, the loss associated with the use of the Global
component of the Modular centrality instead of the standard Betweenness ranges
from 4% to 11%.
In these networks, communities are densely connected and there are few links
lying between the communities. Therefore, in most cases, contagious areas are
found in the core of the communities and the spread of the epidemic may stop be-
fore even reaching the community perimeter. Thus, there is a low probability that
a bridge (inter-community link) propagates the epidemic to the other communities.
This is the reason why the Local component of the Modular centrality performs
always better than the Global component. Furthermore, we can also notice on
20
Zakariya Ghalmane et al.
Fig. 3 The relative difference of the outbreak size ∆r as a function of the fraction of initial
spreaders f0. The Degree (a), Betweenness (b), Closeness (c) and Eigenvector (d) centrality
measures derived from the Modular centrality are compared to the standard counterpart de-
signed for networks with no community structure. Real-world networks with strong community
structure (ego-Facebook, Power-Grid and ca-GrQc networks) are used. The estimated values
of their mixing coefficient is equal respectively to 0.03, 0.034 and 0.095.
0246810121416f0 (%) 20020406080Ɗr (%)ego-Facebook0246810121416f0 (%) 20020406080Ɗr (%)Power-Grid(a) Degree centrality0246810121416f0 (%) 20020406080Ɗr (%)ca-GrQcLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%) 20020406080Ɗr (%)ego-Facebook0246810121416f0 (%) 20020406080Ɗr (%)Power-Grid(b) Betweeness centrality0246810121416f0 (%) 20020406080Ɗr (%)ca-GrQcLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%) 20020406080Ɗr (%)ego-Facebook0246810121416f0 (%) 20020406080Ɗr (%)Power-Grid(c) Closeness centrality0246810121416f0 (%) 20020406080Ɗr (%)ca-GrQcLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%) 20020406080Ɗr (%)ego-Facebook0246810121416f0 (%) 20020406080Ɗr (%)Power-Grid(d) Eigenvector centrality0246810121416f0 (%) 20020406080Ɗr (%)ca-GrQcLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure Centrality in Modular Networks
21
Fig. 4 The relative difference of the outbreak size ∆r as a function of the fraction of initial
spreaders f0. The Degree (a), Betweenness (b), Closeness (c) and Eigenvector (d) central-
ity measures derived from the Modular centrality are compared to the standard counterpart
designed for networks with no community structure. Real-world networks with medium com-
munity structure (Princeton, Email-Eu-core and Caltech) are used. The estimated values of
their mixing coefficient is equal respectively to 0.354, 0.42 and 0.44.
0246810121416f0 (%)01020304050Ɗr (%)Princeton0246810121416f0 (%)01020304050Ɗr (%)Email-Eu-core(a) Degree centrality0246810121416f0 (%)01020304050Ɗr (%)CaltechLocal meas reGlobal meas retan(Δrg)Mod l sWeighted mod lar meas re0246810121416f0 (%)01020304050Ɗr (%)Princeton0246810121416f0 (%)01020304050Ɗr (%)Email-Eu-core(b) Betweeness centrality0246810121416f0 (%)01020304050Ɗr (%)CaltechLocal meas reGlobal meas retan(Δrg)Mod l sWeighted mod lar meas re0246810121416f0 (%)01020304050Ɗr (%)Princeton0246810121416f0 (%)01020304050Ɗr (%)Email-Eu-core(c) Closeness centrality0246810121416f0 (%)01020304050Ɗr (%)CaltechLocal meas reGlobal meas retan(Δrg)Mod l sWeighted mod lar meas re0246810121416f0 (%)01020304050Ɗr (%)Princeton0246810121416f0 (%)01020304050Ɗr (%)Email-Eu-core(d) Eigenvector centrality0246810121416f0 (%)01020304050Ɗr (%)CaltechLocal meas reGlobal meas retan(Δrg)Mod l sWeighted mod lar meas re 22
Zakariya Ghalmane et al.
Fig. 5 The relative difference of the outbreak size ∆r as a function of the fraction of initial
spreaders f0. The Degree (a), Betweenness (b), Closeness (c) and Eigenvector (d) centrality
measures derived from the Modular centrality are compared to the standard counterpart de-
signed for networks with no community structure. A real-world networks with weak community
structure (Georgetown) is used. The estimated value of its mixing coefficient is equal to 0.522.
Figure 3 that when the mixing parameter value increases, (i.e, the community
structure gets weaker), the Local component of the Modular centrality gets less
efficient while the Global component performs better. This is due to the fact that
the Global component increases with the number of inter-community links.
b. Medium community structure strength
Figure 4 shows the relative difference of the epidemic outbreak size between the
community based measures and the standard centrality measure for real-world
networks with medium community structure strength. For that purpose, Prince-
ton, Email-Eu-core and Caltech networks are used since their estimated mixing
parameter values range from 0.354 to 0.448. Results are very clear. In all the situa-
tions, both the local and the Global component of the Modular centrality measures
outperform the standard centrality measure. In addition, there is still a slight ad-
vantage for the Global component on the Local component. To set these ideas on a
simple example, let us consider the Betweenness centrality with an initial fraction
of infected nodes equal to 8%. The Global component of the modular Betweenness
measure is more efficient than the traditional Betweenness with a gain of 12% for
the Princeton, 13% for Email-Eu-core and 15% for the Caltech network. These fig-
0246810121416f0 (%)−10010203040 r (%)DegreeLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%)−10010203040 r (%)BetweennessLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%)−10010203040 r (%)ClosenessLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure0246810121416f0 (%)−10010203040 r (%)EigenvectorLocal measureGlobal measuretan(Δrg)ModulusWeighted modular measure Centrality in Modular Networks
23
ures need to be compared to a gain of 8% for Princeton, 7% for Email-Eu-core and
3% for Caltech using the Local component of the Modular Betweenness instead
of the classical Betweenness centrality. One can also notice that the gap between
their respective performance gets bigger as the value of the mixing parameter in-
creases. Indeed, as the community structure gets weaker, the relative influence of
the Global component of the Modular centrality becomes more and more impor-
tant. In these networks, nodes have approximately as many internal links as there
are external links. Therefore, the epidemic can spreads easily to all the communi-
ties in the network through the large number of inter-community links. This is the
reason why the Global component of the Modular centrality outperforms always
the Local component. In addition to that, the community structure of the network
is still well preserved, which explains that the Modular centrality is more efficient
than the classical centrality.
c. Weak community structure strength
Figure 5 shows the relative difference of the epidemic outbreak size between the
modular based centrality measures and the standard centrality for the Georgetown
network. With a mixing parameter value equal to 0.522, this network is classified
as a network with a weak community structure. In all circumstances, the standard
measure performs better than the Local component of the Modular centrality
and it performs worse than its Global component. On average, there is a gain of
around 10% for the Global component compared to a loss of 5% for the Local
component of the four centrality measures under test. In this type of networks,
the inter-community links predominate, which translates into a greater influence of
the Global component of the Modular centrality. Indeed, the epidemic can spread
more easily into the various communities of the network through the big amount of
external links. Additionally, we notice that the relative difference of the outbreak
size between the Global component of the Modular centrality measure and the
standard measure decreases as compared to networks with a medium community
structure. Indeed, there is less and less topological differences between networks
with a weak community structure and networks that have no community structure
as the value of the mixing parameter increases.
5.2.2 Evaluation of the ranking methods of the Modular centrality
Figures 3 to 5 report also the relative difference of the epidemic outbreak size be-
tween the Modular centrality ranking methods and the standard centrality mea-
sures. The results are clear evidence of the efficiency of the Modular centrality.
Whatever the ranking strategy of the Modular centrality adopted, it outperforms
in all the situations the local and Global component of the Modular centrality
and the classical centrality. The improvements in terms of performance compared
to the classical centrality are quite significant. For instance, with a fraction of
initial spreaders equal to 8% the modulus of the Betweenness Modular centrality
allows a gain of 45% on the ego-Facebook network, 28% on Princeton and 24%
on Georgetown. As the ranking strategies use both the local and the global in-
formation of each node, they are more efficient than measures relying on either
local or global information taken separately. Furthermore, the Weighted Modular
measure is usually the most efficient measure in most cases. It uses the fraction of
24
Zakariya Ghalmane et al.
inter-community links as an additional information to target the most influential
spreaders in each community. It can give more or less weight to the local and the
Global component according to the individual community structure strength. This
explains its superiority over the other ranking measures. To summarize, these ex-
periments reveal that combining the components of the Modular centrality, allows
to design efficient ranking methods. In addition, using more relevant information
about the community structure at the community level allows to design even more
efficient ranking methods. Moreover, the ranking measures exhibit their best re-
sults in networks with strong community structure.
6 Conclusion
In this paper, we propose a generic definition of centrality measures in networks
with non-overlapping community structure. It is based on the fact that the intra-
community and inter-community links should be considered differently. Indeed, the
intra-community edges contribute to the diffusion in localized densely connected
areas of the network, while the inter-community links allows the global propaga-
tion to the various communities of the network. Therefore, we propose to represent
the centrality of modular networks by a two dimensional vector, where the first
component measures the local influence of a node in its community and the sec-
ond component quantifies its global influence on the other communities. Based
on this assumption, centrality measures defined for networks with no-community
structure can be easily extended to modular networks. Considering the most influ-
ential centrality measures as typical examples, we defined their modular extension.
Experiments based on an epidemic spreading scenario using both synthetic and
real-world networks have been conducted in order to better understand the in-
fluence of the two components of the Modular centrality. First of all, results on
synthetic and real-world networks are quite consistent. It appears that the Local
component is more effective in networks with a strong community structure while
the Global component takes the lead as the community structure gets weaker.
Comparison with the classic centrality always turns to the advantage of the Mod-
ular centrality. More precisely, in networks with strong community structure, the
Local component of the Modular centrality outperforms the Global component
and the standard centrality, while in networks with medium or weak community
structure the Global component performs better than its alternatives. Moreover,
it is also observed that combining both components of the Modular centrality in
order to rank the nodes according to their influence is always more efficient than to
use a single component. Furthermore, further gain can be obtained if the ranking
strategy incorporates more information about the community structure strength.
References
1. Wang, Z., Moreno, Y., Boccaletti, S., Perc, M.: Vaccination and epidemics in networked
populations, an introduction. Elsevier (2017)
2. Medo, M., Zhang, Y.-C., Zhou, T.: Adaptive model for recommendation of news. EPL
(Europhysics Letters) 88(3), 38005 (2009)
Centrality in Modular Networks
25
3. Zhang, Z.-K., Liu, C., Zhan, X.-X., Lu, X., Zhang, C.-X., Zhang, Y.-C.: Dynamics of
information diffusion and its applications on complex networks. Physics Reports 651, 1-34
(2016)
4. Fortunato, S., Hric, D.: Community detection in networks: A user guide. Physics Reports
659, 1-44 (2016)
5. Jebabli, M., Cherifi, H., Cherifi, C., Hamouda, A.: User and group networks on youtube:
A comparative analysis. In: Computer Systems and Applications (AICCSA), 2015
IEEE/ACS 12th International Conference Of, pp. 1-8 (2015). IEEE
6. Jebabli, M., Cherifi, H., Cherifi, C., Hammouda, A.: Overlapping community structure
in co-authorship networks: A case study. In: U-and e-Service, Science and Technology
(UNESST), 2014 7th International Conference On, pp. 26-29 (2014). IEEE
7. Salath´e, M., Jones, J.H.: Dynamics and control of diseases in networks with community
structure. PLoS computational biology 6(4), 1000736 (2010)
8. Saxena, R., Kaur, S., Bhatnagar, V.: Social centrality using network hierarchy and com-
munity structure. arXiv preprint arXiv:1806.08964 (2018)
9. Kumar, M., Singh, A., Cherifi, H.: An efficient immunization strategy using overlapping
nodes and its neighborhoods. In: Companion of the The Web Conference 2018 on The
Web Conference 2018, pp. 1269-1275 (2018). International World Wide Web Conferences
Steering Committee
10. Chakraborty, D., Singh, A., Cherifi, H.: Immunization strategies based on the overlapping
nodes in networks with community structure. In: International Conference on Computa-
tional Social Networks, pp. 62-73 (2016). Springer
11. Taghavian, F., Salehi, M., Teimouri, M.: A local immunization strategy for networks with
overlapping community structure. Physica A: Statistical Mechanics and its Applications
467, 148-156 (2017)
12. Lu, L., Chen, D., Ren, X.-L., Zhang, Q.-M., Zhang, Y.-C., Zhou, T.: Vital nodes identifi-
cation in complex networks. Physics Reports 650, 1-63 (2016)
13. Gupta, N., Singh, A., Cherifi, H.: Community-based immunization strategies for epidemic
control. In: Communication Systems and Networks (COMSNETS), 2015 7th International
Conference On, pp. 1-6 (2015). IEEE
14. Gong, K., Tang, M., Hui, P.M., Zhang, H.F., Younghae, D., Lai, Y.-C.: An efficient im-
munization strategy for community networks. PloS one 8(12), 83489 (2013)
15. Zhao, Z., Wang, X., Zhang, W., Zhu, Z.: A community-based approach to identifying in
uential spreaders. Entropy 17(4), 2228-2252 (2015)
16. Nadini, M., Sun, K., Ubaldi, E., Starnini, M., Rizzo, A., Perra, N.: Epidemic spreading in
modular time-varying networks. Scientific reports 8(1), 2352 (2018)
17. Chan, S.Y., Leung, I.X., Li`o, P.: Fast centrality approximation in modular networks. In:
Proceedings of the 1st ACM International Workshop on Complex Networks Meet Infor-
mation & Knowledge Management, pp. 31-38 (2009). ACM
18. Yoshida, T., Yamada, Y.: A community structure-based approach for network immuniza-
tion. Computational Intelligence 33(1), 77-98 (2017)
19. Gupta, N., Singh, A., Cherifi, H.: Centrality measures for networks with community struc-
ture. Physica A: Statistical Mechanics and its Applications 452, 46-59 (2016)
20. Ghalmane, Z., El Hassouni, M., Cherifi, H.: Identifying in uential spreaders of epidemics
on community networks. arXiv preprint arXiv:1806.05637 (2018)
21. Luo, S.-L., Gong, K., Kang, L.: Identifying in uential spreaders of epidemics on community
networks. arXiv preprint arXiv:1601.07700 (2016)
22. Kitromilidis, M., Evans, T.S.: Community detection with metadata in a network of bi-
ographies of western art painters. arXiv preprint arXiv:1802.07985 (2018)
23. Lancichinetti, A., Fortunato, S., Radicchi, F.: Benchmark graphs for testing community
detection algorithms. Physical review E 78(4), 046110 (2008)
24. Orman, G.K., Labatut, V., Cherifi, H.: Towards realistic artificial benchmark for commu-
nity detection algorithms evaluation. International Journal of Web Based Communities
9(3), 349-370 (2013)
25. Orman, G.K., Labatut, V., Cherifi, H.: On accuracy of community structure discovery
algorithms. arXiv preprint arXiv:1112.4134 (2011)
26. Orman, G.K., Labatut, V., Cherifi, H.: Comparative evaluation of community detection
algorithms: a topological approach. Journal of Statistical Mechanics: Theory and Experi-
ment 2012(08), 08001 (2012)
27. Leskovec, J., Mcauley, J.J.: Learning to discover social circles in ego networks. In: Advances
in Neural Information Processing Systems, pp. 539-547 (2012)
26
Zakariya Ghalmane et al.
28. Traud, A.L., Mucha, P.J., Porter, M.A.: Social structure of facebook networks. Physica A:
Statistical Mechanics and its Applications 391(16), 4165-4180 (2012)
29. Moreno, Y., Pastor-Satorras, R., Vespignani, A.: Epidemic outbreaks in complex hetero-
geneous networks. The European Physical Journal B-Condensed Matter and Complex
Systems 26(4), 521-529 (2002)
30. Wang, W., Liu, Q.-H., Zhong, L.-F., Tang, M., Gao, H., Stanley, H.E.: Predicting the
epidemic threshold of the susceptible-infected-recovered model. Scientific reports 6, 24676
(2016)
|
1710.07231 | 1 | 1710 | 2017-10-19T16:22:27 | Modeling Graphs Using a Mixture of Kronecker Models | [
"cs.SI",
"physics.soc-ph"
] | Generative models for graphs are increasingly becoming a popular tool for researchers to generate realistic approximations of graphs. While in the past, focus was on generating graphs which follow general laws, such as the power law for degree distribution, current models have the ability to learn from observed graphs and generate synthetic approximations. The primary emphasis of existing models has been to closely match different properties of a single observed graph. Such models, though stochastic, tend to generate samples which do not have significant variance in terms of the various graph properties. We argue that in many cases real graphs are sampled drawn from a graph population (e.g., networks sampled at various time points, social networks for individual schools, healthcare networks for different geographic regions, etc.). Such populations typically exhibit significant variance. However, existing models are not designed to model this variance, which could lead to issues such as overfitting. We propose a graph generative model that focuses on matching the properties of real graphs and the natural variance expected for the corresponding population. The proposed model adopts a mixture-model strategy to expand the expressiveness of Kronecker product based graph models (KPGM), while building upon the two strengths of KPGM, viz., ability to model several key properties of graphs and to scale to massive graph sizes using its elegant fractal growth based formulation. The proposed model, called x-Kronecker Product Graph Model, or xKPGM, allows scalable learning from observed graphs and generates samples that match the mean and variance of several salient graph properties. We experimentally demonstrate the capability of the proposed model to capture the inherent variability in real world graphs on a variety of publicly available graph data sets. | cs.SI | cs | Modeling Graphs Using a Mixture of Kronecker Models
Suchismit Mahapatra
Computer Science and Engineering
State University of New York at Buffalo
[email protected]
Varun Chandola
Computer Science and Engineering
State University of New York at Buffalo
[email protected]
privacy [4]. Applications include understanding malware
propagation in social networks [19], understanding fraud
in healthcare [5], etc. Graph generative models also allow
researchers to produce realistics simulations at desired scale
which are vital to understand issues such as handling scal-
ability challenges and modeling temporal evolution.
7
1
0
2
t
c
O
9
1
]
I
S
.
s
c
[
1
v
1
3
2
7
0
.
0
1
7
1
:
v
i
X
r
a
though stochastic,
Abstract -- Generative models for graphs are increasingly
becoming a popular tool for researchers to generate realistic
approximations of graphs. While in the past, focus was on
generating graphs which follow general
laws, such as the
power law for degree distribution, current models have the
ability to learn from observed graphs and generate synthetic
approximations. The primary emphasis of existing models
has been to closely match different properties of a single
observed graph. Such models,
tend to
generate samples which do not have significant variance in
terms of the various graph properties. We argue that in
many cases real graphs are sampled drawn from a graph
population (e.g., networks sampled at various time points,
social networks for individual schools, healthcare networks for
different geographic regions, etc.). Such populations typically
exhibit significant variance. However, existing models are not
designed to model this variance, which could lead to issues
such as overfitting. We propose a graph generative model that
focuses on matching the properties of real graphs and the
natural variance expected for the corresponding population.
The proposed model adopts a mixture-model strategy to expand
the expressiveness of Kronecker product based graph models
(KPGM), while building upon the two strengths of KPGM,
viz., ability to model several key properties of graphs and to
scale to massive graph sizes using its elegant fractal growth
based formulation. The proposed model, called x-Kronecker
Product Graph Model, or xKPGM, allows scalable learning
from observed graphs and generates samples that match the
mean and variance of several salient graph properties. We
experimentally demonstrate the capability of the proposed
model to capture the inherent variability in real world graphs
on a variety of publicly available graph data sets.
I. INTRODUCTION
Data occurs as graphs and networks in a wide variety of
applications, ranging from social sciences to biology. Graph
analysis methods are required to understand the structural
properties of graphs. One important class of graph analysis
methods deal with finding generative mechanisms and mod-
els that generate graphs with such structural properties. A
crucial application of these models is to generate synthetic
graphs which "match" the structural properties of real world
graphs. Given the limited availability of real world graph
data, mainly because of high cost and privacy concerns, such
synthetic (and anonymized) graphs are a valuable resource
for researchers to understand network behavior in domains
such as systems analysis (e.g., the Internet), bioinformatics,
and social sciences, while allowing for anonymity and
What should be the salient characteristics of such a
generative model? First, it should be able to capture the
properties of real-world graphs. Traditionally the focus has
been on general "laws" that are expected to be obeyed
by real world graphs, such as power laws for degree
distributions, small diameters, communities, etc. However,
to accurately represent the real world graph, matching on
local graph properties such as edges, transitive triangles,
etc.,
the model should be able to
scale to massive graph sizes, to be applicable in domains
where graphs tend to be big. Third, to allow generation
(simulation) of any sized synthetic graphs, the model should
be parametric and should allow learning of the parameters
from one or more observed graphs.
is important. Second,
Table I compares several existing graph generative models
in terms of which of the above characteristics. Note that
several models satisfy all three characteristics. However, we
argue that there is an additional characteristic that needs to
be incorporated into the generation models -- the ability to
model the natural variance in a population of graphs [14].
Real world graphs can be thought of as being generated
from a natural process. Graph models emulate this process to
generate similar synthetic graphs. We argue that graphs that
are generated by the same process exhibit a natural variance
in terms of the structural properties, and hence synthetic
graphs generated from a model should mirror the same
variance. Examples of such populations include, graphs
collected at different times, social networks for different
groups of people (e.g., schools), healthcare networks for
different spatial regions, etc. Moreno et al., empirically
demonstrated such variance in several social and email net-
work graph populations [14]. Figure 1 illustrates the variance
in the degree power law coefficient (α) for a population of
Autonomous Systems (AS) graphs. Observe that α exhibits
bounded variance and the empirical distribution is close to
the normal distribution. Similar behavior is observed for
many other graph properties (See Figure 2).
1. Learnable
2. Scalable Learning
3. Scalable Generation
4. Match Local Properties
5. Capture Variance
PA [3]
×
--
×
×
×
ERGM [25]
√
√
×
×
×
CL [1]
See note
√
--
×
×
BTER [20]
See note
√
--
√
×
KPGM [12]
√
√
√
√
×
mKPGM [16]
√
×
√
√
√
xKPGM
√
√
√
√
√
Table I: Comparison of existing graph generation models with the proposed xKPGM model. PA - Preferential Attachment,
ERGM - Exponential Random Graph Models, CL - Chung-Lu, BTER- Block Two Erdos-R´enyi, KPGM- Kronecker Product
Graph Model, mKPGM- Mixed KPGM, xKPGM- Proposed Model. Note: The CL and BTER models can be used to generate
a synthetic graph from a given real graph with approximately the same number of nodes, however it does not allow generation
of arbitrary sized synthetic graphs.
in the generative model. The tied-KPGM (tKPGM) and
the mixed-KPGM (mKPGM) are two variants proposed
to alleviate the issues with KPGM [14]. However, as we
demonstrate experimentally, these next generation models
are also not expressive enough to model the natural variance
in the data. Moreover,
the estimation process using the
simulated method of moments approach is expensive and not
scalable to learning from massive graphs.
To alleviate the lack of variation in existing models, we
propose a mixture based approach for graph generation
using Kronecker product. In particular, this paper makes the
following specific contributions:
Figure 1: Illustration of the variance in power law coefficient
(α) for a population of over 700 Autonomous Systems (AS)
graphs sampled at different time points. Few anomalous
samples were removed from the original population.
In summary, the research community needs a scalable
graph generative model which matches properties of real
graphs, including the variance exhibited by a population of
graphs. However, as shown in Table I, most existing graph
generators lack in at least one of the desired characteristics
listed above.
In particular,
the Kronecker Product Graph Model
(KPGM) [12] has been shown to satisfy some of the above
properties. The multiplicative nature of the model allows for
fast sampling of massive sized graphs and has been shown,
both analytically and empirically, to generate graphs that fol-
low many power-law characteristics for several global graph
properties. Moreover, a method of moments based approach
was proposed for parameter learning with KPGM [7], which
allows it to scale to massive graphs. However, recent papers
have identified few shortcomings of KPGM.
The single biggest argument made against KPGM is
that they lack the ability to capture the natural variabil-
ity observed in real world graphs [14], [21]. Researchers
have shown that the synthetic graphs sampled from KPGM
show little variation in terms of several graph properties.
The problem is attributed to the fractal nature of growth
1) We propose a mixture-model based Kronecker product
graph model (xKPGM). We also show that the existing
models are particular instantiations of the proposed
xKPGM model.
2) We analytically examine the expected properties of the
graph generated by the proposed model.
3) We derive expressions for the expectactions for several
salient graph properties for our model. Using these
expressions, we provide a method of moments based
parameter learning algorithm for the proposed model.
4) We propose a random subgraph based method to eval-
uate generative models for graphs.
5) We show, analytically and empirically, that xKPGM
matches the real graph properties and captures the nat-
ural variability in graphs more effectively than existing
Kronecker models and other graph generative models.
II. RELATED WORK
Generative models for graphs has been a widely studied
field for many decades. One of the earliest ones is the
Erdos-R´enyi model [18] which has obvious shortcomings
as it fails to capture properties of real-world graphs. More
recent models have focused on the preferential attachment
property of nodes, i.e., new nodes tend to form edges with
nodes with greater degree [3]. However, in such models,
the graph is grown one node at a time, which makes them
inherently serial and unscalable. Lately, there has been an
emphasis on models which can be learnt from observed
graphs. These include the Exponential Random Graph
Models (ERGM) (also referred to as p* models) [25]. ERGM
essentially defines a log linear model over all possible graphs
G, p(Gθ) ∝ exp(θT s(G), where G is a graph, and s is a
set of functions, that can be viewed as summary statistics for
the structural features of the network. Another popular and
well-known network models are Stochastic Block Models
(SBM) [24] in which each node belongs to a cluster and the
relationships between nodes are determined by their cluster
membership. While conventional SBM are defined for non-
overlapping community assignments, many overlapping or
mixed variants have also been introduced [2]. Another
relevant model is the Chung-Lu (CL) model [1] in which
the probability of an edge is proportional to the product
of the degrees of its end vertices. While CL model effec-
tively captures the degree distribution, it performs poorly
for other properties such as clustering coefficient. A recent
extension to CL model, called the Block Two-Level Erdos-
R´enyi (BTER) model [20] is shown to match both the degree
distribution and clustering coefficient on several graphs.
However the BTER model is not truly generative as it only
allows for creation of a synthetic graph which is exactly the
same size as the observed graph. For generating arbitrary
sized graphs one needs to provide parameters instead of
learning them.
Kronecker product based generative models are increas-
ingly becoming popular [12]. However, given their limited
expressiveness [14], [21], several variants have been pro-
posed [14], [15]. The most promising extension, mKPGM, is
able to capture the variance in a graph population, however,
the parameter estimation phase is expensive. We will be
discussing the original model and two variants in more detail
in the paper. There also have been other related papers
that improve KPGM in other ways, e.g., the Multiplicative
Attribute Graph Models [11].
III. BACKGROUND
In this section we introduce the various Kronecker product
based models that have been proposed for modeling large
graphs. For clarity, we will use the notation used in this
section for the subsequent sections of the paper.
A. Kronecker Product Graph Model
KPGM [12] is a fractal growth model. A graph G with N
nodes is obtained by first generating a N × N stochastic
matrix (containing entries between 0 and 1). The edge
between node i and j is independently (from other edges)
generated using a Bernoulli trial with the ijth entry from
the matrix. The generative algorithm starts with an initial
matrix P1 = Θ, which is b × b; typically b is set to 2 or 3,
e.g.:
(cid:34)
(cid:35)
P1 = Θ =
θ11
θ21
θ12
θ22
(1)
The algorithm takes repeated Kronecker product of P1 with
itself to generate a larger matrix. For example,
Pn = P1 ⊗ P1 ⊗ . . . ⊗ P1
(cid:123)(cid:122)
n times
(cid:125)
(cid:124)
(2)
(1−Pn[σ(u), σ(v)])
Note that the matrix Pn will have bn rows (and columns).
This matrix is then used to generate a graph G = (V, E)
with nodes V = {1, 2, . . . , N}. For each pair (u, v), where
1 ≤ u, v ≤ bn, a Bernoulli sample is generated with
parameter Pn[u, v]. If the sample is 1 (success), edge (u, v)
is added to E. We denote this process as a realization of Pn
to get the adjacency matrix, A (= R(Pn)).
The KPGM model can assign a probability to a given
graph G as long as the correspondence of nodes in G to
the rows of Pn are given (denoted by σ). Given the KPGM
model (P1 = Θ) and the correspondence function σ, the
probability of an observed graph G = (V, E) is given by:
P (GΘ, σ) =
Pn[σ(u), σ(v)]
(cid:89)
(cid:89)
(u,v)∈E
(u,v) /∈E
where Pn is derived from Θ using (2).
(3)
Given an observed graph G, two approaches for learning
the model parameters (Θ) have been proposed. The first is
an MLE approach that approximates computes the gradient
of the log-likelihood of the graph (See (3)) [12]. Since
the correspondence function σ is unknown, the learning
algorithm searches over the factorial possible permutations
using Metropolis-Hastings sampling and then uses a gradient
descent approach to update the parameters in Θ. The second
learning approach is based on method of moments [7]. This
method avoids the issue of searching over the factorial space
of permutations.
It has been shown that KPGM generates graphs which
match real networks in terms of several properties such
as skewed degree distribution, short network path length,
etc. Moreover, the model allows very fast sampling of large
graphs (O(E)). However, recent work has identified several
limitations with KPGM. Seshadri et al. [21] have shown
that graphs generated from KPGM have 50-75% isolated
vertices. Moreno et al. [14] have observed, both analytically
that graphs generated from KPGM do
and empirically,
not capture the variability observed in the real networks
expected to be generated from the same source. They have
attributed this shortcoming to: (i) use of independent edge
probabilities, (ii) fractal expansion, and (iii) small number of
model parameters. In fact, after several Kronecker products,
most entries in the eventual stochastic matrix tend to become
homogeneous. This results in a lack of variance since all
sampled graphs appear similar.
Moreno et al. [14] investigated several simple approaches
to induce variability in KPGM generated graphs. These
include using larger initiator matrices and sampling Θ from
a distribution instead of using a point estimate. None of
these variations induced significant variability in the sampled
graphs. The same authors make the following key observa-
tion:
Observation 1: For a graph G = (V, E) generated
by KPGM, V ar(E) ≤ E[E], independent of n.
The authors empirically observed that in real networks, the
estimated variance of the number of edges across multiple
variations of the same graph (across time) is significantly
greater than the mean. Thus KPGM cannot generate graphs
with large variance.
B. Tied Kronecker Product Graph Model
Moreno et al. [14] proposed a generalization of KPGM
that allows larger variance in the properties of the graphs
sampled from the model by inducing edge dependence
in the generation process. In the generative process, an
adjacency matrix is realized after each iteration from the
current matrix Pt, denoted as R(Pt). The next stochastic
matrix, Pt+1 is obtained by performing a Kronecker product
of the realized adjacency matrix and the initiator matrix,
i.e., Pt+1 = R(Pt) ⊗ P1. Note that for KPGM, only one
realization is done (after the nth iteration). Thus, the final
adjacency matrix can be obtained using recursive realizations
and Kronecker products, i.e.,
A = R(. . . R(R(P1) ⊗ P1) . . .)
(4)
(cid:124)
(cid:123)(cid:122)
n realizations
(cid:125)
The above model is called tied KPGM or tKPGM since the
realization at every step can also be thought of as tying
the Bernoulli parameters. The authors show analytically that
the variance for the expected number of edges in graphs
generated using tKPGM is higher than KPGM, in fact the
variance was higher than desired which motivated the variant
discussed next.
C. Mixed Kronecker Product Graph Model
The mixed KPGM or mKPGM variation allows the graph
to grow as KPGM (without realizations) for first l steps and
then acts as tKPGM for remaining n− l steps. mKPGM has
an additional parameter l which denotes the extent of tying
in the model. Note that for l = 1, mKPGM is equivalent to
tKPGM and for l = n, mKPGM is equivalent to KPGM.
IV. NATURAL VARIABILITY IN REAL GRAPHS
A vital requirement for any graph generative model is the
ability of the model to capture the variability across multiple
observed samples. In fact, both tKPGM and mKPGM were
motivated due to the limited ability of previous models to
capture this variability. Most generative graph models have
focused on capturing the properties observed for a single
instance. But such models are not equipped to model a set
of graphs, assumed to be sampled from the same statistical
distribution. In many real world applications, one can obtain
Name
as [23]
ca-astroPh [23]
elegans [6]
hep-ph [23]
netscience [17]
protein [10]
Description
CAIDA AS Relationship
Graph
Collaboration
network
of Arxiv Astro Physics
C.
elegans metabolic
network
Citation network from
Arxiv HEP-PH
Coauthorship
of scientists
Protein interaction net-
work for Yeast
network
Nodes
Edges
6,474
13,233
18,772
396,160
453
4,596
34,546
421,578
1,589
5,484
1,870
4,480
Table II: Publicly available graph data sets used in the
evaluation. More details available at http://www.cse.buffalo.
edu/∼chandola/research/bigdata2015graphs/info.html.
multiple samples of graphs which would exhibit variance in
the different properties. For example, a set of social networks
of students for different colleges, or a set of citation networks
collected for different decades or disciplines. We show one
example in Figure 1 where we consider daily AS networks
across several years. In general, obtaining such populations
to study the variance is challenging.
In this paper, we address this challenge by generating
random subgraphs of a given graph. For a given large graph,
the idea is to extract subgraphs using a sampling strategy
such that the properties of the original graph are preserved
in the subgraph. Several such methods have been proposed in
the literature [9], [13], [22]. In this paper we use a variant of
the forest fire model to generate subgraphs. In this method, a
random node is chosen and the neighborhood of the chosen
node is traversed in a breadth first approach. Each outgoing
edge from the node is added to the sample with a certain
probability. The nodes at the end of the chosen edges are
then further "burnt" and the "fire" is further spread.
To illustrate the variability we consider several publicly
available real world graphs listed in Table II. For each graph
in Table II, we sample 200 subgraphs (number of nodes were
typically 4 times less than the original number of nodes)
using the forest-fire approach and measure characteristics
of each subgraph. The empirical distribution for several
characteristics are shown in Figure 2. We study the following
five important characteristics: (i) power law coefficient (α) of
the degree distribution (ii) number of edges, (iii) number of
triangles, (iv) average path length, and (v) average clustering
coefficient. The empirical distributions show that for most
of the characteristics, the population of subgraphs exhibit
significant variation. Moreover, the variation itself is not
completely arbitrary, but in most cases, resembles standard
distributions such as Gaussian or exponential.
V. PROPOSED xKPGM MODEL
Existing Kronecker product based models are inadequate
in capturing the natural variance that exists in graph popu-
·10−3
1.5
1
0.5
as
·10−4
1.5
1
0.5
2
2.5
3
3.5
4
1,800 2,000 2,200 2,400 2,600 2,800 3,000 3,200
0.2
0.4
Degree Coefficient
Edges
0.6
0.8
Triangles
1
1.2
·104
·10−5
ca-astroPh
·10−6
4
3
2
1
0
1
0.5
0
2
2.2
2.4
2.6
Average Path Length
4
2
0
2
1.5
1
0.5
0
4
6
Degree Coefficient
5
2.4
2.8
2.6
3.2
Average Path Length
3
7
3
4
5
6
Edges
7
8
·104
0.5
1.5
1
Triangles
2
·106
2.2
·10−3
8
6
4
2
0
elegans
·10−3
1
0.5
0
40
30
20
10
0
3
3.5
4.5
Degree Coefficient
4
5
100
150
200
250
Edges
300
350
0
500
1,000
1,500
Triangles
2,000
2,500
1.78
1.8
1.82
1.84
1.86
Average Path Length
·10−4
hep-ph
·10−5
4
3
Degree Coefficient
5
6
2
1.5
1
0.5
0
2
1
0
1.5
1
0.5
0
0.4
0.6
0.8
1
Edges
1.2
1.4
·104
0
0.5
1
1.5
Triangles
2
2.5
·105
·10−2
netscience
·10−3
3
2
1
0
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
s
e
l
p
m
a
S
f
o
n
o
i
t
c
a
r
F
0.8
0.6
0.4
0.2
0.6
0.4
0.2
0
1
0.5
0
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1.2
1
0.8
0.6
0.4
0.2
0
3
2
2.5
3
3
3
2
1
0
30
20
10
0
0.5
3.4
0.3
0.2
Average Clustering Coefficient
0.4
0.6
0.5
0.7
0.54
0.52
0.6
Average Clustering Coefficient
0.56
0.58
0.5
0.4
0.9
Average Clustering Coefficient
0.6
0.7
0.8
0.3
0.25
Average Clustering Coefficient
0.35
0.45
0.4
0.5
0.55
0.7
0.76
Average Clustering Coefficient
0.72
0.74
4
3
2
1
0
10
8
6
4
2
0
30
20
10
0
20
15
10
5
3
2.5
4
Average Path Length
3.5
1
0.8
0.6
0.4
0.2
0
2
0.8
0.6
0.4
0.2
0
1,400 1,500 1,600 1,700 1,800 1,900 2,000
1
Triangles
3
2
Average Path Length
4
5
5
4
Degree Coefficient
6
7
560
580
600
Edges
620
·10−2
1.5
1
0.5
protein
·10−3
8
6
4
2
1.5
1
0.5
4
6
Degree Coefficient
5
7
300
320
340
Edges
360
380
50
100
150
Triangles
200
4
4.5
5
5.5
Average Path Length
4 · 10−26 · 10−28 · 10−2
Average Clustering Coefficient
0.12
0.1
0.14
Figure 2: Empirical distribution of graph characteristics for the graphs listed in Table II. The histogram is obtained by taking
multiple random sub-graph samples and measuring five properties of the sub-graph.
lations. To address this issue, we propose a mixture-model
based approach which allows capturing this variance. We
call the proposed model xKPGM, where the 'x' signifies that
all variants of KPGM discussed in Section III are specific
instantiations of xKPGM.
The key idea behind xKPGM is to use k (k ≥ 1) initiator
matrices of possibly different sizes. The graph generation
follows similar iterative expansion as seen in other models
discussed in Section III with the difference that at any itera-
tion one of the k initiator matrices is randomly chosen for the
Kronecker product by drawing a sample from a Multinomial
distribution parameterized by a k length probability vector
i πi = 1). Similar to mKPGM, there is a parameter
l (1 ≤ l ≤ k) such that for first l iterations no realization
(or tying) is performed while from iteration l + 1, parameter
tying through realizations is employed.
π ((cid:80)k
To illustrate the impact of using multiple initiator matrices
instead of a single matrix, we generate multiple synthetic
graphs using different generative models. We use the fol-
lowing two initiator matrices:
(cid:33)
(cid:32)
Θ1 =
0.95 0.60
0.60 0.20
Θ2 =
0.99 0.20
0.20 0.70
(5)
(cid:32)
(cid:33)
These exact matrices have been used in previous papers [14]
for conducting empirical simulations for mKPGM and other
variants. The mean and the variance for the number of edges
across the simulations is shown in Figure 3 which illustrates
that the variability (in terms of the number of edges) for
samples generated by mKPGM is low for higher values of
l (≥ 6). In fact, the variance becomes negligible for KPGM
(l = 10). For very low values of l the variability is very
high. On the other hand, xKPGM generates samples with
much more stable variability across l. We observe that for
xKPGM, the variance is generally higher when π = 0.5,
though it also depends on the individual initiator matrices.
We observe similar behavior for other properties (clustering
coefficient, etc.) as well.
A. xKPGM Sampling Algorithm
The steps for sampling a graph in xKPGM are shown in
Algorithm 1. Note that the algorithm needs following inputs:
k initialization matrices, Θ1, Θ2, . . . , Θk, a k length vector
π, the number of iterations n, and the level of tying l. The
number of rows (and columns) for each matrix Θi is denoted
as bi.
B. Interchangeability in Kronecker Products
We make use of the following important result [8] for
Kronecker matrices. For two matrices A and B:
A ⊗ B = M (B ⊗ A)N
(6)
where M and N are permutation matrices which merely
change the order of rows and columns, respectively. In other
words, the above result means that the Kronecker product
2
s
e
g
d
E
f
o
r
e
b
m
u
N
e
g
a
r
e
v
A
s
e
g
d
E
f
o
r
e
b
m
u
N
e
g
a
r
e
v
A
6,000
4,000
2,000
8,000
6,000
4,000
2,000
0
2
mKPGM (Θ1)
mKPGM (Θ2)
4
8
Tying Parameter l
6
10
xKPGM (π = 0.5)
xKPGM (π = 0.8)
4
8
Tying Parameter l
6
10
Figure 3: Variation of number of edges for synthetic graphs
with 210 nodes generated by mKPGM and xKPGM for
varying levels of tying (l). Note that l = 1 indicates purely
tied model (tKPGM) and l = 10 indicates purely untied
model (KPGM). The error bars are computed by generating
2000 samples for each value of l.
is commutative as long as we allow changing of labels of
nodes, which is irrelevant in our case. Since in Algorithm 1,
the order of rows and columns is not important, one can use
this result in the purely untied setting.
Observation 2: In Algorithm 1, the graph generated
using any arbitrary sequence of initialization matrices
is equivalent to the following canonical sequence:
(Θ1 ⊗ Θ1 . . .)
⊗ . . . ⊗ (Θk ⊗ Θk . . .)
⊗ (Θ2 ⊗ Θ2 . . .)
(cid:125)
(cid:124)
(cid:123)(cid:122)
(cid:125)
(cid:124)
(cid:123)(cid:122)
nk times
(cid:125)
(cid:124)
(cid:123)(cid:122)
n1 times
n2 times
The above observation allows us to arrive at several im-
portant results later, because it simplifies the analysis, es-
pecially, in the untied setting for xKPGM, where arbitrary
sequences of initialization matrices can be represented using
the above canonical sequence. In fact, the next result follows
from this observation.
C. Generating Graphs of Arbitrary Sizes
i
.
i=1 bnπi
The input n determines the eventual size of the graph in
terms of the number of nodes (V). For a simple case where
all initialization matrices are of the same size (bi = b,∀i),
the size of the graph will be bn. In general, expected number
of nodes will be(cid:81)k
Note that existing Kronecker based models are limited to
generating graphs with number of nodes as exact powers of
b (typically set to 2 or 3). On the other hand, xKPGM can
generate graphs with number of nodes, which in principle,
need not be restricted to the exact powers of b. Thus, by
using initialization matrices of varying sizes, xKPGM can
generate graphs with more number of possibilities for the
number of nodes. For example, if we use two initialization
matrices of sizes 2 and 3 and if N can be factorized as 2a3b,
a+b and generate a
then one can set π1 = a
graph using Algorithm 1 with N expected number of nodes.
a+b and π2 = b
Algorithm 1 Graph Generation Algorithm for xKPGM
Input: Θ1, Θ2, . . . , Θk, π, n, l
Output: Adjacency matrix A
1 P ← 1
2 foreach t = 1 to l do
3
// Untied Phase
i ∼ M ultinomial(π)
P ← P ⊗ Θi
// Tied Phase
4 foreach t = l + 1 to n do
A ← R(P);
5
i ∼ M ultinomial(π)
6
P ← A ⊗ Θi
7 A ← R(P)
return A
// R - Realization
D. Relationship to Existing Models
If k = 1 and l = 1, xKPGM is equivalent to tKPGM.
Similarly, for k = 1 and l = n, xKPGM is equivalent
to KPGM. For k = 1 and for all other values of l
(1 ≤ l < n) xKPGM is equivalent to mKPGM. Thus all
existing Kronecker product based graph models are specific
instantiations of xKPGM.
Kronecker product models have been analyzed in terms
of the expected number of edges for the generated samples.
In Section VI, we provide similar expressions for other
graph properties as well. In Table III, we list the expressions
for the expectation and variance for the number of edges
for xKPGM and compare them with other models. The
derivation is omitted in the interest of space though the
key is to use the Observation 2 along with the fact that
the expectation of an edge between a pair of nodes is equal
to the corresponding value in the final stochastic matrix.
VI. PROPERTIES OF xKPGM GRAPHS
In this section we provide expressions for expectations
of four key properties for graphs, viz., i) edges, ii) hairpins
(or 2-stars), iii) tripins (or 3-stars), and iv) triangles. These
four properties provide an understanding of local structure of
the samples. Moreover, we will be using the expressions to
estimate the model parameters in Section VII. For simplicity
we assume that all k initiator matrices are of size 2 × 2
and are of the following form: Θk =
. The
(cid:32)
(cid:33)
ak
bk
bk
ck
derivation for expected values in the purely untied (l = n)
setting closely follows the line of argument taken for pure
KPGM [7] (Section 3) and makes use of Observation 2. The
expressions are provided in Table IV.
Note that the expression for E[E] here is slightly different
from a similar expression in Table III. The reason is that in
Table IV we do not count the self-loops and count each edge
only once.
VII. PARAMETER ESTIMATION
The parameters of xKPGM are the elements in the k
initiator matrices and the vector π = π1, π2, . . . , πk. We
describe a method of moments based approach to estimate
these parameters from a given observed graph. The graph
is characterized using a set of statistics (or moments). In
this paper we have used the same moments as used in
the past for KPGM estimation (edges, hairpins, tripins, and
triangles) [7]. The idea is to find the model parameters such
that the expected moments for the model match closely
with the moments computed from the observed graph within
an optimization procedure. Each moment is denoted as Fi
and the computed moment for an observed graph G is
denoted as F ∗
i . The estimation method searches for the
parameters Θ1, Θ2, . . . , π which minimize the following
objective function:
f (Θ, F∗) =
wi
F(cid:88)
i=1
(cid:18) F ∗
i − E[FiΘ]
F ∗
i
(cid:19)2
(7)
where wi indicates a weight assigned to the ith moment. In
this paper we have used equal weights, however, they can be
used to assign more emphasis on certain moments. Note that
the above objective function can be easily extended to learn
from multiple graphs by replacing the observed moment Fi
in (7) with the average moment value across all p observed
graphs, i.e., (cid:101)Fi = 1
p
(cid:80)G FiG.
For xKPGM in a purely untied setting (l = n), we have
expressions for each of the moments as functions of the
parameters (See Table IV). We plugin these expressions into
a gradient based optimizer (e.g., fmincon function available
in Matlab) to get the optimal parameter values.
Note that the above estimation algorithm can be extended
to more general settings of purely untied xKPGM such as
Model
KPGM
tKPGM
mKPGM
xKPGM (purely tied)
E[E]
Sn
Θ
Sn
Θ
Sn
Θ
(π1SΘ1 + π2SΘ2 )n
xKPGM (untied)
Sπ1n
Θ1
Sπ2n
Θ2
var[E]
Θ − Sn
Sn
Θ2
Θ − 1)
Sn−1
Θ (Sn
Θ − 1)
Sn−1
Θ (Sn−l
(π1(SΘ1 − SΘ2
− Sπ1n
SΘ2 )2)(cid:80)2(n−1)
j=1
Sπ2n
Θ2
Sπ1n
Θ1
Θ2
1
SΘ−SΘ2
SΘ−1
SΘ−SΘ2
Θ − Sl
SΘ−1 + (Sl
) + π2(SΘ2 − SΘ2
(π1SΘ1 + π2SΘ2 )j
1
2
Sπ2n
Θ2
2
Θ2 )S2(n−l)
) + π1π2(SΘ1 −
Θ
1
2
and SΘ2
denote the sum of squares of values of Θ, Θ1 and Θ2.
Table III: Comparison of E[e] and var[e] for Kronecker models. SΘ, SΘ1 and SΘ2 denote the sum of values of Θ, Θ1 and
Θ2. SΘ2, SΘ2
2E[E] = (cid:81)n
2E[H] = (cid:81)n
6E[T] = (cid:81)n
− 3(cid:81)n
+ 5(cid:81)n
6E[∆] = (cid:81)n
i=1(ai + 2bi + ci)πin −(cid:81)n
i=1((ai + bi)2 + (bi + ci)2)πin − 2(cid:81)n
i=1((ai + bi)3 + (bi + ci)3)πin − 3(cid:81)n
i (ai + ci))πin + 4(cid:81)n
i )πin − 3(cid:81)n
i=1(ai(ai + bi) + ci(ci + bi))πin −(cid:81)n
i ))πin − 6(cid:81)n
i ))πin + 2(cid:81)n
i=1(ai(ai + bi)2 + ci(bi + ci)2)πin
i + 2b3
i )πin + 2(cid:81)n
i )πin + 2(cid:81)n
i + c3
i + c3
i + 3b2
i + c2
i + bi(a2
i + b2
i (ai + ci) + c3
i=1(a3
i=1(ai(a2
i )πin
i + b3
i (ai + ci) + 2b3
i=1(ai + ci)πin
i + b2
i ) + ci(b2
i + c2
i=1(a3
i + c2
i + c3
i + bi(a2
i )πin
i=1(a3
i=1(a2
i + 2b2
i + c2
i=1(a2
i + c2
i )πin
i=1(a3
i=1(a3
i ) + b2
i=1(a3
i=1(a3
i )πin
i + c3
i + c3
Table IV: Expressions for E - num. edges, H - num. hairpins, T - num. tripins and ∆ - num. of triangles
using initiator matrices of larger sizes (> 2 and using initia-
tor matrices with different sizes, by deriving the appropriate
expressions similar to the ones listed in Table IV. For the
tied setting (l < n), obtaining such closed form expressions
is challenging. However, one could use the simulated method
of moments approach which was originally proposed for
mKPGM [16].
IX. EXPERIMENTAL RESULTS
We present results on several publicly available network
data sets (See Table II). The evaluation has two objectives.
We first show that the parameters learnt for xKPGM using
the algorithm sketched in Section VII allow a more accurate
modeling of the observed graph, compared to other Kro-
necker based models.
VIII. COMPUTATIONAL COMPLEXITY
A. Parameter Estimation
Generating a graph of size N using the learnt initiator
matrices using Algorithm 1 requires n = log2 N Kronecker
products and l
intermediate realization followed by N 2
bernoulli trials for the final realization. In the original KPGM
paper [12], the authors propose a sampling strategy which
is linear in the number of edges, however, that method
requires one to specify the number of edges desired in the
graph. For xKPGM the generation is dominated by the final
O(N 2) realization step. However, given that the realizations
are independent, efficient distributed implementations can be
devised.
The parameter estimation using the method of moments
allows scalable training. The only compute intensive phase
is the computation of the moments for the observed graph,
which is done once. The objective function can be computed
in O(1) time, independent of the size of the observe graph.
In contrast,
the simulated method of moments, used by
mKPGM, requires computation of the moments for the
generated graph multiple times at each iteration of the
gradient descent, making it unscalable to large graphs. The
number of iterations required by the fmincon minimizer will
depend on the number of model parameters, which increases
linearly with the number of initiator matrices (k).
For each graph in Table II, we estimate parameters using
Matlab fmincon with 5000 initial values with constraints that
elements of the initiator matrices are between 0 and 1 and
ai > ci. We train xKPGM with two initiator matrices of
size 2 × 2 and other models (KPGM, tKPGM, mKPGM)
using one initator matrix of size 2 × 2, using the objective
function in (7) and the four moments listed in Table IV. For
mKPGM and tKPGM, we use implementation provided by
the authors [16]1. For KPGM, we used the direct optimiza-
tion method [7].
B. Matching Moments
The objective function values (see (7)) for the different
methods are shown in Figure 4. For every data set, xKPGM
gives the best estimation in terms of the objective function,
especially for ca-astroPh, netscience, protein, and hep-ph
data sets. The relative performance of other models varies
across the different data sets. In general, we observe that
the objective function value improves as the tying level (l)
decreases and is best when l = 1, i.e., tKPGM.
1https://www.cs.purdue.edu/homes/smorenoa/mKPGM.zip
KPGM tKPGM mKPGM xKPGM
2.5
2
1.5
1
0.5
0
n
o
i
t
c
n
u
F
e
v
i
t
c
e
j
b
O
as
ca-astroPh
elegans
hep-Ph
netscience
protein
Figure 4: Comparing xKPGM with other models in terms of
the objective function value obtained after training.
C. Impact of Number of Initiator Matrices
To understand how the choice of k, i.e., the number of
initiator matrices, impacts the performance of the model,
we compared the objective function obtained by training
the xKPGM model using different values of k. The results
for three graphs are shown in Figure 5. In all three cases,
increasing k from 1 to 2 results in a significant improvement
in the objective function. For two of the graphs (netscience
and protein), increasing k beyond 2 does not show any
further improvements. However, for the as graph, higher
values of k show improvement in the objective function.
as
netscience
protein
n
o
i
t
c
n
u
F
e
v
i
t
c
e
j
b
O
1.5
1
0.5
0
1
2
3
4
Number of Initiator Matrices - k
Figure 5: Performance of xKPGM using different number of
initiator matrices (k). Note that k = 1 is KPGM.
(a) Power Law Coefficient of Degree Distribution
(b) Average Path Length
(c) Average Clustering Coefficient
Figure 6: Distribution of network characteristics for samples
using the models trained on cit-hepPh data. The first plot
in each row is the distribution for the subgraph population
discussed in Section IV.
D. Variability of Generated Samples
The second set of experiments attempt to understand the
variability in the samples generated by the xKPGM and other
Kronecker models. For each model, we use the estimated
parameters and generate 200 samples with the same sample
size that was used to generate the subgraphs in Section IV.
For each sample we measure five salient properties similar
to Figure 2 and plot the distribution across the samples.
The distributions for cit-hepPh is shown in Figure 62. In the
results shown in Figure 6, both KPGM and mKPGM exhibit
almost negligible variability for almost all of the network
properties. In fact, it is evident that the variability increases
as we decrease l from n (KPGM) to 1 (tKPGM). xKPGM
shows most variability, and in many cases approximately
matches the variability of the subgraph population (degree
distribution, average clustering coefficient).
E. Comparison with BTER
As noted earlier, the BTER model allows generating a
synthetic graph which approximates an observed graph.
However, by design, BTER can generate graphs of approx-
imately the same size (nodes and edges) as the observed
graph. Using the available implementation3, we observed
that while BTER matches the degree distribution and the
2Results for other data sets available at: http://www.cse.buffalo.edu/
∼chandola/research/bigdata2015graphs/allresults.pdf due to lack of space.
3http://www.sandia.gov/∼tgkolda/feastpack
clustering coefficient, the generated graphs show minimal
variation for a set of samples generated using the same
observed graph. For example, for the ca-astroPH graph, 200
simulations using BTER resulted in graphs with number of
edges in the range: 196156± 208 and the average clustering
coefficient in the range: 0.3265 ± 0.0013.
X. CONCLUSIONS
We propose a new generative model, xKPGM, for graphs
based on mixture of Kronecker models. The model induces
variability by incorporating multiple initiator matrices strung
together using a random multinomial trial, while retaining
the strong features of KPGM, such as scaling to massive
graphs. We provide analytical expressions for various char-
acteristics of the generated graphs which allows us to esti-
mate the parameters using a method of moments approach.
The mixture approach allows us to devise a scalable method
of moments based learning method (similar to KPGM) while
achieving better variance (similar to mKPGM).
We show, both analytically and through experiments, that
xKPGM is able to learn a model which best matches the
moments of the observed graph (See Figure 4) and also
induces variability (See Figure 6) which aligns with the
natural variability observed in real graphs (See Figure 2).
To better evaluate the variability of generative graph models
we have come up with a sub-graph based method. Using this
method we show that the proposed model provides a robust
variance across multiple graph properties. We also provide
comparisons with the state of art generative models and show
that xKPGM outperforms them both in terms of matching
the graph properties and the variance in the population.
ACKNOWLEDGEMENTS
This work was supported by NSF grant CNS:1409551.
Access to computing resources was facilitated by an AWS
in Education Grant award.
REFERENCES
[1] W. Aiello, et. al., "A random graph model for power law
graphs," Experiment. Math., 10(1), 53 -- 66, 2001.
[2] E. M. Airoldi, et. al. Mixed membership stochastic block-
models. J. Mach. Learn. Res., 9:1981 -- 2014, 2008.
[3] A.-L. Barabasi and R. Albert. Emergence of scaling in
random networks. Science, 286(5439):509 -- 512, 1999.
[4] J. Casas-Roma, et al., "An algorithm for k-degree anonymity
on large networks," in ASONAM, 2013.
[5] V. Chandola, et. al., "Knowledge discovery from massive
healthcare claims data," in KDD, 2013.
[7] D. F. Gleich and A. B. Owen. Moment-based estimation of
stochastic kronecker graph parameters. Internet Mathematics,
8(3):232 -- 256, 2012.
[8] H. V. Henderson and S. R. Searle. The vec-permutation
matrix, the vec operator and kronecker products: a review.
Linear and Multilinear Algebra, 9(4):271 -- 288, 1981.
[9] C. Hubler, et. al. Metropolis algorithms for representative
subgraph sampling. In ICDM, 2008.
[10] H. Jeong, et. al. Lethality and centrality in protein networks.
Nature, 411:41 -- 42, 2001.
[11] M. Kim and J. Leskovec. Multiplicative attribute graph model
Internet Mathematics, 8(1-2):113 --
of real-world networks.
160, 2012.
[12] J. Leskovec, et. al. Kronecker graphs:An approach to mod-
eling networks. J. Mach. Lear. Res., 11:985 -- 1042, 2010.
[13] J. Leskovec and C. Faloutsos. Sampling from large graphs.
In Proceedings of the SIGKDD pages 631 -- 636, 2006.
[14] S. Moreno, et. al. Tied kronecker product graph models to
capture variance in network populations. In IEEE Communi-
cation, Control, and Computing (Allerton), 2010.
[15] S. Moreno, et. al. Block kronecker product graph model. In
Workshop on Mining and Learning from Graphs, 2013.
[16] S. I. Moreno, et. al. Learning mixed kronecker product graph
In SIGKDD,
models with simulated method of moments.
pages 1052 -- 1060, 2013.
[17] M. E. J. Newman. Finding community structure in networks
using the eigenvectors of matrices. Phys. Rev. E, 74:36 -- 104,
2006.
[18] P. Erdos and A. R´enyi. On the evolution of random graphs.
In Publication of the Mathematical Institute of the Hungarian
Academy of Sciences, pages 17 -- 61, 1960.
[19] A. Sanzgiri, A. Hughes, and S. Upadhyaya, "Analysis of
malware propagation in Twitter," in IEEE SRDS, 2013.
[20] C. Seshadhri, et. al., "Community structure and scale-free
collections of erdos-r´enyi graphs," CoRR, 2011.
[21] C. Seshadhri, et. al. An in-depth analysis of stochastic
Kronecker graphs. Journal of the ACM, 60(2), 2013.
[22] H. Sethu and X. Chu. A new algorithm for extracting a small
representative subgraph from a very large graph. CoRR, 2012.
[23] SNAP: Stanford network analysis platform.
stanford.edu/data/index.html.
http://snap.
[24] T. A. Snijders and K. Nowicki. Estimation and prediction for
stochastic blockmodels for graphs with latent block structure.
Journal of Classification, 14(1):75 -- 100, 1997.
[6] J. Duch and A. Arenas. Community detection in complex
networks using extremal optimization. Physical Review E,
72:027104, 2005.
[25] S. Wasserman and P. Pattison. Logit models and logistic
regressions for social networks. Psychometrika, 61(3):401 --
425, 1996.
|
1702.05354 | 2 | 1702 | 2017-10-23T18:46:51 | Algorithms for Online Influencer Marketing | [
"cs.SI"
] | Influence maximization is the problem of finding influential users, or nodes, in a graph so as to maximize the spread of information. It has many applications in advertising and marketing on social networks. In this paper, we study a highly generic version of influence maximization, one of optimizing influence campaigns by sequentially selecting "spread seeds" from a set of influencers, a small subset of the node population, under the hypothesis that, in a given campaign, previously activated nodes remain "persistently" active throughout and thus do not yield further rewards. This problem is in particular relevant for an important form of online marketing, known as influencer marketing, in which the marketers target a sub-population of influential people, instead of the entire base of potential buyers. Importantly, we make no assumptions on the underlying diffusion model and we work in a setting where neither a diffusion network nor historical activation data are available. We call this problem online influencer marketing with persistence (in short, OIMP). We first discuss motivating scenarios and present our general approach. We introduce an estimator on the influencers' remaining potential -- the expected number of nodes that can still be reached from a given influencer -- and justify its strength to rapidly estimate the desired value, relying on real data gathered from Twitter. We then describe a novel algorithm, GT-UCB, relying on upper confidence bounds on the remaining potential. We show that our approach leads to high-quality spreads on both simulated and real datasets, even though it makes almost no assumptions on the diffusion medium. Importantly, it is orders of magnitude faster than state-of-the-art influence maximization methods, making it possible to deal with large-scale online scenarios. | cs.SI | cs | Algorithms for Online Influencer Marketing
Paul Lagr´ee1, Olivier Capp´e2, Bogdan Cautis1, and Silviu Maniu1
1LRI, Universit´e Paris-Sud, Universit´e Paris-Saclay
2LIMSI, CNRS, Universit´e Paris-Saclay
7
1
0
2
t
c
O
3
2
]
I
S
.
s
c
[
2
v
4
5
3
5
0
.
2
0
7
1
:
v
i
X
r
a
Abstract
Influence maximization is the problem of finding influential users, or nodes, in a graph so as
to maximize the spread of information. It has many applications in advertising and marketing
on social networks. In this paper, we study a highly generic version of influence maximization,
one of optimizing influence campaigns by sequentially selecting "spread seeds" from a set of
influencers, a small subset of the node population, under the hypothesis that, in a given campaign,
previously activated nodes remain "persistently" active throughout and thus do not yield further
rewards. This problem is in particular relevant for an important form of online marketing,
known as influencer marketing, in which the marketers target a sub-population of influential
people, instead of the entire base of potential buyers. Importantly, we make no assumptions
on the underlying diffusion model and we work in a setting where neither a diffusion network
nor historical activation data are available. We call this problem online influencer marketing
with persistence (in short, OIMP). We first discuss motivating scenarios and present our general
approach. We introduce an estimator on the influencers' remaining potential – the expected
number of nodes that can still be reached from a given influencer – and justify its strength to
rapidly estimate the desired value, relying on real data gathered from Twitter. We then describe
a novel algorithm, GT-UCB, relying on upper confidence bounds on the remaining potential.
We show that our approach leads to high-quality spreads on both simulated and real datasets,
even though it makes almost no assumptions on the diffusion medium. Importantly, it is orders
of magnitude faster than state-of-the-art influence maximization methods, making it possible to
deal with large-scale online scenarios.
Keywords:Influencer marketing, information diffusion, online social networks, influence maxi-
mization, online learning, multi-armed bandits.
1
Introduction
Advertising based on word-of-mouth diffusion in social media has become very important in the
digital marketing landscape. Nowadays, social value and social influence are arguably the hottest
concepts in the area of Web advertising and most companies that advertise in the Web space must
have a "social" strategy. For example, on widely used platforms such as Facebook or Twitter,
promoted posts are interleaved with normal posts on user feeds. Users interact with these posts by
actions such as "likes" (adoption), "shares" or "reposts" (network diffusion). This represents an
unprecedented utility in advertising, be it with a commercial intent or not, as products, news, ideas,
movies, political manifests, tweets, etc, can propagate easily to a large audience [40, 41].
Motivated by the need for effective viral marketing strategies, influence estimation and influence
maximization have become important research problems, at the intersection of data mining and
1
social sciences [12]. In short, influence maximization is the problem of selecting a set of nodes from a
given diffusion graph, maximizing the expected spread under an underlying diffusion model. This
problem was introduced in 2003 by the seminal work of Kempe et al. [21], through two stochastic,
discrete-time diffusion models, Linear Threshold (LT) and Independent Cascade (IC). These models
rely on diffusion graphs whose edges are weighted by a score of influence. They show that selecting
the set of nodes maximizing the expected spread is NP-hard for both models, and they propose
a greedy algorithm that takes advantage of the sub-modularity property of the influence spread,
but does not scale to large graphs. A rich literature followed, focusing on computationally efficient
and scalable algorithms to solve influence maximization. The recent benchmarking study of Arora
et al. [2] summarizes state-of-the-art techniques and also debunks many influence maximization
myths. In particular, it shows that, depending on the underlying diffusion model and the choice of
parameters, each algorithm's behavior can vary significantly, from very efficient to prohibitively slow,
and that influence maximization at the scale of real applications remains an elusive target.
Importantly, all the influence maximization studies discussed in [2] have as starting point a
specific diffusion model (IC or LT), whose graph topology and parameters – basically the edge
weights – are known. In order to infer the diffusion parameters or the underlying graph structure,
or both, [15, 16, 14, 18, 33, 11] propose offline, model-specific methods, which rely on observed
information cascades. In short, information cascades are time-ordered sequences of records indicating
when a specific user was activated or adopted a specific item.
There are however many situations where it is unreasonable to assume the existence of relevant
historical data in the form of cascades. For such settings, online approaches, which can learn the
underlying diffusion parameters while running diffusion campaigns, have been proposed. Bridging
influence maximization and inference, this is done by balancing between exploration steps (of yet
uncertain model aspects) and exploitation ones (of the best solution so far), by so called multi-armed
bandits techniques, where an agent interacts with the network to infer influence probabilities [38, 10,
42, 37]. The learning agent sequentially selects seeds from which diffusion processes are initiated in
the network; the obtained feedback is used to update the agent's knowledge of the model.
Nevertheless, all these studies on inferring diffusion networks, whether offline or online, rely on
parametric diffusion models, i.e., assume that the actual diffusion dynamics are well captured by
such a model (e.g., IC). This maintains significant limitations for practical purposes. First, the
more complex the model, the harder to learn in large networks, especially in campaigns that have a
relatively short timespan, making model inference and parameter estimation very challenging within
a small horizon (typically tens or hundreds of spreads). Second, it is commonly agreed that the
aforementioned diffusion models represent elegant yet coarse interpretations of a reality that is much
more complex and often hard to observe fully. For examples of insights into this complex reality, the
topical or non-topical nature of an influence campaign, the popularity of the piece of information
being diffused, or its specific topic were all shown to have a significant impact on hashtag diffusions
in Twitter [11, 19, 32].
Our contribution Aiming to address such limitations, we propose in this paper a large-scale
approach for online and adaptive influence maximization, in which the underlying assumptions for
the diffusion processes are kept to a minimum (if, in fact, hardly any).
We argue that it can represent a versatile tool in many practical scenarios, including in particular
the one of influencer marketing, which, according to Wikipedia, can be described as follows: "a
form of marketing in which focus is placed on influential people rather than the target market as a
whole, identifying the individuals that have influence over potential buyers, and orienting marketing
activities around these influencers". For instance, influential users may be contractually bound by a
sponsorship, in exchange for the publication of promoted posts on their online Facebook or Twitter
2
accounts. This new form of marketing is by now extensively used in online social platforms, as is
discussed in the marketing literature [13, 1, 6].
More concretely, we focus on social media diffusion scenarios in which influence campaigns consist
of multiple consecutive trials (or rounds) spreading the same type of information from an arbitrary
domain (be it a product, idea, post, hashtag, etc)1. The goal of each campaign is to reach (or activate)
as many distinct users as possible, the objective function being the total spread. These potential
influencees – the target market – represent the nodes of an unknown diffusion medium / online
network. In our setting – as, arguably, in many real-world scenarios – the campaign selects from a
known set of spread seed candidates, so called influencers, a small subset of the potentially large and
unknown target market. At each round, the learning agent picks among the influencers those from
which a new diffusion process is initiated in the network, gathers some feedback on the activations,
and adapts the subsequent steps of the campaign. Only the effects of the diffusion process, namely
the activations (e.g., purchases or subscriptions), are observed, but not the process itself. The agent
may "re-seed" certain influencers (we may want to ask a particular one to initiate spreads several
times, e.g., if it has a strong converting impact). This perspective on influence campaigns imposes
naturally a certain notion of persistence, which is given the following interpretation: users that were
already activated in the ongoing campaign – e.g., have adopted a product or endorsed a political
movement – remain activated throughout that campaign, and thus will not be accounted for more
than once in the objective function.
We call this problem online influencer marketing with persistence (in short, OIMP). Our solution
for it follows the multi-armed bandit idea initially employed in Lei et al. [24], but we adopt instead
a diffusion-independent perspective, whose only input are the spread seed candidates, while the
population and underlying diffusion network – which may actually be the superposition of several
networks – remain unknown. In our bandit approach, the parameters to be estimated are the values
of the influencers – how good is a specific influencer –, as opposed to the diffusion edge probabilities of
a known graph as in [24]. Furthermore, we make the model's feedback more realistic by assuming that
after each trial, the agent only gathers the set of activated nodes. The rationale is that oftentimes,
for a given "viral" item, we can track in applications only when it was adopted by various users,
but not why. A key difference w.r.t. other multi-armed bandit studies for influence maximization
such as [38, 10, 42, 37] is that these look for a constant optimal set of seeds, while the difficulty with
OIMP is that the seemingly best action at a given trial depends on the activations of the previous
trials (and thus the learning agent's past decisions).
The multi-armed bandit algorithm we propose, called GT-UCB, relies on a famous statistical
tool known as the Good-Turing estimator, first developed during WWII to crack the Enigma machine,
and later published by Good in a study on species discovery [17]. Our approach is inspired by the
work of Bubeck et al. [8], which proposed the use of the Good-Turing estimator in a context where
the learning agent needs to sequentially select experts that only sample one of their potential nodes at
each trial. In contrast, in OIMP, when an influencer is selected, it may have a potentially large spread
and may activate many nodes at once. Our solution follows the well-known optimism in the face of
uncertainty principle from the bandit literature (see [7] for an introduction to multi-armed bandit
problems), by deriving an upper confidence bound on the estimator for the remaining potential for
spreading information of each influencer, and by choosing in a principled manner between explore
and exploit steps.
In Section 6 we evaluate the proposed approach on publicly available graph datasets as well a
large snapshot of Twitter activity we collected. The proposed algorithm is agnostic with respect to
the choice of influencers, who in most realistic applications will be selected based on a combination of
1Repeated exposure, known also as the "effective frequency", is a crucial concept in marketing strategies, online or
offline.
3
graph-based and external considerations. We describe however in Section 6.1 several heuristics that
may be used to automatically extract good candidates for the influencer set, when a social network
graph is available. This choice of influencers also makes it possible to compare the results obtained by
our method to those of the baseline methods from the literature (see above), which require knowledge
of the graph and of the diffusion model and, in some cases, of the influence probabilities.
Comparison with previous publication We extend in this article a preliminary study published
in Lagr´ee et al. [22], introducing the following new contributions, which allow us to give a complete
picture on our model and algorithmic solutions:
• A more detailed discussion on the motivation behind our work, in relation to new forms of
online marketing, such as influencer marketing.
• An empirical analysis over Twitter data, which comes to support the assumptions behind our
choice of estimators for the remaining potential in each influencer.
• A detailed theoretical analysis and justification for the upper confidence bounds on which our
algorithm GT-UCB relies.
• Theoretical guarantees on the performance of GT-UCB, formulated in terms of waiting time,
a notion that is better suited in our bandit framework than the usual one of regret.
• An adaptation of GT-UCB (denoted Fat-GT-UCB) and the corresponding theoretical analysis,
for scenarios in which influencers may experience fatigue, i.e., a diminishing tendency to activate
their user base as they are re-seeded throughout a marketing campaign.
• A broader experimental analysis involving the two datasets previously used in [22], as well as
a new set of experimental results in a completely different scenario, involving real influence
spreads from Twitter, for both GT-UCB and Fat-GT-UCB.
To the best of our knowledge, our approach is the first to show that efficient and effective influence
maximization can be done in a highly uncertain or under-specified social environment, along with
formal guarantees on the achieved spread.
2 Setting
The goal of the online influencer marketing with persistence is to successively select (or activate)
a number of seed nodes, in order to reach (or spread to) as many other nodes as possible. In this
section, we formally define this problem.
2.1 Background
Given a graph G = (V, E), the traditional problem of influence maximization is to select a set of
seed nodes I ⊆ V , under a cardinality constraint I = L, such that the expected spread – that is,
the number of activated nodes – of an influence cascade starting from I is maximized. Formally,
denoting by the random variable S(I) the spread initiated by the seed set I, influence maximization
aims to solve the following optimization problem:
E[S(I)].
arg max
I⊆V,I=L
4
Figure 1: Three influencers with associated activation probabilities pk(u).
As mentioned before, a plethora of algorithms have been proposed to solve the influence maximiza-
tion problem, under specific diffusion models. These algorithms can be viewed as full-information
and offline approaches: they choose all the seeds at once, in one step, and they have the complete
diffusion configuration, i.e., the graph topology and the influence probabilities.
In the online case, during a sequence of N (called hereafter the budget) consecutive trials, L seed
nodes are selected at each trial, and feedback on the achieved spread from these seeds is collected.
2.2 Influence maximization via influencers
The short timespan of campaigns makes parameter estimation very challenging within small horizons.
In other cases, the topology – or even the existence – of a graph is too strong an assumption. In
contrast to [24], we do not try to estimate edge probabilities in some graph, but, instead, we assume
the existence of a known set of spread seed candidates – in the following referred to as the influencers
– who are the only access to the medium of diffusion. Formally, we let [K] := {1, . . . , K} be a set
of influencers up for selection; each influencer is connected to an unknown and potentially large
base (the influencer's support) of basic nodes, each with an unknown activation probability. For
illustration, we give in Figure 1 an example of this setting, with 3 influencers connected to 4, 5, and
4 basic nodes, respectively.
Now, the problem boils down to estimating the value of the K influencers, which is typically much
smaller than the number of parameters of the diffusion model. The medium over which diffusion
operates may be a diffusion graph but we make no assumption on that, meaning that the diffusion
may also happen in a completely unknown environment. Finally, note that by choosing K = V
influencers, the classic influence maximization problem can be seen as a special instance of our
setting.
We complete the formal setting by assuming the existence of K sets Ak ⊆ V of basic nodes such
that each influencer k ∈ [K] is connected to each node in Ak. We denote pk(u) the probability for
influencer k to activate the child node u ∈ Ak. In this context, the diffusion process can be abstracted
as follows.
Definition 1 (Influence process). When an influencer k ∈ [K] is selected, each basic node u ∈ Ak is
sampled for activation, according to its probability pk(u). The feedback for k's selection consists of
all the activated nodes, while the associated reward consists only of the newly activated ones.
Remark Limiting the influence maximization method to working with a small subset of the node
base may allow to accurately estimate their value more rapidly, even in a highly uncertain environment,
hence the algorithmic interest. At the same time, this is directly motivated by marketing scenarios
involving marketers who may not have knowledge of the entire diffusion graph, only having access
5
pk(u)pk(u)uInfluencersBasicNodes to a few influential people who can diffuse information (the influencers in our setting), or may
simply prefer such a two-step flow of diffusion for various reasons, such as establishing credibility.
Moreover, despite the fact that we model the social reach of every influencer by 1-hop links to the
to-be-influenced nodes, these edges are just an abstraction of the activation probability, and may
represent in reality longer paths in an underlying unknown real influence graph G.
2.3 Online influencer marketing with persistence
We are now ready to define the online influencer marketing with persistence task.
Problem 1 (OIMP). Given a set of influencers [K] := {1, . . . , K}, a budget of N trials, and a
number 1 ≤ L ≤ K of influencers to be activated at each trial, the objective of the online influencer
marketing with persistence (OIMP) is to solve the following optimization problem:
(cid:91)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) .
arg max
In⊆[K],In=L,∀1(cid:54)n(cid:54)N
E
S(In)
1(cid:54)n(cid:54)N
As noticed in [24], the offline influence maximization can be seen as a special instance of the
online one, where the budget is N = 1. Note that, in contrast to persistence-free online influence
maximization – considered, e.g., in [38, 42] – the performance criterion used in OIMP displays
the so-called diminishing returns property: the expected number of nodes activated by successive
selections of a given seed is decreasing, due to the fact that nodes that have already been activated
are discounted. We refer to the expected number of nodes remaining to be activated as the remaining
potential of a seed. The diminishing returns property implies that there is no static best set of
seeds to be selected, but that the algorithm must follow an adaptive policy, which can detect that
the remaining potential of a seed is small and switch to another seed that has been less exploited.
Our solution to this problem has to overcome challenges on two fronts: (1) it needs to estimate the
potential of nodes at each round, without knowing the diffusion model nor the activation probabilities,
and (2) it needs to identify the currently best influencers, according to their estimated potentials.
Other approaches for the online influence maximization problem rely on estimating diffusion
parameters [24, 38, 42] – generally, a distribution over the influence probability of each edge in the
graph. However, the assumption that one can estimate accurately the diffusion parameters – and
notably the diffusion probabilities – may be overly ambitious, especially in cases where the number
of allowed trials (the budget) is rather limited. A limited trial setting is arguably more in line with
real-world campaigns: take as example political or marketing campaigns, which only last for a few
weeks.
In our approach, we work with parameters on nodes, instead of edges. More specifically, these
parameters represent the potentials of remaining spreads from each of the influencer nodes. We
stress that this potential can evolve as the campaign proceeds. In this way, we can go around the
dependencies on specific diffusion models, and furthermore, we can remove entirely the dependency
on a detailed graph topology.
3 Algorithm
In this section, we describe our UCB-like algorithm, which relies on the Good-Turing estimator to
sequentially select the seeds to activate at each round, from the available influencers.
6
3.1 Remaining potential and Good-Turing estimator
A good algorithm for OIMP should aim at selecting the influencer k with the largest potential for
influencing its children Ak. However, the true potential value of an influencer is a priori unknown to
the decision maker.
In the following, we index trials by t when referring to the time of the algorithm, and we index
trials by n when referring to the number of selections of the influencer. For example, the t-th spread
initiated by the algorithm is noted S(t) whereas the n-th spread of influencer k is noted Sk,n.
Definition 2 (Remaining potential Rk(t)). Consider an influencer k ∈ [K] connected to Ak basic
nodes. Let S(1), . . . , S(t) be the set of nodes that were activated during the first t trials by the seeded
influencers. The remaining potential Rk(t) is the expected number of new nodes that would be
activated upon starting the t + 1-th cascade from k:
(cid:88)
u∈Ak
(cid:40)
u /∈ t(cid:91)
i=1
(cid:41)
Rk(t) :=
1
S(i)
pk(u),
where 1{·} denotes the indicator function.
Definition 2 provides a formal way to obtain the remaining potential of an influencer k at a given
time. The optimal policy would simply select the influencer with the largest remaining potential at
each time step. The difficulty is, however, that the probabilities pk(u) are unknown. Hence, we have
to design a remaining potential estimator Rk(t) instead. It is important to stress that the remaining
potential is a random quantity, because of the dependency on the spreads S(1), . . . , S(t). Furthermore,
due to the diminishing returns property, the sequence (Sk,n)n≥1 is stochastically decreasing.
Following ideas from [17, 8], we now introduce a version of the Good-Turing statistic, tailored to
our problem of rapidly estimating the remaining potential. Denoting by nk(t) the number of times
influencer k has been selected after t trials, we let Sk,1, . . . , Sk,nk(t) be the nk(t) cascades sampled
independently from influencer k. We denote by Uk(u, t) the binary function whose value is 1 if node
u has been activated exactly once by influencer k – such occurrences are called hapaxes in linguistics
– and Zk(u, t) the binary function whose value is 1 if node u has never been activated by influencer k.
The idea of the Good-Turing estimator is to estimate the remaining potential as the proportion of
hapaxes in the nk(t) sampled cascades, as follows:
(cid:88)
(cid:89)
l(cid:54)=k
Rk(t) :=
1
nk(t)
u∈Ak
Uk(u, t)
Zl(u, t).
Albeit simple, this estimator turns out to be quite effective in practice.
If an influencer is
connected to a combination of both nodes having high activation probabilities and nodes having low
activation probabilities, then successive traces sampled from this influencer will result in multiple
activations of the high-probability nodes and few of the low-probability ones. Hence, after observing
a few spreads, the influencer's potential will be low, a fact that will be captured by the low proportion
of hapaxes. In contrast, estimators that try to estimate each activation probability independently
will require a much larger number of trials to properly estimate the influencer's potential.
To verify this assumption in reality, we conducted an analysis of the empirical activation proba-
bilities from a Twitter dataset. Specifically, we used a collection of tweets and re-tweets gathered
via crawling in August 2012. For each original tweet, we find all corresponding retweets, and, for
each user, we compute the empirical probability of a retweet occurring – this, in our case, is a
proxy measure for influence probability. Specifically, for every user v "influenced" by u, i.e., v
7
Figure 2: (left) Twitter empirical retweet probabilities. (right) Sample of 50 empirical retweet
probabilities.
retweeted at least one original tweet from u – we compute the estimated diffusion probability:
pu,v = u's tweets retweeted by v /tweets by u. In Fig. 2 (left), we show the survival function of
resulting empirical probabilities in a log-log plot. We can see that most probabilities are small – the
9th decile has value 0.045.
In Fig. 2 (right), we simulated the activation probabilities of a set of 50 nodes whose activation
probabilities are chosen randomly from the Twitter empirical probabilities. Most of the sampled
values are low, except a few relatively high ones. Using this sample as the activation probabilities of
an hypothetical influencer node, we observe on Fig. 3 (left) the cumulative influence spread. The
curve first shows a steep increase until approximately 20 rounds, where users with high probabilities
of conversion have already been activated, while remaining ones are difficult to activate.
In Fig. 3 (right), we compare the Good-Turing estimator to a Bayesian estimator that maintains
a posterior (through a Beta distribution) on the unknown activation probabilities, updating the
posterior after each trial, similarly to [24]. In the Bayesian approach, the remaining potential can be
estimated by summing over the means of the posterior distributions corresponding to nodes that
have not been activated so far. On Fig. 3 (right), the curves are averaged over 200 runs, and the
shaded regions correspond to the 95% quantiles. Clearly, the Good-Turing estimator is much faster
than its Bayesian counterpart in estimating the actual remaining potential. Varying the number
of nodes – here equal to 50 – shows that the time needed for the Bayesian estimator to provide a
reliable estimate of the remaining potential is proportional to the number of nodes, whereas it grows
only sub-linearly for the Good-Turing estimator.
Remark While bearing similarities with the traditional missing mass concept, we highlight one
fundamental difference between the remaining potential and the traditional missing mass studied
in [8], which impacts both the algorithmic solution and the analysis. Since at each step, after selecting
an influencer, every node connected to that influencer is sampled, the algorithm receives a larger
feedback than in [8], whose feedback is in [0, 1]. However, on the contrary to [8], the hapaxes of
p(u), which
corresponds to the expected number of basic nodes an influencer activates or re-activates in a cascade,
will prove to be a crucial ingredient for our problem.
an influencer (Uk(u, t))u∈Ak are independent. Interestingly, the quantity λk :=(cid:80)
u∈Ak
8
10−410−310−210−1100Retweetprobability10−510−410−310−210−1100Survivalfunction00.0250.050.0750.10.1250.15Retweetprobability Figure 3: (left) Influence spread against number of rounds. (right) Bayesian estimator against
Good-Turing estimator.
3.2 Upper confidence bounds
Following principles from the bandit literature, the GT-UCB algorithm relies on optimism in the
face of uncertainty. At each step (trial) t, the algorithm selects the highest upper-confidence bound
on the remaining potential – denoted by bk(t) – and activates (plays) the corresponding influencer
k. This algorithm achieves robustness against the stochastic nature of the cascades, by ensuring
that influencers who "underperformed" with respect to their potential in previous trials may still be
selected later on. Consequently, GT-UCB aims to maintain a degree of exploration of influencers, in
addition to the exploitation of the best influencers as per the feedback gathered so far.
ALGORITHM 1: – GT-UCB (L = 1)
Require: Set of influencers [K], time budget N
1: Initialization: play each influencer k ∈ [K] once, observe the spread Sk,1, set nk = 1
2: For each k ∈ [K]: update the reward W = W ∪ Sk,1
3: for t = K + 1, . . . , N do
4:
5:
6:
7:
8:
9: end for
10: return W
Compute bk(t) for every influencer k
Choose k(t) = arg maxk∈[K] bk(t)
Play influencer k(t) and observe spread S(t)
Update cumulative reward: W = W ∪ S(t)
Update statistics of influencer k(t): nk(t)(t + 1) = nk(t)(t) + 1 and Sk,nk(t) = S(t).
influencer is chosen at each step.
Algorithm 1 presents the main components of GT-UCB for the case L = 1, that is, when a single
The algorithm starts by activating each influencer k ∈ [K] once, in order to initialize its Good-
Turing estimator. The main loop of GT-UCB occurs at lines 3-9. Let S(t) be the observed spread at
trial t, and let Sk,s be the result of the s-th diffusion initiated at influencer k. At every step t > K,
we recompute for each influencer k ∈ [K] its index bk(t), representing the upper confidence bound
on the expected reward in the next trial. The computation of this index uses the previous samples
Sk,1, . . . , Sk,nk(t) and the number of times each influencer k has been activated up to trial t, nk(t).
Based on the result of Theorem 1 – whose statement and proof are delayed to Section 4 –, the upper
9
01020304050Trial0510152025#Activations5101520Trial03691215#nodesRemainingpotentialGTestimatorBayesestimator confidence bound is set as:
(cid:17)(cid:115)
√
(cid:16)
bk(t) = Rk(t) +
where Rk(t) is the Good-Turing estimator and λk(t) :=(cid:80)nk(t)
nk(t)
1 +
2
s=1
spread from influencer k.
λk(t) log(4t)
+
log(4t)
3nk(t)
,
(1)
Sk,s
nk(t) is an estimator for the expected
Then, in line 5, GT-UCB selects the influencer k(t) with the largest index, and initiates a cascade
from this node. The feedback S(t) is observed and is used to update the cumulative reward set W .
We stress again that S(t) provides only the Ids of the nodes that were activated, with no information
on how this diffusion happened in the hidden diffusion medium. Finally, the statistics associated to
the chosen influencer k(t) are updated.
3.3 Extensions for the case L > 1
Algorithm 1 can be easily adapted to select L > 1 influencers at each round. Instead of choosing
the influencer maximizing the Good-Turing UCB in line 5, we can select those having the L largest
indices. Note that k(t) then becomes a set of L influencers. A diffusion is initiated from the associated
nodes and, at termination, all activations are observed. Similarly to [37], the algorithm requires
feedback to include the influencer responsible for the activation of each node, in order to update the
corresponding statistics accordingly.
4 Analysis
In this section, we justify the upper confidence bound used by GT-UCB in Eq. 1 and provide a
theoretical analysis of the algorithm.
4.1 Confidence interval for the remaining potential
In the following, to simplify the analysis and to allow for a comparison with the oracle strategy, we
assume that the influencers have non intersecting support. This means that each influencer's remaining
potential and corresponding Good-Turing estimator does not dependent on other influencers. Hence,
for notational efficiency, we also omit the subscript denoting the influencer k. After selecting the
n . We note that
the non-intersecting assumption is for theoretical purposes only – our experiments are done with
influencers that can have intersecting supports.
influencer n times, the Good-Turing estimator is simply written Rn =(cid:80)
λ =(cid:80)
The classic Good-Turing estimator is known to be slightly biased (see Theorem 1 in [28] for
example). We show in Lemma 1 that our remaining potential estimator adds an additional factor
u∈A p(u) to this bias:
u∈A
Un(u)
Lemma 1. The bias of the remaining potential estimator is
(cid:20)
(cid:21)
− λ
n
, 0
.
E[Rn] − E[ Rn] ∈
10
Proof.
E[Rn]−E[ Rn] =
u∈A
(cid:88)
(cid:88)
(cid:34)(cid:88)
u∈A
E
u∈A
= − 1
n
= − 1
n
(cid:104)
p(u)(1 − p(u))n−1(cid:105)
p(u)(1 − p(u))n − n
n
p(u) × np(u)(1 − p(u))n−1
(cid:35)
(cid:20)
(cid:80)
∈
−
(cid:21)
, 0
u∈A p(u)
n
p(u)Un(u)
Since λ is typically very small compared to A, in expectation, the estimation should be relatively
accurate. However, in order to understand what may happen in the worst-case, we need to characterize
the deviation of the Good-Turing estimator:
Theorem 1. With probability at least 1 − δ, for λ =(cid:80)
u∈A p(u) and βn :=(cid:0)1 +
2(cid:1)(cid:113) λ log(4/δ)
n
√
+
1
3n log 4
δ , the following holds:
−βn − λ
n
≤ Rn − Rn ≤ βn.
Note that the additional term appearing in the left deviation corresponds to the bias of our
estimator, which leads to a non-symmetrical interval.
Proof. We prove the confidence interval in three steps: (1) Good-Turing estimator deviation, (2) re-
maining potential deviation, (3) combination of these two inequalities to obtain the final confidence
interval.
Here, the child nodes are assumed to be sampled independently, which is a simplification compared
to the classic missing mass concentration results that relies on negatively associated samples [28, 27].
On the other hand, since we may activate several nodes at once, we need original concentration
arguments to control the increments of both Rn and Rn.
(1) Good-Turing deviations. Let Xn(u) := Un(u)
(cid:88)
u∈A
v :=
E[Xn(u)2] =
1
n2
(cid:88)
n . We have that
E[Un(u)] ≤ λ
n
.
u∈A
Applying Bennett's inequality (Theorems 2.9, 2.10 in [5]) to the independent random variables
Moreover, clearly the following holds: Xn(u) ≤ 1
n .
{Xn(u)}u∈A yields
(cid:32)
(cid:114)
(cid:33)
P
Rn − E[ Rn] ≥
2λ log(1/δ)
n
+
log(1/δ)
3n
≤ δ.
(2)
The same inequality can be derived for left deviations.
Rn =(cid:80)
(2) Remaining potential deviations. Remember that Zn(u) denotes the indicator equal
to 1 if u has never been activated up to trial n. We can rewrite the remaining potential as
u∈A Zn(u)p(u). Let Yn(u) = p(u)(Zn(u) − E[Zn(u)]) and q(u) = P(Zn(u) = 1) = (1 − p(u))n.
11
For some t > 0, we have next that
−t (cid:89)
etYn(u)(cid:105)
E(cid:104)
P(Rn − E[Rn] ≥ ) ≤ e
(cid:16)
= et (cid:89)
−t (cid:89)
u∈A
≤ e
u∈A
q(u)etp(u)(1−q(u)) + (1 − q(u))e
−tp(u)q(u)(cid:17)
exp(p(u)t2/(4n)) = exp(cid:0)−t + t2/(4n)λ(cid:1) .
The first inequality is well-known in exponential concentration bounds and relies on Markov's
inequality. The second inequality follows from [4] (Lemma 3.5).
Then, choosing t = 2n
λ , we obtain
u∈A
(cid:32)
(cid:114)
(cid:33)
P
Rn − E[Rn] ≥
λ log(1/δ)
n
≤ δ.
(3)
We can proceed similarly to obtain the left deviation.
(3) Putting it all together. We combine Lemma 1 with Eq. (2), (3), to obtain the final result.
4 to ensure that both the left and right bounds for the Good-Turing
Note that δ is replaced by δ
estimator and the remaining potential are verified.
4.2 Theoretical guarantees
We now provide an analysis of the waiting time (defined below) of GT-UCB, by comparing it to the
waiting time of an oracle policy, following ideas from [8]. Let Rk(t) be the remaining potential of
influencer k at trial number t. This differs from Rk,n, which is the remaining potential of influencer
k once it has been played n times.
Definition 3 (Waiting time). Let λk = (cid:80)
p(u) denote the expected number of activations
obtained by the first call to influencer k. For α ∈ (0, 1), the waiting time TU CB(α) of GT-UCB
represents the round at which the remaining potential of each influencer k is smaller than αλk.
Formally,
u∈Ak
TU CB(α) := min{t : ∀k ∈ [K], Rk(t) ≤ αλk}.
The above definition can be applied to any strategy for influencer selection and, in particular,
to an oracle one that knows beforehand the α value that is targeted, the spreads (Sk,s)k∈[K],1≤s≤t
sampled up to the current time, and the individual activation probabilities pk(u), u ∈ Ak. A policy
having access to all these aspects will perform the fewest possible activations on each influencer. We
denote by T ∗(α) the waiting time of the oracle policy. We are now ready to state the main theoretical
property of the GT-UCB algorithm.
λmin ≥ 13, for any α ∈(cid:2) 13
λmin , 1(cid:3), if we define τ∗ := T ∗(cid:0)α − 13
λmin
(cid:1), with probability at least 1 − 2K
λmax
Theorem 2 (Waiting time). Let λmin := mink∈[K] λk and let λmax := maxk∈[K] λk. Assuming that
the following holds:
TUCB(α) ≤ τ
∗ + Kλmax log(4τ
∗ + 11Kλmax) + 2K.
The proof of this result is given in Appendix B. Unsurprisingly, Theorem 2 says that GT-UCB
must perform slightly more activations of the influencers than the oracle policy. With high probability
12
– assuming that the best influencer has an initial remaining potential that is much larger than the
number of influencers – the waiting time of GT-UCB is comparable to T ∗(α(cid:48)), up to factor that is
only logarithmic in the waiting time of the oracle strategy. α(cid:48) is smaller than α – hence T ∗(α(cid:48)) is
larger than T ∗(α)– by an offset that is inversely proportional to the initial remaining potential of the
worst influencer. This essentially says that, if we deal with large graphs, and if the influencers trigger
reasonably large spreads, our algorithm is competitive with the oracle.
5 OIMP with influencer fatigue
In our study of the OIMP problem so far, a key assumption has been that the influencers have
a constant tendency to activate their followers. This hypothesis may not be verified in certain
situations, in which the moment influencers promote products that do not align with their image
(misalignement) or persist in promoting the same services (weariness). In such cases, they can expect
their influence to diminish [34, 23]. To cope with such cases of weariness, we propose in this section
an extension to GT-UCB that incorporates the concept of influencer fatigue.
In terms of bandits, the idea of our extension is similar in spirit to the one of Levine et al. [25]: a
new type of bandits – called rotting bandits – where each arm's value decays as a function of the
number of times it has been selected. We also mention the work of Louedec et al. [26] in which the
authors propose to take into account the gradual obsolescence of items to be recommended while
allowing new items to be added to the pool of candidates. In this latter work, the item's value is
modeled by a decreasing function of the number of steps elapsed since the item was added to the
pool of items, whereas in our work – and in that of [25] –, the value is a function of the number of
times the item has been selected.
5.1 Model adaptation
The OIMP problem with influencer fatigue can be defined as follows.
Problem 2 (OIMP with influencer fatigue). Given a set of influencers [K], a budget of N trials,
a number 1 ≤ L ≤ K of influencers to be activated at each trial, the objective of online influencer
marketing with persistence (OIMP) and with influencer fatigue is to solve the following optimization
problem:
(cid:91)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ,
arg max
In⊆[K],In=L,∀1(cid:54)n(cid:54)N
E
S(In)
1(cid:54)n(cid:54)N
knowing that, at the s-th selection of an influencer k ∈ [K], the probability that k activates some
basic node u is:
for γ : N∗ → (0, 1] a known non-increasing function and p(u) ∈ [0, 1].
ps(u) = γ(s)p(u),
Our initial OIMP formulation can be seen as a special instance of the one with influencer fatigue,
where the non-increasing function γ – referred to as the weariness function in the following – is the
constant function n (cid:55)→ 1. We follow the same strategy to solve this new OIMP variant, by estimating
the remaining potential of a given influencer by an adaptation of the Good-Turing estimator. What
makes the problem more complex in this setting is the fact that our hapax statistics must now take
into account the round at which they occured.
13
5.2 The Fat-GT-UCB algorithm
As we did previously, to simplify the analysis, we assume that the influencers have non intersecting
support. We redefine the remaining potential in the setting with influencer fatigue as
Rk(t) :=
1{u never activated}γ(nk(t) + 1)p(u),
(cid:88)
u∈Ak
where p(u) is the probability that the influencer activates node u, independently of the number of
spreads initiated by the influencer. Again, the remaining potential is equal to the expected number
of additional conversions upon starting the t + 1-th cascade from k. The Good-Turing estimator
adapted to the setting with influencer fatigue is defined as follows:
(cid:88)
U γ
nk(t)(u),
Rk(t) =
1
nk(t)
u∈Ak
k,n(u) :=(cid:80)
(cid:17)(cid:115)
(cid:16)
√
1 +
2
γ(i)
1≤i≤n
1{Xk,1(u) = . . . = Xk,i−1(u) = Xk,i+1(u) = . . . = Xk,n(u) = 0, Xk,i(u) =
where U γ
1} γ(n+1)
. In short, if i is the round at which a hapax has been activated, we reweight it by the
factor γ(n + 1)/γ(i) since we are interested in its contribution at the n + 1-th spread initiated by the
influencer. We provide a formal justification to this estimator by computing its bias in Appendix C.
Following the same strategy and principles from the bandit literature, the Fat-GT-UCB adapta-
tion of GT-UCB selects at each step (trial) t the highest upper-confidence bound on the remaining
potential – denoted by bk(t) – and activates (plays) the corresponding influencer k. The upper
confidence bound can now be set as follows (the full details can also be found in Appendix C – see
Theorem 4):
bk(t) = Rk(t) +
λk(t) log(4t)
nk(t)
+
log(4t)
3nk(t)
,
(4)
where Rk(t) is the Good-Turing estimator and
λk(t) :=
γ(nk(t) + 1)
nk(t)
nk(t)(cid:88)
s=1
Sk,s
γ(s)
is an estimator for the expected spread from influencer k.
6 Experiments
We conducted experiments on two types of datasets: (i.) two graphs, widely-used in the influence
maximization literature, and (ii.) a crawled dataset from Twitter, consisting of tweets occurring
during August 2012. All methods are implemented 2 in C++ and simulations are done on an Ubuntu
16.04 machine with an Intel Xeon 2.4GHz CPU 20 cores and 98GB of RAM.
6.1 Extracting influencers from graphs
GT-UCB does not make any assumptions about the topology of the nodes under the scope of
influencers. Indeed, in many settings it may be more natural to assume that the set of influencers is
2The code is available at https://github.com/smaniu/oim.
14
given and that the activations at each trial can be observed, while the topology of the underlying
graph G remain unknown. In other settings, we may start from an existing social network G, in
which case we need to extract a set of K representative influencers from it. Ideally, we should choose
influencers that have little intersection in their "scopes of influence" to avoid useless seed selections.
While this may be interpreted and performed differently, from one application to another, we discuss
next some of the most natural heuristics for selecting influencers which we use in our experiments.
MaxDegree. This method selects the K nodes with the highest out-degrees in G. Note that by
this criterion we may select influencers with overlapping influence scopes.
Greedy MaxCover. This strategy follows the well-known greedy approximation algorithm for
selecting a cover of the graph G. Specifically, the algorithm executes the following steps K times:
1. Select the node with highest out-degree
2. Remove all out-neighbors of the selected node
To limit intersections among influencer scopes even more, nodes reachable by more than 1 hops may
be removed at step (2).
DivRank [29]. DivRank is a PageRank-like method relying on reinforced random walks, with
the goal of producing diverse high-ranking nodes, while maintaining the rich-gets-richer paradigm.
We adapted the original DivRank procedure by inverting the edge directions. In doing so, we get
influential nodes instead of prestigious ones. By selecting the K highest scoring nodes as influencers,
the diversity is naturally induced by the reinforcement of random walks. This ensures that the
influencers are fairly scattered in the graph and should have limited impact on each other.
Influence maximization approximated algorithms. The fourth method we tested in our
experiments assigns uniformly at random a propagation probability to each edge of G, assuming the
IC model. Then, a state-of-the-art influence maximization algorithm – PMC in our experiments – is
executed on G to get the set of K influencers having the highest potential spread.
6.2 Graph datasets
Similarly to [24], we tested our algorithm on HepPh and DBLP, two publicly available collaboration
networks. HepPh is a citation graph, where a directed edge is established when an author cited at
least one paper of another author. In DBLP undirected edges are drawn between authors which have
collaborated on at least one indexed paper. The datasets are summarized in Table 1. We emphasize
that we kept the datasets relatively small to allow for comparison with computation-heavy baselines,
even though GT-UCB easily scales to large data, as will be illustrated in Section 6.3.
Table 1: Summary of the datasets.
Dataset
HepPh DBLP Twitter
# of nodes
# of edges
34.5K 317K 11.6M
2.1M 38.4M
422K
Diffusion models. In the work closest to ours, Lei et al. [24] compared their solution on the
Weighted Cascade instance of IC, where the influence probabilities on incoming edges sum up to
1. More precisely, every edge (u, v) has weight 1/dv where dv is the in-degree of node v. In this
experimental study, and to illustrate that our approach is diffusion-independent, we added two other
diffusion scenarios to the set of experiments. First, we included the tri-valency model (TV), which
associates randomly a probability from {0.1, 0.01, 0.001} to every edge and follows the IC propagation
15
(a) HepPh (WC – Impact of K)
(b) HepPh (WC – Influencer extrac-
tion)
(c) DBLP (WC – Impact of K)
(d) DBLP (WC – Influencer extrac-
tion)
Figure 4: Impact of K and the influencer extraction criterion on influence spread.
model. We also conducted experiments under the Linear Threshold (LT) model, where the edge
probabilities are set like in the WC case and where thresholds on nodes are sampled uniformly from
[0, 1].
Baselines. We compare GT-UCB to several baselines. Random chooses a random influencer
at each round. MaxDegree selects the node with the largest degree at each step i, where the
degree does not include previously activated nodes. Finally, EG corresponds to the confidence-bound
explore-exploit method with exponentiated gradient update from [24]; it is the state-of-the-art method
for the OIMP problem (code provided by the authors). We use this last baseline on WC and TV
weighted graphs and tune parameters in accordance to the results of their experiments: Maximum
Likelihood Estimation is adopted for graph update and edge priors are set to Beta(1, 20). Note that
EG learns parameters for the IC model, and hence is not applicable for LT. These baselines are
compared to an Oracle that knows beforehand the diffusion model together with probabilities. At
each round, it runs an influence maximization approximated algorithm – PMC for IC propagation,
SSA for LT. Note that previously activated nodes are not counted when estimating the value of a
node with PMC or SSA, thus, making Oracle an adaptive strategy.
All experiments are done by fixing the trial horizon N = 500, a setting that is in line with many
real-world marketing campaigns, which are fairly short and do not aim to reach the entire population.
Choice of the influencers. We show in Fig. 4b and 4d the impact of the influencer extraction
criterion on HepPh and DBLP under WC model. We can observe that the spread is only slightly
affected by the extraction criterion: different datasets lead to different optimal criteria. On the
16
100200300400500Trial0.00.20.50.81.01.21.51.8In(cid:131)uenceSpread×104K102050100200100200300400500Trial0.00.20.50.81.01.21.5In(cid:131)uenceSpread×104MaxCoverPMCDivRankMaxDegree100200300400500Trial0.01.02.03.04.0In(cid:131)uenceSpread×104K102050100200100200300400500Trial0.01.02.03.04.0In(cid:131)uenceSpread×104MaxCoverMaxDegreePMCDivRank HepPh network, DivRank clearly leads to larger influence spreads. On DBLP, however, the extraction
method has little impact on resulting spreads. We emphasize that on some other graph and model
combinations we observed that other extraction routines can perform better than DivRank. In
summary, we note that GT-UCB performs consistently as long as the method leads to influencers
that are well spread over the graph. In the following, for each graph, we used DivRank as the
influencer extraction criterion in accordance with these observations.
In Fig. 4a and 4c, we measure the impact of the number of influencers K on the influence spread.
We can observe that, on DBLP, a small number of influencers is sufficient to yield high-quality results.
If too many influencers (relative to the budget) are selected (e.g. K = 200), the initialization step
required by GT-UCB is too long relative to the full budget, and hence GT-UCB does not reach its
optimal spread – some influencers still have a large remaining potential at the end. On the other
hand, a larger amount of influencers leads to greater influence spreads on HepPh: this network is
relatively small (34.5K nodes), and thus half of the nodes are already activated after 400 trials. By
having more influencers, we are able to access parts of the network that would not be accessible
otherwise.
Figure 5: DBLP (WC) – Execution time.
GT-UCB vs. baselines. We evaluate the execution time of the different algorithms in Fig. 5.
As expected, GT-UCB largely outperforms EG (and Oracle). The two baselines require the
execution of an approximated influence maximization algorithm at each round. In line with [2], we
observed that SSA has prohibitive computational cost when incoming edge weights do not sum up
to 1, which is the case with both WC and TV. Thus, both Oracle and EG run PMC on all our
experiments with IC propagation. GT-UCB is several orders of magnitude faster: it concentrates
most of its running time on extracting influencers, while statistic updates and UCB computations
are negligible.
In Fig. 6, we show the growth of the spread for GT-UCB and baselines. For each experiment,
GT-UCB uses K = 50 if L = 1 and K = 100 if L = 10. First, we can see that MaxDegree is a
strong baseline in many cases, especially for WC and LT. GT-UCB results in good quality spreads
across every combination of network and diffusion model. Interestingly, on the smaller graph HepPh,
we observe an increase in the slope of spread after initialization, particularly visible at t = 50 with
WC and LT. This corresponds to the step when GT-UCB starts to select influencers maximizing
bk(t) in the main loop. It shows that our strategy adapts well to the previous activations, and chooses
good influencers at each iteration. Interestingly, Random performs surprisingly well in many cases,
17
100200300400500Trial10−1100101102103104Runningtime(s)OracleEG-CBRandomMaxDegreeGT-UCB (a) HepPh (WC – L = 1)
(b) DBLP (WC – L = 1)
(c) DBLP (WC – L = 10)
(d) HepPh (TV – L = 1)
(e) DBLP (TV – L = 1)
(f) DBLP (TV – L = 10)
(g) HepPh (LT – L = 1)
(h) DBLP (LT – L = 1)
(i) DBLP (LT – L = 10)
Figure 6: Growth of spreads against the number of rounds.
especially under TV weight assignment. However, when certain influencers are significantly better
than others, it cannot adapt to select the best influencer unlike GT-UCB. EG performs well on
HepPh, especially under TV weight assignment. However, it fails to provide competitive cumulative
spreads on DBLP. We believe that EG tries to estimate too many parameters for a horizon T = 500.
After reaching this time step, less than 10% of all nodes for WC, and 20% for TV, are activated.
This implies that we have hardly any information regarding the majority of edge probabilities, as
most nodes are located in parts of the graph that have never been explored.
6.3 Experiments on Twitter
We continue the experimental section with an evaluation of GT-UCB on the Twitter data, introduced
as a motivating example in Section 3. The interest of this experiment is to observe actual spreads,
instead of simulated ones, over data that does not provide an explicit influence graph.
18
100200300400500Trial0.00.51.01.52.0In(cid:131)uenceSpread×104OracleEG-CBRandomMaxDegreeGT-UCB100200300400500Trial0.01.02.03.04.0In(cid:131)uenceSpread×104100200300400500Trial0.00.20.50.81.01.21.51.8In(cid:131)uenceSpread×105100200300400500Trial0.01.02.03.04.05.0In(cid:131)uenceSpread×103OracleEG-CBRandomMaxDegreeGT-UCB100200300400500Trial0.00.20.40.60.81.0In(cid:131)uenceSpread×105100200300400500Trial0.00.20.40.60.81.01.2In(cid:131)uenceSpread×105100200300400500Trial0.00.51.01.52.0In(cid:131)uenceSpread×104OracleRandomMaxDegreeGT-UCB100200300400500Trial0.01.02.03.04.05.0In(cid:131)uenceSpread×104100200300400500Trial0.00.51.01.52.0In(cid:131)uenceSpread×105 Figure 7: Twitter spread against rounds: (left) L = 1 (right) L = 10.
From the retweeting logs, for each active user u – a user who posted more than 10 tweets – we
select users having retweeted at least one of u's tweets. By doing so, we obtain the set of potentially
influenceable users associated to active users. We then apply the greedy algorithm to select the
users maximizing the corresponding set cover. These are the influencers of GT-UCB and Random.
MaxDegree is given the entire reconstructed network (described in Table 1), that is, the network
connecting active users to re-tweeters.
To test realistic spreads, at each step, once an influencer is selected by GT-UCB, a random cascade
initiated by that influencer is chosen from the logs and we record its spread. This provides realistic,
model-free spread samples to the compared algorithms. Since Twitter only contains successful
activations (re-tweets) and not the failed ones, we could not test against EG, which needs both kinds
of feedback.
In Fig. 7, we show the growth of the diffusion spread of GT-UCB against MaxDegree and
Random. Again, GT-UCB uses K = 50 if L = 1 and K = 100 if L = 10. We can see that GT-UCB
outperforms the baselines, especially when a single node is selected at each round. We can observe
that MaxDegree performs surprisingly well in both experiments. We emphasize that it relies on
the knowledge of the entire network reconstructed from retweeting logs, whereas GT-UCB is only
given a set of (few) fixed influencers.
6.4 Influencer fatigue
We conclude the experimental section with a series of experiments on Twitter data and taking into
account influencer fatigue.
In a similar way to Section 6.3, we compute the set of potentially influenceable users (the support)
associated to all active users – the set of all users who retweeted at least one tweet from the active
user. We then choose 20 influencers as follows: we take the 5 best influencers, that is, the 5 active
users with the largest support; then, the 51st to 55th best influencers, then, the 501st to 505th best
influencers, and finally the 5 worst influencers. By doing so, we obtain a set of 20 influencers with
diverse profiles, roughly covering the possible influencing outcomes. Ideally, a good algorithm that
takes into account influencer fatigue, would need to focus on the 5 best influencers at the beginning,
but would need to move to other influencers when the initially optimal ones start to lose influence
due to fatigue.
We compare Fat-GT-UCB to GT-UCB, which does not use the influence fatigue function, and
19
100200300400500Trial0.01.02.03.04.0In(cid:131)uenceSpread×104RandomMaxDegreeGT-UCB100200300400500Trial0.00.51.01.52.02.53.03.5In(cid:131)uenceSpread×105RandomMaxDegreeGT-UCB Figure 8: Fat-GT-UCB vs competitors on Twitter logs with (left) γ1, (right) γ2.
to the Random baseline. As in Section 6.3, when an algorithm selects an influencer, we can choose a
random spread from the logs (belonging to the selected influencer), and we can now simulate the
fatigue by removing every user in the spread with probability γ(n), where n is the number of times
the influencer has already been played.
√
We show the results of this comparison in Fig. 8. We tested with two different weariness functions,
namely γ1(n) = 1/n and γ2(n) = 1/
n. We can see that, in both scenarios, Fat-GT-UCB performs
the best, showing that our UCB-like approach can effectively handle the notion of influencer fatigue
in the OIMP problem. Unsurprisingly, GT-UCB performs better with weariness function γ2 than is
does with γ1: the former has a lower diminishing impact and thus, the penalty of not incorporating
fatigue is less problematic with γ2.
7 Other related work
We have already discussed in Section 1 the main related studies in the area of influence maximization.
For further details, we refer the interested reader to the recent survey in [2], which discusses the
pros and cons of the best known techniques for influence maximization. In particular, the authors
highlight that the Weighted Cascade (WC) instance of IC, where the weights associated to a node's
incoming edges must sum to one, leads to poor performance for otherwise rather fast IC algorithms.
They conclude that PMC [31] is the state-of-the-art method to efficiently solve the IC optimization
problem, while TIM+ [36] and IMM [35] – later improved by [30] with SSA – are the best current
algorithms for WC and LT models.
Other methods have been devised to handle the prevalent uncertainty in diffusion media, e.g.,
when replacing edge probability scores with ranges thereof, by solving an influence maximization
problem whose robust outcome should provide some effectiveness guarantees w.r.t. all possible
instantiations of the uncertain model [20, 9].
Methods for influence maximization that take into account more detailed information, such as
topical categories, have been considered in the literature [11, 3, 39]. Interestingly, [32] experimentally
validates the intuition that different kinds of information spread differently in social networks, by
relying on two complementary properties, namely stickiness and persistence. The former can be
seen as a measure of how viral the piece of information is, passing from one individual to the next.
The latter can be seen as an indicator of the extent to which repeated exposures to that piece of
20
100200300400500Trial0.01.02.03.04.05.0Cumrewards×103Fat-GT-UCBGT-UCBRandom100200300400500Trial0.00.20.40.60.81.01.2Cumrewards×104Fat-GT-UCBGT-UCBRandom information impact its adoption, and it was shown to characterize complex contagions, of controversial
information (e.g., from politics).
8 Conclusion
We propose in this paper a diffusion-independent approach for online and adaptive influencer
marketing, whose role is to maximize the number of activated nodes in an arbitrary environment,
under the OIMP framework. We focus on scenarios motivated by influencer marketing, in which
campaigns consist of multiple consecutive trials conveying the same piece of information, requiring
as only interfaces with the "real-world" the identification of potential seeds (the influencers) and the
spread feedback (i.e., the set of activated nodes) at each trial. Our method's online iterations are
very fast, making it possible to scale to very large graphs, where other approaches become infeasible.
The efficiency of GT-UCB comes from the fact that it only relies on an estimate of a single quantity
for each influencer – its remaining potential. This novel approach is shown to be very competitive on
influence maximization benchmark tasks and on influence spreads in Twitter. Finally, we extend our
method to scenarios where, during a marketing campaign, the influencers may have a diminishing
tendency to activate their user base.
Acknowledgments
This work was partially supported by the French research project ALICIA (grant ANR-13-CORD-
0020).
References
[1] In Duncan Brown and Nick Hayes, editors, Influencer Marketing. Butterworth-Heinemann,
Oxford, 2008.
[2] A. Arora, S. Galhotra, and S. Ranu. Debunking the myths of influence maximization: An
in-depth benchmarking study. In SIGMOD. ACM, 2017.
[3] N. Barbieri, F. Bonchi, and G. Manco. Topic-aware social influence propagation models. Knowl.
Inf. Syst., 37(3):555–584, 2013.
[4] D. Berend and A. Kontorovich. On the concentration of the missing mass.
In Electronic
Communications in Probability, pages 1–7, 2013.
[5] S. Boucheron, G. Lugosi, P. Massart, and M. Ledoux. Concentration inequalities : a nonasymp-
totic theory of independence. Oxford university press, 2013.
[6] D. Brown and S. Fiorella. Influence Marketing: How to Create, Manage, and Measure Brand
Influencers in Social Media Marketing. Always learning. Que, 2013.
[7] S. Bubeck and N. Cesa-Bianchi. Regret analysis of stochastic and nonstochastic multi-armed
bandit problems. Foundations & Trends in ML, 2012.
[8] S. Bubeck, D. Ernst, and A. Garivier. Optimal discovery with probabilistic expert advice: finite
time analysis and macroscopic optimality. Journal of Machine Learning Research, 14(1):601–623,
2013.
21
[9] W. Chen, T. Lin, Z. Tan, M. Zhao, and X. Zhou. Robust influence maximization. In SIGKDD,
pages 795–804, 2016.
[10] W. Chen, Y. Wang, Y. Yuan, and Q. Wang. Combinatorial multi-armed bandit and its extension
to probabilistically triggered arms. JMLR, 17(1), 2016.
[11] N. Du, L. Song, H. Woo, and H. Zha. Uncover topic-sensitive information diffusion networks. In
AISTATS, pages 229–237, 2013.
[12] D. Easley and J. Kleinberg. Networks, Crowds, and Markets - Reasoning About a Highly
Connected World. Cambridge University Press, 2010.
[13] Paul Gillin. The New Influencers: A Marketer's Guide to the New Social Media. Quill Driver
Books, Sanger, CA, 2007.
[14] M. Gomez-Rodriguez, D. Balduzzi, and B. Scholkopf. Uncovering the temporal dynamics of
diffusion networks. In ICML, pages 561–568, 2011.
[15] M. Gomez-Rodriguez, J. Leskovec, and A. Krause. Inferring networks of diffusion and influence.
ACM Trans. Knowl. Discov. Data, 5(4), February 2012.
[16] M. Gomez-Rodriguez, J. Leskovec, and B. Scholkopf. Structure and dynamics of information
pathways in online media. In WSDM, pages 23–32, 2013.
[17] I. J. Good. The population frequencies of species and the estimation of population parameters.
Biometrika, 40(3-4):237, 1953.
[18] A. Goyal, F. Bonchi, and L. Lakshmanan. Learning influence probabilities in social networks.
In WSDM, pages 241–250, 2010.
[19] P. Grabowicz, N. Ganguly, and K. Gummadi. Distinguishing between topical and non-topical
information diffusion mechanisms in social media. In ICWSM, pages 151–160, 2016.
[20] X. He and D. Kempe. Robust influence maximization. In SIGKDD, pages 885–894, 2016.
[21] D. Kempe, J. Kleinberg, and ´E. Tardos. Maximizing the spread of influence through a social
network. In SIGKDD, pages 137–146. ACM, 2003.
[22] Paul Lagr´ee, Olivier Capp´e, Bogdan Cautis, and Silviu Maniu. Effective large-scale online
influence maximization. In ICDM, 2017.
[23] Kamiu
Lee.
Influencer
marketing
2.0:
Key
trends
in
2017.
https://influence.bloglovin.com/influencer-marketing-2-0-key-trends-in-2017-a5fb97424cd,
2017.
[24] S. Lei, S. Maniu, L. Mo, R. Cheng, and P. Senellart. Online influence maximization. In SIGKDD,
2015.
[25] Nir Levine, Koby Crammer, and Shie Mannor. Rotting bandits.
In Advances in Neural
Information Processing Systems 30 (NIPS), 2017.
[26] Jonathan Louedec, Laurent Rossi, Max Chevalier, Aur´elien Garivier, and Josiane Mothe.
Algorithme de bandit et obsolescence : un mod`ele pour la recommandation. 2016.
22
[27] D. McAllester and L. Ortiz. Concentration inequalities for the missing mass and for histogram
rule error. JMLR, 4:895–911, 2003.
[28] D. McAllester and R. Schapire. On the convergence rate of good-turing estimators. In COLT,
pages 1–6, 2000.
[29] Q. Mei, J. Guo, and D. Radev. Divrank: The interplay of prestige and diversity in information
networks. In SIGKDD, 2010.
[30] 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, 2016.
[31] N. Ohsaka, T. Akiba, Y. Yoshida, and K. Kawarabayashi. Fast and accurate influence maxi-
mization on large networks with pruned monte-carlo simulations. In AAAI, 2014.
[32] D. Romero, B. Meeder, and J. Kleinberg. Differences in the mechanics of information diffusion
across topics: Idioms, political hashtags, and complex contagion on twitter. In WWW, pages
695–704, 2011.
[33] K. Saito, R. Nakano, and M. Kimura. Prediction of information diffusion probabilities for
independent cascade model. In KES, pages 67–75, 2008.
[34] Cole
Sletten.
fatigue.
https://www.forbes.com/sites/onmarketing/2017/01/24/how-to-prepare-brands-for-influencer-
fatigue, 2017.
influencer
brands
for
How
to
prepare
[35] Y. Tang, Y. Shi, and X. Xiao. Influence maximization in near-linear time: A martingale approach.
In SIGMOD, pages 1539–1554, 2015.
[36] Y. Tang, X. Xiao, and Y. Shi. Influence maximization: Near-optimal time complexity meets
practical efficiency. In SIGMOD, pages 75–86, 2014.
[37] S. Vaswani, B. Kveton, Z. Wen, M. Ghavamzadeh, L. Lakshmanan, and M. Schmidt. Diffusion
independent semi-bandit influence maximization. In ICML, 2017.
[38] S. Vaswani, V.S. Lakshmanan, and M. Schmidt. Influence maximization with bandits. In
Workshop NIPS, NIPS '15, 2015.
[39] S. Wang, X. Hu, P. Yu, and Z. Li. Mmrate:
inferring multi-aspect diffusion networks with
multi-pattern cascades. In SIGKDD, pages 1246–1255, 2014.
[40] D. Watts. Six Degrees: The Science of a Connected Age. W. W. Norton, NY, 2003.
[41] D. Watts and P. Dodds.
Influentials, networks, and public opinion formation. Journal of
Consumer Research, 34(4):441–458, 2007.
[42] Z. Wen, B. Kveton, and M. Valko. Influence maximization with semi-bandit feedback. Technical
report, 2016.
23
Lemma 2 (Bennett's inequality (Theorem 2.9 and 2.10 [5])). Let X1, . . . , Xn be independent
random variables with finite variance such that Xi ≤ b for some b > 0 for all i ≤ n. Let
A Useful Lemmas
S :=(cid:80)n
This implies that, P(cid:16)
i=1 (Xi − E[Xi]) and v :=(cid:80)n
(cid:112)2v log 1/δ + b
E[X 2
log E(cid:2)etS(cid:3) ≤ v
S >
i=1
3 log 1/δ
i ]. Writing φ(u) = eu − u − 1, then for all t > 0,
b2 φ(bt) ≤
(cid:17) ≤ δ.
2(1 − bt/3)
vt2
.
Lemma 3 (Lemma 7 – [4]). Let n ≥ 1, λ ≥ 0, p ∈ [0, 1] and q = (1 − p)n. Then,
qeλp(1−q) + (1 − q)e
−λpq ≤ exp(pλ2/(4n))
qeλp(q−1) + (1 − q)eλpq ≤ exp(pλ2/(4n))
(5)
(6)
B Analysis of the Waiting Time of GT-UCB Algorithm
Proof. Denote by Xs(x) := Us−1(x)
can easily verify that
Lemma 4. For any s ≥ 3, P(cid:16) Rs ≤ Rs−1 − λ
s−1 − Us(x)
e(s−2) −(cid:113) 2λ
s−1 . We can rewrite Rs−1 − Rs =(cid:80)
(cid:18) 1
v(x) := E(cid:2)Xs(x)2(cid:3) = p(x)(1 − p(x))s−2
(cid:32)
(cid:114) 2λ
Rs−1 − Rs ≥ E(cid:104) Rs−1 − Rs
s−1 log(1/δ) − 1
(cid:19)
− 1 − p(x)
≤ p(x)
s − 1
s ≤ 1
Let t > 0. By applying Lemma2, one obtains
s − 1
(cid:33)
(cid:105)
3(s−1) log(1/δ)
(cid:17) ≤ δ.
≤ δ.
P
λ
+
3(s − 1)
e(s−2) , that is, E[ Rs−1 − Rs] ≤
We conclude remarking that E[Xs(x)] = p(x)2(1 − p(x))s−2 ≤ p(x)
e(s−2) .
log(1/δ) +
s − 1
log(1/δ)
λmin ≥ 13. Then, for any α ∈(cid:2) 13
λmin , 1(cid:3), if we define τ∗ := T ∗(cid:0)α − 13
Theorem 3 (Waiting time). Denote λmin := mink∈[K] λk and λmax := maxk∈[K] λk. Assume that
1 − 2K
λmax ,
(cid:1), with probability at least
λmin
x∈A Xs(x) and
(7)
1
s
.
TUCB(α) ≤ τ
∗ + Kλmax log(4τ
∗ + 11Kλmax) + 2K.
Proof. Let us define the following confidence bounds:
(cid:114)
(cid:114)
b+
k,s(t) := (1 +
−
k,s(t) := (1 +
b
√
√
2)
2)
−
k,t(t) :=
c
λ
e(s − 2)
+
3λk log(2t)
s
3λk log(2t)
(cid:114) 6λk log(t)
s
s − 1
+
+
log(2t)
s
,
log(2t)
s
+
λk
s
, and
+
log(t)
s − 1
.
24
Let S > 0. Using these definitions, we introduce the following events:
(cid:110)∀k ∈ [K],∀t > S,∀s ≤ t, Rk,s − b
(cid:110)∀k ∈ [K],∀s ≥ S, Rk,s ≥ Rk,s−1 − c
(cid:111)
−
k,s(t)
,
F :=
G :=
E := F ∩ G.
k,s(t) ≤ Rk,s ≤ Rk,s + b+
−
k,s(t)
(cid:111)
,
Using Theorem 1, Lemma 4 and a union bound, one obtains P(E) ≥ 1 − 2K
S (by setting δ ≡ 1
t3 ).
Indeed,
P(cid:0) ¯E(cid:1) ≤ P( ¯F) + P( ¯G) ≤ 2
K(cid:88)
(cid:88)
(cid:88)
k=1
t>S
s≤t
(cid:88)
t>S
1
t3 = 2K
1
t2 ≤ 2K
S
.
In the following, we work on the event E. Recall that we want to control TU CB(α), the time at
which every influencer attains a remaining potential smaller than α following GT-UCB strategy. We
aim at comparing TU CB(α) to T ∗(α), the same quantity following the omniscient strategy. With
that in mind, one can write:
TU CB(α) = min(cid:8)t : ∀k ∈ [K], Rk,Nk(t) ≤ αλk
(cid:9) ,
∗(α) =
T
∗
k (α), where T
T
k (α) = min{s : Rk,s ≤ αλk} .
∗
K(cid:88)
k=1
Following ideas from [8], we can control TU CB(α) by comparing it to U (α) defined below, and
which replaces the remaining potential by an upper bound on the estimator of the remaining potential
(the Good-Turing estimator). Indeed, recall that we can control this on event F.
(cid:110)
t ≥ 1 : ∀k ∈ [K], Rk,Nk(t) + b+
(cid:111)
k,Nk(t)(t) ≤ αλk
.
U (α) = min
Let S(cid:48) ≥ S. On event E, one has that TU CB(α) ≤ max(S(cid:48), U (α)). If U (α) ≥ S(cid:48), one has
Rk,Nk(U (α)) ≥ Rk,Nk(U (α)) − b
−
k,Nk(U (α))(U (α))
k,Nk(U (α))(U (α)) − c
≥ Rk,Nk(U (α))−1 − b
−
−
k,Nk(U (α))(U (α))
k,Nk(U (α))(U (α)) − c
−
−
k,Nk(U (α))(U (α))
(we are on event F and U (α) > S(cid:48) ≥ S)
(where are on event G)
k,Nk(U (α))−1(U (α))
(cid:17) − b
≥(cid:16)
αλk − b+
The third inequality's justification is more evolved. Let t be the time such that Nk(t) = Nk(U (α))− 1
and Nk(t + 1) = Nk(U (α)). This implies that k is the chosen expert at time t, that is, the one
maximizing the GT-UCB index. Moreover, since t < U (α), one knows that this index is greater
than αλk.
If Nk(U (α)) ≥ S(cid:48) + 2, some basic calculations lead to
Rk,Nk(U (α)) ≥ αλk − 11
λk log(2U (α))
S(cid:48)
− 3 log(2U (α))
S(cid:48)
− 3λk
2S(cid:48)
We denote by λmax := maxk λk. If we take S(cid:48) = λmax log(2U (α)), we can rewrite the previous
inequality as
(cid:114)
Rk,Nk(U (α)) ≥ αλk − 11 − 3
λmax − 3
2
25
Thus, by definition of T ∗
k (α), and if λmax > 6, one gets
α − 13
λk
Nk,U (α) ≤ T
∗
k
(cid:18)
Finally, if we denote by λmin = mink λk, we obtain that
U (α) ≤ K(S
(cid:48) + 2) + T
We now apply Lemma 5. We obtain that
(cid:19)
∗(cid:18)
(cid:48) + 2.
+ S
(cid:19)
.
α − 13
λmin
∗ + Kλmax log (8K + 4τ
U (α) ≤ 2K + τ
We conclude with TU CB(α) ≤ max(S(cid:48), U (α)).
Lemma 5 (Lemma 3 from [8]). Let a > 0, b ≥ 0.4, and x ≥ e, such that x ≤ a + b log x. Then one
has
∗ + 11Kλmax) + 2K.
∗ + 10Kλmax) ≤ τ
∗ + Kλmax log (4τ
Moreover, we add that if b ≥ 3, then x ≤ a + b log(2a + 5b).
x ≤ a + b log(2a + 4b log(4b)).
C Confidence intervals in the influencer fatigue setting
In this section, we consider a single influencer and omit its index k. We recall that we make the
assumption that influencers have non-intersecting support. Thus, after selecting the influencer n
times, the remaining potential can be rewritten
(cid:88)
u∈A
Rn =
1{u never activated }pn+1(u),
– n + 1 because this is the remaining potential for the n + 1th spread – and the corresponding
Good-Turing estimator is
n (u) =(cid:80)n
where U γ
1{X1 = . . . = Xi−1 = Xi+1 = . . . = Xn = 0, Xi = 1} γ(n+1)
γ(i)
.
i=1
(cid:88)
u∈A
1
n
Rn =
U γ
n (u),
Estimator bias. Lemma 6 shows that the estimator of the remaining potential for the influencer
fatigue setting is hardly biased.
Lemma 6. Denoting λ =(cid:80)
u∈A p(u), the bias of the remaining potential estimator is
E[Rn] − E[ Rn] ∈
−γ(n + 1)
(cid:20)
(cid:21)
λ
n
, 0
.
Proof. We have that
E[U γ
n (u)] =
n(cid:88)
i=1
pi(u)
(cid:89)
j(cid:54)=i
(1 − pj(u))
γ(n + 1)
γ(i)
= pn+1(u)
26
n(cid:88)
(cid:89)
i=1
j(cid:54)=i
(1 − pj(u)).
We now can compute the bias of the estimator:
E[Rn] − E[ Rn] =
=
1
n
1
n
(1 − pj(u)) − n(cid:88)
i=1
(cid:89)
j(cid:54)=i
(1 − pj(u))
n(cid:88)
n(cid:89)
(cid:89)
n(cid:88)
n(cid:88)
j(cid:54)=i
i=1
i=1
j=1
i=1
pn+1(u)
u∈A
(cid:88)
(cid:88)
(cid:88)
(cid:34)(cid:88)
u∈A
u∈A
E
u∈A
= − 1
n
= − 1
n
pn+1(u)
(1 − pj(u))[1 − pi(u) − 1]
(cid:89)
(cid:35)
j(cid:54)=i
∈
pn+1(u)
pi(u)
pn+1(u)Un(u)
(1 − pj(u))
(cid:80)
(cid:20)
−
n
u∈A pn+1(u)
(cid:21)
, 0
Note that the random variable Un(u) correspond to the hapax definition given in the original OIMP
problem, that is, Un(u) = 1{u activated exactly once}.
Unsurprisingly, we obtain the same bias for the case where γ is constant equal to 1 (no fatigue).
Confidence Intervals. To derive an optimistic algorithm, we need confidence intervals on the
remaining potential. We operate in three steps:
u∈A
U γ
n (u)
n . We have next that
=
i=1
i=1
(1 − pj(u))
pi(u)
E[U γ
pn+1(u)
n (u)2] =
(1 − pj(u))
(cid:89)
(cid:89)
j(cid:54)=i
1. Good-Turing deviations: Remember that Rn =(cid:80)
Thus, we have that v :=(cid:80)
n(cid:88)
n(cid:88)
E(cid:104) U γ
(cid:32)
Rn − E(cid:104) Rn
(cid:105) ≥
where λn := γ(n)(cid:80)
2. Remaining potential deviations: Remember that Rn =(cid:80)
The same inequality can be derived for left deviations.
j(cid:54)=i
≤ np(u)γ(n + 1).
2λn+1 log(1/δ)
(cid:105) ≤
u∈A p(u).
u∈A pn+1(u)
(cid:114)
n (u)2
n2
1
3n
(cid:80)
u∈A
P
n
.
n
γ(n + 1)2
γ(i)2
γ(n + 1)
γ(i)
(cid:33)
Applying Bennett's inequality to the independent random variables {X γ
n(u)}u∈A yields
+
log(1/δ)
≤ δ,
(8)
u∈A Zn(u)pn+1(u) where Zn(u) =
1{u never activated} = 1{X1(u) = ··· = Xn(u) = 0}. We denote Yn(u) = pn+1(u)(Zn(u) −
27
u∈A
P (Rn − E[Rn] ≥ ) ≤ e
i=1(1 − ps(u)). We have that
E[Zn(u)]) and qn(u) = P(Zn(u) = 1) =(cid:81)n
etYn(u)(cid:105)
E(cid:104)
(cid:16)P(Zn(u) = 1)etpn+1(u)(1−qn(u)) + P(Zn(u) = 0)e
−tpn+1(u)qn(u)(cid:17)
(cid:16)
(cid:18) pn+1(u)t2
(cid:80)
u∈A pn+1(u) ),
−t (cid:89)
−t (cid:89)
−t (cid:89)
−t (cid:89)
qn(u)etpn+1(u)(1−qn(u)) + (1 − qn(u))e
Minimizing on t, we obtain (for t =
(cid:19)
≤ e
u∈A
u∈A
u∈A
= e
= e
exp
4n
2n
−tpn+1(u)qn(u)(cid:17)
(by Eq. 9 in Lemma 7)
P (Rn − E[Rn] ≥ ) ≤ exp
(cid:18)
(cid:80)
−2n
u∈A pn+1(u)
(cid:19)
.
We can proceed similarly to obtain the left deviation.
Putting it all together, we obtain the following confidence intervals in Theorem 4, which can be
used in the design of the optimistic algorithm Fat-GT-UCB.
Theorem 4. With probability at least 1 − δ, for λn = γ(n)(cid:80)
(cid:113) λn+1 log(4/δ)
+ 1
3n log 4
δ , the following holds:
n
u∈A p(u) and βn := (cid:0)1 +
2(cid:1) ×
√
Lemma 7 (Adaptation of Lemma 3.5 in [4]). Let n ≥ 1, p ∈ [0, 1], γ : N → [0, 1] a non-increasing
function and t ≥ 0. We denote pn = γ(n)p and qn =(cid:81)
i≤t(1 − pi).
−βn − λ
n
≤ Rn − Rn ≤ βn.
(a) qnetpn(1−qn) + (1 − qn)e
−tpnqn ≤ exp
(b) qnetpn(qn−1) + (1 − qn)etpnqn ≤ exp
(cid:19)
(cid:18) pnt2
(cid:19)
(cid:18) pnt2
4n
4n
(9)
(10)
(cid:19)
.
t2p2
n
n = (1 − pn)n. Clearly, qn ≤ q(cid:48)
n.
Proof. Let q(cid:48)
(a) Using Theorem 3.2 in [4] with p ≡ qn and t ≡ tpn, we have that,
1 − 2qn
(cid:18)
qnetpn(1−qn) + (1 − qn)e
−tpnqn ≤ exp
4 log((1 − qn)/qn)
So it suffices to show that
(1 − 2qn)t2p2
n/4 log((1 − qn)/qn) ≤ pnt2/4n,
or equivalently,
(1 − 2qn)pn log((1 − qn)/qn) ≤ log(1 − pn)/ log(q
(cid:48)
n).
28
Rearranging, we obtain,
L(qn, q
(cid:48)
n) :=
(1 − 2qn) log(1/q(cid:48)
n)
log((1 − qn)/qn)
≤ 1/ log(1 − pn)
pn
:= R(pn).
As is [4], we show that L ≤ 1 ≤ R. The second inequality is true (see [4]).
the left-hand side is negative and thus, L(qn, q(cid:48)
right-hand side as L(qn, q(cid:48)
in [4]. This concludes the proof.
The left-hand side can be written L(qn, q(cid:48)
n) ≤ (1−2qn) log(1/qn)
log((1−qn)/qn)
n). Clearly, L2(q(cid:48)
n) ≥ 0. If L1(qn) ≤ 0,
n) = L1(qn)L2(q(cid:48)
n) ≤ 1. Else, L1(qn) ≥ 0, we can upper bound the
(because qn ≤ q(cid:48)
n), which is proven to be less than 1
(b) It is shown in the proof Lemma 3.5 (b) in [4] that
(cid:105) ≤ 1
t2p2
n
t2pn
4 log qn/ log(1 − pn)
=: R(qn).
(cid:104)
L(t) :=
1
t2p2
n
log
−tpn(1−qn) + (1 − qn)etpnqn
qne
It suffices to show that R(qn) ≤ R(q(cid:48)
n) to obtain the desired inequality. This is true because
0 ≤ log(1 − pn)
log qn
≤ log(1 − pn)
log q(cid:48)
n
.
29
|
1707.05213 | 5 | 1707 | 2018-12-23T00:38:21 | Balance of thrones: a network study on Game of Thrones | [
"cs.SI"
] | TV dramas constitute an important part of the entertainment industry, with popular shows attracting millions of viewers and resulting in significant revenues. Finding a way to explore formally the social dynamics underpinning these show has therefore important implications, as it would allow us not only to understand which features are most likely to be associated with the popularity of a show, but also to explore the extent to which such fictional world have social interactions comparable with the real world. To begin tackling this question, we employed network analysis to systematically and quantitatively explore how the interactions between noble houses of the fantasy drama TV series Game of Thrones change as the show progresses. Our analysis discloses the invisible threads that connected different houses and shows how tension across the houses, as measure via structural balance, changes over time. To boost the impact of our analysis, we further extended our analysis to explore how different network features correlate with viewers engagement and appreciation of different episodes. This allowed us to derive an hierarchy of features that are associated with the audience response. All in all, our work show how network models may be able to capture social relations present in complex artificial worlds, thus providing a way to qualitatively model social interactions among fictional characters, hence allowing a minimal formal description of the unfolding of stories that can be instrumental in managing complex narratives. | cs.SI | cs | Balance of thrones: a network study on Game of
Thrones
Dianbo Liu1,2,3,* and Luca Albergante2,4,5,6,7
1Computer Science and Artificial Intelligence Lab, Massachusetts Institute of Technology, Cambridge,MA,USA
2School of Life Sciences, University of Dundee, Dow Street, Dundee, UK
3The Broad Institute of MIT and Harvard,Cambridge,MA,USA
4Computational and Systems Biology of Cancer, Institut Curie, Paris, France
5PSL Research University, Paris, France
6Institut National de la Sant´e et de la Recherche M´edicale U900, Paris, France
7Mines ParisTech, Paris, France
*Corresponds to: [email protected]
ABSTRACT
TV dramas constitute an important part of the entertainment industry, with popular shows attracting millions of viewers and
resulting in significant revenues. Finding a way to explore formally the social dynamics underpinning these show has therefore
important implications, as it would allow us not only to understand which features are most likely to be associated with the
popularity of a show, but also to explore the extent to which such fictional world have social interactions comparable with the
real world. To begin tackling this question, we employed network analysis to systematically and quantitatively explore how the
interactions between noble houses of the fantasy drama TV series Game of Thrones change as the show progresses. Our
analysis discloses the invisible threads that connected different houses and shows how tension across the houses, as measure
via structural balance, changes over time. To boost the impact of our analysis, we further extended our analysis to explore
how different network features correlate with viewers engagement and appreciation of different episodes. This allowed us to
derive an hierarchy of features that are associated with the audience response. All in all, our work show how network models
may be able to capture social relations present in complex artificial worlds, thus providing a way to qualitatively model social
interactions among fictional characters, hence allowing a minimal formal description of the unfolding of stories that can be
instrumental in managing complex narratives.
8
1
0
2
c
e
D
3
2
]
I
S
.
s
c
[
5
v
3
1
2
5
0
.
7
0
7
1
:
v
i
X
r
a
1
Introduction
Network modeling is a powerful tool to represent interactions between different agents (usually called nodes), which can
represent either physical or abstract entities1, 2. Once, a network has been constructed, it can be used to formally explore the
different features observed, and hence to quantitatively assess the modes of interaction among the agents. Social sciences have
a long history of using networks to explore human interactions3, with the classical experiment of Milgram, paving the way to
popular concepts like social hubs and six degrees of separation between individuals.
While many theoretical approaches focused on studying the structures of networks with only one type of interaction (e.g.,
network of friends), in a few cases authors considered more rich models, where each interaction is associated with one of
more interaction types. One of such theories is behind the concept of structural balance4 -- 7. In this theory, the nodes represent
individuals that can form positive or negative relationships with the others. Using a theoretical approach it is also possible to
show that the system will converge toward two ideal states: either all the individual display a positive attitude or the individuals
forms two camps displaying a negative attitude towards each other8.
Network analysis has been extensively applied in the field of social sciences9, but its application to the study of dynamical
changes of a social network over time are more limited, partly due to the complexity of controlling confounding factors such as
uncontrollable external event. We therefore sought to explore the potential of network modeling to explore the isolated fictional
world where the popular American fantasy drama TV series Game of Thrones takes place. The series has been adapted from
George Martin's novel A Song of Ice and Fire10 and has multiple interconnected plot lines and we decided to focus on the story
arc unfolding on the fictional continent "Westeros" where noble houses of the "Seven Kingdoms" fight for the "Iron Throne" or
independence. In employing network analytics to explore the structures and dynamics of the popular TV show, we assessed
whether network theory could be used to quantify the changes associated with the evolution of the story and to gain insight into
the reaction of the viewers to such changes11 -- 14.
Results
0.1 Constructing a social network of houses
Instead of focusing on individual characters of the drama, we decided to model the interactions among large meta-entities, such
as major noble houses, armies, religious groups and other groups with significant political influences. Networks of relationships
among the meta-entities were built for each episode in the first six seasons of the Game of Thrones TV series, in such a way to
explore the changes episode by episode. To provide a more focused interpretation, the analysis was restricted to interactions
that took places in 'Westeros', a frictional isolated continent where most of the actions are localized. Hence, events that took
places beyond 'the wall' and to the east of 'the Narrow Sea' were not taken into consideration.
The complicated relationships among the meta-entities were abstracted and represented using a signed networks. Each node
indicates a meta-entity and each edge represents a political relationship between two entities (Figure 1). A edge was marked as
either positive (colored in blue in the figures) -- if the two entities were on friendly terms during the episode -- or negative
(colored in red in the figures) -- if the relationship was hostile. Only political relationship involving actual interactions were
taken into account. If two entities were indifferent to each other or did not have real interactions, no edge was drawn between
the two nodes.
0.2 Large-scale properties of the networks
As a first step in exploring structural patterns in the relationships among the meta-entities, large-scale network properties were
considered, starting with degree-based measures. The degree of a node in a network is the number of direct edges connecting
this note to the others and is related with the importance of a node in the network15, as node with larger degree interact with
more entities. When the degree of all of the nodes was derived for the network associated with each episode (Figure 2A), we
found that the political entities taking part in most of the story lines, such as 'House of Stark' and 'House of Lannister', are
associated with nodes possessing a very high degrees throughout a large part of the episodes considered. This is intuitively
understandable as a highly active house, or army, is more likely to have interactions with others.
To further this analysis, and find potentially less obvious connections between the houses, we considered the correlation
of degrees among different entities(Figure 2B) and across episodes(Figure 2C). Degree correlation among houses suggested
similarity in dynamics of importances in the story. First of all, it is interesting to note a predominance of positive correlations
in both representations. From the meta-entities points of view, this supports the ideas that most meta-entities tend to adjust
the number of interaction in a coordinated way, suggesting the idea that the proportion of total number of interactions remain
somewhat stable over the episodes. While from the episode point of view, this indicates that, across episodes, the number of
interactions distribute in a somewhat proportional way.
Looking at the dendrograms accompanying the plots it is also possible to observe a number of strongly defined clusters.
One example is the strong correlation between the degrees of the 'House of Frey' and 'House of Tully'. In the show, these
two houses were geologically located close to each other and had complicated political interactions, this justifying the finding.
Another interesting observation is the high correlation between the "the Sparrows" and "Laddy Melisandre", both of which
were representatives of religious group and were heavily involved in political conflicts, which suggested that some common
patterns of how religions influenced on the progresses of the story.
Looking at the correlation episode-wise, our analysis highlights how the different episodes from the same season tend to be
closed together in term of clustering (see row annotations of Figure 2C). Overall, this suggests that different seasons mostly
contained episodes with coherent importance groups, as expected from the flow of the story. Looking at the data from a more
general point of view, we can also observe cross-season connections. For example, season 3 to 5 were quite similar to each
other (see the lower half of the heatmap) while season 1 occupy a rather independent place.
Network theory employs the notion of centrality to measure the extent to which a node is situated at the center of the network
and hence is structurally positioned to influence, either directly or indirectly, a large part of the network. One classical measures
of centrality is betweenness centrality, which measures the number of shortest path passing trough a node15. According to our
analysis, major political entities in the story had high betweenness centrality scores throughout the episodes, as expected. On
the other hand political entities that are geologically isolated by 'the narrow sea' or 'the wall' generally had lower betweenness
centrality scores, e.g. 'the Night's Watch'(Figure 2D). As we did for the degree, we computed the correlation of the betweenness
centrality score between each house and each episode (Figure 2E and 2F). Two medium-size clusters emerged from the analysis
with the bigger one containing the 'House of Lannister' and the political entities that had strong interactions with it (Figure 2E).
Additionally, most episodes in season 3 to 5 formed again a big cluster, consistent with degree analysis.
Assortativity measures the tendency of nodes to connect to similar nodes in the network15, 16 and can be considered for
different properties. For example, Assortativity by degree is the tendency of a node to be directly connected to others with
similar degrees and assortativity by betweenness centrality is the tendency of a node to be directly connected to others with
similar betweenness centrality. In both scenarios, negative assortative scores were observed in most of the episodes (Figure 2G
and Figure2H), indicating that nodes with similar properties tended not to be connected. Not all the entities were present in all
the episodes of the first six seasons. Presence and absence of the entities were summarized in figure 2I.
0.3 Changes of relationships correlated with number of votes and ratings from viewers
Figures 2A and 2D suggested that substantial changes were present in the topological structure of the network as the story
progressed, and hence that the interactions among the meta-entities changed perceivably across episodes. To further investigate
this issue and better interpret the results of our analysis, we classified network changes from one episode to the next into three
categories: 1) 'relationship establishment' (a new edges was added, indicating the establishment of a new relationship), 2)
'relationship flipping' (the sign of a edges was changed, indicating that allies became enemies or enemies became allies), and 3)
'relationship disruption' (an edge was removed, indicating a previously present relationship was removed due to either two
entities ceasing to interact or to the removal of one, or both, of the interacting entities). The establishment of a new relationship
was the most frequent change, followed by flipping and disruption (3A and C). It is also worth mentioning that while most new
relationships were formed between existing political entities (87.5% of the changes), relationship disruption was mostly due to
the elimination of entities (Figure 3 3B).
0.4 Dynamics of structural balance of the network were different from real social networks
Although a network representation is a powerful way to describe the structure of a system of interacting entities, the actual
dynamics of the system depends on many specificities that are abstracted away when a network is produced. Nonetheless, it
is sometime possible to associate certain types of dynamics with particular network structures17 -- 20. In the context of social
interaction, the theory of structural balance has been used to formally explore how positive (e.g., friendship) and negative (e.g.,
hate) interactions can produce situations that are more or less balanced due to the psychological stress of the interacting agents.
Structural balance is grounded in social psychology theories pioneered by Heider in 1940s4 and has been previously used to
study the evolution of human interactions in different contexts8.
Structural balance is based on classifying triads (i.e., a groups of three entities that interact with each other), into balanced
and unbalanced depending on the sign of the interactions. More formally, given two meta-entities of 'Game of Thrones' i and j,
we associate a value to the edge Si j connecting the two entities so that Si j = 1 if they are allies and Si j = −1 if they are enemy.
Then, a triad consisting of three meta-entities i , j and k is balanced if Si j · S jk · Sik = 1 and imbalanced if Si j · S jk · Sik = −1. The
idea of balanced is based on the intuitive notion that if all of the members of the triad are allied (all edge are positive) or two
members of the triad are allied against the third one (exactly one edge is positive), then the situation is balanced and there is not
pressure for the members of the triad to change mode of interaction. In all the other cases, at least one of the member is under a
pressure that will produce an imbalance and hence a likely change8. A network is said to be balanced if each constituent triad is
balanced (Figure 4).
It was suggested that in social and political networks, relationships change to minimize the number of imbalanced triads8.
However, such trend seems to be absent in the network of Game of Thrones. The number and percentage of imbalanced
3/11
triads through the first 6 seasons seemed to fluctuate over time with an average value of 4.75 which corresponded to about 30
percentages of all triads (Figure 5A and 5B). This phenomenon can be interpreted as a strategy by the author, directors and
screenwriters to keep the story engaging, since in a balanced network, no political entity would be pressured to change and the
story would potentially stall. Interestingly, a temporary tendency towards balance was observed at the beginning of the show
when two dominating alliances initiated a total war against each other ('the War of Five Kings'). However this balance was
broken as with the story progresses (or perhaps was broken to allow the story to progress).
The change of friendly relationship to hostile or the introduction of hostile interaction between previous non-interacting
entities were the most common expedients to introduce imbalance in the networks (Figure 5B), while the level of imbalance
was reduced mainly by eliminating hostile interactions via the disappearance of political entities (Figure 5C).
The average numbers of imbalanced triads each house was contributing to over the evolution of the story were also
considered (Figure 5D). Hub of imbalance, and thus thus instability, can be identified by our analysis. These hubs were
associated with houses that are involved in many stories lines of the TV series, such as 'Stark' and 'Lannister'. To better assess
this result, we also explored to which extent, the result observed was compatible with randomized networks (i.e. network
obtained by shuffling interaction between the entities while maintaining the degree of each entity unvaried). This analysis
suggests that most of the entities were taking part in a number of imbalanced triads that is lower than expected by chance alone
and hence that the story introduced a controlled mix of balanced and imbalanced triads, in such a way to prevent the system
from being overwhelmed by tensions, due to a large number of imbalanced triads, and to avoid creating a completely balanced
systems, which would have no pressure to change.
0.5 Strategy to introduce unpredictability
Unpredictable twists in the storyline of a TV show are common expedients to keep viewers engaged and introduce new plot
elements. Using our analytical framework, we sought to quantify the extent to which this is happening in Game of Thrones. To
this end, we classified triads into different types according to types of interaction present (Figure 4). Establishment, elimination
and change of type of all triads present in at least one episode of the show are summarized in Figure 6A. Triads of type 3,
corresponding to balanced situations where two allies are hostile towards a common enemy, were the most frequently affected
by changes, with ¿47% of triad formation, ¿86% of triad state changes and ¿40% of triad disappearance.
One possible explanation for this finding is that triads of type 3 can be transformed into any of the two imbalanced triads
by changing the type of only one edge. Therefore, this type of triads provides a balance starting point that can subsequently
be used to introduce imbalance in a more soft way. Hence type 3 triads can be used to start a story arc and potentially to let
it die. Indeed, we found that a large number of type 3 triads get destroyed, which was a good proxy for the end of a subplot.
Moreover, we can see that many triads changed from, and into, type 3, indicating that the interaction mode supported by this
triad was a common intermediate step (Figure 6A).
Since triads of type 2 and 3 can be both changed into two different types, we reasoned that they could be used as proxy of
the unpredictability of the story (as in both cases two choices are possible, in addition to the removal of the triad). To better
quantify the level of unpredictability as previously defined, we introduced an unpredictability score defined as
U =
T3 + T2
∑i Ti
where Ti is the number of triads of state i. U takes values between 0 and 1, indicating, respectively, low and high unpredictability.
Our analysis suggests that unpredictability peaks in season 2 during "the War of Five Kings" which were two major wars
happened in Game of Thrones.
0.6 Correlation between network properties and viewers' responses
Up to this point, a number of network topological and dynamic features of the Game of Thrones network have been explored.
One question that could be asked is whether these network properties could be used to dissect viewers' responses to the TV
show and therefore estimate how interesting the story is.
Viewers' votes and ratings were both taken into consideration in this study. Number of votes and viewers' ratings are
organized for each episode (Figure ?? a and c). Spearman partial correlation was conducted to understand potential correlation
between votes or ratings with a variety of network properties. Number of nodes, number of edges , number of edge changes ,
percentage of imbalanced triads , number of triads and assortativity were included into this analysis. P values of association of
votes with different network properties are 0.66, 0.99,0.79 ,0.29, 0.68, 0.0017 respectively. The correlation coefficients are
-0.059,0.00041 ,-0.036 ,0.14 ,-0.056 ,0.41 (Figure ?? b). This indicated that network assortativity has a significant positive
correlation with number of vote for each episode. P values of association of ratings with the network properties are 0.064,
0.42,0.88 ,0.98,0.93, 0.14 respectively. The correlation coefficients are -0.25 ,0.11, 0.020, -0.0033, 0.011,0.20 (Figure ?? d).
This shows that larger numbers of nodes of the interaction networks likely correlates with lower ratings.
4/11
1 Discussion
Structures and dynamics of complex systems or stories consisting of many entities and relationships are usually difficult to
understand and simplified model can be very effective in exploring the key properties of such systems. In this article we
followed this philosophy and employed network modeling to quantify structural changes over the story arc of the Game of
Thrones TV show by capturing both topological and dynamics changes in a systematic way.
The relationship network was abstracted at the level of noble houses or political entities to rely on a high level representation
that limits the uncertainty associated with single characters. Degree distribution and centrality of the network were shown
to be an efficient way to identify important political entities. The fact that the major houses of the story, such as "House of
Lannister" and "House of Stark", had much higher numbers of connections than others is consistent with findings in other
networks and provides a formal way to asses their importance in the unfolding of the story. The negative assortativity observed
in the network has structural similarities to natural systems such as marine food web, with minor entities gravitating around
major players. Such structure is also quite dissimilar to human systems such as co-authorship networks. Overall, this seems
to suggest a somewhat fractal organization of the story that allows the authors to concentrate on story arcs tightly associated
with major players, while allowing a large number of minor plots entangled with the mayor story. Hence allowing a multilevel
story-telling that can be engaging for different viewer types and may contribute to boost the interest in the show.
In addition to static topologies, the dynamics of the network was tracked through the 60 episodes composing the first six
seasons of the show. The positive correlation identified between number of edges changes per episode and viewer's votes
suggested that dramatic changes in relationships between houses can make the series more attractive to audiences, perhaps as a
consequence of the reorganization in the balance of powers.
Furthermore, since the edges of the network can be associated with an positive or negative interaction between the entities,
the theory of structural balance can be employed to explore how different types of three-body interactions contribute to the
evolution of the story. This dynamic analysis allows us to explore the preferred story-telling structures in a formally sound
way, highlighting the different importances of the various interaction modes and suggesting effective ways to unfold story in a
natural, but still engaging, way. To our best knowledge, this is the very first time structural balance theory was used to explore
the dynamics of fiction and TV series.
The results from partial correlation analysis suggest that assortivity and number of nodes correlated with viewer's responses.
The assortativity coefficient measures the tendency of nodes to connect to similar nodes in the network. Its correlation with
number of votes possibly reflects clustering of similar houses in sense of interaction patterns raise population of the show
among audiences. On the other hand, the negative correlation between number of nodes and ratings potentially suggest that an
over complex story is not favored by the viewers.
Writing a novel or a TV show is a complex process that requires skills and imagination. Successful authors are able to create
stories that stimulate the interest of the readers or viewers, by engaging fictional worlds with their own life. By employing a
mathematical approaches to analyze such fictional stories, we were able to show how formal methods can be used to provides
new perspectives that complement more classical text analysis. All in all, our works suggest that the methods developed in the
context of network theory can complement the tool sets already available to story writers to plan and explore how the dynamics
of interaction among single fictional entities contribute to the complex web of relationships that support unforgivable stories.
Methods
1.1 Construction of relationship network
Each major political entity located on 'Westeros' or possessing strong connections with entities on 'Westeros' was modeled
as node in a network. If a house split into more than political entities, they were considered different nodes. One example is
the 'House of Baratheon', that split into one army led by Stannis and another army led by Renny. If two nodes have actually
interactions with each other, a edge is included between them. All the edges are associated with a color indicating the type of
interaction. Friendly relationships were marked as positive (blue) and hostile relationships were marked as negative (red)(Figure
1). If two entities did not interacted directly, no edge was included between the corresponding nodes. For certain edges, where
the relationships were complicated, only the effects of real actions were taken into consideration.
1.2 Dynamics of the the network
Changes of nodes and edges were tracked at the resolution of each episode and statuses of nodes and episodes at the end of each
episode were compared. Network analysis in this article were conducted using igraph package in R environment (version3.4).
1.3 Number of votes and rating from viewers
The votes and rating from viewers were downloaded from IMdb http://www.imdb.com/title/tt0944947/eprate
on August the 2nd, 2017. Both quantities were normalized by dividing over the mean of each season. This was done to account
for potential variation in popularity of the show across seasons.
5/11
1.4 Network randomization
To estimate the expected number of imbalanced triad each node was entangled in, signs of edges in each episode were randomly
shuffled. A total of 30 different randomized networks were combined to calculate the expected number of imbalanced triads.
1.5 Data and code availability
The network data and the code used to write the article are available at https://github.com/kaiyuanmifen/
BalanceOfThrones.
References
1. Newman, M. E. The structure and function of complex networks. SIAM review 45, 167 -- 256 (2003).
2. Newman, M., Barabasi, A.-L. & Watts, D. J. The structure and dynamics of networks (Princeton University Press, 2011).
3. Barabasi, A.-L. Linked: How everything is connected to everything else and what it means. Plume Ed. (2002).
4. Heider, F. Attitudes and cognitive organization. The J. psychology 21, 107 -- 112 (1946).
5. Cartwright, D. & Harary, F. Structural balance: a generalization of heider's theory. Psychol. review 63, 277 (1956).
6. Harary, F. & Kabell, J. A. A simple algorithm to detect balance in signed graphs. Math. Soc. Sci. 1, 131 -- 136 (1980).
7. Kunegis, J. et al. Spectral analysis of signed graphs for clustering, prediction and visualization. In Proceedings of the 2010
SIAM International Conference on Data Mining, 559 -- 570 (SIAM, 2010).
8. Antal, T., Krapivsky, P. L. & Redner, S. Social balance on networks: The dynamics of friendship and enmity. Phys. D:
Nonlinear Phenom. 224, 130 -- 136 (2006).
9. Waumans, M. C., Nicod`eme, T. & Bersini, H. Topology analysis of social networks extracted from literature. PloS one 10,
e0126470 (2015).
10. Martin, G. R. A Song of Ice and Fire (Bantam, 1997).
11. Facchetti, G., Iacono, G. & Altafini, C. Computing global structural balance in large-scale signed social networks. Proc.
Natl. Acad. Sci. 108, 20953 -- 20958 (2011).
12. Marvel, S. A., Kleinberg, J., Kleinberg, R. D. & Strogatz, S. H. Continuous-time model of structural balance. Proc. Natl.
Acad. Sci. 108, 1771 -- 1776 (2011).
13. Antal, T., Krapivsky, P. L. & Redner, S. Dynamics of social balance on networks. Phys. Rev. E 72, 036121 (2005).
14. Beveridge, A. & Shan, J. Network of thrones. Math Horizons 23, 18 -- 22 (2016).
15. Newman, M. Networks: an introduction (Oxford university press, 2010).
16. Newman, M. E. Assortative mixing in networks. Phys. review letters 89, 208701 (2002).
17. Alon, U. An Introduction to Systems Biology: Design Principles of Biological Circuits (Chapman & Hall/CRC Mathemati-
cal and Computational Biology) (Chapman and Hall/CRC, 2006), 1 edn.
18. May, R. M. Qualitative stability in model ecosystems. Ecology 54, 638 -- 641 (1973). DOI 10.2307/1935352.
19. Albergante, L., Blow, J. J. & Newman, T. J. Buffered qualitative stability explains the robustness and evolvability of
transcriptional networks. eLife 3, e02863 (2014). DOI 10.7554/eLife.02863.
20. Liu, D., Albergante, L. & Newman, T. J. Universal attenuators and their interactions with feedback loops in gene regulatory
networks. Nucleic Acids Res. 45, 7078 -- 7093 (2017). DOI 10.1093/nar/gkx485. /oup/backfile/content_public/
journal/nar/45/12/10.1093_nar_gkx485/1/gkx485.pdf.
Acknowledgements
We would like to thank Timothy Newman, Rana Al Mamun, and Sam Palmer for interesting discussions and insights.
Author contributions statement
DL constructed the networks, provided original concepts, developed the code to perform the analysis, and wrote the manuscript.
LA provided original concepts, supervised the work, and wrote the manuscript.
6/11
Figure 1. Network of relationships among meta-entities in Game of Thrones. This network models the last episode of season
six. Only activities on the 'Westeros' continent were taken into consideration.
7/11
Figure 2. Large Scale Properties of the networks. (A) Node degree (connections to other node) of each noble houses/political
entity in season 1-6 of Game of Thrones. (B) Correlation between the node degrees by house across the the episodes. (C)
Correlation between the node degrees by episode across houses. (D) Betweenness centrality of each houses/political entity. (E)
Correlation of betweenness centrality by episode across the the houses. (F) Correlation of betweenness centrality by house
across the the episodes. (G) Changes of assortativity by degree over the first six seasons of the TV series. (H) Changes of
assortativity by betweenness over the first six seasons of the TV series. (I) Presence(red) and absence(blue) of different entities
over time.
Figure 3. Changes of edges in the network overtime.(A) Formation of new edges was more frequent than flipping signs and
removal of edges. (B) Most new edges were formed between existing nodes (blue). Most edge removals were coupled with
elimination of political entities (brown). (C) Changes of edges in each episode.
8/11
Figure 4. Structural balance. The complex network of Game of Thrones can be divided into a large number of constituent
triads. If everyone is allied or if any two members are allied against a common enemy, the triad is balanced. If everyone is
hostile to others or a single entity is allied to two entities that are enemies with each other, then the triad is imbalanced. A
network is balanced if each constituent triad is balanced8
9/11
Figure 5. Structural balance of the network. (A) Number of different types of triads (upper panel) and number of imbalanced
triads over time (lower panel).(B) Percentage of balanced (upper panel) and imbalanced triads (upper and lower panel) over
time. (C) Contribution of different types of edge changes to the increase in the number of imbalanced triads. "+" stands for
positive edge , '-' for negative 'edge', and '0' represents the absence of relationship. (D) Contribution of different types of edge
changes to the decrease in the number of imbalanced triads. (E) Average number of imbalanced triads per episode for each
entity. Red bars indicate the real data extracted from 'Game of Thrones' and blue bars indicate the expected number of
imbalanced triads from randomised networks. Error bars indicate standard deviation.
10/11
Figure 6. Statistics of dynamics of triads.(A)Summary of all triads changes in the first six seasons. A large percentage of
changes involved triads of state 3(B)Changes of levels of unpredictability of the network over time. The level reached 100%
during the "the War of Five Kings"
11/11
|
1812.00076 | 1 | 1812 | 2018-11-30T22:18:45 | Scalable Graph Learning for Anti-Money Laundering: A First Look | [
"cs.SI",
"cs.AI",
"cs.LG"
] | Organized crime inflicts human suffering on a genocidal scale: the Mexican drug cartels have murdered 150,000 people since 2006, upwards of 700,000 people per year are "exported" in a human trafficking industry enslaving an estimated 40 million people. These nefarious industries rely on sophisticated money laundering schemes to operate. Despite tremendous resources dedicated to anti-money laundering (AML) only a tiny fraction of illicit activity is prevented. The research community can help. In this brief paper, we map the structural and behavioral dynamics driving the technical challenge. We review AML methods, current and emergent. We provide a first look at scalable graph convolutional neural networks for forensic analysis of financial data, which is massive, dense, and dynamic. We report preliminary experimental results using a large synthetic graph (1M nodes, 9M edges) generated by a data simulator we created called AMLSim. We consider opportunities for high performance efficiency, in terms of computation and memory, and we share results from a simple graph compression experiment. Our results support our working hypothesis that graph deep learning for AML bears great promise in the fight against criminal financial activity. | cs.SI | cs |
Scalable Graph Learning for Anti-Money
Laundering: A First Look
Mark Weber
Jie Chen
Toyotaro Suzumura
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
IBM Research
IBM Research
IBM Research
[email protected]
[email protected]
[email protected]
Aldo Pareja
Tengfei Ma
Hiroki Kanezashi
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
IBM Research
IBM Research
IBM Research
[email protected]
[email protected]
[email protected]
Tim Kaler
Charles E. Leiserson
Tao B. Schardl
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
MIT-IBM Watson AI Lab
MIT CSAIL
[email protected]
MIT CSAIL
[email protected]
MIT CSAIL
[email protected]
Abstract
Organized crime inflicts human suffering on a genocidal scale: the Mexican drug
cartels have murdered 150,000 people since 2006; upwards of 700,000 people per
year are "exported" in a human trafficking industry enslaving an estimated 40 mil-
lion people. These nefarious industries rely on sophisticated money laundering
schemes to operate. Despite tremendous resources dedicated to anti-money laun-
dering (AML) only a tiny fraction of illicit activity is prevented. The research
community can help. In this brief paper, we map the structural and behavioral
dynamics driving the technical challenge. We review AML methods, current and
emergent. We provide a first look at scalable graph convolutional neural networks
for forensic analysis of financial data, which is massive, dense, and dynamic. We
report preliminary experimental results using a large synthetic graph (1M nodes,
9M edges) generated by a data simulator we created called AMLSim. We con-
sider opportunities for high performance efficiency, in terms of computation and
memory, and we share results from a simple graph compression experiment. Our
results support our working hypothesis that graph deep learning for AML bears
great promise in the fight against criminal financial activity.
1 Anti-Money Laundering in 2018
Anti-money laundering (AML) is the task of preventing criminals from moving illicit funds through
the financial system. AML is generally viewed through the lens of regulatory compliance because
the burden of forensic analysis falls primarily on financial institutions, which are responsible for
meeting Know Your Customer (KYC) standards, monitoring transactions, shutting down or restrict-
ing accounts deemed suspect, and submitting timely Suspicious Activities Reports (SARs) to law
enforcement agencies. In the United States, these requirements are primarily stipulated in the Bank
Secrecy Act (1970) and the Patriot Act (2001), with enforcement chiefly by the Financial Crimes En-
NeurIPS 2018 Workshop on Challenges and Opportunities for AI in Financial Services: the Impact of Fairness,
Explainability, Accuracy, and Privacy, Montréal, Canada. This is a non-archival publication - the authors may
submit revisions and extensions of this paper to other publication venues.
forcement Network (FinCEN). Global costs of AML compliance run in the tens of billions of dollars
and have been growing at about 15% per year since 2004 [KMPG, 2014]. Penalties for failed AML
compliance are severe: in 2018, the Commonwealth Bank of Australia unhappily set a new record
with a USD $530 million fine for AML violations involving its "Intelligent Deposit Machines" (a
cautionary reminder that calling things "intelligent" does not make them so) [BBC, 2018].
Meanwhile we are rapidly becoming a cashless society engaged in global economic exchange. The
advent of cryptocurrency has catalyzed a paradigm shift in peer-to-peer transactions and extrana-
tional financial governance. Cryptocurrencies present new challenges but also new opportunities for
AML, and these vary across type (e.g. forensic insight into Bitcoin's public ledger is much better
than trying to track traditional cash, but more privacy-preserving cryptocurrencies like Zcash are far
more difficult to monitor).
Thus, we observe two rival, converging forces: 1) the increasing difficulty of AML in the age of
globalization and cryptocurrency; 2) the acceleration of advanced forensic tools in the age of big
data and artificial intelligence. Our aim is to ensure the second outpaces the first.
1.1 A pressing societal problem
Amidst the legal and econometric detail, the real societal problem too often fades into the back-
ground: tens of millions of innocent people are suffering under the malice of drug cartels, human
trafficking rings, and terrorist organizations. Moreover, the resulting AML efforts of governments
become a force for financial exclusion as financial institutions pass along high compliance costs to
customers and turn more people away; low-income people, immigrants, and refugees are hit the
hardest [Mitchell, 2012].
1.2 The bad guys are winning
Today, we are woefully ineffective at preventing money laundering. Europol estimates only 1% of
illicit funds are confiscated [Europol, 2017]. To understand why, we consider both the technical and
behavioral challenges, as well as the current methods for addressing them.
Technically, we have a needle-in-a-haystack problem of entity classification and hidden pattern dis-
covery in massive, dynamic, high dimensional, time-series transaction data sets with high noise-to-
signal ratios, combinatorial complexity, and non-linearity. Data sets are often fragmented, inaccu-
rate, incomplete, and/or inconsistent, within as well as across organizations. Synthesizing informa-
tion from multimodal data streams is difficult to automate, and so it falls to resource-constrained
human analysts. Behaviorally, the heavy hand of compliance (i.e. the cost asymmetry of a false pos-
itive versus a false negative) motivates over-reporting, leading to a secondary needle-in-a-haystack
problem for resource-constrained government law enforcement.
A recent Reuters survey of c-suite executives at 2,373 large global organizations provides color
on how these challenges play out for stakeholders. An estimated 5% of reviewed transactions are
filed as Suspicious Activities Reports (SARs) and only 10% of those receive a meaningful investi-
gation by law enforcement. Thus 0.5% of criminal activity alerts lead to action. Each organization
transacts with thousands of third-party vendors and millions of clients, with finance and insurance
companies comprising the upper fractiles of the distribution. Internal monitoring is just as impor-
tant; respondents estimated 58% of financial crimes involving human trafficking were internal to
their own organizations [Reuters, 2018].
2 Review of current methods
AML procedure follows five key steps: (1) installation of a compliance organization within the
company including formal AML training for employees, often with external support from special-
ized firms such as Promontory; (2) development and execution of Know Your Customer (KYC)
on-boarding and profile maintenance procedures; (3) account activity oversight and constraints via
transaction monitoring systems (TMS); (4) manual review of flagged accounts and transactions; (5)
filing of Suspicious Activities Reports (SARs) to law enforcement and corresponding restrictive
action against suspect accounts. We consider Steps 2 and 3: transaction monitoring and evaluation.
2
2.1 Transaction Monitoring
Transaction monitoring systems are predominantly rules-based thresholding protocols tuned for vol-
ume and velocity of transactions with tiered escalation procedures. For example, a U.S. bank might
outsource Level 1 (our term) transaction monitoring to a company in India, which flags all trans-
actions above $10,000 for review, including "near misses" such as a transaction of $9,999 and/or
transaction series such as five transactions of $2,000 each in a 24 hour window. A Level 2 analyst
performs a preliminary review of each flag and decides whether or not to escalate to the bank's
headquarters back in the U.S. There, a Level 3 analyst, in coordination with the compliance team,
will conduct a more thorough analysis using sophisticated techniques to determine whether or not a
SAR need be filed and the account in question suspended.
2.2 Forensic Analysis
Suspiciousness heuristics govern the initial review of a flagged account. These include political
exposure, geographic dynamics, round numbers, transaction type and properties (e.g. if the transac-
tion is a purchase of an asset, how liquid is the asset and how stable is it as a store of value?), and
behavioral logic (e.g., something does not add up).
For example, "Account A" is flagged for a transaction of $9,500, a "near-miss" on the $10,000
rule. The Level 2 analyst pulls up the KYC profile and sees the account belongs to U.S. senator
(a Politically Exposed Person or "PEP" in AML parlance) has been credited by a foreign account
in the Bahamas, a money-laundering hotspot. The analyst pulls up more transaction activity and
also sees a $4,000 credit from XYZ Corporation, which the analyst cannot find record of in a quick
Google search (why is a PEP on a mystery corporate's payroll?). The Level 2 analyst deems this
transaction/account suspicious and escalates to the Level 3 analyst for further review.
In this toy scenario, suspiciousness is obvious; reality is far more difficult. First, the aforementioned
needle-in-the-haystack problem, compounded by the threat of fines, produces high false positive
rates. Second, the manual suspiciousness heuristics above do not scale well with time and human
resource constraints (analysts are often significantly backlogged). Third, criminals can be quite
sophisticated in masking the true nature of their transactions with complicated account layering,
multi-hop transactions, and other techniques. More sophisticated AML methods are needed to meet
these challenges.
Graph analytics have emerged as an increasingly important tool for AML analysis because money
laundering involves cash flow relationships between entities i.e. network structures. Multiple graph
types can be constructed. For example, a graph can be formulated where a single account is repre-
sented as a vertex and a single transaction between two accounts is represented as an edge. Another
approach is to represent a group of accounts as a vertex (e.g. such as those under a holding company,
or such as those inferred to share an owner via clustering) and define an edge as the aggregate trans-
action volume with a neighboring node over a period of time. The latter is the predominant method
for forensic analysis of cryptocurrency activity, involving a combination of 1) address clustering
using multiple-input and change heuristics, and 2) address tagging with attribution data gathered
from public sources [Paquet-Clouston et al., 2018]. In the traditional financial system, where KYC
rules assist with entity identification, classical graph algorithms such as cycle detection, PageRank,
degree distribution, egonet, and label propagation are used to identify anomalies, relationships, and
other features of the network that may be deemed suspicious. Working with private bank data, we
have found these methods can reduce false positives by 20 -- 30% when performed by expert analysts.
Natural Language Processing (NLP) is a newer tool with emergent potential for AML. Han et al.
[2018] recently demonstrated how NLP processing of unstructured, heterogeneous data from online
sources can be used in combination with transaction monitoring systems and private bank data to
produce suspiciousness scores and visualizations to augment human forensic analysis. In this pre-
liminary framework, the authors tap open data streams such as news articles, social media, financial
reports, and fraud databases. Then, cross-referencing with the private bank data such as KYC and
transaction activity, they use a sequential combination of sentiment analysis, name entity recogni-
tion, and relation extraction to build knowledge graphs and relation graphs pertinent to the flagged
account. The authors report positive stakeholder feedback on their stated objective to reduce aver-
age investigation times by 30%. We find this work compelling, but in need of more powerful and
efficient graph methods to support the effective and efficient analysis of such graphs at scale.
3
Graph learning using deep neural networks is a new method with great promise for the complex
pattern and feature discovery central to the AML challenge, as well as for the aforementioned mul-
timodal constructions of various graph types. We explore this in more detail in the next section.
3 Deep learning for massive, dynamic graph data
The proliferation of graph analytics in AML illuminates the potential for ever more powerful graph
methods such as graph learning, an emerging sub-field of machine learning.
3.1 Scalable Graph Convolutional Networks (GCN)
Deep learning has achieved remarkable success on Euclidean data, including images, audio, and
video. Deep learning for graph structured data, however, has lagged due to the scalability challenges
of dealing with combinatorial complexity and non-linearity inherent to graphs of any meaningful
size and density. But this is changing.
Recently, a proliferation of graph neural network models have emerged [Bruna et al., 2014,
Defferrard et al., 2016, Li et al., 2016, Hamilton et al., 2017, Gilmer et al., 2017], and several have
made major strides toward scalability with promising experimental efficiency results [Chen et al.,
2018, Ying et al., 2018]. We prioritize FastGCN [Chen et al., 2018] in the next section, as it was
demonstrated on benchmark data sets to surpass peer methods GCN and GraphSAGE by as much
as two orders of magnitude without sacrificing accuracy.
3.2 Graph learning for AML
The challenge of AML involves high dimensional, massive graph data mapping billions of relation-
ships (edges) between millions of entities (nodes). In AML transaction monitoring, a node entity
might be a single account or a set of associated accounts (either known or inferred via clustering to
operate as one). A node could also represent another graph from a previous step in a time series.
The known properties are those explicitly defined in the data, such as information collected in stan-
dard Know Your Customer (KYC) processes or multi-modal data mined from public or partner
information streams, as well as the observable transactions and any associated flags or SARs filed.
The unknown, hidden properties (such as covert criminal relationships and activity) are the subject
of the forensic inquiry; i.e., what more can we ascertain about an entity than the human eye or clas-
sical graph analytics can detect? In the language of deep learning, we begin with an initial, often
quite limited, attribute vector of a node entity, and we aim to arrive at a much more characteristic
feature representation with a higher degree of confidence.
The principle challenge is scalability as an inverse function of graph size, and in AML the graph size
is massive and aways increasing (new accounts and interactions are added constantly, producing an
ever-growing graph, or graphs within graphs).
3.3 Preliminary testing of graph convolutional networks for AML
We have conducted preliminary experiments using graph learning models for AML. We detail one
such experiment here:
Synthetic data: Since transaction data and customer profiles are privacy sensitive, we created a pro-
totype data simulator called AMLSim (to be described in detail in a forthcoming paper). AMLSim
is a multi-agent simulation platform tailored for an AML problem, where each agent behaves as a
bank account transferring money to other agent accounts, and where a small number of agents con-
duct nefarious activity modeled on real-world patterns. AMLSim constructs data in two steps: (1)
generate a graph via NetworkX1 based on a given degree distribution; (2) generate a time-series of
transactions via PaySim [Lopez-Rojas et al., 2016], based on transaction distributions and dynamics
observed in real data. We draw on domain expertise to define for the simulator a list of known sus-
picious patterns. This procedure allows one to generate a massive dynamic graph of arbitrary size
that contains semi-realistic suspicious activities. While meaningful accuracy measures require real
1https://networkx.github.io
4
data, the structurally congruous synthetic data from AMLSim provides a helpful sandbox for agile
research on scalability, explainability, and other requirements.
Graph topology: For this experiment, we used AMLSim to simulate an AML graph with 1M nodes
and 9M edges. Each vertex represents an account with attributes including account number, account
type (e.g. business type), owner name, and date/time created. Each edge has a transaction ID,
amount, and time stamp. The data is sparsely labeled with flagged transactions (i.e. transactions that
violate volume and velocity rules) and SARs (full analysis has confirmed suspiciousness).
Deep learning task: Using the SARs and escalated alerts as labels, the task is semi-supervised
learning of a predictive model to (a) predict the suspiciousness of a given target node, and (b) identify
other potential bad actors in the transaction network via direct or indirect connections to nodes
known to be suspicious. To do this, we conduct feature extraction via graph embedding, then convert
these features into downstream machine learning tasks.
is a workstation with In-
Preliminary experimental results: The evaluation environment
tel(R) Xeon(R) CPU E5-2697 v4 @ 2.30GHz,
total 72 cores (18 cores, 4 sockets) and
1TB DDR4 2400MHz memory. We ran the original, still popularly used deep graph model,
GCN [Kipf and Welling, 2017], with batched training. GCN took 611 seconds to converge in 32
epochs. Next we ran an improved variant, FastGCN [Chen et al., 2018], which reduces the training
cost through neighborhood sampling. FastGCN took 386 seconds to converge in the same number of
epochs. The hyperpameters we used include 128 hidden units, a 0.01 learning rate, and 400 samples
in FastGCN. Overall, FastGCN was nearly twice as fast as GCN for this synthetic graph. Projected
with a linear scaling, with billion-scale AML graphs (e.g., 1B nodes and 9B edges) the training time
would estimate to a few days. We hypothesize the development of high-performance code could im-
prove the training time by additional orders of magnitude for faster turnover given graph dynamism
with frequent transaction updates.
Table 1: Graph learning on AMLSim data (1M nodes, 9M edges)
Method
Epochs Training time
GCN
FastGCN
32
32
611 seconds
386 seconds
Workstation: Intel(R) Xeon(R) CPU E5-2697 v4 @ 2.30GHz, 72 cores, 1TB DDR4 2400MHz memory
3.4 Optimizing for speed and storage with high-performance computing
Deep learning methods for identifying money laundering must be efficient in order to be useful.
Financial institutions handle million transactions per second and require timely identification of
suspicious activities. The efficiency is explored from both the computation and the memory angle.
Sparse computation on graphs: Systems performing inference using deep neural networks for de-
tecting suspicious activities need the ability to update the graph and node representations sparsely in
response to the addition of new transactions. Without dynamic sparse recomputation, a newly added
transaction cannot be classified until the entire graph is recomputed. Such a long delay is unaccept-
able for latency-sensitive applications. Efficient techniques for performing sparse dynamic compu-
tation have been developed for a variety of common graph algorithms including PageRank, loopy
belief propagation, coordinate descent, co-EM, alternating least squares, singular-value decomposi-
tions, and matrix factorizations. For each of these algorithms, sparse dynamic recomputation can
significantly enhance performance in practice.
Memory-efficient graph representations: Storagewise, the parallelism and locality of graph-based
deep learning algorithms can benefit from graph reordering in a variety of ways. These algorithms
typically perform iterative updates on data associated with a vertex, based on data associated with
adjacent vertices. An effective reordering of the graph vertices can improve the data locality of
these algorithms, allowing them to use the machine's memory system more efficiently.
In ad-
dition, reordering schemes can improve the compressibility of the graph using difference-coding
techniques [Blandford and Blelloch, 2002]. Finally, many efficient and scalable techniques for
processing graphs in parallel, including chromatic scheduling [Kaler et al., 2016], DAG schedul-
ing [Hasenplaugh et al., 2014, Gruevski et al., 2018], and serial blocking [Gruevski et al., 2018],
5
benefit from effective graph reorderings, because reordered graph layouts can be leveraged to dra-
matically reduce synchronization overheads.
Preliminary experimental results: We used the graph compressor, Ligra+ [Shun et al., 2015], and
experimented with the simulated AML graph as well as a few deep learning benchmark graphs
used in Kipf and Welling [2017]. Ligra+ is a compression system that achieves reduced space usage
while maintaining competitive or even improved performance with respect to running the algorithms
on uncompressed graphs on multicore machines. The compression results are reported in Table 2.
We see a compression ratio as high as 2x, although larger graphs appear to be more challenging to
compress. This finding opens up the opportunity for memory-efficient deep neural network training
and inference leveraging compression and reordering techniques.
Table 2: Graph compression with Ligra+
Graph
Size
Compressed Size
Cora
59 Kb
61 Kb
Citeseer
Pubmed
461 Kb
AMLSim 21 Mb
28 Kb
29 Kb
252 Kb
13 Mb
4 Summary
In summary, we have motivated the challenge of anti-money laundering not only in terms of legal
compliance, but in terms of meaningful societal impact. The human suffering from organized crime
such as drug cartel violence and human trafficking is immense, as are the higher order negative ef-
fects such as costly penalties for banks and financial exclusion for innocent people on the margins.
We have encouraged the technical research community to help financial institutions and law enforce-
ment harness artificial intelligence technologies to fight this problem. We have reviewed current
AML methods and provided a first look at deep learning for transaction monitoring and analysis.
We have shared preliminary experimental results using synthetic AML data generated from AML-
Sim. Lastly, we have reviewed some computational considerations involved in implementing graph
learning for AML at scale.
6
References
BBC. Commonwealth bank offers to pay record fine in laundering case, June 2018.
D. Blandford and G. Blelloch. Index compression through document reordering. In Data Compres-
sion Conference, 2002.
Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally
connected networks on graphs. In ICLR, 2014.
Jie Chen, Tengfei Ma, and Cao Xiao. Fastgcn: Fast learning with graph convolutional networks via
importance sampling. In ICLR, 2018.
Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on
graphs with fast localized spectral filtering. In NIPS, 2016.
Europol. From suspicion to action -- converting financial intelligence into greater operational impact.
Financial Intelligence Group, 2017.
Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural
message passing for quantum chemistry. In ICML, 2017.
Pregrad Gruevski, William Hasenplaugh, David Lugato, and James Thomas. Laika: Efficient in-
place scheduling for 3d mesh graph computations. In Proceedings of the 30th ACM Symposium
on Parallelism in Algorithms and Architectures, 2018.
William L. Hamilton, Rex Ying, and Jure Leskovec.
Inductive representation learning on large
graphs. In NIPS, 2017.
Jingguang Han, Utsab Barman, Jeremiah Hayes, Jinhua Du, Edward Burgin, and Dadong Wan.
Nextgen aml: Distributed deep learning based language technologies to augment anti money
laundering investigation.
In Proceedings of ACL 2018, System Demonstrations, pages 37 -- 42.
Association for Computational Linguistics, 2018.
William Hasenplaugh, Tim Kaler, Tao B. Schardl, and Charles E. Leiserson. Ordering heuristics for
parallel graph coloring. In Proceedings of the 26th ACM symposium on Parallelism in algorithms
and architectures, 2014.
Tim Kaler, William Hasenplaugh, Tao B. Schardl, and Charles E. Leiserson. Executing dynamic
data-graph computations deterministically using chromatic scheduling. ACM Trans. Parallel Com-
put., 3(1), 2016.
Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional net-
works. In ICLR, 2017.
KMPG. Global anti-money laundering survey 2014, 2014.
Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural
networks. In ICLR, 2016.
Edgar Lopez-Rojas, Ahmad Elmir, and Stefan Axelsson. Paysim: A financial mobile money simu-
lator for fraud detection. In 28th European Modeling and Simulation Symposium, 2016.
Daniel J. Mitchell. World bank study shows how anti-money laundering rules hurt the poor. Forbes,
April 2012.
Masarah Paquet-Clouston, Bernhard Haslhofer, and Benoit Dupont. Ransomware payments in the
bitcoin ecosystem. CoRR, abs/1804.04080, 2018. URL http://arxiv.org/abs/1804.04080.
Reuters. Revealing the true cost of financial crime: 2018 survey report, 2018.
Julian Shun, Laxman Dhulipala, and Guy E. Blelloch. Smaller and faster: Parallel processing of
compressed graphs with ligra+. In Proceedings of the 2015 Data Compression Conference, 2015.
Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec.
Graph convolutional neural networks for web-scale recommender systems. In KDD, 2018.
7
|
1604.08507 | 1 | 1604 | 2016-04-28T16:41:29 | Graph Decompositions Analysis and Comparison for Cohesive Subgraphs Detection | [
"cs.SI",
"cs.DM",
"cs.DS"
] | Massive networks have shown that the determination of dense subgraphs, where vertices interact a lot, is necessary in order to visualize groups of common interest, and therefore be able to decompose a big graph into smaller structures. Many decompositions have been built over the years as part of research in the graph mining field, and the topic is becoming a trend in the last decade because of the increasing size of social networks and databases. Here, we analyse some of the decompositions methods and also present a novel one, the Vertex Triangle k-core. We then compare them and test them against each other. Moreover, we establish different kind of measures for comparing the accuracy of the decomposition methods. We apply these decompositions to real world graphs, like the Collaboration network of arXiv graph, and found some interesting results. | cs.SI | cs | Graph Decompositions Analysis and Comparison
for Cohesive Subgraphs Detection
Etienne CALLIES
X2011 - École Polytechnique
[email protected]
Tomás YANY ANICH
X2011 - École Polytechnique
[email protected]
6
1
0
2
r
p
A
8
2
]
I
S
.
s
c
[
1
v
7
0
5
8
0
.
4
0
6
1
:
v
i
X
r
a
Abstract-Massive networks have shown that the de-
termination of dense subgraphs, where vertices interact
a lot, is necessary in order to visualize groups of com-
mon interest, and therefore be able to decompose a big
graph into smaller structures. Many decompositions
have been built over the years as part of research in
the graph mining field, and the topic is becoming a
trend in the last decade because of the increasing size
of social networks and databases. Here, we analyse
some of the decompositions methods and also present a
novel one, the Vertex Triangle k-core. We then compare
them and test them against each other. Moreover, we
establish different kind of measures for comparing
the accuracy of the decomposition methods. We apply
these decompositions to real world graphs, like the
Collaboration network of arXiv graph, and found some
interesting results.
Index Terms-Graph decomposition, massive net-
works, community detection and evaluation, k-core,
triangle k-cores, vertex triangle k-core.
I. INTRODUCTION
Graphs representing real data are nowadays evolv-
ing at a speed we can hardly control. They are
bigger than before, they grow up faster and they
are more complex as well. The information boom
in the last decade has been a crucial motivation
for understanding and analysing in a more efficient
way large graphs. Social networks, for example, are
represented with graphs that can go up to billions
of vertices, and more. Therefore, it is compulsory
to build new tools that would allow us to identify
smaller structures, denser subgraphs or at least more
connected vertices that could represent communities
inside this big pool: the graph itself.
In fact, clustering is an important problem in graph
mining. How to determine if there are, and which
are they, groups inside a large graph? Learning this
represent a considerable issue in two senses: how to
identify and calculate in an effective and efficient way
this dense subgraphs, and how to determine if the
results are accurate as we would expect, i.e. if they in
fact reveal real communities. Once we have answered
this, we can no longer preoccupy ourselves with the
big and complex graph, but more with its smaller
decompositions which represent better and denser
communities. This could be useful, for instance, for
targeted advertising, news clustering and others.
Many methods, that we will review, have been
presented over the last few years. We no longer found
graph's description in the vertices' degree1. Today,
we need better tools that could make the difference
between kinds of communities –if we are talking
about social graphs– or clusters, in the general case.
In this paper we will review some of the principal
modern decompositions techniques, analyse their
algorithms, compare them and evaluate them. The
computing time will be important in our task since
we need algorithms not only to be good, but also to
be efficient. Our contributions lie on the regrouping
and side by side comparison between existing, and
an innovative, decomposition techniques, and on a
novel metric for evaluation of the detected important
subgraphs.
We will be working with a simple undirected graph
G = (V,E) where V is the set of vertices and E the
set of edges. Moreover, we will consider the degree
dG(v) of a vertex v ∈ V to be the number of edges
e ∈ E incident to the vertex itself. Some others
assumptions and definitions will be explained when
needed.
1It can be shown that high degree nodes can coexist with low
degree nodes.
2
II. GRAPH DECOMPOSITIONS
We would begin with some important definitions
that, although are typical among the computer science
community, are always necessary to clarify.
Definition 1. A graph G is said to be complete when
every of its vertices is connected to each other, i.e.,
∀(x,y) ∈ V 2,x (cid:54)= y,∃e ∈ E with e = {x,y}.
When we are thinking of finding the most con-
nected subgraph inside a graph, we are somehow
thinking on the clique problem, i.e., finding cliques
in a graph G. The definition of a clique might in
fact vary according to the literature, so we prefer to
establish the one we think it is the most used.
Definition 2. Let G(cid:48) ⊂ G be a subgraph of a graph
G. We say that G(cid:48) is a clique if it is a complete
subgraph of G. We say that G(cid:48) is a clique of degree
k if it is a clique where each vertex has degree k
Our quest should be then to find maximal size
cliques in a graph G, in a way, to solve the clique
problem. This presents two majors issues, as said by
[1], cliques of small size are too numerous for having
any kind of special interest, and those cliques which
are big enough for being interesting are too rare to
find. The second problem is the complexity of the
clique problem: listing every possible size clique –if
there are any– can be computed in exponential time.
Therefore, finding other kind of interesting sub-
graphs of a graph, which could represent dense and
almost complete structure, without being a clique,
appear to be an important task, and a useful one,
especially if we can show that this task can be
computed in polynomial time. We will here review
some of the most known decomposition one can find
in literature.
We will use Figure 1 as our canonic example
through the whole paper.
A. k-core
As defined in [2], we use the following notations.
Definition 3. Let G be a simple undirected graph. We
define ∆(G) = minv∈V dG(v), the minimum degree
of a vertex in G.
Figure 1. Simple undirected graph G0
Figure 2. k-core decomposition of G0
We can immediately notice that this approach is
more relaxed than the clique's one. In fact, k-cores
are restrained subgraphs, but not as much as cliques.
We ask them to be as connected as we want them
to –varying the k– but we don't ask them to be the
most connected. Nevertheless, k-cores represent very
connected structures, as said in [3], they are "seedbeds,
within which cohesive subsets can precipitate out".
Therefore their study become important as we will
see in the following parts of this paper. Moreover,
some interesting analysis can be done for each vertex
of a graph when a k-core decomposition is made.
Definition 5. Let v ∈ V be a vertex of a graph G.
The core number of v is the highest k for which v
belongs to the k-core of G.
This last definition will be very important in our
examples because it will allow us to considerate what
we will call levels in a decomposition.
Definition 6. Given a graph decomposition, we call
level the set of vertices having the same core number.
Having this definition set, we can define what a
k-core is.
Definition 4. Let G(cid:48) ⊂ G be a subgraph of a graph
G. We say that G(cid:48) is the k-core of G if it is the
maximum size subgraph of G where ∆(G(cid:48)) ≥ k.
Figure 2 shows a k-core decomposition over the
initial graph G0. The numbers next to the vertices
represent the core number. Different colours here
represent different levels.
Remark 7. Some interesting assets can be detected
in k-cores. In fact, as shown in [4], k-cores have the
following properties:
• k-cores are nested, i.e., let k < k(cid:48) and G,G(cid:48)
k-core and k'-core of a graph G ⇒ G(cid:48) ⊆ G.
• k-cores can have more than one connected
component.
Proposition 8. If we consider a connected k-core
G(cid:48) = (V (cid:48),E(cid:48)) of a graph G = (V,E) and we define
n =(cid:12)(cid:12){v,v ∈ V (cid:48)}(cid:12)(cid:12) then, if k = n − 1, G(cid:48) is a clique.
Proof: If k = n − 1 then each vertex on the
k-core is connected to all of its neighbours. This is
the definition of a clique.
Remark 9. It is clear that if the k-core has more than
one connected component then the result can be apply
to each one of them.
This last proposition show us that a k-core is very
close to a clique, without being specifically a clique
itself. Knowing this, we will develop in the following
parts of this paper some measures for the difference
between a deduced subgraph and a clique, and see
how close were we to find a proper clique. Actually,
with the definition of a clique of degree k given before,
it is trivial that, if such clique of degree k exists in a
graph G, then it will be a subset of a k-core G(cid:48) of
G.
B. Generalized cores
The idea behind k-cores can be extended. When we
talked about k-cores we were considering the function
∆(G) as the minimum degree of a vertex in a graph
G. What if we consider another different function to
which we ask some properties? This generalization
as seen in [4] can allow us to define more general
decompositions than k-cores.
Definition 10. Let Z = {z : z ⊆ V }. Let p : V ×
Z → R be a function. We say that p is a vertex
property function, or a p-function.
We can now define the notion of p-core at level t,
p = (V t
which come to extend the idea of k-core.
Definition 11. Let Gt
a graph G. We say that Gt
if and only if
• ∀v ∈ V t
• V t
p : t ≤ p(v,V t
p )
p is a maximal subset with this property
p ,Et
p) be a subgraph of
p is a p-core at level t ∈ R
Remark 12. We can notice that the function used for
the k-core decomposition is in fact a particular p-
3
function, where for v ∈ V,U ⊆ V we have p(v,U ) =
minv∈U dU (v).
As said before, we need to ask some particular
properties to this p function so it can be coherent
with what we expect from decompositions.
Definition 13. Let p be a vertex property function. We
say that p is monotone if and only if for U ⊂ U(cid:48) ⊂ V
and ∀v ∈ U we have p(v,U ) ≤ p(v,U(cid:48)).
This definition allow us to enunciate an important
property for decompositions.
Proposition 14. Let p be a monotone function. p-
cores are nested, i.e., for t2 < t1 ⇒ V t1
p ⊆ V t2
p
p are the vertices of the p-core at
where V t1
level t1 and t2 respectively.
p and V 2
Proof: Let t2 < t1,v ∈ V t1
p ). But, by definition, V t2
p
p . By definition we
p ). Besides, t2 < t1. Therefore,
is the maximal
p ⊆
have t1 ≤ p(v,V t1
t2 < p(v,V t1
subset U such that t2 ≤ p(v,U ). Therefore V t1
p .
V t2
Example 15. Many functions can be chosen for
constructing p-core at level t graphs.
get the k-cores defined before.
• p1(v,V (cid:48)) = dG(cid:48)(v). With this function one can
• p2(v,V (cid:48)) = number of different cycles of length
3 (triangles) going through v. We will analyse
this particular function in a future part of this
paper.
following p-functions: p3(v,V (cid:48)) = din
p4(v,V (cid:48)) = dout
• For a directed graph G we can consider the
G(cid:48)(v) and
G(cid:48) (v).
C. Triangle k-core and k-truss
Triangle k-cores and k-trusses –which we will show
are the same– are very important decompositions for
our analysis. In fact, latest studies [5], [6] have shown
that in order to look up for connected subgraphs inside
a graph we need to search for vertices that not only
are connected between them, but that have common
neighbours. This last feature is actually represented
through triangle k-cores and k-trusses.
Definition 16. We call a triangle a cycle of length 3
inside a graph. Let G(cid:48) be a subgraph of a graph G.
We say that G(cid:48) is triangle k-core of G if each edge
of G(cid:48) is contained within at least k triangles in G(cid:48)
itself.
We define as well what a k-truss is.
4
Figure 3. Triangle k-core decomposition of G0
Figure 4. Vertex triangle k-core decomposition of G0
Definition 17. Let G(cid:48) be a subgraph of a graph G.
We say that G(cid:48) is a k-truss of G if each edge of G(cid:48)
connects two vertices who have at least k−2 common
neighbours.
Figure 3 shows the triangle k-core decomposition
in our graph G0.
Proposition 18. A k-truss is a triangle (k − 2)-core.
Proof: In a k-truss, in order to have k − 2
common neighbours, every edge must be reinforced
by at least k − 2 pairs of edges: they form therefore
k − 2 triangles.
Although some literature has been written for one
and the other kind of decomposition, we have just
shown that they get the same result at the end. Because
of this, we will be working the whole time with the
triangle k-core term.
Proposition 19. Triangle k-core are nested sub-
graphs, i.e., k2 < k1 ⇒ triangle k1-core ⊆ triangle
k2-core.
Proof: Let e ∈ Ek1, i.e. e is contained within k1
triangles. Since k2 < k1, e is also contained in k2
triangles. So e ∈ Ek2.
Proposition 20. If G has a clique of degree k > 1
then it is contained within a (k + 1)-truss, i.e. within
a triangle (k − 1)-core.
Proof: Let C = (VC,EC) be a clique of degree
k > 1 and K = (Vk,Ek) a (k+1)-truss within a graph
G = (V,E). Since each vertex in C has degree k, then
VC = k + 1. Therefore, for each e ∈ EC, e = {u,v},
u and v will be neighbours with every other vertex in
C, in particular, they will share (k + 1) − 2 common
neighbours. With this we have EC ⊆ Ek. Since a
graph can be reduced to its edges set, when isolated
vertices don't exist, we have just shown that C ⊆ K.
D. Vertex triangle k-core
We wanted to find another decomposition following
the philosophy behind the triangle k-core decom-
position. Therefore, instead of considering triangles
around an edge, we will look for triangles around a
vertex. With this, we have the same idea as mentioned
before: we try to identify dense subgraphs in terms
of vertices having common neighbours.
Definition 21. Let G(cid:48) be a subgraph of a graph G.
We say that G(cid:48) is a vertex triangle k-core if every
vertex in G(cid:48) is contained in at least k triangles.
Proposition 22. Vertex triangle k-core are nested
subgraphs.
Proof: Using the proposition 14 we just have to
show that p(v,U ) = number of triangles through v,
is a monotone function.
Let U1 ⊆ U2 ⊆ V . Since U2 has more vertices than
U1 we have two options: whether this new vertices
are creating triangles going through v and therefore
p(v,U1) < p(v,U2), whether they are not creating any
new triangle going through v and therefore p(v,U1) =
p(v,U2). In any case, we have p(v,U1) ≤ p(v,U2).
Figure 4 shows the vertex triangle k-core decom-
position of G0. We notice immediately that this
decomposition is different from the (edge) triangle
k-core decomposition.
Remark 23. The maximum core number for vertex
triangle decomposition is (n−1)(n−2)
whereas for the
(edge) triangle decomposition it goes up to n − 2.
Moreover, for the k-core decomposition it goes only
up to n − 1.
2
III. ALGORITHMS
Once we have described the possible decomposi-
tions one can apply to a graph, we would like to
review the principal algorithms that are associated
with these subgraphs detection. We chose to do it
in a separate part because although definitions and
properties are always valid, algorithms can change
during time and can in particular be improved (in
the best scenario). We will review algorithms for
k-core, triangle k-core and vertex triangle k-core.
For any further analysis we could make, these three
decomposition are the only ones to think of.
The general procedure for each algorithm will be
the same. In fact, since we will be asking to vertices
–or edges– to have some properties, the method will
almost always be to delete those vertices or edges that
do not fulfil the condition we are asking. Repeating
this process will get us to the final subgraph we are
looking for. But the idea for us will be to actually
memorize the core number for each vertex, therefore
our result will not be a subgraph but a numbering
over the graph vertices.
In our study we consider the complexity as the
evolution of the number of elementary operations in
relation to the number of vertices in the input graph
G = (V,E). The complexity is said quadratic (resp.
cubic) if the number of operations grows like n2 (resp.
n3) where n = V is the number of vertices.
A. k-core
One of the great advantages of k-core decompo-
sition is it easy way to compute it. The idea of
the algorithm will be to take a graph G and delete
recursively all vertices of degree less than k. With
them, we delete as well every edge incident with the
concerned vertices.
In order to evaluate the complexity of the algo-
rithm, we will first consider the initialization, then
the emptying of unprocessedVertices which contain
vertices unprocessed by default, and also neighbours
vertices added after a remove.
Initialization. The calculation of degree is
quadratic in terms of vertices, because it is linear
in terms of edges.
Copied in unprocessedVertices. The k number
goes from 0 to n − 1. There are therefore in all no
more than n2 vertices copied in unprocessedVertices.
Unprocessed after a remove. When a vertex is
removed, all its neighbours are added to unprocessed-
Vertices. There are therefore in total no more than n2
vertices added to unprocessedVertices after a remove.
⇒ Global complexity: O(n2)
5
Algorithm 1 k-core decomposition algorithm
forall the v ∈ V do
d(v) ← # of edges that contain v
end
k ← 0
while V (cid:54)= ∅ do
k ← k + 1
unprocessedV ertices = copy(V )
while unprocessedV ertices (cid:54)= ∅ do
Vertex v = pop(unprocessedV ertices)
if d(v) < k then
forall the neighbours w of v do
d(w) ← d(w) − 1
unprocessedV ertices.add(w)
end
v.coreNumber← k − 1
remove v from V
end
end
end
B. Triangle k-core
Triangle k-cores are a bit more difficult to compute
than k-cores since we have to considerate triangles
through edges. Nevertheless, the general form of the
algorithm remains similar.
Initialization. The calculation of degree is
quadratic, and the calculation of triangle(f) for an
edge f is linear, so globally cubic complexity.
Copied in unprocessedEdges. The k number goes
from 0 to n − 2. There are therefore in total no more
than n3 edges copied in unprocessedEdges.
Unprocessed after a remove. When an edge is
removed, all other edges which formed a triangle with
the removed edge are added to unprocessedEdges.
The number of such edges is not more than twice the
number of vertices. There are therefore in total no
more than n3 edges added to unprocessedEdges after
a remove.
⇒ Global complexity: O(n3)
C. Vertex triangle k-core
Vertex triangle k-core are almost as difficult to
compute as (edge) triangle k-core, with the advantage
we only go through vertices and never look for edges.
Initialization. The calculation of triangle(v) is
quadratic for a vertex v, so globally cubic complexity.
Copied in unprocessedVertices. The k number
goes from 0 to (n−1)(n−2)
. There are therefore in total
no more than n3 edges copied in unprocessedVertices.
2
Algorithm 2 Triangle k-core decomposition algo-
rithm
forall the v ∈ V do
d(v) ← # of edges that contain v
Algorithm 3 Vertex triangle k-core decomposition
algorithm
forall the v ∈ V do
triangle(v) ← # of triangles that contain v
6
end
forall the e ∈ E do
triangle(e) ← # of triangles that contain e
end
k ← 0
while E (cid:54)= ∅ do
k ← k + 1
unprocessedEdges = copy(E)
while unprocessedEdges (cid:54)= ∅ do
Edge e = pop(unprocessedEdges)
if triangle(e) < k then
forall the triangle t of e do
for both other edges f of t do
triangle(f ) ← triangle(f ) − 1
unprocessedEdges.add(f )
end
end
k ← 0
while V (cid:54)= ∅ do
k ← k + 1
unprocessedV ertices = copy(V )
while unprocessedV ertices (cid:54)= ∅ do
Vertex v = pop(unprocessedV ertices)
if triangle(v) < k then
forall the triangle v of t do
for both other vertices w of t do
triangle(w) ← triangle(w)− 1
unprocessedV ertices.add(w)
end
end
v.coreNumber← k − 1
remove v from V
What should a "good" graph decomposition look
like? If you consider the core with the highest k, that
we will call the best community, we would expect it
to be limited in terms of vertices. The cohesiveness of
the best community should be close to the one of the
clique composed by the same vertices. But finding a
good community is not all, the graph decomposition
should list wider communities –not as good and small
as the best one–, but still interesting enough for
being analysed. The speed and the regularity of the
increase (resp. decrease) of the number of vertices
while decreasing (resp. increasing) the level number,
is fundamental for our analysis.
In our study, we focused on these two aspects : the
best community and the core size decrease. A few
examples of graph decompositions is enough to notice
the most important point of this whole study. No graph
decomposition –of the three we are trying to compare–
is better than the other in any case. We cannot assert
anything in absolute terms about the best community
size, neither can we about the core size decrease. For
a big set of graphs we calculate several individual
measures. Then average values of these metrics are
analysed and compared. The measures we made are
detailed in the corresponding paragraphs. First we
precise on which graphs we run our decompositions
end
for both vertices v of e do
d(v) ← d(v) − 1
if d(v) = 0 then
v.coreNumber← k − 1
end
remove e from E
end
end
end
end
end
end
end
Unprocessed after a remove. When a vertex is
removed, all other vertices which formed a triangle
with the removed vertex are added to unprocessed-
Vertices. There are therefore in total no more than n3
vertices added to unprocessedVertices after a remove.
⇒ Global complexity: O(n3)
IV. EXPERIMENTAL EVALUATION
How to compare two graph decompositions? What
do we measure? How can we assert that a decom-
position is better than another one? Once we have
decided our comparison criteria, on which and on how
many graphs do we run tests? In this part, we propose
to answer this questions according to certain criteria
we chose. This doesn't necessarily mean that we got
the best and unique answer –which we think doesn't
actually exists– but that we analysed the problem
within a perspective.
7
best community size increases. It has to be mentioned
that the classical density, i.e. the number of vertices
divided by the number of edges, is not relevant for
this study, because with this definition a clique could
have a lower density than a subgraph in which it is
included, which is a situation we do not want.
Figure 5. Best level size comparison
in order to compare them.
A. Graph samples
Exhaustive Graph. The first possibility is to
2) =
generate all graphs with n vertices. There are 2(n
2n(n−1)/2 graphs with n vertices. We made it for 5
vertices –1024 graphs– and for 6 vertices –32,768
graphs– but for 7 vertices –2,097,152 graphs– it
exceeds the capacity of our software.
(cid:1) = n(n−1)
of vertices n, there are(cid:0)n
Random Graph. To avoid memory issues, we
decided to test on random graph. Given a number
possible edges.
We add each edge to the graph with a probability of
1/2. We construct 10,000 graphs this way. We did the
study for many different n, but we kept the results
for 9, 15, 25 and 45 vertices.
2
2
Real Graph. Real graphs are as "random" as those
we built. For example, in social network there is a high
probability of having the friend of someone's friend be
their friend. In other words, there is a high probability
of a triangle. We have tested our algorithms on several
real graphs. We had some memory issues, due to the
kind of Java objects we use in our algorithms. The
biggest real graph we managed to decompose in the
three shown ways, and the only one we analysed, is
called ca-HepPh which can found in [8]. It is the graph
of Collaboration network of arXiv in High Energy
Physics, it contains 12,008 nodes and 118,489 edges.
B. Best community
Figure 6. Best level clique-density comparison
In order to discuss about graph decomposition
algorithm efficiency for finding the best community,
we define two kinds of accuracy measures.
Definition 24. We define the best level size as the
number of vertices with the highest core number.
Remark 25. For the three decompositions of the
preceding example G0, the best level size is 5.
Definition 26. We define best level clique-density
as the number of edges divided by the number of
possible edges in the best community, i.e.
.
n
(n−1
2 )
Example 27. In the example G0, the best community
is a clique, so the density is 1.
Expected. A "good" graph decomposition should
have a little best level size mean and higher best level
density. It is obvious that both ideas are equivalent.
The best communities given by different decompo-
sition are nested. Thus the density decreases as the
Results. As seen in Figure 5, the vertex triangle
decomposition seems to be the decomposition that
finds the shortest community, on average, in very
small graphs; triangle k-core in wider but still small
graphs; finally vertex triangle again. The k-core
decomposition finds bigger communities in small
graphs but the gap is not as significant when the
size increases. Unfortunately, on all tests we have
done on real graphs, the best community was the
same with the three decompositions, and was always
a clique. We tried to increase the number of vertices
in random tests, but the hierarchy on finding the best
community on average was always the same: first the
vertex triangle, then the triangle k-core, and then k-
core at the end. As for the clique-density (Figure 6), it
is higher when the community size is smaller, which
is the case for the vertex triangle decomposition.
Regarding the search of the best communities, both
triangle decompositions are better than k-core. But
it is still questionable to assert that vertex triangle
k-core is better than (edge)-triangle k-core.
the number of not empty levels increases. For these
two reasons we calculate RMS, and we expected it
to be low for "good" decomposition.
8
C. Core size decrease
After considering the best community, we should
observe the rest of the vertices.
Definition 28. We call a core the set of vertices that,
given an integer k, have core number ≤ k.
In how many and which levels are held the
remaining vertices? Are they well distributed into
all k-levels or abruptly bonded together in only few
levels? Which graph decomposition enables the most
regular core size decrease? We decided to proceed
with the following measures :
Highest core number. In the preceding example
it is 4 for k-core (Figure 2), 3 for triangle k-core
(Figure 3) and 6 for vertex triangle (Figure 4).
Definition 29. We define the level number as the
quantity of not empty levels.
Remark 30. In the example for G0 it is 3 for k-core, 4
for triangle k-core and 5 for vertex triangle. It can be
seen graphically as the number of different colours.
Definition 31. We define the Root Mean Square
(RMS) as the quadratic mean of the distribution of
the vertices in the different levels
n(k)2
n2
(cid:115)(cid:88)
k
Figure 7. Highest core number comparison
Figure 8. Level number comparison
where n(k) is the number of vertices in the k-core
subgraph and n is the total number of vertices.
Remark 32. In the example for G0 it is 0.63 for
k-core, 0.61 for triangle k-core and 0.59 for vertex
triangle.
Expected. The highest core number is not an
effective measure. The reason is that
the vertex
triangle decomposition lets a lot of empty levels. If
you consider only not empty levels, the core number
for these levels rises quadratically. However, the
level number tells about the possibility of vertex
segmentation. More levels means more different ways
to choose the wideness of communities. It is quite
useless to have very small levels or very big levels,
in terms of size. Ideally, the levels have all the same
size. On the one hand, the RMS is minimal when it is
the case. On the other hand, the RMS is lower when
Figure 9. Root Mean Square comparison
Results. As provided, the highest core number, as
seen in Figure 7, is bigger on average in the vertex
triangle decomposition than the other two, because
the best core number evolution is quadratic for it
and linear for the other two. What was not obvious
is that the highest core number for triangle k-core
is significantly lower than the k-core. As for the
level number, as seen in Figure 8, it is clearly the
vertex triangle decomposition which has the best. If
the triangle k-core is better in that way than the k-
core in small graphs, the real graph confirms that
the tendency goes on the other way while graph size
increases. Triangle k-core decomposition offers far
less different levels than the k-core decomposition.
Our study has not made it clear about the root
mean square. Considering the random tests, it appears
that both triangle decompositions are fighting for the
first place for RMS, as seen in Figure 9, with a
slight advantage for vertex triangle. This tendency
is confirmed when random tests are done with more
than 45 vertices. The real graph shows it differently.
Vertex triangle is still leading, but challenged by k-
core, and triangle k-core RMS values are far bigger.
Given these results we ran test on other real graphs
to check that it was not an isolated characteristic
of the chosen graph. Maybe the fact that triangle k-
core decomposition has less levels, explains its bigger
RMS.
Figure 10. Core size decrease for random 15 vertices graph
9
the average over 10,000 graphs. Figure 11 show as
well the core size decrease as k increases, for the
real arXiv graph, with the difference that we don't
considerate k if there are no vertices with exactly k
as their core number.
Considering these curves of core size decrease, the
vertex triangle decomposition (green curve) seems to
have the slowest and most regular decrease, on both
random and real graphs. As for k-core and triangle
k-core, it is difficult to tell. The core size for k-core
is bigger for little k but decreases abruptly. It is more
useful to decrease slowly for high k, in order to
distinguish good and very good communities rather
than bad and very bad ones.
Regarding core size decrease, the best decomposi-
tion is the vertex triangle decomposition in various
aspects. It has more levels, quite regularly filled,
which enables to select a community with a more
precise size. The k-core and triangle k-core are both
bad at it, but the triangle k-core is more regular in
its core size decrease.
D. Algorithm execution time
In order to verify the complexity of the three
algorithms, we ran them on random graphs with
different number of vertices. The way of constructing
random graph is the same as described in IV-A. For
a number of vertices, the program creates randomly
1000 graphs and the three algorithms calculate the
decompositions. The mean is done on the 1000
execution time. These means established for number
of vertices from 10 to 40, are reported in Figure 12.
Figure 11. Distinct core size decrease for arXiv graph
Figure 12. Execution time for the three algorithms
Figure 10 show the decrease of core size while
k increases. For this last, values are considered as
As expected, both triangle decompositions take
far more time than k-core. To check whether the
complexity is cubic for triangle decompositions and
quadratic for k-core decomposition, the figures 13 and
14 show the evolution of execution times of algorithms
divided by their complexity (O(n2) or O(n3)).
10
triangle k-core, better than the (edge) triangle k-core:
core size decrease for vertex triangle k-core is more
regular and it converges to a smaller community. In
a future, some new p-functions for generalized core
decomposition could perhaps give us better results.
ACKNOWLEDGMENT
We thank our professor Michalis Vazirgiannis,
and his postdoctoral researcher Fragkiskos Malliaros,
who proposed the subject and provided insight and
expertise that greatly assisted the research, although
they may not agree with all of the interpretations of
this paper.
REFERENCES
[1] J. Cohen, "Trusses: Cohesive Subgraphs for Social Network
Analysis", Technical Report (National Security Agency),
2009.
[2] C. Giatsidis, D.M. Thilikos & M.Vazirgiannis, "Evaluating
Cooperation in Communities with the k-core Structure",
ASONAM 2011, pp. 87–93.
[3] S.B. Seidman, "Network Structure and Minimum Degree",
Social Networks, vol. 5, pp. 269–287, 1983.
[4] V. Batagelj & M. Zaveršnik, "Generalized Cores", ACM,
2002.
[5] J. Wang & J. Cheng, "Truss Decomposition in Massive
Networks", VLDB Conf., 2012.
[6] Y. Zhang & S. Parthasarathy, "Extracting Analysing and
Visualizing Triangle k-core Motifs within Networks", ICDE
Conf., 2012.
[7] V. Batagelj & M. Zaveršnik, "An O(m) Algorithm for Cores
Decomposition of Networks", 2002.
[8] arXiv.
(2003, April). High Energy Physics
- Phe-
nomenology collaboration network [Online]. Available:
http://snap.stanford.edu/data/ca-HepPh.html
Figure 13. Quadratic Verification
Figure 14. Cubic Verification
The k-core decomposition algorithm is definitely
quadratic, but both triangle algorithms we used are
a little more than cubic. It can be explained by
the fact we used ArrayList Java objects for a more
comfortable implementation, whose basic operations
are not executed in constant time.
V. CONCLUSIONS
We have shown that k-core decomposition are
efficient but not very precise: it can be computed
in a quadratic time –versus cubic time for the others
decompositions– but the core size decrease is not very
regular. Sacrificing complexity for results, we got a
far more useful decomposition using the novel vertex
|
1908.09120 | 3 | 1908 | 2019-09-02T05:50:59 | Intellectual and social similarity among scholarly journals: an exploratory comparison of the networks of editors, authors and co-citations | [
"cs.SI",
"cs.DL"
] | This paper explores, by using suitable quantitative techniques, to what extent the intellectual proximity among scholarly journals is also a proximity in terms of social communities gathered around the journals. Three fields are considered: statistics, economics and information and library sciences. Co-citation networks (CC) represent the intellectual proximity among journals. The academic communities around the journals are represented by considering the networks of journals generated by authors writing in more than one journal (interlocking authorship: IA), and the networks generated by scholars sitting in the editorial board of more than one journal (interlocking editorship: IE). For comparing the whole structure of the networks, the dissimilarity matrices are considered. The CC, IE and IA networks appear to be correlated for the three fields. The strongest correlations is between CC and IA for the three fields. Lower and similar correlations are obtained for CC and IE, and for IE and IA. The CC, IE and IA networks are then partitioned in communities. Information and library sciences is the field where communities are more easily detectable, while the most difficult field is economics. The degrees of association among the detected communities show that they are not independent. For all the fields, the strongest association is between CC and IA networks; the minimum level of association is between IE and CC. Overall, these results indicate that the intellectual proximity is also a proximity among authors and among editors of the journals. Thus, the three maps of editorial power, intellectual proximity and authors communities tell similar stories. | cs.SI | cs |
Intellectual and social similarity among scholarly journals: an
exploratory comparison of the networks of editors, authors and
co-citations
Alberto Baccini1 ∗
Lucio Barabesi1
Mahdi Khelfaoui2
Yves Gingras2
1Department of Economics and Statistics, University of Siena, Italy
2Université du Quebec à Montreal, Canada
Abstract
This paper explores, by using suitable quantitative techniques, to what extent the intellectual
proximity among scholarly journals is also a proximity in terms of social communities gathered
around the journals. Three fields are considered: statistics, economics and information and
library sciences. Co-citation networks (CC) represent the intellectual proximity among journals.
The academic communities around the journals are represented by considering the networks
of journals generated by authors writing in more than one journal (interlocking authorship:
IA), and the networks generated by scholars sitting in the editorial board of more than one
journal (interlocking editorship: IE). For comparing the whole structure of the networks, the
dissimilarity matrices are considered. The CC, IE and IA networks appear to be correlated for
the three fields. The strongest correlations is between CC and IA for the three fields. Lower
and similar correlations are obtained for CC and IE, and for IE and IA. The CC, IE and IA
networks are then partitioned in communities.
Information and library sciences is the field
where communities are more easily detectable, while the most difficult field is economics. The
degrees of association among the detected communities show that they are not independent. For
all the fields, the strongest association is between CC and IA networks; the minimum level of
association is between IE and CC. Overall, these results indicate that the intellectual proximity
is also a proximity among authors and among editors of the journals. Thus, the three maps of
editorial power, intellectual proximity and authors communities tell similar stories.
Keywords: Interlocking editorship network; Gatekeeper; Interlocking authorship network;
Co-citation network; generalized distance correlation; communities in network.
∗Alberto Baccini is the recipient of a grant by the Institute For New Economic Thinking Grant ID INO17-
00015. The funders had no role in study design, data collection and analysis, decision to publish, or preparation
of the manuscript. The authors are very grateful to the two reviewers for their very useful comments.
1
1
Introduction
The main objects analyzed in this paper are scholarly journals and communities gathered around
them. Scholarly journals have grown in relevance as outlet for communicating research results in
the social sciences and humanities (Kulczycki, Engels, Pölönen, Bruun, Dušková, Guns, Nowot-
niak, Petr, Sivertsen, Istenič Starčič, and Zuccala 2018), following a trend that began in the
natural sciences a century earlier (Csiszar 2018). Over the last two decades, in the context of
the publish-or-perish environment, where academic careers of scholars depend more and more
on the "quality" of the journals in which they have published their articles, journals have gained
a new importance as brands(Heckman and Moktan 2018). It is therefore hardly surprising that
the interest of scientometric scholars for journals mainly focused on the building of indicators,
such as the Impact Factor, to be used for evaluative purposes (Todeschini and Baccini 2016).
The analysis of scholarly journals as social institutions of science appears a bit less developed.
Indeed, scholarly journals connect members of academic communities (Potts, Hartley, Mont-
gomery, Neylon, and Rennie 2017). Editorial boards of journals constitute a first layer of such a
community. They act as gatekeepers of science: they are, directly or indirectly, responsible for
the refereeing processes, they decide which papers are worth publishing in their journals (Crane
1967; Hoenig 2015). The stronger the link between the prestige of journals and the career ad-
vancement of scholars, the stronger the academic power exercised by the members of an editorial
board. From this point of view, it is possible to consider editorial boars as engines of academic
power. A possible way for studying the role of editors consists in observing the presence of
the same editors on the boards of different journals. The network of journals generated by the
presence of the same person on the editorial board of more than one journal is called an Inter-
locking Editorship network (IE) (Baccini, Barabesi, and Marcheselli 2009; Baccini 2009; Baccini
and Barabesi 2011; Baccini and Barabesi 2014). Thus, if two journals share the same persons
on their editorial boards, it can be assumed that they have at least similar or complementary
editorial policies, since they are managed by similar groups of scholars (Baccini, Barabesi, and
Marcheselli 2009). From another perspective, editors have the power to drift the paper selection
processes toward decisions favoring department colleagues, or disciples, and so on (Laband and
Piette 1994; Klein and DiCola 2004). In this sense the IE network can be used to try to identify
some kind of favoritism in the refereeing processes (Erfanmanesh and Morovati 2017), or for
illustrating the self-referentiality of national communities of scholars (Baccini 2009).
A second social community gathered around scholarly journals is constituted by the authors
of the published articles. While many studies exist about authorship and co-authorship, only a
few are focused on the communities of authors of specific journals (Potts, Hartley, Montgomery,
Neylon, and Rennie 2017).
In turn, it is possible to work analogously to the IE network by
considering the journal network generated by the scholars authoring papers in different journals.
The network among journals generated by the crossed presence of the same authors in different
journals could be called the Interlocking Authorship (IA) network. To the best of our knowledge,
this kind of network has been rarely explored (Brogaard, Engelberg, and Parsons 2014; Ni,
Sugimoto, and Cronin 2013; Ni, Sugimoto, and Jiang 2013). In the IA network, the proximity
2
between two journals can be considered proportional to the number of common authors. Such a
proximity is, in some sense, intellectual since it is based on the choices made by authors on where
to publish their papers, and on decisions of the editors to accept or not to publish those papers.
The community of authors around a journal thus reflects to a certain degree the contents of the
journal and the activity of the gatekeepers of the journal. If two journals are in proximity, it
can be supposed that they have similar contents and that their editorial policies are similar or
complementary.
Scholarly journals contribute to the definition of the intellectual landscapes of research fields.
Co-citation analysis is probably the best known instrument for studying the intellectual proxim-
ity among authors, papers and journals (Small 1973). For instance, if two authors are frequently
cited together in many different papers, this suggest that these two persons are somehow intel-
lectually connected by the topic or methodology of their work. Similarly, two different journals
often cited together in the same paper suggest that these journals are connected. The more
often they are cited together the stronger the link between these authors or journals. We thus
obtain a network connecting the journals based on their being often cited together. Let us call
this network CC as it is based on a different measure than those obtained through IE and IA.
In this paper we consider the IE, IA, CC networks of journals summarily described above
and we compare the degree of proximity of journals in the three networks. The first intuitive
question is to what extent these three networks are similar. If two journals are well connected
in the CC network, that is if they have a strong intellectual proximity, does a similar proximity
exists in the IA or IE network? The basic idea is to explore to what extent the social proximity
among journals observed in the network of the editorial boards is similar to the social/intellectual
proximity observed in the IA network and in the intellectual proximity in the CC network.
This question is explored by considering the IE, IA and CC networks in three fields: eco-
nomics (EC), statistics (STAT), and information and library science (ILS). Two reasons justify
the choice of the three fields. The first one is practical: for the three fields data on the edi-
torial boards of journals were already available because they had been collected by two of the
authors in a previous research project. Data on editorial boards have to be collected by hand.
Hence, their availability is a big advantage. The second reason is that scholars in the three fields
differ in the way they use scholarly journals as outlet for publishing research results. While
in statistics journals articles are largely dominant, scholars in economics and in information
and library sciences continue to write book chapters and books (Kulczycki, Engels, Pölönen,
Bruun, Dušková, Guns, Nowotniak, Petr, Sivertsen, Istenič Starčič, and Zuccala 2018). Hence
the similarity analysis considered three different scholarly communication contexts.
For each field we compare the three networks as a whole by using suitable statistical tech-
niques. Subsequently, for each field, we partition the three networks in "communities of journals"
and we analyze the coherence of these communities between pairs of networks.
The paper is organized as follow. Section 2 contains the technical definitions used in the rest
of the paper and the description of the methodology adopted for empirical analysis. Section 3
contains the outcome of the analysis. Section 4 discusses the results and Section 5 concludes by
3
suggesting further steps of the present research.
2 Journal networks data
The journal networks considered here are all one-mode (Wasserman and Faust 1994). In an IE
network, nodes are scholarly journals and the edge between two journals indicates that at least
one scholar sits in the board of the two. Each edge can be weighted by the number of common
editors between the linked journals. Analogously, in the IA networks, the edges between journals
are generated by common authors and the weight of the edge is the number of common authors.
Finally in a CC network, the edge between journals is generated by the fact that the two journals
are cited together at least in one article; the weight of the edge is the number of articles citing
the two journals together.
We have constructed the three networks (IE, IA, CC) for the three fields for a total of nine
networks. For IE networks, as anticipated, we used three existing databases, each containing
the journal editorial boards in a given year. Details on their collection and normalization can
be found in the papers referenced below. Moreover, IA and CC networks were constructed by
using Web of Science (WoS) data for a five years time-period starting from the year for which
the IE was recorded. The raw data for the nine one-mode networks can be downloaded from
https://doi.org/10.5281/zenodo.3350797.
For economics, we considered a set of 169 journals listed in the EconLit database, and indexed
in the Journal Citation Reports for the year 2006. The IE network (Figure 1) was extracted
from the database collected by Baccini and Barabesi (2010) for the year 2006. The IA (Figure
2) and CC (Figure 3) networks for economic journals were built on WoS data, by considering
respectively the authors of and the references in the papers published in the journals in the years
2006-2010.
For the field of statistics, the set includes the 79 journals listed in the category "Statistics
and probability" of the Journal Citation Reports for the year 2005. IE data (Figure 4) are the
ones collected in Baccini, Barabesi and Marcheselli (2009) for the year 2006. Similarly, for the
discipline of statistics, IA (Figure 5) and CC (Figure 6) networks were built using WoS data, by
considering papers published in the years 2006-2010.
Finally, for the domain of information and library sciences, the set includes the 59 journals
listed in the category "Information science and library science" of the Journal Citation Reports
for the year 2008. IE data (Figure 7) are the ones collected in Baccini and Barabesi (2011) for
the year 2010. Again, IA (Figure 8) and CC (Figure 9) networks were built on WoS data, by
considering papers published in the years 2010-2014.
In Figures 1-9 the size of a node is proportional to its degree and the width of an edge
is proportional to the value of the link.
In the IE network, for example, the size of a node
is proportional to the number of journals to which it is linked; the width of the link between
two nodes is proportional to the number of their common editors. For each field, the visual
comparison of the three networks is hardly informative. For instance, it is apparent that for
all three fields, the IE networks are less connected than the IA and CC networks. Also, in the
4
Table 1: Generalized distance correlations between networks
Networks
co-citation vs editor
co-citation vs author
editor vs author
Rd
√
P-value
√
P-value
√
P-value
Rd
Rd
Statistics
Information and library sciences
Economics
0.5947
0.00001
0.6431
0.00001
0.5985
0.00001
0.5386
0.00058
0.6389
0.00001
0.4969
0.00382
0.5228
0.00001
0.7518
0.00001
0.5112
0.00001
center of the networks there are not always the same journals; and a journal may have a different
size in the three networks. We therefore need a better way of comparing networks.
3 Dissimilarities among networks
For each network, it is possible to build a pseudo-measure of the distance among journals by
calculating a matrix of dissimilarities. The Jaccard index was adopted as a dissimilarity measure
(for more details on the Jaccard index, see e.g. Levandowsky and Winter (1971)). More precisely,
if A and B represent the sets containing the members of the editorial boards of two journals,
the Jaccard dissimilarity is defined as
J(A, B) =
A ∪ B − A ∩ B
A ∪ B
(1)
As an example, in the IE network, the similarity among journals is proportional to the number
of common editors in their boards. Hence, the minimum dissimilarity J(A, B) = 0 is reached
when two journals have exactly the same editorial board, i.e. all the editors of a journal are
also the editors of the other and vice versa. The maximum dissimilarity J(A, B) = 1 is reached
when two journals have no editors in common.
In order to compare the three dissimilarity
matrices arising from co-citation, editorial board and author networks for each discipline, we
adopt the generalized distance correlation Rd suggested by Omelka and Hudecová (2013) on
the basis of the seminal proposal by Székely et al. (2007). It should be remarked that such a
correlation index avoids the drawbacks emphasized by Dutilleul et al. (2000) when the classical
Mantel coefficient is assumed instead (Omelka and Hudecova 2013). Hence, we considered the
three possible couples of networks and we computed the corresponding values of
Rd for each
discipline. It is worth noting that Rd is somehow similar to the squared Pearson correlation
coefficient - and hence
Rd should be interpreted as a generalization of the usual correlation
coefficient. More precisely, Rd is defined in the interval [0, 1], in such a way that values close
to zero indicate no or very weak association, while larger values suggest a stronger association,
which is perfect for Rd = 1 - and similar considerations obviously hold for
Rd (for more details,
see e.g. Omelka and Hudecová (2013)). The generalized distance correlation was evaluated in
the R-computing environment (R Core Team 2018) by using the R function dcor in the package
energy (Rizzo and Székely 2018). These values of
√
Rd are reported in Table 1.
√
√
√
5
√
From the analysis of Table 1, the dependence between the considered dissimilarity matrices
is apparent. Indeed, the observed values of
Rd are greater than (or nearly equal to) the value
0.5 for each combination of networks in the three disciplines. Moreover, the permutation test for
assessing independence, as proposed by Omelka and Hudecová (2013), was also carried out. The
statistical details of the permutation test are rather involved, even if they are clearly explained
by Omelka and Hudecová (2013). Loosely speaking, the rationale behind the test stems from the
fact that, under the null hypothesis of independence, the generalized distance correlation should
not be affected by a random permutation of the rows and the corresponding columns of the
"centred" distance matrices. The permutation principle is widely adopted in order to carry out
nonparametric inference, since assumptions are minimal and practical implementation is often
straightforward (see e.g. Lehmann and Romano (2005), Section 10). The permutation test of
independence was in turn implemented by using the package energy (Rizzo and Székely 2018).
The significance of the test statistic was computed by means of the R function dcov.test (for more
details Omelka and Hudecová (2013)). On the basis of the achieved P-values given in Table 1,
the independence hypotheses can be rejected at the significance level α = 0.01. Since the three
statistical tests within each discipline are obviously dependent, we also consider the Bonferroni
procedure in order to control the familywise error rate (for more details, see e.g. Bretz, Hotorn
and Westfall (2011)). Thus, by assuming such a procedure and a global significance level given by
α = 0.01, the marginal independence hypotheses may be rejected if the corresponding P-values
are less than α/3 = 0.0033, which is the case for all the considered tests - except the editorial
board and author networks for information and library sciences. However, it is worth remarking
that - even in this case - the corresponding P-value is just slightly larger than the threshold.
Hence, the co-citation, editorial board and author networks display structures which may be
considered associated for each considered discipline - at least on the basis of the considered
dissimilarity matrices.
4 Correlations among communities of journals
The proximity among journal networks can be explored by focusing on communities of journals.
The first step consists in detecting communities inside each network; the second in verifying
the degree of association between the communities detected in different networks of the same
field. A non-overlapping community of nodes of a network is a set of nodes densely connected
internally and only sparsely connected with external nodes. Each network is partitioned in
communities by using the Louvain algorithm (Blondel, Guillaume, Lambiotte, and Lefebvre
2008) as implemented in the software Pajek (de Nooy, Mrvar, and Batagelj 2018). It consists
in the optimization of the modularity of the network (Newman 2004; Newman and Girvan
2004). The quality of the partition is quantitatively measured by modularity values. Table 2
reports the values of modularities and the resolution parameters adopted for optimization. The
resolution parameter is used to control the size of the communities detected; higher values of
the parameter produce larger number of communities and viceversa. Table 2 also reports the
number of communities detected.
6
Table 2: Main features of networks and communities
Information and library sciences
Economics
Density
Average degree
Isolated journals
Resolution
Modularity
n. communities
E-I unweighted
E-I weighted
n. non-isolated communities
Statistics
CC
0.671
52.379
IA
0.91
70.962
IE
0.0935
5.423
CC
0.644
37.356
0
1
3
3
0
1
4
4
9
1
14
5
0
1
3
3
IE
0.121
9.443
4
1
0.4
10
6
0.108
0.171
0.528
0.266
0.444
0.09
0.218
-0.04
-0.309
0.22
-0.141
0.41
-0.02
-0.425
-0.651
0.201
-0.355
0.108
-0.202
0.322
0.131
0.435
0.038
IE
0.07
11.751
CC
0.566
95.053
IA
0.744
125.001
7
1
16
9
0
1
4
4
0
1
5
5
IA
0.764
44.339
0
0.8
0.329
3
3
0.2
-0.328
For all the pairs of the networks inside each research field, the association between the result-
ing communities is then analyzed by using statistical techniques as available in Pajek (de Nooy,
Mrvar, and Batagelj 2018). All the indicators considered are adopted under an exploratory
approach. χ2 statistics provide an index aiming to assess the degree of independence of the par-
titions of each pair of networks. Cramér's V is a measure of association giving a value between 0
(no association) and +1 (perfect association) (Cramér 1946). Rajski's coherence (Legendre and
Legendre 1998) is presented in three variants, all defined in [0,1] range: a symmetrical version in-
dicating the coherence between each pair of classification; and two asymmetrical versions called
in Table 3 "Rajski's right" and "Rajski's left". When the communities in the IE-CC networks
are considered, Rajski's left indicates the extent to which the first communities classification
IE is able to predict the second communities classification CC; Rajski's right indicates instead
the extent to which the second classification is able to predict the first. Finally, the adjusted
Rand index measures the degree of association between partitions and is bounded between ±1
(Hubert and Arabie 1985). All indices are reported in Table 3.
For the three fields analyzed here, we observe that the IE is the least dense network and
the network with the lowest average degree. For the three fields, the CC networks are in the
intermediate position for density and average degree, and finally the IA networks have the highest
values of density (0.91 for statistics) and average degree (Wasserman and Faust 1994).
In general, the community detection algorithm was more successful in sparser networks: for
the three fields, the values of modularity are indeed the highest for the IE network, intermediate
for CC and lowest for IA. In the IE networks many detected communities are actually isolated
journals, i.e.
journals with no common editors with other considered journals. In every case,
the number of communities detected in the IE networks is always bigger than the number of
communities detected in the other networks.
Information and library sciences is the field where the communities are more easily detectable
and more clearly defined, as shown by the highest modularity values and by the lowest values
of the E-I indices (Table 2) (de Nooy, Mrvar, and Batagelj 2018).
In particular for the IA
network, communities were detected by adopting a resolution value of 0.8. This resolution was
preferred to the value of 1 adopted for all the other networks, because the resulting communities
exhibited better E-I indices1. The E-I index was calculated as the difference between the number
1With a value of resolution of 1 the IA network is partitioned in 4 communities, with modularity 0.257, E-I
unweighted= 0.255 and E-I weighted=−0.083.
7
s
e
i
t
i
n
u
m
m
o
c
n
e
e
w
t
e
b
s
e
x
e
d
n
i
n
o
i
t
a
i
c
o
s
s
A
:
3
e
l
b
a
T
A
I
-
C
C
9
3
6
.
0
9
2
2
.
0
1
5
3
.
0
8
9
3
.
0
5
6
2
.
0
s
c
i
m
o
n
o
c
E
A
I
-
E
I
5
1
6
.
0
7
0
2
.
0
1
3
4
.
0
4
8
2
.
0
2
.
0
)
2
1
(
9
2
.
7
0
2
)
0
6
(
4
3
.
5
5
2
)
5
4
(
8
5
.
9
9
1
C
C
E
I
-
7
2
6
.
0
1
7
1
.
0
8
9
3
.
0
2
3
2
.
0
7
4
1
.
0
s
e
c
n
e
i
c
s
y
r
a
r
b
i
l
d
n
a
n
o
i
t
a
m
r
o
f
n
I
A
I
-
C
C
)
4
(
8
6
.
8
6
A
I
-
E
I
C
C
E
I
-
)
6
2
(
1
5
.
0
8
)
6
2
(
5
4
.
5
7
3
6
7
.
0
4
3
4
.
0
7
1
6
.
0
4
9
5
.
0
3
0
5
.
0
6
2
8
.
0
1
0
3
.
0
9
7
6
.
0
1
5
3
.
0
3
0
3
.
0
9
9
7
.
0
9
7
2
.
0
4
2
6
.
0
6
3
3
.
0
6
9
2
.
0
A
I
-
C
C
)
6
(
5
8
.
0
8
5
1
7
.
0
2
1
3
.
0
4
2
5
.
0
5
3
4
.
0
4
4
4
.
0
A
I
-
E
I
s
c
i
t
s
i
t
a
t
S
8
0
7
.
0
1
9
2
.
0
7
2
5
.
0
4
9
3
.
0
6
4
3
.
0
C
C
E
I
-
9
5
6
.
0
7
0
2
.
0
8
4
4
.
0
8
7
2
.
0
9
4
2
.
0
)
7
2
(
8
8
.
8
1
1
)
8
1
(
9
5
.
8
6
)
.
f
.
d
(
2
χ
V
s
'
r
é
m
a
r
C
s
'
i
k
s
j
a
R
t
h
g
i
r
s
'
i
k
s
j
a
R
t
f
e
l
s
'
i
k
s
j
a
R
x
e
d
n
i
d
n
a
R
.
j
d
A
8
of edges within communities and the number of edges between communities; that difference is
then divided for the total number of edges of the network. The weighted version of the index is
calculated by considering the value of the edges. The range of the index is between -1 (all edges
are inside communities) and 1 (all edges are between communities). The χ2 values show that the
detected communities for the three networks are not independent. The association between the
partitions of communities as measured by Cramér's V is high. The highest level of association as
measured by the adjusted Rand's index is found between communities detected in the CC and
IA networks. Rajski's right indicates that the communities detected in the IE network predict
well the communities detected in the other networks.
The field of statistics is in an intermediate position: the values of modularity are very low
for CC and IA networks, nevertheless the resulting partitions have negative values of the E-I
weighted indices. Communities in the IE network are more easily detectable and more clearly
defined than in the IA and especially in the CC network. Also in this field, the χ2 values show
that the detected communities for the three networks are not independent. The association
between the partitions of communities is a bit higher between CC and IA than for the other
pairs of networks. Also in this case Rajski's right indicates that the communities of the IE
network predict well the communities in the other two networks.
For the case of economics, community detection is particularly problematic, i.e. small changes
of the value of the resolution parameter changed substantially the number of detected commu-
nities and the values of the indicators considered. For CC and IA, the community detection
procedure results in very low values of modularity and in positive values of E-I. Only for the IE
network the modularity is around 0.5 and the value of the E-I weighted index is less than zero.
Also in this field the χ2 values show that detected communities for the three networks are not
independent. The association between the partitions of communities is the lowest of the three
fields analyzed in this paper. Rajski's right indicates also for economics that the communities in
the IE network predict the communities in the other two networks, but the values of the index
are the lowest of the three.
5 Discussion and conclusions
The main aim of this paper was to explore, by using suitable quantitative techniques, to what
extent the intellectual proximity among scholarly journals is also a proximity in terms of social
communities gathered around the journals.
For representing the intellectual proximity among journals we have used the CC network.
For having information about the academic communities around journals, we have considered
the networks of journals generated by authors writing in more than one journal as well as the
networks generated by scholars sitting in the editorial board of more than one journal. The first
step of the exploratory analysis consisted of comparing the whole structure of the networks on
the basis of dissimilarity matrices. The CC, IE and IA networks appear to be associated for all
the three considered fields. The second step consisted of partitioning the IE, IA and CC networks
in communities and then in verifying the degree of association among the detected communities.
9
The results of that analysis show that the communities detected in the three networks are not
independent for the three research fields considered. The results of both approaches are coherent
in showing that the strongest correlations between networks is between CC and IA for the three
fields. Lower and similar correlations were obtained for CC and IE, and for IE and IA. When
communities are considered, the strongest association between communities is between CC and
IA networks; the minimum level of association is between IE and CC.
To the best of our knowledge, the only similar analysis was performed by Ni, Sugimoto and
Cronin (2013) in their investigation of scholarly communication. They focused on information
and library sciences, by considering networks of journals generated by common authors, co-
citation, common topics and common editors. They descriptively compared clusters of journals
between networks and calculated a correlation between pairs of matrices by using the quadratic
assignment procedure. Their results appears to be coherent with the ones presented here since
they estimated statistically significant correlations for networks of journals based on authors,
co-citation and editors.
Overall, the results of our analysis show that the intellectual proximity is also a proximity
among authors and, more surprisingly, among editors of the journals. This leads to the question
of whether the structures obtained could ever be independent if the same set of people were
predominantly involved in the editorial boards, the publishing of papers, and the citing of
papers. In that case the structures are just a consequence of the existence of a publishing and
gatekeeping élite in the considered research fields.2 This is a topic worth to be investigated by
considering the dual-networks that we used for generating the nine one mode networks analyzed
in this paper. At the current state of knowledge, it is only possible to affirm that the map of
editorial power, the map of intellectual proximity and the map of author communities tell similar
stories. The fact that the results are comparable for the three fields studied suggests that the
method presented here is more generally applicable to any scientific field and that there should
be in general a coherence among journals at the three scales of 1) editorial boards, 2) authors
choice of publications and 3) co-citations.
References
Baccini, A. (2009). Italian economic journals. a network-based ranking and an exploratory
analysis of their influence on setting international professional standards. Rivista Italiana
degli Economisti 14 (3), 491 -- 511.
Baccini, A. and L. Barabesi (2010). Interlocking editorship. a network analysis of the links
between economic journals. Scientometrics 82 (2), 365 -- 389.
Baccini, A. and L. Barabesi (2011). Seats at the table: The network of the editorial boards
in information and library science. Journal of Informetrics 5 (3), 382 -- 391.
Baccini, A. and L. Barabesi (2014). Gatekeepers of economics: the network of editorial boards
in economic journals. In A. Lanteri and J. Vromen (Eds.), The Economics of Economists,
2We owe this observation to one of the two reviewers.
10
pp. 104 -- 150. Cambridge: Cambridge University Press.
Baccini, A., L. Barabesi, and M. Marcheselli (2009). How are statistical journal linked? a
network analysis. Chance 22 (3), 34 -- 43.
Blondel, V. D., J.-L. Guillaume, R. Lambiotte, and E. Lefebvre (2008, oct). Fast unfolding
of communities in large networks. Journal of Statistical Mechanics: Theory and Experi-
ment 2008 (10), P10008.
Bretz, F., T. Hothorn, and P. Westfall (2011). Multiple Comparisons Using R. Boca Raton:
CRC Press.
Brogaard, J., J. Engelberg, and C. A. Parsons (2014). Networks and productivity: Causal
evidence from editor rotations. Journal of Financial Economics 111 (1), 251 -- 270.
Cramér, H. (1946). Mathematical Methods of Statistics. Princeton: Princeton University Press.
Crane, D. (1967). The gatekeepers of science: Some factors affecting the selection of articles
of scientific journals. American Sociologist 32 (2), 195 -- 201.
Csiszar, A. (2018). The Scientific Journal. Authorship and Politicis of Knwoledge in the Nine-
teenth Century. Chicago: Chicago University Press.
de Nooy, W., A. Mrvar, and V. Batagelj (2018). Exploratory Social Network Analysis with Pa-
jek. Revised and expanded edition for updated software. Cambridge: Cambridge University
Press.
Dutilleul, P., J. D. Stockwell, D. Frigon, and P. Legendre (2000). The mantel test versus pear-
son's correlation analysis: Assessment of the differences for biological and environmental
studies. Journal of Agricultural, Biological, and Environmental Statistics 5 (2), 131 -- 150.
Erfanmanesh, M. and M. Morovati (2017). Interlocking editorships in scientific journals. Sci-
ence and Engineering Ethics 2017.
Heckman, J. J. and S. Moktan (2018). Publishing and promotion in economics: The tyranny
of the top five. National Bureau of Economic Research Working Paper Series No. 25093.
Hoenig, B. (2015). Gatekeepers in social science. In J. D. Wright (Ed.), International En-
cyclopedia of the Social and Behavioral Sciences (Second Edition), pp. 618 -- 622. Oxford:
Elsevier.
Hubert, L. and P. Arabie (1985). Comparing partitions. Journal of Classification 2 (1), 193 --
218.
Klein, D. B. and T. DiCola (2004). Institutional ties of journal of development economics
authors and editors. Econ Journal Watch 1 (2), 319 -- 330.
Kulczycki, E., T. C. E. Engels, J. Pölönen, K. Bruun, M. Dušková, R. Guns, R. Nowotniak,
M. Petr, G. Sivertsen, A. Istenič Starčič, and A. Zuccala (2018). Publication patterns in
the social sciences and humanities: evidence from eight european countries. Scientomet-
rics 116 (1), 463 -- 486.
11
Laband, D. N. and M. J. Piette (1994). Favoritism versus search for good papers: Empirical ev-
idence regarding the behavior of journal editors. The Journal of Political Economy 102 (1),
194 -- 203.
Legendre, P. and L. Legendre (1998). Numerical Ecology. Amsterdam: Elsevier Science.
Lehmann, E. and J. Romano (2005). Testing Statistical Hypotheses. New York, Springer.
Levandowsky, M. and D. Winter (1971). Distance between sets. Nature 234 (5323), 34 -- 35.
Newman, M. E. J. (2004, Jun). Fast algorithm for detecting community structure in networks.
Phys. Rev. E 69, 066133.
Newman, M. E. J. and M. Girvan (2004, Feb). Finding and evaluating community structure
in networks. Phys. Rev. E 69, 026113.
Ni, C., C. R. Sugimoto, and B. Cronin (2013). Visualizing and comparing four facets of
scholarly communication: producers, artifacts, concepts, and gatekeepers. Scientomet-
rics 94 (3), 1161 -- 1173.
Ni, C., C. R. Sugimoto, and J. Jiang (2013). Venue-author-coupling: A measure for identifying
disciplines through author communities. Journal of the American Society for Information
Science and Technology 64 (2), 265 -- 279.
Omelka, M. and S. Hudecova (2013). A comparison of the mantel test with a generalised
distance covariance test. Environmetrics 24 (7), 449 -- 460.
Potts, J., J. Hartley, L. Montgomery, C. Neylon, and E. Rennie (2017). A journal is a club: a
new economic model for scholarly publishing. Prometheus 35 (1), 75 -- 92.
R Core Team (2018). R: A Language and Environment for Statistical Computing. Vienna,
Austria.
Rizzo, M. and G. Székely (2018). E-statistics: Multivariate inference via the energy of data.
R package 'energy' version 1.7-5.
Small, H. (1973). Co-citation in the scientific literature: A new measure of the relationship
between two documents. Journal of the American Society for Information Science 24,
265 -- 269.
Székely, G. J., M. L. Rizzo, and N. K. Bakirov (2007). Measuring and testing dependence by
correlation of distances. The Annals of Statistics 35 (6), 2769 -- 2794.
Todeschini, R. and A. Baccini (2016). Handbook of Bibliometric Indicators. Quantitative Tools
for Studying and Evaluating Research. Weinheim (Germany): Wiley-VCH.
Wasserman, S. and K. Faust (1994). Social Network Analysis: Method and Application. Cam-
bridge: Cambridge University Press.
12
s
l
a
n
r
u
o
j
c
i
m
o
n
o
c
e
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
t
i
d
e
g
n
i
k
c
o
l
r
e
t
n
I
:
1
e
r
u
g
i
F
13
s
l
a
n
r
u
o
j
c
i
m
o
n
o
c
e
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
h
t
u
a
g
n
i
k
c
o
l
r
e
t
n
I
:
2
e
r
u
g
i
F
14
s
l
a
n
r
u
o
j
c
i
m
o
n
o
c
e
f
o
k
r
o
w
t
e
n
n
o
i
t
a
t
i
c
-
o
C
:
3
e
r
u
g
i
F
15
s
l
a
n
r
u
o
j
l
a
c
i
t
s
i
t
a
t
s
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
t
i
d
e
g
n
i
k
c
o
l
r
e
t
n
I
:
4
e
r
u
g
i
F
16
s
l
a
n
r
u
o
j
l
a
c
i
t
s
i
t
a
t
s
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
h
t
u
a
g
n
i
k
c
o
l
r
e
t
n
I
:
5
e
r
u
g
i
F
17
s
l
a
n
r
u
o
j
l
a
c
i
t
s
i
t
a
t
s
f
o
k
r
o
w
t
e
n
n
o
i
t
a
t
i
c
-
o
C
:
6
e
r
u
g
i
F
18
s
l
a
n
r
u
o
j
s
e
c
n
e
i
c
s
y
r
a
r
b
i
l
d
n
a
n
o
i
t
a
m
r
o
f
n
i
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
t
i
d
e
g
n
i
k
c
o
l
r
e
t
n
I
:
7
e
r
u
g
i
F
19
s
l
a
n
r
u
o
j
s
e
c
n
e
i
c
s
y
r
a
r
b
i
l
d
n
a
n
o
i
t
a
m
r
o
f
n
i
f
o
k
r
o
w
t
e
n
p
i
h
s
r
o
h
t
u
a
g
n
i
k
c
o
l
r
e
t
n
I
:
8
e
r
u
g
i
F
20
s
l
a
n
r
u
o
j
s
e
c
n
e
i
c
s
y
r
a
r
b
i
l
d
n
a
n
o
i
t
a
m
r
o
f
n
i
f
o
k
r
o
w
t
e
n
n
o
i
t
a
t
i
c
-
o
C
:
9
e
r
u
g
i
F
21
|
1501.05921 | 1 | 1501 | 2015-01-23T19:19:19 | The Spatial Structure of Transnational Human Activity | [
"cs.SI",
"physics.soc-ph"
] | Recent studies have shown that the spatial structures of animal displacements and local-scale human motion follow L\'{e}vy flights. Whether transnational human activity (THA) also exhibits such a pattern has however not been thoroughly examined as yet. To fill this gap, this article examines the planet-scale spatial structure of THA (a) across eight types of mobility and communication and (b) in its development over time. Combining data from various sources, it is shown that the spatial structure of THA can indeed be approximated by L\'{e}vy flights with heavy tails that obey power laws. Scaling exponent and power-law fit differ by type of THA, being highest in refuge-seeking and tourism and lowest in student exchange. Variance in the availability of resources and opportunities for satisfying associated needs appears to explain these differences. Over time, the L\'{e}vy-flight pattern remains intact and remarkably stable, contradicting the popular idea that socio-technological trends lead to a "death of distance." Longitudinal change occurs only in some types of THA and predominantly at short distances, indicating regional shifts rather than globalization. | cs.SI | cs | The Spatial Structure of
Transnational Human Activity∗
Emanuel Deutschmann
Bremen International Graduate School of Social Sciences
School of Humanities and Social Sciences, Jacobs University Bremen
[email protected]
Draft version, November 8, 2018
Abstract
Recent studies have shown that the spatial structures of animal displacements and local-scale
human motion follow Lévy flights. Whether transnational human activity (THA) also exhibits
such a pattern has however not been thoroughly examined as yet. To fill this gap, this article
examines the planet-scale spatial structure of THA (a) across eight types of mobility and
communication and (b) in its development over time. Combining data from various sources,
it is shown that the spatial structure of THA can indeed be approximated by Lévy flights with
heavy tails that obey power laws. Scaling exponent and power-law fit differ by type of THA,
being highest in refuge-seeking and tourism and lowest in student exchange. Variance in the
availability of resources and opportunities for satisfying associated needs appears to explain
these differences. Over time, the Lévy-flight pattern remains intact and remarkably stable,
contradicting the popular idea that socio-technological trends lead to a "death of distance."
Longitudinal change occurs only in some types of THA and predominantly at short distances,
indicating regional shifts rather than globalization.
Keywords: transnationalism, mobility, communication, Lévy flight, power law, globalization
1.
Introduction
A hungry shark searching for prey in the ocean will frequently move short distances, interrupted
by random changes of direction that are only occasionally followed by moves over longer
distances. Sorting the displacement lengths covered in this journey by size and frequency
results in a distribution with a long tail that obeys a power law1. Such mobility patterns,
so-called Lévy flights2, have been shown to occur not only in the foraging movements of sharks
and other marine predators like sea turtles and penguins (Sims et al., 2008), but also in the
motion of smaller species like plankton (Bartumeus et al., 2003). For humans, Lévy-like patterns
have been found in local (Gonzalez, Hidalgo and Barabasi, 2008; Song et al., 2010; Rhee et al.,
2011), and nation-wide mobility (Brockmann, Hufnagel and Geisel, 2006). Whether planet-scale
human activity also follows a Lévy flight has however not been thoroughly examined to date.
While this natural-scientific strand of research on Lévy flights has largely omitted the global
sphere as yet, the social-scientific literature that deals with the global sphere has conversely
not taken the findings of the Lévy-flight debate into account. On the contrary, it is dominated
by the idea that technological and socioeconomic trends have led to a diminishing or even
vanishing role of physical distance in structuring human activity. Symptomatic are shibboleths
∗I am grateful to the participants of the BIGSSS Colloquium at Jacobs University Bremen for invaluable comments
on an earlier version of this paper, to the World Tourism Organization for generously granting me access to their data
on global tourism flows and to Lea Kliem for assistance in compiling the online friendship matrix used in this study.
5
1
0
2
n
a
J
3
2
]
I
S
.
s
c
[
1
v
1
2
9
5
0
.
1
0
5
1
:
v
i
X
r
a
like "demise of geography" (Toffler, 1970), "time-space compression" (Harvey, 1989), "end of
geography" (O'Brien, 1992), "collapse of space" (Kirsch, 1995), "death of distance" (Cairncross,
1997), and "flat world" (Friedman, 2007). Although a third stream of research has argued that
such statements are exaggerated and that geographic space (still) structures social relations,
a rigorous empirical analysis of the association between physical distance and planet-scale
human activity is also missing in the social sciences.
The persistence of this research gap and the stark contrast between the natural- and social-
scientific positions is astonishing given that many problems humanity is facing today, from
health epidemics like H5N1 and Ebola to climate change and terrorism, are heavily intertwined
with transnational human mobility and communication. Driven by the hope that a better
understanding of how people move around the world can contribute to solving such pressing
social problems, a group of researchers has recently challenged the scientific community to
get active and "to collect large-scale human mobility traces" (Hui et al., 2010). The aim of
this article is to respond to this call and to fill in some missing pieces, namely to add the
global-scale analysis to the natural-scientific Lévy-flight debate, and the Lévy-flight analysis
to the social-scientific globalization debate. In specific, we search for answers to the following
research questions:
1. Does the spatial structure of transnational human activity follow a Lévy-flight?
2. Does the spatial structure differ by type of activity, and if yes -- why?
3. Has the spatial structure of transnational human activity changed over time?
To do so, data on five types of mobility (asylum-seeking, migration, refuge-seeking, student mo-
bility, tourism) and three types of communication (phone calls, online friendships, remittances)
from a range of sources was standardized to cover the same set of 196 sending and receiving
countries (i.e. 38,220 country dyads). In total, the structure of approximately 1.8 million valued
ties across 47 network matrices at various time points between 1960 and 2010 are analyzed,
involving, inter alia, 1.7 billion tourist trips and 240 billion international phone call minutes.
While the online friendship dataset for instance is entirely new and thus original in itself, the
major innovation of this study is to link the hitherto disconnected natural- and social-scientific
debates by conducting the first encompassing comparative analysis of the spatial structure of
transnational human activity (a) at a planet-wide scale, (b) across various types of mobility and
communication, and (c) over time.
This article is organized as follows: in the next section, we review the existing literature and
theories concerning the spatial structure of human (and non-human) activity. We then set out
our own conceptual approach. Next, the data and methods are described. Finally, we present
the results, focusing first on the current spatial structure, then move to examining longitudinal
trends, and finally compare our results to theoretical scenarios and results from other studies.
We conclude with a summary of our findings and a discussion of their implications.
2. Theory and state of research
The literature on the spatial structure of human activity can be described as consisting of
three separate streams: (a) the Lévy-flight debate, (b) the geography-is-dead debate, and
(c) the distance-decay debate. In the following, we briefly review these streams of research,
highlighting their achievements and deficits.
2.1. The Lévy-flight debate
The Lévy-flight debate is carried out mainly by natural and complexity scientists and began
with research that showed that the motion patterns of many non-human species, from mammals
2
like spider monkeys (Ramos-Fernández et al., 2004) to marine predators like sharks, sea turtles
and penguins (Sims et al., 2008) and much smaller organisms like plankton (Bartumeus et al.,
2003), follow power laws. While some of the early alleged Lévy-flight diagnoses (e.g. for
albatrosses) turned out not to hold upon closer scrutiny (Buchanan, 2008), the debate remains
vital and was recently expanded to human mobility. Lévy-like mobility patterns have now been
found in studies that followed the GPS traces of volunteers on university campuses, theme
parks, state fairs, and in cities (Rhee et al., 2011; Noulas et al., 2012), the movements of mobile
phones (Gonzalez, Hidalgo and Barabasi, 2008; Song et al., 2010), and the traces of banknotes in
the United States (Brockmann, Hufnagel and Geisel, 2006).
The theoretical argument for the Lévy-flight pattern is -- at least for animal motion -- the
optimization of food search. It has been shown that in environments where food is scarce,
a power-law distribution of flight lengths with a scaling exponent β=2 is the optimal search
strategy (Viswanathan et al., 1999). For human mobility, though, the existing papers contain
little theoretical reflection as to why the Lévy flight should apply (e.g. whether the goals humans
pursue in their everyday life are readily equitable with foraging). Another critical point is that
in contrast to the random walk that underlies the Lévy-flight model, humans do not actually
move into random directions. It is more likely that the directions are determined by needs
humans seek to fulfil (Noulas et al., 2012) or the information they have about their environment
(Miller, 1972). Yet, while this discrepancy has repeatedly been recognized (Rhee et al., 2011;
Song et al., 2010), little has been done to overcome it. The biggest shortcoming of the Lévy-flight
debate however are its current scale and data restrictions: US-dollar bills for instance can only
be used in the United States and "it remains unclear whether the observed properties are
specific to the US or whether they represent universal features" (Brockmann and Theis, 2008,
p. 33). Two recent studies (Cheng et al., 2011; Noulas et al., 2012) have shown that log-ins to
location sharing services (LSS) like Foursquare follow a Lévy-flight pattern on a global scale.
Yet, LSS require smartphone access and are used only by a small, well-off minority of the world
population. Mainstream forms of planet-scale mobility and communication (including poverty-
driven ones) are disregarded by the Lévy-flight debate as yet, as is potential longitudinal change.
2.2. The geography-is-dead debate
The idea of a decline and eventual vanishing of the role of space in structuring human interaction
has a long history and takes many facets. Already in the mid-19th century, Marx and Engels
argued that "[i]n place of the old local and national seclusion and self-sufficiency, we have
intercourse in every direction, universal interdependence of nations" (1948, p. 12). Later, Marx
coined the term "annihilation of space by time" to describe the tendency of capital (and in its
wake all parts of society) to "tear down every spatial barrier to intercourse" (1973, p. 539). In
1962, McLuhan argued that the "new electronic interdependence recreates the world in the
image of a global village" (1962, p. 31). A decade later, Toffler coined the term "demise of
geography" arguing that in contrast to the nomads of the past who were bound by place, "the
new nomads of today leave the physical structure behind" (1970, p. 91). Similarly, O'Brien
argued for an "end of geography" (1992) in financial markets, while Cairncross maintained
that "[t]he death of distance as a determinant of the cost of communicating will probably be
the single most important force shaping society in the first half of the next century" (1997, p.
1). Friedman stated that we recently entered a new era of globalization which "is shrinking
the world from a size small to a size tiny and flattening the playing field" (2007, p. 10). The
central characteristic of this new era, he argued, was "the newfound power for individuals
to collaborate and compete globally" (Ibid). Numerous other terms, such as "time-space
compression" (Harvey, 1989), "collapse of space" (Kirsch, 1995), and "shrinking world" (Allen
and Hamnett, 1995) have been proposed to describe this phenomenon.
The recurrent argument in this debate is that technological innovations, declining costs
3
and massive growth of transport and communication infrastructure lead to a world in which
physical distance plays a smaller or even no role at all anymore, first and foremost for flows
of money and messages, but also for the movement of commodities and individuals. The
geography-is-dead debate in its current and past forms has largely been driven by theoretical
arguments with little empirical grounding. Yet its arguments have been highly influential in the
social sciences and popular in the public sphere. It is completely detached from the Lévy-flight
debate, but another stream of more empirics-oriented research, the distance-decay debate, has
-- explicitly and implicitly -- referred to it critically.
2.3. The distance-decay debate
This third stream of research acknowledges the (continuing) salience of distance in structuring
human activity, however usually without analyzing the precise shape of this relation as in the
Lévy-flight debate. Most studies do not go beyond the notion that physical proximity (still)
matters. A typical statement, which became known as Tobler's First Law of Geography, is:
"Everything is related to everything else, but near things are more related than distant things"
(Tobler, 1970, p. 236). Other central terms are the "principle of least effort" (Zipf, 1949) and
"distance decay". "Distance decay" is used in a wide range of fields from criminology, where it
is known that offenders tend to commit crimes in proximity to their residency (Rengert, Piquero
and Jones, 1999), to eco-geography, where studies found that biological similarity decreases
with geographical distance (Soininen, McDonald and Hillebrand, 2007). In military science, the
continued importance of the "loss of strength gradient" has been emphasized (Webb, 2007).
Several studies on international trade and transport costs have explicitly criticized the idea of a
"death of distance" as exaggerated (Kano, Kano and Takechi, 2013), unfounded (Leamer and
Levinsohn, 1995), or even maintained the opposite: that instead of diminishing or disappearing,
the effect of distance on trade rather increased during the 20th century (Disdier and Head,
2008).3
With respect to human mobility, Ravenstein (1885) already studied the "laws of migration"
using census data from England in the 19th century. Although his contemporaries thought that
the term "laws" was inappropriate (Ibid, p. 233), he realized that the large majority of migrants
only moved short distances while few migrants moved long distances. Later, Stouffer (1940)
presented graphically a spatial distribution of family movements in Cleveland that featured a
heavy tail. Concerning transnational human activity, the central structuring role of geographical
distance has been recognized in a number of empirical studies. For transactions like diplomatic
exchange and mail flows, geographic proximity was described as "the most compelling force
of attraction in the international system" (Brams, 1966, p. 889), "a very pronounced influence
and constraint on the pattern of West European communications and interactions" (Clark and
Merritt, 1987, p. 226) and "one of the major factors in global communication networks" (Choi,
1995, p. 181). McKercher, Chan and Lam (2008) used the concept of "distance decay" in relation
to international tourism. Focusing on 41 countries, they show that 80 percent of all international
travel occurs to countries within 1,000km distance. Pertaining to the supposedly placeless
digital world, a recent study on the spatial structure of the internet found an effect of physical
distance in the form of a power law with a cutoff (Tranos and Nijkamp, 2013), while another
study only reported that the structure of transnational Facebook friendships was "apparently
influenced by geography" (Ugander et al., 2011, p. 13). Takhteyev, Gruzd and Wellman (2012,
p. 78) showed a distribution of Twitter messages by geographical distance in which distance
clearly mattered, however again without analyzing this relation further.
Two explanations have been put forward as to why human mobility is affected by geographic
space (cf. Miller, 1972; Noulas et al., 2012). The first one, the gravity hypothesis, is inspired by
Newton's law of gravity and states that costs connected to distance itself are responsible for
fewer long distance movements (cf. Deutsch and Isard, 1961; Zhou, 2011). The second one, the
4
intervening opportunities hypothesis, argues that it is not the costs of distance itself that matter,
but intervening opportunities that allow to fulfill one's needs already at close distances, making
long-distance mobility unnecessary (cf. Stouffer, 1940; Freymeyer and Ritchey, 1985).
While in many ways more differentiated than the geography-is-dead debate, an encompass-
ing analysis of the precise shape of the relation between distance and various types of human
activity on a global scale is also missing in this stream of research. Moreover, to our knowledge
no study has tracked change in the spatial structure of human mobility and communication
over time.
3. Conceptual approach
Our aim is to analyze and compare the spatial structure of various types of transnational
human activity (THA). We use THA as an umbrella term for transnational human mobility
(THM), which denotes activities in which individuals cross nation-state borders physically, and
transnational human communication (THC), which comprises communicative acts across nation-
state borders that do not necessarily involve physical mobility. Starting from the Lévy-flight
debate, we first test to which extent various types of THA feature a probability distribution
whose tail follows a power-law function of the form
P(r) ∝ r−β,
(1)
where P(r) is the probability of a displacement length r to occur and β is the scaling
exponent. A larger β means that the curve is steeper, i.e. relatively more activity occurs at
short distances, whereas a smaller β denotes a flatter relation, i.e. relatively more activity takes
place over longer distances. We expect β and the fit of this power-law function (measured
as R2) to vary by type of THA. For THM, the logic of our argument, which is inspired by
the distance-decay debate (and in specific the intervening opportunities hypothesis), is the
following: if we accept the assumptions that
1. THM is associated with type-specific goals,
2. the availability of opportunities for goal-attainment varies by goal
3. the average amount of resources available to attain goals varies by type of THM,
4. humans aim at spending as little of their resources as necessary to attain their goals,
then the spatial structure of a specific type of THM i can be expected to be determined by
two factors: the availability of opportunities for attainment of the goals associated with i and
the resources available on average to the individuals engaging in i. The broader the availability
of opportunities for goal-attainment, the higher is the likelihood for individuals to stop their
movement at closer locations (and thus the higher R2 and β), because their needs are already
fulfilled to a satisfying extent and any further movement would only diminish the stack of
resources without leading to additional benefits. Similarly, the higher the average amount of
resources available to the group of people engaging in a particular type of THM, the less they
will be physically bound by the costs of THM (and thus the lower R2 and β), ceteris paribus.
The real world is of course more complex, but still this simple model may help to obtain a first
explanation for systematic differences between spatial structures of various THM types.
Our considerations allow to delineate specific expectations for the five types of THM under
study (Table 1). Major goals commonly associated with refuge- and asylum-seeking are mere
survival, fulfillment of basic needs and security. These goals can usually be fulfilled in many
places, oftentimes already in neighboring countries just outside a warzone. The resource
stock available to refugees and asylum-seekers tends to be rather low. As a result, the spatial
structure of refuge- and asylum-seeking is expected to feature a high power-law fit and scaling
5
Table 1: Theoretical expectations concerning THM
Major goal(s)
Ubiquity
Resources Power-
High
High
fit
law
(R2)
High
Low
Medium Medium-
high
Scaling
exponent
(β)
Large
Medium-
high
Refugees
and
asylum-seekers
Tourists
Migrants
Students
Survival, fulfilment of
basic needs, security
Entertainment, pleas-
ant environment/ busi-
ness opportunities
Economic well-being
Education, distinction
Medium Medium Medium Medium
Low
Small
High
Low
exponent. Tourists are often interested in a pleasant, entertaining environment (if on holidays)
or in business opportunities (if on business trips). Both are widely available in many countries
around the world (and easier to pursue in closer ones, e.g. due to cultural similarity), but tourists
are likely to possess more resources on average than refugees. Therefore, R2 and β should still
be high for tourism but slightly lower than for refuge-seeking. Migrants are often interested in
improving their economic well-being, which is quite stratified globally and although moving
short distances may already result in relative improvements, moving a bit further may in many
cases still lead to additional benefits. Migrants are also likely to possess more resources on
average than refugees, so that overall we expect a medium R2 and β. International students
tend to aim for excellent education and social distinction, which is best available only in a
small number of institutions in a select number of countries, as the global university system is
highly stratified (Barnett and Wu, 1995). In theoretical terms, this stratification leads to a lack
of intervening opportunities: students cannot just go to neighboring countries but have to reach
England or the US to attain their goals. Therefore -- and because international students will
also be comparatively well-situated on average -- we expect low R2- and β-values for student
exchange.
For THC, our argument is inspired by the geography-is-dead debate. First of all, the spatial
structure of THC will to some extent be a function of THM, as people will often communicate
with friends, kin or business partners that have gone abroad. Yet, over and above, there may
be variance inherent to the specific type of THC that depends on its technological standard
and usage cost structure. This argument can easily be explained by comparing phone calls
with online friendships. The two networks are similar in that both should to a certain extent
resemble THM. Yet, they are different in that the monetary costs of international telephone
communication are relatively high and increase with distance (Cairncross, 1997, p. 6), while
the material costs of online friendships are low and independent of distance: for someone
from Switzerland, having an online friendship with an Austrian is as cheap as having one
with an Australian, but calling Australia by phone is far more expensive than calling Austria.
Accordingly, we expect the spatial structure of online friendships to be "flatter" and more
detached from the power-law pattern (reflected in lower β and R2 values) than the spatial
structure of phone calls. Remittances are special in that they are additionally influenced by
economic power and we refrain from formulating a specific hypothesis about its spatial structure
ex ante.
In the longitudinal part of the analysis, we build on the geography-is-dead debate's argu-
ments in expecting the world to become "flatter" with globalization, and thus the type-specific
R2- and β-values to decline over time. In a final analytical step, we contrast the empirical THA
distribution with three ideal geography-is-dead scenarios.
6
Table 2: Types of THA used
THA Type
THM Asylum seekers
Migrants
Refugees
Students
Tourists
Individual
weight
(%)
0.1
16.9
0.8
0.2
82.0
THC Online friendships
Phone calls
Remittances
33.3
33.3
33.3
Combined
weight
(%)
Available
years
Data source(s)
60
40
UNHCR (2013)
2010
1960-2010 World Bank, UN (2012)
2000-2010 UNHCR (2013)
1960-2010
1995-2010 UNWTO (2014)
PINA, UNESCO (2013)
n.d.
1983-1995
2010
Facebook (2012)
PINA, ITU
World Bank
4. Research design
4.1. Data
This study is based on directed network matrices for eight types of THA at various time
points between 1960 and 2010 as well as an undirected matrix for geographic distance, all
of which were standardized4 to cover the same set of 196 sending and receiving countries
(full list available upon request). Each of the 53 THA matrices thus contains information
on 196 × 196 − 196 = 38, 220 country dyads. For the measurement of geographic distance we
draw on CEPII's GeoDist dataset (Mayer and Zignago, 2011). We use the weighted geodesic
distance (distwces), which provides the average distance between countries based on the spatial
distribution of the population in the countries' 25 largest cities. Table 2 provides a summary of
the eight types of THA under study.
Of the eight types of THA, five involve physical mobility (THM): asylum-seeking, migration,
refuge-seeking, student exchange, and tourism. Data on asylum- and refuge-seeking was obtained
from UNHCR. According to the 1951 Refugee Convention (as broadened by a 1967 Protocol), a
refugee is defined as a person who:
"owing to well-founded fear of being persecuted for reasons of race, religion,
nationality, membership of a particular social group or political opinion, is outside
the coun-try of his nationality and is unable or, owing to such fear, is unwilling to
avail himself of the protection of that country; or who, not having a nationality and
being outside the country of his former habitual residence as a result of such events,
is unable or, owing to such fear, is unwilling to return to it." (UNHCR, 2014a)
An asylum-seeker in turn is "someone who says he or she is a refugee, but whose claim has
not yet been definitively evaluated" (UNHCR, 2014b). Data on migration was extracted from
the World Bank's Global Bilateral Migration Dataset for the years 1960 to 2000 (Özden et al.,
2011), supplemented by United Nations data for the year 2010 (UN, 2012). The latter source
defines migrants as "foreign-born" persons, or, where data on place of birth is unavailable, as
"foreign citizens" (UN, 2012, p. 3). Information on transnational student mobility was obtained
from Princeton's International Networks Archive (PINA)5 for the years 1960 to 1998 and from
UNESCO for the years 2000 to 2010. UNESCO defines international students as "[s]tudents
who have crossed a national or territorial border for the purposes of education and are now
7
enrolled outside their country of origin" (UNESCO, 2010, p. 264). Data on tourism, available
from 1995 to 2010, was obtained from the World Tourism Organization (UNWTO), according
to which "[a] visitor (domestic, inbound or outbound) is classified as a tourist (or overnight
visitor) if his/her trip includes an overnight stay" (UNWTO, 2008). Here, we are interested in
"arrivals of non-resident tourists at national borders, by country of residence"6. Note that this
definition does not premise any specific visiting purpose and may thus include business travel
just as holiday trips.
Three types of THA under study represent indirect communication (THC): online friend-
ships, phone calls, and remittances. Online friendships are based on Facebook data which was
retrieved from an interactive graph that is available online (Facebook, 2012) and converted into
a network matrix. For each country c, this matrix contains the five countries with which c's
population has most Facebook friendships with, ranked from 5 (highest number of Facebook
friendships) to 1 (fifth-highest number of Facebook friendships). Data on international phone
calls (measured in million minutes) from 1983 to 1995 originates from the International Telecom-
munication Union (ITU) and was retrieved from PINA (cf. Louch, Hargittai and Centeno,
1999). Information on remittances was obtained from the World Bank (cf. Ratha and Shaw,
2007). Remittances can be defined as "current private transfers from migrant workers who
are considered residents of the host country to recipients in the workers' country of origin"
(WorldBank, 2011, p. xvi). We regard remittances as a type of THC because they are transfers
between individuals that "often involve related persons" (IMF, 2005, p. 75) and can thus be
understood as expressions of support or solidarity, and ultimately as a form of communication.
In addition to analyzing these eight activity types individually, we are also interested in
getting an idea of what the spatial structure of THA looks like as a whole. The multiplexity of
human mobility and communication, i.e. the variety of ways in which people interact, needs to
be addressed as concentrating on single activity types alone may lead to biases (Stopczynski
et al., 2014). To do so, we link the activity types in three indices. First, a THM index, in which
the cell values of the 2010 matrices of the five types of mobility are added up. This simple
procedure is reasonable because all mobility networks are based on the same unit of analysis
(individuals moving between countries). As shown in Table 2, the weight in the THM index
differs drastically by mobility type, with tourists and migrants making up for 82.0 and 16.9
percent, respectively, whereas asylum seekers, refugees7, and students taken together account
for only 1.1 percent of all THM. Second, a THC index is created from the latest available matrix
of the three forms of communication under study. This is less straightforward as the units
differ between the types of THC (remittances are in US-Dollar, phone calls in minutes, etc.). We
deal with this issue by normalizing the units and calculating the average value across the three
types of THC, giving each of them the same weight. Third, we create a THA index by adding
the standardized values of THM and THC, giving a weight of .6 to the former and a weight
of .4 to the latter. The purpose of these factors is to account for the fact that physical mobility
requires more effort than indirect communication and should therefore receive more weight.
The overall indices should be understood as only providing a tentative impression of THM,
THC, and THA as a whole, because (a) we do not include all conceivable activity types, (b) the
size of the weighting factors in the latter two indices is to a certain extent arbitrary, and (c) not
all elements date from the same year (although our finding of long-term stability [see below]
indicates that older data can readily be used as a proxy). Despite these shortcomings, we think
that our indices constitute a significant first step to covering the multiplex nature of THA.
4.2. Methods
In order to determine the spatial structure of THA, we compare the empirically observed
probability density functions of distances r (in km) to the ideal pattern of a power-law distri-
bution and describe to which extent the two are similar. To do so, we use the curvefit module in
8
Figure 1: Probability density functions of THA and their power-law fit
9
Stata, which provides a goodness-of-fit measure (R2) and other relevant parameters like the
scaling exponent β.
While the potential step lengths and directions are continuous in real physical space, here
they are limited to a specific set based on the grid of the world's nation-states. In order to solve
the issue that not every conceivable distance actually exists in this real-world set of country
dyads (leading to spurious gaps) while some country pairs coincidentally feature the very same
distance (resulting in spurious spikes), individual observations had to be binned first, i.e. step
lengths that lie close to each other needed to be aggregated (cf. Buchanan, 2008). After trial
computations with lower and higher step lengths, we decided to use a bin width of 500km as a
reasonable compromise between inflating the variance and overly flattening the distribution.
The binning results in a reduction of data points, i.e. the 38,220 original observations in each
matrix are pooled into 39 meta-observations. In fitting the power law, we focus on the tail of
the binned distribution (i.e. the part to the right of the global maximum, which lies at 500km
for most activity types). This is necessary because activity at short distances (i.e. left of the
global maximum) is artificially restricted as we focus on transnational activity alone and cannot
capture the huge amount of activity occurring within countries.
As a robustness check, we partially repeated the analysis with Kernel-weighted local poly-
nomial smoothing using an Epanechnikov-kernel function and the ROT-bandwidth estimator,
which minimizes the conditional weighted mean integrated squared error. This procedure, in
which the number of observations remains at 38,220, leads to similarly high power-law fits
(results available upon request). Due to space restrictions, we only present the findings from
the more parsimonious binning method.
5. Results
We first examine the spatial structure of THA today, then turn to the analysis of longitudinal
trends, and finally compare our findings to theoretical scenarios and results from other studies.
5.1. The spatial structure of transnational human activity today
Fig. 1 shows the probability density distributions of all types of THA under study at the latest
available point in time. The dots represent binned empirical observations, whereas the solid
lines depict power-law curves that are fitted to the tails of the empirical distributions. The
goodness-of-fit is indicated by the R2 in the upper right corner (0=no fit, 1=perfect fit), by
which the subgraphs are sorted. The scaling exponent β, placed below the R2, indicates the
steepness of the power-law curve. To additionally illustrate the power-law fit, the insets show
the distributions and curves on logarithmic axes (on which the power law takes the form
of a straight line along which the observations should cluster), for displacement lengths of
r ≤ 15, 000km.
Concerning THM (Fig. 1A), the power-law fit is highest for refuge-seeking (R2=.996), and
just slightly lower for tourism (R2=.982), asylum-seeking (R2=.951), and migration (R2=.931).
Only for student exchange does the power law not fit the empirical distribution well (R2=.799).
The spatial structure of student exchange differs from that of the other THA types in that there
is a second peak at middle-range distances (approx. 7,000-14,000km). A closer look at the data
reveals that this peak results mainly from large flows between three country pairs (China -- US,
India -- US, and Australia -- UK), which fits our assumption about the role of the global university
system's heavy stratification and the related lack of intervening opportunities. The scaling
exponent β is highest for refuge-seeking (β=2.13), medium-high for tourism (β=1.62), medium
for migration (β=1.47), and small for asylum-seeking (β=1.19) and student exchange (β=0.73).
Apart from the unexpected low β for asylum-seekers8, the theoretically expected order holds,
indicating that type-specific goal-attainment opportunities and available resources help predict
10
spatial structures of human mobility across nation-state borders. THM as a whole also clearly
follows a power law (R2=.977, β=1.51).
Regarding THC (Fig. 1B), the power-law fit and scaling exponent are highest for online
friendships (R2=.968, β=1.19), lower for phone calls (R2=.897, β=1.13), and smallest for re-
mittances (R2=.886, β=0.95). The smaller scaling exponent for analogue phone calls in 1995
compared to digital Facebook friendships in the late 2000s is clearly at odds with the popular
argument that lower communication costs lead to a "death of distance", or at least to a flatter
probability density distribution -- surprisingly, our data suggests the opposite.
With respect to THA as a whole (Fig. 1C), we again find an excellent power-law fit (R2=.974,
β=1.21), indicating that most human activity beyond nation-state borders today only spans rela-
tively short distances and genuinely global mobility and communication is still extremely rare.
Contrary to popular notions about the detachment of human activity from spatial constraints
in the age of globalization in the geography-is-dead debate, the forces of physical space seem
to be intact. But is THA today, compared to the past, at least somewhat more globalized? To
answer this question, we now turn to analyzing trends over time.
5.2. Developments over time
To investigate longitudinal trends, we look at four different indicators: the overall amount of
THA (Fig. 2A), the percentage of THA that takes place at relatively short distances (Fig. 2B), the
power-law fit (Fig. 2C), as well as the scaling exponent β (Fig. 2D).
Fig. 2A illustrates that almost all types of THA have seen massive growth over the years.
The number of transnationally mobile students (depicted by triangles) grew exponentially from
255,000 in 1960 to 2.9 million in 2010, an 11.5-fold increase. Transnational phone-call minutes
(depicted by crosses9) increased from 12.4 billion in 1983 to 56.6 billion in 1995, rising 4.6-fold in
little more than a decade. Similarly, the number of transnational tourists (depicted by diamonds)
rapidly grew from 457.6 million in 1995 to 987.3 million in 2010, an increase of 115.8 percent.
The number of migrants (depicted by squares) was estimated to be 91.2 million in 1960 and
more than doubled (plus 123.6 percent) to 204.0 million in 2010. Only the number of refugees
(depicted by circles) does not follow a clear direction over time. All in all, the amount of THA
has increased dramatically10 over the last decades.
By contrast, Fig. 2B-C show a remarkable degree of over-time stability when it comes to
the spatial structure of THA. Fig. 2B illustrates the percentage of THA that takes place at
displacement lengths of r ≤ 5, 000km, i.e. at distances shorter than about 1⁄4 of the largest
possible distance between two countries. For comparison: the driving distance between Seattle
and Miami is 5,400km. We are thus talking about relatively short distances at the planet scale.
The graph reveals two things: (a) the percentage of mobility and communication that takes
place below this threshold varies between types of THA, and (b) it remains strongly attached to
these type-specific levels over time. For students, the percentage of stays abroad at distances of
r ≤ 5, 000km remained at about 50.3 percent (standard deviation σ = 2.6 percent) over a period
of fifty years. Similarly, in every decade from 1960 to 2010, about 77.7 percent (σ = 1.8 percent)
of migrants moved these relatively short distances. At nine measured time points between
1995 and 2010, the fraction of tourists that went such short distances varied closely around 86.2
percent (σ = 0.9 percent). For refugees, the share was about 90.2 percent (σ = 2.5 percent). Only
phone calls (mean: 71.9 percent, σ = 2.7 percent) appear to witness a slow but constant drop --
yet the development of refuge-seeking during the 2000s shows how quickly such a trend can
reverse. Overall, the consistently low standard deviations highlight the robustness of the spatial
structure of THA over time. Similarly, the power-law fit (Fig. 2C) remains remarkably stable.
For phone calls, it remains solid at about .886 (σ = .014). In the case of migration, it varies
steadily around .950 (σ = .023). Tourism and refuge-seeking remain at the ceiling, with fits of
.986 (σ = .005) and .986 (σ = .009) respectively. For student exchange there are some ups and
11
e
m
i
t
r
e
v
o
s
d
n
e
r
T
:
2
e
r
u
g
i
F
12
Figure 3: Log-log plots of THA over time
downs, but all in all the power-law fit oscillates around .768 (σ = .052). Thus, the power-law fit
is another indicator of the startling long-term stability of the spatial structure of THA.
Per contra, Fig. 2D shows a mixed picture when it comes to the development of the scaling
exponent β, i.e. the "flatness of the world". For tourists and migrants (which together represent
98.9 percent of all THM) a steady drop in the size of β can be observed. For migration, this
decrease is first visible in the 1980s, after two decades of absolute stability. With respect to
phone calls, we see only a slight gradual decrease over time. Concerning refugees, β first drops
between 2000 and 2004, but then increases again until 2010. For students on the other hand, the
scaling exponent remains more or less stable at a low level. This finding of partial stability and
partial flattening can be examined more closely in log-log plots (Fig. 3).
The graphs in Fig. 3 show the best-fitting power-law lines (which form straight lines here, as
both axes are logarithmic) for the spatial structures of the four types of THA with a clear β-value
trend, at various points in time. To visualize both absolute growth and relative shifts, the
y-axes now show the actual amount of mobility and communication taking place. Remarkable
differences become apparent: for students and phone calls (Fig. 3A-B), the fitted power-law
13
curves form parallel lines that just move upwards on the y-axes as the number of students and
phone-call minutes increases over time. In other words, student exchange and phone calls grow
at more or less the same rate at all distances over time, thus retaining their spatial structures'
shapes. For tourists and migrants (Fig. 3C-D) however, the picture looks different. Here, the
lines still move upwards, but also become flatter over the years. This difference shows that no
universal trends hold for all types of THA, underlining the fruitfulness of our comparative
approach.
At first sight, there appears to be a contradiction between the over-time stability observed
in Fig. 2B and the shifts unveiled in Fig. 3C-D. Are at least some types of THA "globalizing"
over time or not? The seeming paradox can be resolved by reminding oneself that the straight
lines in Fig. 3 are only imaginary -- in reality (i.e. on regular axes) they form heavily bended
curves. Shifts that are meaningful in size only occur within the short-distance range (i.e. at
displacement lengths of r ≤ 5, 000km), leaving the share of long-distance THA practically unaf-
fected. Therefore, these changes should rather be taken as signs of regionalization (i.e. gradual
extensions of the spatial reach of THA at a regional scale) than as evidence for globalization. At
the planet-wide scale, stability prevails -- even for tourism and migration (as visible in Fig. 2B).
5.3. Relation of our results to theoretical scenarios and other studies
So far, we have mainly adhered to the Lévy-flight debate's approach in our analysis and only
made allusions to the geography-is-dead debate en passant in the longitudinal analysis by
assuming a "flattening" world. In this last analytical step, we examine what would happen if the
geography-is-dead debate's statements were understood in a stricter sense and distance played
no role at all anymore. To do so, Fig. 4 compares the empirically observed spatial structure of
THA (our overall index, depicted by the solid line) to various hypothetical scenarios.
The first one, the strict geography-is-dead scenario, shows what the distribution would look like
if the probability of activity to occur would literally be the same at all distances (a "flat world"
in Friedman's terms). The second one, the distance scenario, refines this assumption by taking
the actual geographic distribution of country dyads in the world into account. As short and
long distances are empirically less common than middle-range distances, the relation would be
reverse-u-shaped if the amount of activity was the same between each of the 38,220 existing
country pairs (reminiscent of Marx and Engels's "universal interdependence of nations"). The
third one, the population scenario, is a further refinement that takes into consideration that
countries differ by population. It shows what would happen if each individual in each country
had the same chance of going to any other country independent of its geographic location (akin
to Toffler's new nomads who "leave the physical structure behind").
It immediately becomes apparent that none of the three scenarios comes anywhere near the
actual data. In reality, far more THA occurs at short distances and far less at longer distances
than the geography-is-dead hypothesis in its strict sense would suggest. The dotted power-law
line, which we include here again for comparison, shows that the Lévy-flight is much better
at realistically representing the spatial structure of THA than any geography-is-dead scenario.
Again, we find that distance is not dead at all.
Finally, we are also able to compare our findings to those of other studies at lower geographic
scales and for different species. Concerning local human mobility, Gonzalez, Hidalgo and
Barabasi (2008) report a scaling exponent of β=1.75, while Brockmann, Hufnagel and Geisel
(2006) specify β=1.59 for national human mobility in the US. Our value for transnational human
mobility is only slightly lower with β=1.51 and practically identical to the β=1.50 reported by
Noulas et al. (2012) for the global structure of LSS log-ins, which may be taken as an indicator
of robustness. For a set of marine predators, Sims et al. (2008) found β=2.12, whereas for spider
monkeys, Ramos-Fernández et al. (2004) name β=2.18. This consistent ranking of the scaling
exponents could indicate that while mobility follows a Lévy-flight pattern at all geographical
14
Figure 4: The relation of empirical THA to theoretical scenarios
scales and across many species, the scaling exponent is smaller (i.e.
longer distances are
relatively more common) in humans than in non-human species and at higher geographic scales
than at lower geographic scales. With due caution not to over-interpret these numbers (which
are known to be susceptible to fluctuation), we leave this impression to further scrutiny by
future research.
6. Summary and conclusion
Starting from contradicting propositions of largely disconnected debates in the natural and social
sciences, this article comparatively analyzed the worldwide spatial structure of transnational
human activity (THA) across eight different types of mobility and communication and in its
development over time. Four findings should be highlighted:
1. Contrary to popular social scientific accounts, geography is not "dead". The large majority
of THA still occurs at short distances and truly global mobility and communication con-
tinues to be scarce. The spatial structure of THA is similar to that of animal displacements
and local-scale human motion in that it can be approximated by Lévy flights with heavy
tails that obey power laws.
2. Scaling exponent and power-law fit differ by type of mobility, being highest in refuge-
seeking and tourism and lowest in student exchange. This pattern suggests that the
availability of opportunities for attaining type-specific goals as well as the resource stock
disposable on average to the individuals engaging in a specific mobility type play a role
in determining the precise shape of the spatial structure.
3. For communication, we observe that distance is stronger in determining online friend-
ships in the late 2000s than analogous phone calls more than a decade earlier. This
finding is completely at odds with Cairncross's popular argument that declining costs of
communication lead to a "death of distance".
4. Despite dramatic increases in the absolute amount of transnational mobility and com-
munication over the years, the Lévy-flight pattern remains intact and remarkably stable
over time. Longitudinal change occurs only for some types of THA and predominantly at
15
short distances, indicating shifts at the regional level rather than globalization.
Overall, our findings suggest that geography still shapes the patterns of planet-scale human
activity -- be it physical mobility or indirect communication. Where humans interact across
nation-state borders, they are very likely to do so with neighboring countries and within the
world regions they live in. The discrepancy between our findings and popular accounts of
globalization -- from McLuhan's "global village" to Friedman's "flat world" -- may indicate
that we as humans tend to overestimate our capability to transcend nature. Despite the profound
technological, socio-economic and infrastructural revolutions that took place during the last
half century -- from the end of the bipolar world system after the fall of the iron curtain to the
rise of mass tourism and the dawn of the internet age -- the overall patterns of human mobility
remain largely unchanged, bound to the same natural laws as the motion of all kinds of species,
from plankton to penguin.
With this study, we aimed at filling a gap in the Lévy-flight debate by conducting an encom-
passing analysis of the spatial structure of planet-scale human activity, and at complementing
the geography-is-dead debate by subjecting some of its propositions to a rigorous empirical test.
Now we can say that our findings fit the Lévy-flight model well, whereas the death-of-distance
arguments turned out not to hold under closer scrutiny for individual human mobility and
communication. This result underlines the need for sociologists to engage in research trends
beyond the boundaries of their own discipline. We deem further exchange concerning this topic
across natural- and social-scientific camps highly desirable.
There are many paths future empirical research could take: for one thing, it should be
gainful to abandon the global perspective to comparatively examine the spatial structure by
world region or separately for the world system's center and periphery. Another productive
next step would be to compare the over-time stability of THA with the longitudinal systemic
stability found in other domains, e.g. in the tripartite global wealth structure (Babones, 2005).
Furthermore, future research should aim at going beyond the Lévy-flight debate's simplistic
assumption of a random walker. As the directions of human activity are actually not random
(Rhee et al., 2011; Song et al., 2010), it needs to be scrutinized which factors (economic, political,
legal, cultural, etc.) influence where humans go. Our analysis of the role of opportunities for
type-specific goal-attainment and the average amount of available resources may be seen as
a sketchy first step into this direction that needs to be further expanded.11 Finally, it would
be desirable to link our insights with other fields of research that are heavily intertwined with
THA, from global food supplies and transportation to electricity transmission and the spread of
epidemics. We share the hope of Hui et al. (2010) that new insights into how we as humans
move and communicate worldwide may eventually help tackle such pressing global problems.
Notes
1. A power law describes the relationship between two quantities, where one quantity varies
at the power of the other. Mathematically, the power-law relation between two quantities y
and x can be defined as y = ax−β, where a is a prefactor and β is the scaling exponent. Power
laws are common in the social and natural world and have been shown to exist in phenomena
as diverse as the gamma-ray intensity of solar flares (Clauset, Shalizi and Newman, 2009),
the sizes of strike waves (Biggs, 2005), and the distribution of accusations between detainees
(Deutschmann, 2014).
16
2. A Lévy flight is a random walk (i.e. a succession of steps into random directions) in which
the step-lengths feature a heavy-tailed probability distribution. For a precise mathematical
definition, cf. Shlesinger and Klafter (1986). In empirical research, Lévy flights are diagnosed
by "showing that the power-law distribution holds" (Buchanan, 2008, p. 715).
3. In contrast to the Lévy-flight debate's assumption of a non-linear power-law relation, these
economic studies seem to assume a linear relation (e.g., when describing the influence of
distance on trade in the form of a single "distance coefficient" or "elasticity"), as such mean
effect sizes would otherwise be rather meaningless.
4. The set of 196 countries in the standardized matrices was obtained by excluding countries
that were not contained in the original datasets of all types of THM under study. For some types
of THC, the original set of countries was smaller and the matrices were artificially expanded. We
decided to follow this inclusive approach as a more exclusive one would have resulted in a small,
limited remainder of countries that would hardly allow speaking of a planet-scale analysis
anymore. We set missing values to zero in all matrices, thereby following the example of Reyes
(2013, p. 154), who showed the robustness of this approach for the UNWTO tourism dataset and
finds that alternative procedures like multiple imputation lead to similar outcomes. In order
to make the student and phone-call matrices from earlier years comparable with recent ones,
historic states were replaced with their currently existing "equivalents". For example, Dahomey
was equated with Benin, Upper Volta with Burkina Faso, and so on (full list of equations
available upon request). This procedure does obviously not do justice to the complexity of
historic developments, but what matters for the purposes of this article is a similar geographic
location, not precise historic equivalence. The World Bank migration data, which also goes back
to 1960, already comes in the form of currently existing states for all years in its original format.
5. Available at http://www.princeton.edu/~ina/, accessed 13/08/2013.
6. For a few countries this category is unavailable. In order not to lose these countries, the
category "arrivals of non-resident visitors at national borders" was used in these instances. In
cases where both these categories are missing, the category "arrivals of tourists in all types of
accommodation establishments" was used instead.
7. In some countries' census statistics, refugees are also counted as migrants, while in others
they are not (UN, 2012). The UN tries to account for this by including refugees in "most
developing countries" as migrants in the migration dataset (Ibd). It is therefore possible that
refugees are sometimes counted twice. However their small relative weight (0.8 percent) shows
that this issue has little practical consequences for the overall THM index.
8. The unexpected values for asylum-seeking should not be overinterpreted as the number
of asylum-seekers is relatively low, making the structure susceptible to rather meaningless
fluctuations.
9. For phone calls, post-1995 values are extrapolated by fitting an exponential trend line
(R2=.998) of the form y = 1E + 10e0.13x, where x is 1 in 1983, 2 in 1984, etc. This assumption of
exponential growth, which leads to an estimate of 380.9 billion minutes for 2010, is conservative
given that TeleGeography (2014, p. 2) estimates the international call volume to exceed 400
billion minutes in 2010.
10. One may want to consider that the world population also increased over time: in the 196
countries under study, it grew from 3.0 billion in 1960 to 6.9 billion in 2010 (short-dashed line in
Fig. 2A). Student mobility, tourism and phone calls grew at a faster rate, but, contrary to what
catchphrases like "age of migration" (Castles and Miller, 2005) suggest, the relative strength of
migrants as a share of the world population decreased during the 1960s, '70s, '80s, and '90s,
and only in 2010 reached the level it had in 1960.
11. In fact, this paper is part of a larger project in which this question is tackled.
17
References
Allen, John and Chris Hamnett. 1995. A Shrinking World?: Global Unevenness and Inequality.
Oxford: Oxford University Press.
Babones, Salvatore J. 2005. "The country-level income structure of the world-economy." Journal
of World-Systems Research 11(1):29 -- 55.
Barnett, George A and Reggie Yingli Wu. 1995. "The international student exchange network:
1970 & 1989." Higher Education 30(4):353 -- 368.
Bartumeus, Frederic, Francesc Peters, Salvador Pueyo, Celia Marrasé and Jordi Catalan. 2003.
"Helical Lévy walks: adjusting searching statistics to resource availability in microzooplank-
ton." Proceedings of the National Academy of Sciences 100(22):12771 -- 12775.
Biggs, Michael. 2005. "Strikes as forest fires: Chicago and Paris in the late nineteenth century."
American Journal of Sociology 110(6):1684 -- 1714.
Brams, Steven J. 1966. "Transaction flows in the international system." The American Political
Science Review 60(4):880 -- 898.
Brockmann, Dirk and Fabian Theis. 2008. "Money circulation, trackable items, and the emer-
gence of universal human mobility patterns." Pervasive Computing, IEEE 7(4):28 -- 35.
Brockmann, Dirk, Lars Hufnagel and Theo Geisel. 2006. "The scaling laws of human travel."
Nature 439(7075):462 -- 465.
Buchanan, Mark. 2008. "The mathematical mirror to animal nature." Nature 453(7196):714 -- 716.
Cairncross, Frances. 1997. The Death of Distance: How the Communications Revolution is Changing
our Lives. Boston, MA: Harvard Business School Press.
Castles, Stephen and Mark Miller. 2005. The Age of Migration. New York: Palgrave Macmillan.
Cheng, Zhiyuan, James Caverlee, Kyumin Lee and Daniel Z Sui. 2011. "Exploring millions of
footprints in Location Sharing Services." ICWSM 2011:81 -- 88.
Choi, Young. 1995. "The effect of social and physical distance on the global communication
networks." International Communication Gazette 54(2):163 -- 192.
Clark, Cal and Richard Merritt. 1987. European community and intra-European communications:
The evidence of mail flows. In Communication and Interaction in Global Politics, ed. Claudio
Cioffi-Revilla, Richard Merritt and Dina Zinnes. Newbury Park, CA: Sage Publications
pp. 209 -- 235.
Clauset, Aaron, Cosma Rohilla Shalizi and Mark EJ Newman. 2009. "Power-law distributions
in empirical data." SIAM review 51(4):661 -- 703.
Deutsch, Karl W and Walter Isard. 1961. "A note on a generalized concept of effective distance."
Behavioral Science 6(4):308 -- 311.
Deutschmann, Emanuel. 2014. "Between Collaboration and Disobedience: The Behavior of the
Guantánamo Detainees and its Consequences." Journal of Conflict Resolution ahead-of-print,
doi: 10.1177/0022002714545331.
Disdier, Anne-Célia and Keith Head. 2008. "The puzzling persistence of the distance effect on
bilateral trade." The Review of Economics and Statistics 90(1):37 -- 48.
18
Facebook. 2012. "Mapping the World's Friendships.".
URL: http: // www. facebookstories. com/ stories/ 1574
Freymeyer, Robert H and P Neal Ritchey. 1985. "Spatial Distribution of Opportunities and
Magnitude of Migration An Investigation of Stouffer's Theory." Sociological Perspectives
28(4):419 -- 440.
Friedman, Thomas L. 2007. The World is Flat. New York: Picador.
Gonzalez, Marta C, Cesar A Hidalgo and Albert-Laszlo Barabasi. 2008. "Understanding
individual human mobility patterns." Nature 453(7196):779 -- 782.
Harvey, David. 1989. The Condition of Postmodernity. Oxford: Blackwell.
Hui, Pan, Richard Mortier, Michal Piórkowski, Tristan Henderson and Jon Crowcroft. 2010.
Planet-scale human mobility measurement. In Proceedings of the 2nd ACM International Work-
shop on Hot Topics in Planet-scale Measurement. ACM.
IMF. 2005. Balance of Payments and International Investment Position Manual. Fifth Edition. Wash-
ington: International Monetary Fund.
Kano, Kazuko, Takashi Kano and Kazutaka Takechi. 2013. "Exaggerated death of distance:
Revisiting distance effects on regional price dispersions." Journal of International Economics
90(2):403 -- 413.
Kirsch, Scott. 1995. "The incredible shrinking world? Technology and the production of space."
Environment and Planning D 13:529 -- 529.
Leamer, Edward E and James Levinsohn. 1995. International trade theory: the evidence. In
Handbook of International Economics, ed. Gene Grossman and Kenneth Rogoff. New York:
Elsevier pp. 1339 -- 1394.
Louch, Hugh, Eszter Hargittai and Miguel Angel Centeno. 1999. "Phone calls and fax machines:
the limits to globalization." Washington Quarterly 22(2):83 -- 100.
Marx, Karl. 1973. Grundrisse: Introduction to the Critique of Political Economy. New York: Random
House.
Marx, Karl and Friedrich Engels. 1948. Manifesto of the Communist Party. New York: International
Publishers.
Mayer, Thierry and Soledad Zignago. 2011. "Notes on CEPII's distances measures: The GeoDist
database." CEPII Working Paper 2011-25 .
McKercher, Bob, Andrew Chan and Celia Lam. 2008. "The impact of distance on international
tourist movements." Journal of Travel Research 47(2):208 -- 224.
McLuhan, Marshall. 1962. The Gutenberg Galaxy. Toronto: University of Toronto Press.
Miller, Edward. 1972. "A note on the role of distance in migration: Costs of mobility versus
intervening opportunities." Journal of Regional Science 12(3):475 -- 478.
Noulas, Anastasios, Salvatore Scellato, Renaud Lambiotte, Massimiliano Pontil and Cecilia
Mascolo. 2012. "A tale of many cities: universal patterns in human urban mobility." PloS one
7(5):e37027.
O'Brien, Richard. 1992. Global Financial Integration: The End of Geography. New York: Council on
Foreign Relations Press.
19
Özden, Çaglar, Christopher R Parsons, Maurice Schiff and Terrie L Walmsley. 2011. "Where on
earth is everybody? The evolution of global bilateral migration 1960 -- 2000." The World Bank
Economic Review 25(1):12 -- 56.
Ramos-Fernández, Gabriel, José L Mateos, Octavio Miramontes, Germinal Cocho, Hernán
Larralde and Barbara Ayala-Orozco. 2004. "Lévy walk patterns in the foraging movements of
spider monkeys (Ateles geoffroyi)." Behavioral Ecology and Sociobiology 55(3):223 -- 230.
Ratha, Dilip and William Shaw. 2007. South-South Migration and Remittances. Number 102 World
Bank Publications.
Ravenstein, Ernest George. 1885. "The laws of migration." Journal of the Statistical Society of
London pp. 167 -- 235.
Rengert, George F, Alex R Piquero and Peter R Jones. 1999. "Distance decay reexamined."
Criminology 37(2):427 -- 446.
Reyes, Victoria. 2013. "The structure of globalized travel: A relational country-pair analysis."
International Journal of Comparative Sociology 54(2):144 -- 170.
Rhee, Injong, Minsu Shin, Seongik Hong, Kyunghan Lee, Seong Joon Kim and Song Chong.
2011. "On the levy-walk nature of human mobility." IEEE/ACM Transactions on Networking
(TON) 19(3):630 -- 643.
Shlesinger, Michael F and Joseph Klafter. 1986. Lévy walks versus Lévy flights. In On Growth
and Form, ed. Harry Stanley and Nicole Ostrowsky. Dordrecht: Martinus Nijhoff pp. 279 -- 283.
Sims, David W, Emily J Southall, Nicolas E Humphries, Graeme C Hays, Corey JA Bradshaw,
Jonathan W Pitchford, Alex James, Mohammed Z Ahmed, Andrew S Brierley, Mark A Hindell
et al. 2008. "Scaling laws of marine predator search behaviour." Nature 451(7182):1098 -- 1102.
Soininen, Janne, Robert McDonald and Helmut Hillebrand. 2007. "The distance decay of
similarity in ecological communities." Ecography 30(1):3 -- 12.
Song, Chaoming, Tal Koren, Pu Wang and Albert-László Barabási. 2010. "Modelling the scaling
properties of human mobility." Nature Physics 6(10):818 -- 823.
Stopczynski, Arkadiusz, Vedran Sekara, Piotr Sapiezynski, Andrea Cuttone, Mette My Madsen,
Jakob Eg Larsen and Sune Lehmann. 2014. "Measuring large-scale social networks with high
resolution." PloS one 9(4):e95978.
Stouffer, Samuel A. 1940. "Intervening opportunities: a theory relating mobility and distance."
American Sociological Review 5(6):845 -- 867.
Takhteyev, Yuri, Anatoliy Gruzd and Barry Wellman. 2012. "Geography of Twitter networks."
Social Networks 34(1):73 -- 81.
TeleGeography. 2014. "TeleGeography Report: Executive Summary.".
URL:
telegeography-report-database/ index. html
http: // www. telegeography. com/ research-services/
Tobler, Waldo R. 1970. "A computer movie simulating urban growth in the Detroit region."
Economic Geography pp. 234 -- 240.
Toffler, Alvin. 1970. Future shock. Random House LLC.
Tranos, Emmanouil and Peter Nijkamp. 2013. "The Death of Distance Revisited: Cyber-Place,
Physical and Relational Proximities." Journal of Regional Science 53(5):855 -- 873.
20
Ugander, Johan, Brian Karrer, Lars Backstrom and Cameron Marlow. 2011. "The anatomy of
the facebook social graph." arXiv preprint arXiv:1111.4503 .
UN. 2012. "Trends in International Migrant Stock: Migrants by Destination and Origin.
Documentation and Methodology.".
URL: http: // esa. un. org/ MigOrigin/
UNESCO. 2010. Global Education Digest 2010: Comparing Education Statistics Across the World.
Montreal: UNESCO Institute for Statistics.
UNESCO. 2013. "Inbound Internationally Mobile Students by Country of Origin.".
URL: http: // data. uis. unesco. org/ Index. aspx? queryid= 171
UNHCR. 2013. "Population Statistics Reference Database, United Nations High Commissioner
for Refugees.".
UNHCR. 2014a. "About Refugees.".
URL:
article&id= 179&Itemid= 54
http: // unhcr. org. au/ unhcr/ index. php? option= com_ content&view=
UNHCR. 2014b. "Asylum-seekers.".
URL: http: // www. unhcr. org/ pages/ 49c3646c137. html
UNWTO. 2008. International Recommendations for Tourism Statistics 2008. Madrid: UNWTO.
UNWTO. 2014. Compendium of Tourism Statistics Dataset. Madrid: UNWTO, data updated on
12/01/2014.
Viswanathan, GM, Sergey V Buldyrev, Shlomo Havlin, MGE Da Luz, EP Raposo and H Eugene
Stanley. 1999. "Optimizing the success of random searches." Nature 401(6756):911 -- 914.
Webb, Kieran. 2007. "The Continued Importance of Geographic Distance and Boulding's Loss
of Strength Gradient." Comparative Strategy 26(4):295 -- 310.
WorldBank. 2011. Migration and Remittances Factbook 2011, Second Edition. Washington: World
Bank.
Zhou, Min. 2011. "Intensification of geo-cultural homophily in global trade: Evidence from the
gravity model." Social Science Research 40(1):193 -- 209.
Zipf, George Kingsley. 1949. Human Behavior and the Principle of Least Effort. Addison-Wesley
Press.
21
|
1904.04626 | 1 | 1904 | 2019-04-09T12:43:33 | Distributed Computation of Top-$k$ Degrees in Hidden Bipartite Graphs | [
"cs.SI",
"cs.DC"
] | Hidden graphs are flexible abstractions that are composed of a set of known vertices (nodes), whereas the set of edges are not known in advance. To uncover the set of edges, multiple edge probing queries must be executed by evaluating a function $f(u,v)$ that returns either true or false, if nodes $u$ and $v$ are connected or not respectively. Evidently, the graph can be revealed completely if all possible $n(n-1)/2$ probes are executed for a graph containing $n$ nodes. However, the function $f()$ is usually computationally intensive and therefore executing all possible probing queries result in high execution costs. The target is to provide answers to useful queries by executing as few probing queries as possible. In this work, we study the problem of discovering the top-$k$ nodes of a hidden bipartite graph with the highest degrees, by using distributed algorithms. In particular, we use Apache Spark and provide experimental results showing that significant performance improvements are achieved in comparison to existing centralized approaches. | cs.SI | cs |
Distributed Computation of Top-k Degrees
in Hidden Bipartite Graphs
Panagiotis Kostoglou1, Apostolos N. Papadopoulos1, and
Yannis Manolopoulos2
1 Aristotle University of Thessaloniki, Greece,
panakost,[email protected],
2 Open University of Cyprus, Cyprus,
[email protected]
Abstract. Hidden graphs are flexible abstractions that are composed of
a set of known vertices (nodes), whereas the set of edges are not known
in advance. To uncover the set of edges, multiple edge probing queries
must be executed by evaluating a function f (u, v) that returns either true
or false, if nodes u and v are connected or not respectively. Evidently,
the graph can be revealed completely if all possible n(n − 1)/2 probes
are executed for a graph containing n nodes. However, the function f ()
is usually computationally intensive and therefore executing all possible
probing queries result in high execution costs. The target is to provide
answers to useful queries by executing as few probing queries as possible.
In this work, we study the problem of discovering the top-k nodes of
a hidden bipartite graph with the highest degrees, by using distributed
algorithms. In particular, we use Apache Spark and provide experimental
results showing that significant performance improvements are achieved
in comparison to existing centralized approaches.
Keywords: Graph mining · Hidden networks · Top-k degrees.
1
Introduction
Nowadays, graphs are frequently used to model real-life problems in many diverse
application domains, such as social network analysis, searching and mining the
Web, pattern mining in bioinformatics and neuroscience. Graph mining is an
active and growing research area, aiming at knowledge discovery from graph
data [4,1]. In its simplest form, a graph G(V, E) is defined by a set of nodes (or
vertices) V and a set of edges (or links) E. Each edge e ∈ E connects a pair of
nodes. In this work, we focus on simple bipartite graphs, where the set of nodes
is composed of two subsets B (the black nodes) and W (the white nodes), such
as V = B ∪ W and B ∩ W = ∅. Moreover, all edges of G connect nodes from B
and nodes of W , meaning that the two endpoints of an edge belong to different
node subsets.
Bipartite graphs have many interesting applications in diverse fields. For
example, a bipartite graph may be used to represent product purchases by cus-
tomers. In this case, an edge exists between a product p and a customer c when
2
P. Kostoglou et al.
p was purchased by c. As another example, in an Information Retrieval or Text
Mining application a bipartite graph may be used to associate different types of
tokens that exist in a document. Thus, an edge between a document d and a to-
ken t represents the fact that the token t appears in document d. Moreover, there
are many applications of bipartite graphs in systems biology and medicine [7].
Fig. 1. An example of an undirected bipartite graph, with two bipartitions B =
{u, v, w} and W = {a, b, c, d}.
Figure 1 depicts a simple bipartite graph containing seven nodes and five
edges. Among the black nodes, u has the maximum number of neighbors (i.e.,
three). One of the nodes is isolated, i.e., it does not have any neighbors.
Motivation and Contributions. Conventional bipartite graphs are character-
ized by the fact that the sets of vertices B and W as well as the set of edges
E are known in advance. Nodes and edges are organized in a way to enable
efficient execution of fundamental graph-oriented computational tasks, such as
finding nodes with large degree, computing shortest paths, clustering and com-
munity detection. Usually, the adjacency list representation is being used, which
is a good compromise between space requirements and computational efficiency.
However, a concept that has started to gain significant interest recently is that
of hidden graphs. In contrast to a conventional graph, a hidden bipartite graph
is defined as G(B, W, f ()), where B and W are the subsets of nodes and f ()
is a function B × W → {0, 1} which takes as an input two vertex identifiers
and returns true or false if the edge exists or not respectively. Therefore, in a
hidden graph the edge set E is not given explicitly and it is inferred by using
the function f ().
Hidden graphs are very powerful because they are able to represent any type
of relationship among a set of entities. Formally, for n nodes, there is an exponen-
tial number of different graphs that may be generated (the exact number is 2(n
2)).
To materialize all these different graphs demands significant space requirements.
Distributed Computation of Top-k Degrees in Hidden Bipartite Graphs
3
Moreover, it is unlikely that all these graphs will be used eventually. In contrast,
the use of a hidden graph enables the representation of any possible graph by
the specification of the appropriate function f (). It is noted that the function
f () may require the execution of a complex algorithm in order to decide if two
nodes are connected by an edge. Therefore, interesting what-if scenarios may be
tested with respect to the satisfaction of meaningful properties. It is evident,
that the complete graph structure may be revealed if all possible nb × nw edge
probing queries are executed, where nb = B and nw = W . However, this solu-
tion involves the execution of a quadratic number of probes, which is expected
to be highly inefficient, taking into account that the function f () is computa-
tionally intensive and that real-world graphs are usually very large. Therefore,
the target is to provide a solution to a graph-oriented problem by using as few
edge probing queries as possible. The problem we are targeting is the discovery
of the k nodes with the highest degrees. The degree of a node v is defined as the
number of neighbors of v. This problem has been addressed previously by [9] and
[11]. In this work, we are interested in solving the problem by using distributed
algorithms and over Big Data architectures. This enables the analysis of mas-
sive hidden graphs and will provide some baseline for executing more complex
graph mining tasks, thus overcoming the limitations provided by centralized ap-
proaches. In particular, we study distributed algorithms for the discovery of the
top-k degree nodes in Apache Spark [] over YARN and HDFS [10] and we offer
experimental results based on a cluster of 32 physical machines. Performance
evaluation results demonstrate that the proposed techniques are scalable and
can be used for the analysis of massive hidden networks. We note that this is
the first work addressing the problem of hidden graph analysis in a distributed
setting.
Roadmap. The rest of the paper is organized as follows. The next section con-
tains the most representative related work in the area. Section 4 studies our main
contribution in detail. Performance evaluation results are offered in Section 5,
whereas Section 6 concludes our work and presents briefly interesting directions
for future work in the area.
2 Related Research
A hidden graph is able to represent arbitrary relationship types among graph
nodes. A research direction that is strongly related to hidden graphs is graph
property testing [5]. In this case, one is interested in detecting if a graph satisfies
a property or not by using fast approximate algorithms.
Evidently, detecting graph properties efficiently is extremely important. How-
ever, an ever more useful and more challenging task is to detect specific sub-
graphs or nodes that satisfy specific properties, by using as few edge probing
queries as possible.
Another research direction related to hidden graphs focuses on learning a
graph or a subgraph by using edge probing queries using pairs or sets of nodes
4
P. Kostoglou et al.
(group testing) [2]. A similar topic is the reconstruction of subgraphs that satisfy
certain structural properties [3].
An important difference between graph property testing and analyzing hid-
den graphs is that in the first case the graph is known whereas in the second case
the set of edges is unknown and must be revealed gradually. Moreover, property
testing focuses on checking if a specific property holds or not, whereas in hid-
den graph analysis we are interested in answering specific queries exactly. The
problem we attack is the discovery of the top-k nodes with the highest degrees.
This problem was solved by [9] and [11] assuming a centralized environment.
The basic algorithm used in [9] was extended in [8] for unipartite graphs,
towards detecting if a hidden graph contains a k-core or not. Again, the algorithm
proposed in [8] is centralized and its main objective is to minimize the number
of edge probing queries performed.
In this paper, we take one step forward towards the design and performance
evaluation of distributed algorithms for detecting the top-k nodes with the high-
est degrees in hidden bipartite graphs, aiming at the reduction of the overall
computational cost which depends on both the number of edge probing queries
and the level of parallelism used. We note that this is the first work in the
literature to attack the specific problem in a distributed setting.
3 Fundamental Concepts and Problem Definition
In this section, we present some fundamental concepts related to our research and
state the problem formally. Recall that, the detection of the top-k nodes with the
highest degrees in a bipartite graph has been attached in [9] taking a centralized
perspective. In that paper, the authors present the Switch-On-Empty (SOE)
algorithm which shows the best performance and provide an optimal solution
with respect to the number of edge probing queries that are required. Table 1
presents the most frequently used symbols.
Before diving into the details of our proposal, there is a need to describe
briefly the way the SOE algorithm works. SOE receives as input a hidden bipar-
tite graph, with bipartitions B and W . The output of SOE is composed of the
k vertices from B or W with the highest degrees. Without loss of generality, we
are focusing on vertices of B. Edge probing queries are executed as follows:
-- Initially, SOE starts from a vertex b1 ∈ B, selects a vertex w1 ∈ W and
executes f (b1, w1). If the edge (b1, w1) is solid, it continues to perform probes
between b1 and another vertex w2 ∈ W .
-- Upon a failure, i.e., when the probe f (b1, wj ) returns an empty result, the
algorithm applies the same process for another vertex b2 ∈ B. Vertices for
which all the probes have been applied, do not participate in future edge
probes.
-- A round is complete when all vertices of B have been considered. After each
round, some vertices can be safely included in the result set R and they are
removed from B. When a vertex bi must be considered again, we continue
the execution of probes remembering the location of the last failure.
Distributed Computation of Top-k Degrees in Hidden Bipartite Graphs
5
Table 1. Frequently used symbols.
Symbol
Interpretation
G(B, W, f ) a hidden graph
B
W
set of black vertices of G
set of black vertices of G
f (b, w)
true if the edge (b, w) exists, false otherwise
nb, nw
number of black and white vertices
k
number of highest degree vertices requested
N (u)
d(u)
E
m
s(u)
e(u)
probes
set of neighbors of vertex u
degree of vertex u
(unknown) set of edges
(unknown) number of edges (m = E)
number of known existing neighbors of u
number of known non-existing neighbors of u
total number of edge probing queries issued
-- SOE keeps on performing rounds until the upper bound of vertex degrees in
B is less than the current k-th highest degree determined so far. In that case,
R contains the required answer and the algorithm terminates. Note that all
equal-degree vertices are included in R.
In this work, we proceed one step further in order to detect high-degree nodes
in larger graphs using multiple resources. In this respect, our algorithms are
implemented in the Apache Spark engine [12,13] using the Scala programming
language [6]. Spark offers a powerful environment that enables the execution of
distributed applications in massive amounts of data. Apache Spark is a unified
distributed engine with a rich and powerful API for Scala, Python, Java and
R [13]. One of its main characteristics is that (in contrast to Hadoop MapReduce)
it exploits main memory as much as possible, being able to persist data across
rounds to avoid unnecessary I/O operations. Spark jobs are executed based on
a master-slave model in cooperation with a cluster manager such as YARN 3 or
MESOS 4.
4 Proposed Approach
In the following sections, we will focus on the algorithms developed to solve
the distributed detection of top-k nodes in a hidden bipartite graph. Two al-
gorithms are presented that are both inspired by the SOE (Switch-On-Empty)
3 http://hadoop.apache.org/
4 http://mesos.apache.org/
6
P. Kostoglou et al.
algorithm [9]. The first algorithm, Distributed Switch-On-Empty (DSOE), is a
distributed version of the SOE algorithm. The second algorithm, , DSOE∗, is an
improved and more flexible version of DSOE.
4.1 The DSOE Algorithm
Our first solution is very simple but at the same time very effective. The main
mechanism has the following rationale. For a vertex b ∈ B we are executing
edge-probing queries until we get a certain amount of negative results f . The
value of f is relevant to the size of the graph and is being reduced exponentially.
More precisely, the first batch will have f = W /2 the second f = W /4 and so
on. The idea behind this is that, because the majority of the degree distribution
in real life graphs follows a Power law distribution, we do not expect to find
many vertices with a high degree. The closer we get to the point of exhausting
all w ∈ W the smaller we set f , because we want to avoid as many unnecessary
queries as possible. The pseudocode of the routine is given below.
Function routine(v: vertex, f: Int) : void is
while (F alseprobes) ≤ f do
Execute the next query
if edge-probing query==False then
F alseprobes + +
end
end
The routine can be implemented easily to work in a distributed setting of
the form ∀b ∈ B. After each batch of routines, we check for the vertices that
belong to the answer set R. Vertices that belong to R can be recognized easily,
since vertices that have completed all possible probes ∀w ∈ W should be in the
answer. If R ≥ k, then we need one last batch of routines to finalize R.
Algorithm 1: DSOE
Result: Return a set R of vertices that have the highest degree
1 f = 2;
2 while R ≤ k do
3
4
5
map(routine(∀b ∈ B,f )) ;
∀b ∈ B we check if b has probe all the vertices ∀w ∈ W ,each vertices
that did it is being removed from B and is being added to R;
f = f ∗ 2
6 end
7 map(routine(∀b ∈ B,f ));
8 ∀b ∈ B we check if b has probed all the vertices ∀w ∈ W ,each vertices
that did it, is being removed from B and is being added to R;
9 return R;
Distributed Computation of Top-k Degrees in Hidden Bipartite Graphs
7
In this part we will prove that our conditions for the loop and for the vertices
that belong in R are correct. For a vertex b1 ∈ B, if b1 completes all possible
probes from the first routine, then there are only two options for it's degree:
either d(b1) = W or d(b1) = W − 1, because there is a possibility that the last
query could be negative and still b1 could have exhausted all possible probes.
Assume that while we execute the loop we are in the t1 repeat. Also we assume
that after t1 we get a vertex b2 ∈ B that has exhausted all possible edge-probing
queries. For d(b2) again we have two possibilities, either d(b2) = W − t1 or
d(b2) = W − t1 − 1 for the same reason as previously.
From the previous observations we can safely conclude that for a vertex
b3 ∈ B, that has completed all possible queries in t2 with t2 > t1 the best-
case scenario for the degree is d(b3) = W − t2 and the worst case scenario is
d(b2) = W − t2 − 1.
It is evident that d(b2) ≥ d(b3) and the equality is possible only if t2 = t1 + 1.
From this conclusion we are sure that if R ≤ k then if a vertex b ∈ B exhausts
all the queries, it can be inserted into R. Taking into concern the equality we have
to run the batch of routines one last time for the possibility that we may find a
vertex with degree equal to the minimum degree in R. Algorithm 1 contains the
outline of DSOE.
4.2 The DSOE∗ Algorithm
DSOE is the natural extension to the centralized SOE algorithm. However, we
advance one step further in order to improve runtime as much as possible, even
if the number of probes increases. For very large graphs we already expect a
very large number of probes that is inevitable. For that reason trying to improve
the execution time by adding a very small amount of probes in the already
big number of inevitable probes may worth for the overall performance of our
algorithm. The updated algorithm is named DSOE∗.
DSOE treate all vertices equally, so for algorithms we want to make an initial
prediction about the degree of the vertices and after that, to emphasize during
the routines the vertices that we predict to have a large degree by having a
more loose condition in contrast to those that we predict that will have small
degree. The result of this process is to calculate all possible edge-probing queries
for the vertices we predict to have a large degree. With the information of the
exact number of the degree for those vertices we are able to have a better ending
condition for the rest of the vertices. Therefore, we will be able to stop processing
of the vertices with low degrees faster and this way we will have better execution
times. Moreover, for real life graphs, the degrees of the vertices usually follow
a Power law distribution,so we expect that most vertices of the graph will have
way smaller degree in comparison with the degree of the top-k vertices.
More specifically, DSOE∗ initially performs some random edge-probing queries
∀b ∈ B. The number for these queries is set to log(log(W )). After this step,
the course of DSOE∗ is quite similar to that of DSOE. We execute repeatedly
a batch of routines ∀b ∈ B, with the difference that, this time a single routine
for a vertex b ∈ B is complete when N negative edge-probing queries, where N
8
P. Kostoglou et al.
is an outcome of the prediction performed by the sampling process to quickly
detect nodes that potentially have large degree.
When a vertex exhausts all possible queries, then it is added gets added to
a temporary set M . The vertices in the set M may not be completely correct,
so we have to continue with the probing queries. However, the contents of M
provide a threshold T which is set to the minimum degree among the vertices
in M . Probing queries are executed ∀b ∈ B, until d(b) < T . The vertices that
have completed all possible probes are added to M . Finally, the k best vertices
from M with respect to the degree are returned as the final answer. Algorithm
2 contains the pseudocode of DSOE∗.
Function prediction(v: vertex,probes: int) : int is
while counter < probes do
if edge-probing query==True then
prediction++
end
return prediction
end
Function routine(v: vertex) : void is
while negatives < v.prediction do
if edge-probing query==False then
negatives++
end
end
Function exhaust(v: vertex,T: int) : void is
while T ≤ v.degree do
execute the next edge-probing query
end
end
Algorithm 2: DSOE∗
Result: Return a set R of vertices that have the highest degree
10 map(prediction(∀b ∈ B,log(log( W ))));
11 while M ≤ k do
12
13
map(routine(∀b ∈ B));
∀b ∈ B we check if b has probe all the vertices ∀w ∈ W ,each vertices
that did it is being removed from B and is being added to M ;
14 end
15 threshold=min(M);
16 map(exhaust(∀b ∈ B,threshold));
17 ∀b ∈ B we check if b has probe all the vertices ∀w ∈ W ,each vertices that
did it is being removed from B and is being added to M ;
18 R=top-k from M ;
19 return R;
Distributed Computation of Top-k Degrees in Hidden Bipartite Graphs
9
5 Performance Evaluation
In this section, we present performance evaluation results depicting the efficiency
of the proposed distributed algorithms. All experiments have been conducted in
a cluster of 32 physical nodes (machines) running Hadoop 2.7 and Spark 2.1.0.
One node is used as the master whereas the rest 31 nodes are used as workers.
The data resides in HDFS and YARN is being used as the resource manager.
5.1 Datasets
All datasets used in performance evaluation correspond to real-world networks.
The networks used have different number of black and white nodes as well as
different number of edges. More specifically, we have used three networks: DBLP,
YOUTUBE, and WIKI. All networks are publicly available at the Koblenz Net-
work Collection, which is a project to collect large network datasets of dif-
ferent types, in order to assist research in network science and related fields.
The collection is maintained by the Institute of Web Science and Technologies
at the University of KoblenzLandau and it accessible by the following URL:
http://konect.uni-koblenz.de/.
-- DBLP. The DBLP network is the authorship network from the DBLP com-
puter science bibliography. The network is bipartite and a node is either an
author or a publication. Each edge connects an author to one of the publi-
cations. The characteristics of this network are as follows: B = 4, 000, 150,
W = 1, 425, 813, the highest degree for set B is 114, the highest degree for
set W is 951 and the average degree for B is 6.0660 and for W is 2.1622.
-- YOUTUBE1. This is the bipartite network of YouTube users and their
group memberships. Nodes correspond to users and groups, and an edge
between a user and a group denotes a group membership. The properties
of this network are as follows: B = 94, 238 and W = 30, 087, the highest
degree of set B is 1, 035, the highest degree of set W is 7, 591, the average
degree for B is 3.1130 and the average degree of W is 9.7504.
-- YOUTUBE2. This network represents friendship relationships between
YouTube users. Nodes are users and an undirected edge between two nodes
indicates a friendship. Although this graph is unipartite we modified it in
order to become bipartite. We duplicate it's vertices and the set B is one
clone of the nodes whereas W is the other. Two nodes between B and W are
connected if their corresponding nodes are connected in the initial graph.
Evidently, for this graph the statistics for the two sets B and W are exactly
the same: B = W = 1, 134, 890, the maximum degree is 28, 754 for both
sets and the average degree for both sets is 5.2650
-- WIKIPEDIA. This is the bipartite network of English Wikipedia articles
and the corresponding categories they are contained in. The first set of nodes
corresponds to articles and the second corresponds to categories. For this
graph we have B = 182, 947, W = 1, 853, 493, the highest degree for B is
11, 593 and for W is 54, the average degree for B is 2.048 while for set W is
20.748.
10
P. Kostoglou et al.
5.2 Experimental Results
In the sequel, we present some representative experimental results demonstrating
the performance of the proposed techniques. First, we perform a comparison of
DSOE and DSOE∗ with respect to runtime (i.e., time duration to complete
the computation in the cluster) by modifying the number of Spark executors
running. For this experiment, we start with 8 Spark executors and gradually we
keep on increasing their number keeping k = 10. The DBLP graph has been used
in this case. Figure 2(a) shows the runtime of both algorithms by increasing the
number of executors. It is evident that both algorithms are scalable, since there
is a significant speedup as we double the number of executors. Moreover, DSOE∗
shows better performance than DSOE.
DSOE
DSOE*
)
c
e
s
(
e
m
i
t
n
u
r
14000
12000
10000
8000
6000
4000
2000
0
0
20
40
60
80
100
120
140
s
e
b
o
r
p
f
o
r
e
b
m
u
n
9
2.88x10
9
2.86x10
9
2.84x10
9
2.82x10
2.8x10
9
9
2.78x10
9
2.76x10
9
2.74x10
9
2.72x10
DSOE
DSOE*
1
10
100
1000
number of executors
k
(a) runtime vs number of executors
(b) number of probes vs k
Fig. 2. Comparison of DSOE and DSOE∗ with respect to runtime and number of
probes.
Another important measure that we want to examine is the number of probes
that both algorithms execute. For this experiment, the number of executors
and the execution time are irrelevant, so we focus only on different values of
k. Moreover we are running our algorithms in a significantly smaller dataset
(YOUTUBE1) for different values of k (i.e., 1, 10, 100 and 1000). The corre-
sponding results are given in Figure 2(b). The first observation is that both
algorithms perform a significant amount of probes. However, this was expected
taking into account that if the average node degree is very small or very large,
then many probes are required before the algorithms can provide the answer,
as it has been shown in [9]. Also, in general DSOE∗ requires less probes in
comparison to DSOE for small values of k.
One more very interesting aspect that we examine is the performance of our
approaches with respect to the size of the graph. Our motivation is to compare
the impact of sets B and W to the execution time. For this reason we will
use DSOE∗ on the DBLP graph twice: for the first execution we have B =
4, 000, 150 and W = 1, 425, 813 and the for the second time we reverse the
Distributed Computation of Top-k Degrees in Hidden Bipartite Graphs
11
direction of the queries from W to B. This way the graphs that we compare
have the exact same number of edges and vertices but the differ significantly in
the cardinality of B and W . Figure 3(a) presents the corresponding results. It
is observed that in the reverse case the runtime is significantly higher, since for
every node in W there are more options to perform probes on B. In general, the
cost of the algorithm drops if the cardinality if B is larger than that of W .
)
c
e
s
(
e
m
i
t
n
u
r
6500
6000
5500
5000
4500
4000
3500
3000
2500
2000
1500
focusing on B
focusing on W
7000
6000
5000
4000
3000
2000
1000
)
c
e
s
(
e
m
i
t
n
u
r
Reverse
Original
graph size
(a) runtime vs source set
(b) runtime vs cardinality of B and W
0
182947
1134890
4000150
Fig. 3. Performance of DSOE∗ with respect to the source nodes and the size of the
node sets B and W .
The goal of the last experiment is to test the scalability of the algorithms
by increasing the size of the data. First we focus on B and then on W . For
this experiment DBLP, YOUTUBE2 and WIKIPEDIA datasets have been used.
These graphs have almost the same cardinality in one set and they differ on the
cardinality of the other. For all tests we have used 64 Spark executors and the
implementation of the DSOE∗ algorithm with k = 10. The corresponding results
are given in Figure 3(b). It is evident that although the cardinalities of both B
and W have an impact on peformance, execution time is more sensitive on the
cardinality of W when used as the source set.
6 Conclusions
In this work, we study for the first time the distributed detection of the top-
k nodes with the highest degrees in a hidden bipartite graph. Since the set of
edges is not available a-priori, edge probing queries must be applied in order
to be able to explore the graph. We have designed two algorithms to attack
the problem (DSOE and DSOE∗) and evaluate their performance based on real-
world networks. In general the algorithms are scalable, showing good speedup
factors by increasing the number of Spark executors.
There are many interesting directions for future work. By studying the ex-
perimental results, one important observation is that the number of probes is
12
P. Kostoglou et al.
in general large. Therefore, more research is required to be able to significantly
reduce the number of probes preserving the good quality of the answer. Also,
it is interesting to evaluate the performance of the algorithms in even larger
networks containing significantly larger number of nodes and edges.
Acknowledgments
The authors would like to thank the DaSciM group of the LiX laboratory of
Ecole Polytechnique, and specifically Prof. Michalis Vazirgiannis and Dr. Chris-
tos Giatsidis for sharing the cluster to conduct the experimental evaluation.
References
1. Aggarwal, C.C., Wang, H.: Managing and Mining Graph Data. Springer (2010)
2. Alon, N., Asodi, V.: Learning a hidden subgraph. SIAM Journal on Discrete Math-
ematics 18(4), 697 -- 712 (2005)
3. Bouvel, M., Grebinski, V., Kucherov, G.: Combinatorial search on graphs moti-
vated by bioinformatics applications: A brief survey. In: Revised Selected Papers,
31st International Workshop on Graph-Theoretic Concepts in Computer Science
(WG). pp. 16 -- 27. Metz, France (2005)
4. Cook, D.J., Holder, L.B.: Mining Graph Data. John Wiley & Sons, Inc., USA
(2006)
5. Goldreich, O., Goldwasser, S., Ron, D.: Property testing and its connection to
learning and approximation. Journal of the ACM 45(4), 653 -- 750 (1998)
6. Odersky, M., Spoon, L., Venners, B.: Programming in Scala: Updated for Scala
2.12. Artima Incorporation, USA, 3rd edn. (2016)
7. Pavlopoulos, G., Kontou, P., Pavlopoulou, A., Bouyioukos, C., Markou, E., Bagos,
P.: Bipartite graphs in systems biology and medicine: a survey of methods and
applications. GigaScience 7 (02 2018). https://doi.org/10.1093/gigascience/giy014
8. Strouthopoulos, P., Papadopoulos, A.N.: Core discovery in hidden graphs. CoRR
(to appear in Data and Knowledge Engineering) abs/1712.02827 (2017)
9. Tao, Y., Sheng, C., Li, J.: Finding maximum degrees in hidden bipartite graphs. In:
Proceedings ACM International Conference on Management of Data (SIGMOD).
pp. 891 -- 902. Indianapolis, IN (2010)
10. White, T.: Hadoop: The Definitive Guide. O'Reilly Media, Inc., 4th edn. (2015)
11. Yiu, M.L., Lo, E., Wang, J.: Identifying the most connected vertices in hidden
bipartite graphs using group testing. IEEE Transactions on Knowledge & Data
Engineering 25(10), 2245 -- 2256 (2013)
12. Zaharia, M., Chowdhury, M., Das, T., Dave, A., Ma, J., McCauley, M.,
Franklin, M.J., Shenker, S., Stoica, I.: Resilient distributed datasets: A fault-
tolerant abstraction for in-memory cluster computing. In: Proceedings of the
9th USENIX Conference on Networked Systems Design and Implementa-
tion. pp. 2 -- 2. NSDI'12, USENIX Association, Berkeley, CA, USA (2012),
http://dl.acm.org/citation.cfm?id=2228298.2228301
13. Zaharia, M., Xin, R.S., Wendell, P., Das, T., Armbrust, M., Dave, A., Meng, X.,
Rosen, J., Venkataraman, S., Franklin, M.J., Ghodsi, A., Gonzalez, J., Shenker,
S., Stoica, I.: Apache spark: A unified engine for big data processing. Communi-
cations of the ACM 59(11), 56 -- 65 (Oct 2016). https://doi.org/10.1145/2934664,
http://doi.acm.org/10.1145/2934664
|
1804.02981 | 1 | 1804 | 2018-04-09T13:49:38 | Lumping the Approximate Master Equation for Multistate Processes on Complex Networks | [
"cs.SI",
"physics.soc-ph"
] | Complex networks play an important role in human society and in nature. Stochastic multistate processes provide a powerful framework to model a variety of emerging phenomena such as the dynamics of an epidemic or the spreading of information on complex networks. In recent years, mean-field type approximations gained widespread attention as a tool to analyze and understand complex network dynamics. They reduce the model's complexity by assuming that all nodes with a similar local structure behave identically. Among these methods the approximate master equation (AME) provides the most accurate description of complex networks' dynamics by considering the whole neighborhood of a node. The size of a typical network though renders the numerical solution of multistate AME infeasible. Here, we propose an efficient approach for the numerical solution of the AME that exploits similarities between the differential equations of structurally similar groups of nodes. We cluster a large number of similar equations together and solve only a single lumped equation per cluster. Our method allows the application of the AME to real-world networks, while preserving its accuracy in computing estimates of global network properties, such as the fraction of nodes in a state at a given time. | cs.SI | cs |
Lumping the Approximate Master Equation for
Multistate Processes on Complex Networks
Gerrit Grossmann1, Charalampos Kyriakopoulos1, Luca Bortolussi2, and
Verena Wolf1
1 Computer Science Department, Saarland University
2 Department of Mathematics and Geosciences, University of Trieste
Abstract. Complex networks play an important role in human society
and in nature. Stochastic multistate processes provide a powerful frame-
work to model a variety of emerging phenomena such as the dynamics of
an epidemic or the spreading of information on complex networks. In re-
cent years, mean-field type approximations gained widespread attention
as a tool to analyze and understand complex network dynamics. They
reduce the model's complexity by assuming that all nodes with a similar
local structure behave identically. Among these methods the approxi-
mate master equation (AME) provides the most accurate description of
complex networks' dynamics by considering the whole neighborhood of a
node. The size of a typical network though renders the numerical solution
of multistate AME infeasible. Here, we propose an efficient approach for
the numerical solution of the AME that exploits similarities between the
differential equations of structurally similar groups of nodes. We clus-
ter a large number of similar equations together and solve only a single
lumped equation per cluster. Our method allows the application of the
AME to real-world networks, while preserving its accuracy in computing
estimates of global network properties, such as the fraction of nodes in
a state at a given time.
Keywords: Complex Networks, Multistate Processes, AME, Model Re-
duction, Lumping
1
Introduction
Various emerging phenomena of social, biological, technical, or economic nature
can be modeled as stochastic multistate processes on complex networks [1, 3,
24, 26]. Such networks typically consist of millions or even billions of nodes [1,
3], each one being in one of a finite number of states. The state of a node can
potentially change over time as a result of interaction with one of its neighbor-
ing nodes. The interactions among neighbors are specified by rules and occur
independently at random time points, governed by the exponential distribution.
Hence, the underlying process is a discrete-state space Markovian process in
continuous time (CTMC). Its state space consists of all labeled graphs repre-
senting all possible configurations of the complex network. For instance, in the
susceptible-infective (SI) model, which describes the spread of a simple epidemic
process, each node can either be susceptible or infected; infected nodes propagate
the infection to their susceptible neighbors [19, 5].
Monte-Carlo simulations can be carried out only for small networks [11, 19],
as they become very expensive for large networks, due to the large number
of simulation runs which are necessary to draw reliable conclusions about the
network's dynamics.
An alternative and viable approach is based on mean-field approximations, in
which nodes sharing a similar local structure are assumed to behave identically
and can be described by a single equation, capturing their mean behavior [18,
3, 4, 10, 12]. The heterogeneous (also called degree-based) mean-field (DBMF)
approach proposes a system of ordinary differential equations (ODEs) with one
equation approximating the nodes of degree k which are in a certain state [25,
9, 18]. The approximate master equation (AME) provides a far more accurate
approximation of the network's dynamics, considering explicitly the complete
neighborhood of a node in a certain state [16, 17, 14]. However, the corresponding
number of differential equations that have to be solved is of the order O(cid:0)kSmax(cid:1),
where kmax is the network's largest degree and S the number of possible states.
A coarser approximation called pair approximation (PA) can be derived from
AME by imposing the multinomial assumption for the number of neighbors in
a state [16, 17]. Nevertheless, solving PA instead of AME is faster but for many
networks not accurate enough [17].
Lumping is a popular model reduction technique for Markov-chains and sys-
tems of ODEs [21, 6, 28, 7, 8]. It has also been applied to the underlying model
of epidemic contact processes [27, 19] and has recently been shown to be ex-
tremely effective for the DBMF equation as well as for the PA approach [20].
In this work, we generalize the approach of [20] providing a lumping scheme for
the AME, leveraging the observation that nodes with a large degree having a
similar neighborhood structure have also typically very similar behaviors. We
show that it is possible to massively reduce the number of equations of the AME
while preserving the accuracy of global statistical properties of the network. Our
contributions, in particular, are the following: (i) we provide a fully automated
aggregation scheme for the multistate AME; (ii) we introduce a heuristic to find
a reasonable trade-off between number of equations and accuracy; (iii) we evalu-
ate our method on different models from literature and compare our results with
the original AME and Monte-Carlo simulation; (iv) we provide an open-source
tool3 written in Python, which takes as input a model specification, generates
and solves the lumped (or original) AME.
The remainder of this paper is organized as follows: In Section 2 we describe
multistate Markovian processes in networks and formally introduce the AME. In
Section 3 we derive lumped equations for a given clustering scheme and in Section
4 we propose and evaluate a clustering algorithm for grouping similar equations
together. Case studies are presented in Section 5. We draw final conclusions and
identify open research problems in Section 6.
3 https://github.com/gerritgr/LumPyQest
2 The Multistate Approximate Master Equation
In this section, we first define contact processes and introduce our notation and
terminology for the multistate AME.
2.1 Multistate Markovian Processes
state s.
We describe a contact process in a network (G,S, R, L) by a finite undirected
graph G = (V, E), a finite set of states S, a set of rules R, and an initial state
for each agent (node) of the graph L : V → S. We use s, s0, s00 and s1, s2, . . . to
denote elements of S. At each time point t ≥ 0, each node v ∈ V is in a state
s ∈ S. The rules R define how neighboring nodes influence the probability of
state transitions. A rule consists of a consumed state, a produced state, and a
transition rate, which depends on the neighborhood of the node. We use integer
vectors to model a node's neighborhood. For a given set of states S and maximal
degree kmax, the set of all potential neighborhood vectors is M = {m ∈ ZS
≥0
Ps∈S
m[s] ≤ kmax}, where we write m[s] to refer to the number of neighbors in
A rule r ∈ R is a triplet r = (s, f, s0) with s, s0 ∈ S, s 6= s0 and rate function
f : M → R≥0 corresponding to the exponential distribution. A rule r (also
s0) can be applied at every node in state s, and, when applied,
denoted as s
it transforms this node into state s0. Note that this general formulation of a rule
containing the rate function can express all types of rules that are described in
[16, 17, 14] such as spontaneous changes of a node's state (independent rules) or
changes due to the state of a neighbor (contact rules). The delay until a certain
rule is applied is exponentially distributed with rate f (m), with rules competing
in a race condition where the one with the shortest delay is executed. This results
in an underlying stochastic model described by a CTMC.
f−→
In the following, we indicate with Rs+ = {(s0, f, s) ∈ R, s0 ∈ S} all the rules
that change the state of a node into s, and with Rs− = {(s, f, s0) ∈ R, s0 ∈ S}
all rules that change an s-node into a different state.
Example In the SIS model, a susceptible node can become infected by one
of its neighbors. An infected node becomes susceptible again, independently of
λ1·m[I]
−
−−−−→ I and the recovery rule is
λ2−→ S, where m[I] denotes the number of infected neighbors and λ1, λ2 ∈ R≥0
its neighbors. Hence, the infection rule is S
I
are rule-specific rate constants.
2.2 Multistate AME
Here, we briefly present the multistate AME, similarly to [14, 20]. The AME
assumes that all nodes in a certain state and with the same neighborhood struc-
m[s] = k} to be
the subset of neighborhood vectors referring to nodes of degree k. In addition,
ture are indistinguishable. We define Mk = {m ∈ M Ps∈S
1 ,s−
for s1, s2 ∈ S and m ∈ M, we use m{s+
2} to denote a neighborhood vector
where all entries are equal to those of m, apart from the s1-th entry, which is
equal to m[s1] + 1, and the s2-th entry, which is equal to m[s2] − 1.
Let xs,m(t) be the fraction of network nodes that are in state s and have
a neighborhood m at time t, and assume the initial state xs,m(0) is known.
Formally, the AME approximates the time evolution of xs,m with the following
set of deterministic ODEs4:
∂xs,m
∂t
= X(s0,f,s)∈Rs+
+ X(s1,s2)∈S 2
− X(s1,s2)∈S 2
s16=s2
s16=s2
f (m)xs0,m − X(s,f,s0)∈Rs−
f (m)xs,m
βss1→ss2x
s,m{s+
1 ,s−
2 } m{s+
1 ,s−
2}[s1]
(1)
βss1→ss2xs,mm[s1] ,
where, the term βss1→ss2 is the the average rate at which an (s, s1)-edge changes
into an (s, s2)-edge, if s, s1, s2 ∈ S with s1 6= s2.
The first term in the right hand side models the inflow into (s, m) nodes
from (s0, m) nodes, while the second term models the outflow from (s, m) due
to the application of a rule. The other two terms describe indirect effects on a
(s, m) node due to changes in its neighboring nodes, again considering inflow
and outflow (cf. Fig. 1). In particular, a node in the neighborhood m of (s, m),
say in state s1, changes to state s2 by the firing of a rule.
To compute βss1→ss2 we need to define the subset of rules which consume a
s1-node and produce an s2-node: Rs1→s2 = {(s1, f, s2) ∈ R f : M → R≥0}.
Then
βss1→ss2 = Pm∈M
(s1,f,s2)∈Rs1→s2
xs1,mm[s]
P
Pm∈M
f (m)xs1,mm[s]
,
(2)
where in the denominator we normalize dividing by the fraction of (s, s1) edges.
The total number of equations of AME is determined by the number of states
S and the maximal degree kmax, and equals:
(cid:18)kmax + S
S − 1 (cid:19)(kmax + 1) .
(3)
The binomial arises from the number of ways in which, for a fixed degree k, one
can distribute k neighbors into S different states, see [20] for the proof.
As xs,m are fractions of network nodes, the following identity holds for all t:
Xs,m∈S×M
xs,m(t) = 1
(4)
4 we omit t for the ease of notation
node changes
neighborhood changes
2βss→ss'xs,(2,2)
f2(2, 2)xs,(2,2)
s'
s'
s'
xs',(2,2)
s
s'
xs,(2,2)
s
s
s
f1(2, 2)xs',(2,2)
2βss'→ssxs,(2,2)
s'
s'
xs,(1,3)
s'
s
3βss'→ssxs,(1,3)
3βss→ss'xI,(3,1)
s
s'
xs,(3,1)
s
s
Fig. 1: Illustration of how the AME governs the fraction of xs,(2,2) in a two-state
model with rules (s0, f1, s), (s, f2, s0). The inflow and outflow between xs,(2,2)
and xs0,(2,2) is induced by the direct change of a node's state from s to s0 or vice
versa. The inflow and outflow between xs,(2,2) and xs,(3,1), xs,(1,3) is attributed
to the change of state of a node's neighbor.
Moreover, we use xs to denote the global fraction of nodes in a fixed state s,
which we get by summing over all possible neighborhood vectors
xs(t) = Xm∈M
xs,m(t) ,
(5)
again with Ps∈S
xs(t) = 1. Intuitively, xs is the probability that a randomly
chosen node from the network is in state s. This is the value of primary interest
in many applications, e.g. [3, 24, 26]. Finally, the degree distribution P (k) gives
the probability that a randomly chosen node is of degree k (0 ≤ k ≤ kmax). If
we sum up all xs,m which belong to a specific k (i.e. m ∈ Mk), as the network
structure is assumed to be static, we will necessarily obtain the corresponding
degree probability. Hence, for each t ≥ 0, we have
Xs,m∈S×Mk
xs,m(t) = P (k) .
(6)
3 Lumping
The key idea of this paper is to group together equations of the AME which have
a similar structure and to solve only a single lumped equation per group. This
lumped equation will capture the evolution of the sum of the AME variables in
each group.
Therefore, we divide the set {xs,m s ∈ S, m ∈ M} into groups or clusters,
constructing our clustering such that two equations xs,m, xs0,m0 can only end
up in the same group if s = s0 and m is 'sufficiently' similar to m0. This ensures
that the fractions within a cluster as well as their time derivatives are similar,
provided the change in the rate as a function of m is relatively small when m is
large.
C ⊂ 2M andSC∈C
In the sequel, we consider a clustering C defined as a partition over M, i.e.,
C = M and all clusters C are disjoint and non-empty. Before
we discuss in detail the construction of C in Section 4, we derive the lumped
equations for a given clustering C.
First, recall that we want to approximate the global fractions for each state
(cf. Eq. (5)), which can be split into sums over the clusters
xs(t) =XC∈C Xm∈C
xs,m(t) .
(7)
Our goal is now to construct a smaller equation system, where the variables
zs,C approximate the sum over all xs,m with m ∈ C
xs,m(t) .
(8)
(9)
Henceforth, we can approximate the global fractions as
zs,C(t) ≈ Xm∈C
xs(t) ≈XC∈C
zs,C(t) .
The number of equations is then given by S·C. As one might expect, there is
a trade-off between the accuracy of zs,C(t) and the computational cost, propor-
tional to the number of clusters.
3.1 Lumping the Initial State and the Time Derivative
As the initial values of xs,m are given, we define the initial lumped values
zs,C(0) = Xm∈C
xs,m(0) .
(10)
To achieve the criterion in Eq. (8) for the fractions computed at t > 0, we
seek for time derivatives which fulfill
∂zs,C
∂t ≈ Xm∈C
∂xs,m
∂t
.
(11)
Note that an exact lumping is in general not possible as ∂zs,C
is a function of the
∂t
individual xs,m(t). In order to close the equations for zs,C, we need to express
xs,m as an approximate function of zs,C. The naive idea is to assume that the
true fractions xs,m are similar for all m that belong to the same cluster, i.e., if
m, m0 ∈ C then xs,m ≈ xs,m0, leading to an approximation of xs,m as zs,C/C.
This is however problematic, as it neglects the fact that neighbors of nodes
of different degree have different size. In fact, even if for two degrees k1 < k2
in the same cluster we have P (k1) = P (k2) (while typically P (k1) > P (k2)),
the number of possible different neighbors m2 of a k2-node is larger than the
number of different neighbors m1 of a k1-node, Mk1 < Mk2, hence typically
xs,m2 < xs,m1, as the mass of P (k2) has to be split among more variables. In
order to correct for this asymmetry between degrees in each cluster, we introduce
the following assumption:
Assumption: All fractions xs,m inside a cluster C that refer to the same degree
contribute equally to the sum zs,C. Equations of different degree contribute pro-
portionally to their degree probability P (k) and inversely proportionally to the
neighborhood size for that degree.
Based on the above assumption, we define a degree dependent scaling-factor
wC,k ∈ R≥0, which only depends on the corresponding cluster C and degree k.
According to the above assumption wC,k ∝ P (k)
. To ensure that the weights of
Mk
one cluster sum up to one, we define
wC,k =
P (k)
Mk · Xm∈C
P (km)
Mkm!−1
,
(12)
m[s] is the degree of a neighborhood m. We compute approx-
where km =Ps∈S
imations of xs,m based on zs,C as
xs,m ≈ zs,C · wC,km .
(13)
3.2 Building the Lumped Equations
To define a differential equation for the lumped fraction zs,C, we consider again
Eq. (11) and replace ∂xs,m
by the l.h.s. of Eq. (1). Then we substitute every
occurrence of xs,m by its corresponding lumped variable multiplied with the
scaling factor, i.e., zs,C · wC,km, where m ∈ C. Since m ∈ C does generally
not imply that m{s+
2} ∈ C, the substitution of x
2 } is somewhat more
complicated. Let C(m) denote the cluster m belongs to. If m lies "at the border"
1 ,s−2 }) might be different than C(m). The lumped AME
of a cluster then C(m{s+
s,m{s+
1 ,s−
∂t
1 ,s−
takes then the following form:
wC,kmf (m)(cid:17)
∂zs,C
∂t
zs0,C(cid:16) Xm∈C
= X(s0,f,s)∈Rs+
zs,C(cid:16) Xm∈C
wC,kmf (m)(cid:17)
(cid:16) Xm∈C
zs,C(cid:16) Xm∈C
− X(s,f,s0)∈Rs−
+ X(s1,s2)∈S 2
− X(s1,s2)∈S 2
βss1→ss2
L
βss1→ss2
L
C(m{s+
1 ,s−
s16=s2
w
s16=s2
wC,kmm[s1](cid:17) ,
2 }),km
z
s,C(m{s+
1 ,s−
2 })
m{s+
1 ,s−
2}[s1](cid:17)
(14)
where
zs1,C
= PC∈C
βss1→ss2
L
f (m)wC,kmm[s]
(s1,f,s2)∈Rs1→s2 Pm∈C
PC∈C
P
zs1,C Pm∈C
wC,km m[s]
.
(15)
To gain a significant speedup compared to the original equation system, it is
necessary that the lumped equations can be efficiently evaluated. In particular,
we want the number of terms in the lumped equation system to be proportional
to the number of fractions zs,C and not to the number of xs,m. This is possi-
ble for Eq. (14), because each time we have a sum over m ∈ C, for instance
equations and do not have to evaluate it at every step of the ODE solver. The
sum
Pm∈C wC,km f (m), we can precompute this value during the generation of the
X
m∈C
z
s,C(m{s+
1 ,s−
2 })
m{s+
1 ,s−
2}[s1]
can be evaluated efficiently since we only have to consider lumped variables
that correspond to clusters C(m{s+
2}) that are close to C(m), i.e., that can
be reached from a state in C(m) by the application of a rule. The number of
such neighboring clusters is typically small, due to our definition of clusters, see
Section 4.
1 ,s−
Remark 1. For large kmax, the number of neighbor vectors in M, i.e. the size of
the AME, becomes prohibitively large. For instance, for a maximum degree of
the order of 10 thousands, quite common in real networks, the size of M becomes
of the order of 1012. Even summing a number of elements of this order while
generating equations becomes very costly. To overcome this limit, the solution is
to approximate terms involving summations in Eq. (14). Consider for instance
w.r.t. wC,km, we can only evaluate f at the mean neighborhood vector hmiC,
Pm∈C wC,kmf (m). Instead of evaluating f at every m ∈ C and averaging it
where each coordinate is defined as hmiC[s] =Pm∈C wC,kmm[s]. We can then
approximatePm∈C wC,kmf (m) ≈ f(cid:0)hmiC(cid:1). See Appendix A for details.
4 Partitioning of the Neighborhood Set
In this section we describe an algorithm to partition M, and construct the clus-
tering C. Our algorithm builds partitions with a varying granularity to control
the trade-off between accuracy and execution speed. We consider three main
criteria: the similarity of different equations, their impact on the global error,
and how fast is the evaluation of the lumped equations. Furthermore, as the size
of M can be extremely large, we cannot rely on typical hierarchical clustering
algorithms having a cubic runtime in the number of elements to be clustered.
Our solution is to decouple each m into two components: its degree km (encod-
ing its length) and its projection to the unit simplex (encoding its direction).
We cluster these two components independently.
4.1 Hierarchical Clustering for Degrees
Since our clustering is degree-dependent, we first partition the set of degrees
{0, . . . , kmax}. Let K ⊂ 2{0,...,kmax} be a degree partitioning, i.e., the disjoint
union of all K ∈ K is the set of degrees. The goal of the degree clustering is to
merge together consecutive degrees with small probability while putting degrees
with high probability mainly in separate clusters. This is particularly relevant for
the power-law distribution, which is predominantly found in real world networks
[2, 1] as it allows us to cluster a large number of high degrees with low total
probability all together without losing much information.
We use an iterative procedure inspired by bottom-up hierarchical clustering
to determine K. We start by assigning to each degree an individual cluster and
iteratively join the two consecutive clusters that increase the cost function L
by the least amount. The cost function L punishes disparity in the spread of
probability mass over clusters, leading to clusters that have approximately the
same total probability mass. It is defined as
L(K) = XK∈K(cid:16)Xk∈K
P (k)(cid:17)2
.
(16)
Note that L(K) is minimal when allPk∈K P (k) have equal values. The algorithm
needs O(k2
max) comparisons to determine the degree cluster of each element. At
the end of this procedure, each m ∈ M has a corresponding degree-cluster K
with km ∈ K.
4.2 Proportionality Clustering
Independently of K, we partition M along the different components of vectors
m ∈ M. First, observe that if we normalize m by dividing each dimension by
km, we can embed each Mk into the unit simplex in RS. The idea is then
to partition the unit simplex, and apply the same partition to all Mk. More
specifically, we construct such partition coordinate-wise. As each element of the
normalized m takes values in [0, 1], we split the unit interval in p+1 subintervals
(a)
(b)
Fig. 2: Left: Clustering of M for a 2–state model with kmax = 20 and K =
P = 7. Right: Proportionality cluster of a 3–state model with kmax = 50 and
P = 5. Only the plane M50 is shown.
p , 2
p ), [ 1
p ), . . . , [ p−1
P = {[0, 1
p , 1]}. Then, two normalized neighbor vectors are in
the same proportionality cluster if and only if their coordinates all belong to the
subinterval P ∈ P, possibly different for each coordinate.
4.3 Joint Clusters
Finally, we construct C such that two points m, m0 are in the same cluster if
and only if they are in the same degree-cluster (i.e., ∃K ∈ K : km, km0 ∈ K) and
in the same proportionality cluster, (i.e., for each dimension s ∈ S, there exists
a P ∈ P, such that m[s]
The effect of combining degree and proportionality clusters, for a model with
two different states, is shown in Fig. 2a, where the proportionality clustering gives
equally sized triangles that are cut at different degrees by the degree clustering. If
we fix a degree k, each cluster has only two neighbors (one in each direction). In
the 3d-case, the proportionality clustering creates tetrahedra, which correspond
to triangles if we fix a degree k (cf. Fig. 2b).
km0 ∈ P ).
, m0[s]
km
The above clustering admits some advantageous properties: (1) If we fix a
degree, all clusters have approximately the same size and spatial shape; (2) The
number of 'direct spatial neighboring clusters' of each cluster is always small,
which simplifies the identification of clusters in the 'border' cases and eases
the generation and evaluation of the lumped AME. Hence, the clusters can be
efficiently computed even if M is very large. Next, we discuss how to choose the
size of the clusterings K and P.
4.4 Stopping Heuristic
To find an adequate number of clusters, we solve the lumped AME of the model
multiple times while increasing the number of clusters. We stop when the differ-
ence between different lumped solutions converges. The underlying assumption
is that the approximations become more accurate with an increasing number
of clusters and that the respective difference between consecutive lumped solu-
tions becomes evidently smaller when the error starts to level off. Our goal is to
stop when the increase in the number of clusters does not bring an appreciable
increase in accuracy.
Let z0(t), z00(t) be two solution vectors, i.e., containing the fractions of nodes
in each state at time t, of the lumped AME that correspond to two different
clusterings C0 and C00. We define the difference between two such solutions z0,
z00 as their maximal Euclidean distance over time.
(z0, z00) = max
0≤t≤HsXs∈S(cid:16)z0s(t) − z00s (t)(cid:17)2
.
(17)
For the initial clustering we choose K = P = c0. In each step, we increase the
number of clusters by multiplying the previous ci with a fixed constant, thus
ci+1 = brcic (r > 1). We find this to be a more robust approach than increasing
ci by only a fixed amount in each step. We stop when the difference between
two consecutive solutions are smaller than stop > 0. We consistently observe in
all our case studies that (z0, z00) is a very good indicator on the behavior of the
real error (cf. Fig. 6b). For our experiments we set empirically c0 = 10, r = 1.3,
and stop = 0.01.
5 Case Studies
We demonstrate our approach on three different processes, namely the well-
known SIR model, a rumor spreading model, and a SIS model with competing
pathogens [22, 13]. We test how the number of clusters, the accuracy, and the
runtime of our lumping method relate. In addition, we compare the dynamics
of the original and lumped AME with the outcome of Monte-Carlo simulations
on a synthetic network of 105 nodes [15, 23]. We performed our experiments
on an Ubuntu machine with 8 GB of RAM and quad-core AMD Athlon II X4
620 processor. The code is written in Python 3.5 using SciPy's vode 5 ODE
solver. The lumping error we provide is the difference between lumped solutions
(corresponding to different granularities) and the outcome of the original AME.
That is, for the original solution x and a lumped solution z, we define the
lumping errors of z as (x, z). To generate the error curves, we start with P =
K = 5 and increase both quantities by one in each step. Note that we test our
approach on models with comparably small kmax. In general, this undermines
the effectiveness of our lumping approach; however using a larger kmax would
have hindered the generation of the complete error curve.
(a)
(b)
Fig. 3: SIR model. (a): Lumping error and runtime of the ODE solver w.r.t. the
number of clusters. (b): Fractions of S,I,R nodes over time, as predicted by the
original AME (solid line), by the lumped AME (dashed line), and based on
Monte-Carlo simulations (diamonds).
5.1 SIR
First, we examine the well-known SIR model, where infected nodes (I) go through
a recovery state (R) before they become susceptible (S) again:
S
λ1·m[I]
−
−−−−→ I
I
λ2−−→ R
R
λ3−−→ S .
We choose (λ1, λ2, λ3) = (3.0, 2.0, 1.0) and assume a network structure with
kmax = 60 and a truncated power-law degree distribution with γ = 2.5. The
initial distribution is (xI(0), xR(0), xS(0)) = (0.25, 0.25, 0.5).
In this model the lumping is extremely accurate. In particular, we see that the
lumping error of our method becomes quickly very small (Fig. 3a) and that we
only need a few hundred ODEs to get a reasonable approximation of the original
AME. The lumped solution z we get from the stopping heuristic, consisting of less
than 5% of the original equations, is almost indistinguishable from the original
AME solution x and the Monte-Carlo simulation (Fig. 3b). The lumping error is
(x, z) = 0.0015. The lumped solution used here 1791 clusters with a runtime of
235 seconds while solving the original AME we needed 39711 clusters and 7848
seconds.
5.2 Rumor Spreading
In the rumor spreading model [13], agents are either ignorants (I) who do not
know about the rumor, spreaders (S) who spread the rumor, or stiflers (R)
who know about the rumor, but are not interested in spreading it. Ignorants
learn about the rumor from spreaders and spreaders lose interest in the rumor
when they meet stiflers or other spreaders. Thus, the rules of the model are the
following:
I
λ1·m[S]
−
−−−−→ S
S
λ2·m[R]
−−−−−→ R
S
λ3·m[S]
−
−−−−→ R .
https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.integrate.ode.html
5
Runtime (s)
6000
4000
000
02
39711
20000
Number of Clusters
0.075
0.050
0.025
Lumping Error
0.000
0
IRS
1.0
0.5
Fractions
0.0
0
1
3
2
Time
4
5
(a)
(b)
Fig. 4: Rumor spreading model. (a): Lumping error and runtime of the ODE
solver w.r.t. the number of clusters. (b): Fractions of nodes in each state over
time given by the original AME (solid line), by the lumped AME (dashed line),
and based on Monte–Carlo simulations (diamonds).
(a)
(b)
Fig. 5: Competing pathogens dynamics. (a): Fractions of nodes in I, J, S: origi-
nal AME (solid line); lumped AME (dashed line); Monte-Carlo simulations (di-
amonds). (b): Comparison of pair approximation with Monte-Carlo simulation
(diamonds).
We assume (λ1, λ2, λ3) = (6.0, 0.5, 0.5) with kmax = 60 and γ = 3.0. The initial
distribution is set to (xI(0), xR(0), xS(0)) = ( 3
5 ). Again, we find that Monte–
Carlo simulations, original AME, and lumped AME are in excellent agreement
(Fig. 4b). The error curve, however, converges slower to zero than in the SIR
model but it gets fast enough close to it (Fig. 4a). The lumped solution cor-
responds to 1032 clusters with a lumping error of 0.0059 and a runtime of 35
seconds compared to 39711 clusters of the original AME solution the runtime of
which was 1606 seconds.
5 , 1
5 , 1
5.3 Competing Pathogens
We, finally, examine an epidemic model with two competing pathogens [22]. The
pathogens are denoted by I and J and the susceptible state by S:
S
λ1·m[I]
−
−−−−→ I
S
λ2·m[J]
−
−−−−→ J
I
λ3−−→ S
J
λ4−−→ S .
1500
1000
Runtime (s)
00
05
39711
20000
Number of Clusters
0.03
0.02
0.01
Lumping Error
0.00
0
IRS
0.5
Time
1.0
1.0
0.5
Fractions
0.0
0.0
IJS
1.0
0.5
Fractions
0.0
0
1
2
3
5
6
7
8
4
Time
IJS
1.0
0.5
Fractions
0.0
0
1
2
3
5
6
7
8
4
Time
We assume that both pathogens have the same infection rate and differ only in
their respective recovery rates. Specifically, we set
(λ1, λ2, λ3, λ4) = (5.0, 5.0, 1.5, 1.0) and assume network parameters of kmax = 55
and γ = 2.5. The initial distribution is (xI(0), xJ(0), xS(0)) = (0.2, 0.1, 0.7).
This model is the most challenging case study for our approach. AME solution
and naturally lumped AME are not in perfect alignment with Monte Carlo
simulations (Fig. 5a) and our lumping approach needs a, comparably to the
previous cases, larger number of clusters to get a reasonably good approximation
of the AME (Fig. 6a). The computational gain is, however, large as well. The
lumped solution that comes with an approximation error of 0.02 corresponds to
2135 clusters and a runtime of 961 seconds compared to 30856 clusters and 17974
seconds of the original AME solution. Pair approximation approach (Fig. 5b)
even tough faster (40 seconds) than the lumped AME would have here resulted
to a much larger approximation error than our method (cf. Fig. 5b).
At last, the slow convergence of the error curve makes the competing pathogen
model a good test case the for our stopping heuristic. The heuristic evaluates
the model for three different clusterings (509, 986, 2135 clusters). It stops as
the difference between the two last clusterings is smaller than stop, showing
its effectiveness also for challenging models. In Fig. 6b we show the alignment
between the true lumping error and the surrogate error used by the heuristic.
(a)
(b)
Fig. 6: Competing pathogens lumping. (a): Lumping error and runtime of the
ODE solver w.r.t. the number of clusters. (b): Lumping error compared to the
error used by the heuristic.
6 Conclusions and Future Work
In this paper, we present a novel model-reduction technique to overcome the
large computational burden of the multistate AME and make it tractable for
real world problems. We show that it is possible to describe complex global
behavior of dynamical processes using only an extremely small fraction of the
Runtime (s)
10000
20000
15000
Number of Clusters
0
30856
0.2
0.1
Lumping Error
0.0
0
Heuristic Error
0.04
0.02
15000
Number of Clusters
0.00
30856
0.2
0.1
Lumping Error
0.0
0
original equations. Our approach exploits the high similarity among the original
equations as well as the comparably small impact of equations belonging to the
tail of the power-law degree distribution. In addition, we propose an approach
for finding a reasonable trade-off between accuracy and runtime of our method.
Our approach is particularly useful in situations where several evaluations of
the AME are necessary such as for the estimation of parameters or for model
selection.
For future work, we plan to develop a method for on-the-fly clustering, which
joins equations and breaks them apart during integration. This would allow the
clustering to take into account the concrete (local) dynamics and to analyze
adaptive networks with a variable degree distribution.
7 Acknowledgments
This research was been partially funded by the German Research Council (DFG)
as part of the Collaborative Research Center "Methods and Tools for Under-
standing and Controlling Privacy". We thank James P. Gleeson for his comments
regarding the performance of AME on specific models and Michael Backenkohler
for his comments on the manuscript.
References
1. Albert-L´aszl´o Barab´asi. Network science. Cambridge university press, 2016.
2. Albert-L´aszl´o Barab´asi and R´eka Albert. Emergence of scaling in random networks.
science, 286(5439):509–512, 1999.
3. Alain Barrat, Marc Barthelemy, and Alessandro Vespignani. Dynamical processes
on complex networks. Cambridge university press, 2008.
4. Luca Bortolussi, Jane Hillston, Diego Latella, and Mieke Massink. Continuous
approximation of collective system behaviour: A tutorial. Performance Evaluation,
70(5):317–349, 2013.
5. Fred Brauer. Mathematical epidemiology: Past, present, and future. Infectious
Disease Modelling, 2(2):113–127, 2017.
6. Peter Buchholz. Exact and ordinary lumpability in finite markov chains. Journal
of applied probability, 31(1):59–75, 1994.
7. Luca Cardelli, Mirco Tribastone, Max Tschaikowski, and Andrea Vandin. Erode:
a tool for the evaluation and reduction of ordinary differential equations. In Inter-
national Conference on Tools and Algorithms for the Construction and Analysis of
Systems, pages 310–328. Springer, 2017.
8. Luca Cardelli, Mirco Tribastone, Max Tschaikowski, and Andrea Vandin. Syntactic
In Models, Algorithms,
markovian bisimulation for chemical reaction networks.
Logics and Tools, 2017.
9. Claudio Castellano and Romualdo Pastor-Satorras. Thresholds for epidemic
spreading in networks. Physical review letters, 105(21):218701, 2010.
10. Eric Cator and Piet Van Mieghem. Second-order mean-field susceptible-infected-
susceptible epidemic threshold. Physical review E, 85(5):056111, 2012.
11. Wesley Cota and Silvio C Ferreira. Optimized gillespie algorithms for the sim-
ulation of markovian epidemic processes on large and heterogeneous networks.
Computer Physics Communications, 219:303–312, 2017.
12. G Demirel, F Vazquez, GA Bohme, and T Gross. Moment-closure approximations
for discrete adaptive networks. Physica D: Nonlinear Phenomena, 267:68–80, 2014.
13. Nick Fedewa, Emily Krause, and Alexandra Sisson. Spread of a rumor. Society for
Industrial and Applied Mathematics. Central Michigan University, 25, 2013.
14. P. G. Fennell. Stochastic processes on complex networks: techniques and explo-
rations. PhD thesis, University of Limerick, 2015.
15. Bailey K Fosdick, Daniel B Larremore, Joel Nishimura, and Johan Ugander.
Configuring random graph models with fixed degree sequences. arXiv preprint
arXiv:1608.00607, 2016.
16. James P Gleeson. High-accuracy approximation of binary-state dynamics on net-
works. Physical Review Letters, 107(6):068701, 2011.
17. James P Gleeson. Binary-state dynamics on complex networks: Pair approximation
and beyond. Physical Review X, 3(2):021004, 2013.
18. James P Gleeson, Sergey Melnik, Jonathan A Ward, Mason A Porter, and Pe-
ter J Mucha. Accuracy of mean-field theory for dynamics on real-world networks.
Physical Review E, 85(2):026106, 2012.
19. Istvan Z Kiss, Joel C Miller, and P´eter L Simon. Mathematics of epidemics on
networks: from exact to approximate models. Forthcoming in Springer TAM series,
2016.
20. Charalampos Kyriakopoulos, Gerrit Grossmann, Verena Wolf, and Luca Borto-
lussi. Lumping of degree-based mean-field and pair-approximation equations for
multistate contact processes. Physical Review E, 97(1):012301, 2018.
21. Genyuan Li and Herschel Rabitz. A general analysis of approximate lumping in
chemical kinetics. Chemical engineering science, 45(4):977–1002, 1990.
22. Naoki Masuda and Norio Konno. Multi-state epidemic processes on complex net-
works. Journal of Theoretical Biology, 243(1):64–75, 2006.
23. Mark EJ Newman. The structure and function of complex networks. SIAM review,
45(2):167–256, 2003.
24. Romualdo Pastor-Satorras, Claudio Castellano, Piet Van Mieghem, and Alessandro
Vespignani. Epidemic processes in complex networks. Reviews of modern physics,
87(3):925, 2015.
25. Romualdo Pastor-Satorras and Alessandro Vespignani. Epidemic spreading in
scale-free networks. Physical review letters, 86(14):3200, 2001.
26. Mason Porter and James Gleeson. Dynamical systems on networks: A tutorial,
volume 4. Springer, 2016.
27. P´eter L Simon, Michael Taylor, and Istvan Z Kiss. Exact epidemic models on
graphs using graph-automorphism driven lumping. Journal of mathematical biol-
ogy, 62(4):479–508, 2011.
28. James Wei and James CW Kuo. Lumping analysis in monomolecular reaction sys-
tems. analysis of the exactly lumpable system. Industrial & Engineering chemistry
fundamentals, 8(1):114–123, 1969.
A Simplification of Equation Generation
compute a scalar of the formPm∈C wC,kmf (m) . In this section we introduce an
We constructed the lumped equations such that they can be evaluated efficiently
in each step of the ODE solver. However, for very large kmax, the size of M
becomes enormous. This makes the generation of the equations costly because
we iterate multiple times over M. This iteration is necessary, each time we
approximative scheme to generate lumped equations without the computational
burden of looking at each individual m ∈ M.
Our main idea is to only consider the center of each cluster w.r.t. wC,km, and
not at each cluster element. We use hmiC to denote the center of cluster C, each
entry being defined as:
hmiC[s] = Xm∈C
wC,kmm[s] .
(18)
We can efficiently compute hmiC by only considering the direction of a cluster
(which only depends on the associated proportionality cluster) and the mean
degree of a cluster (which can be computed by only considering the degree dis-
tribution).
Next, we approximate the average cluster rate by only evaluating the rate
function of each rule at the cluster mean:
X
m∈C
wC,kmf (m) ≈ f(cid:0)hmiC(cid:1) .
(19)
(20)
(21)
This, of course, only makes sense if the rate function is reasonably smooth (which
is the case in our models).
Likewise, inside the βss1→ss2, we approximate:
X
m∈C
f (m)wC,km m[s] ≈ f(cid:0)hmiC(cid:1) · wC,km · hmiC[s] .
Finally, we approximate the in- and outflow related to the βs. Note that, by
design, our clustering has the property that for given s1, s2 ∈ S and C ∈ C there
is only exactly one neighboring cluster in which probability mass can flow by
adding (resp. subtracting) state s1 (resp. s2) from the neighborhood vector. We
now assume a fixed s1, s2 and use C0 ∈ C to denote this cluster. We define
CN B = {m m{s+
CB = {m m{s+
1 ,s−
1 ,s−
2} ∈ C}
2} /∈ C} = {m m{s+
1 ,s−
2} ∈ C0} .
We see that all m ∈ CN B occur in the third term (inflow) and in the fourth
term (outflow) in the lumped AME. Hence, they cancel out and we can ignore
them. To determine the flow of probability mass between two clusters only CB
is of interest. Since the flow is symmetrical (i.e., the inflow of C0 is the outflow
of C) it is sufficient to approximate one direction. We use
βss1→ss2
L
· zs,C0 · hmiCB [s1] · C B
C
(22)
hmiC0[s1] .
(23)
1 2
to approximate the flow from C to C0. Hence, we add (resp. substract) this
value in the equation corresponding to C0 (resp. C). Note that hmiCB denotes
an approximation of the mean value of C B. As these points lie at the border to
C0, we use:
1 2
hmiCB [s1] =
hmiC[s1] +
CB
is a scaling factor which corresponds to the size of the border between
C
the clusters (the larger the border area, the more probability flows between the
clusters). Note that also the cardinality of C and the cardinality of CB can
be efficiently approximated by combinatoric reasoning without looking at the
individual elements.
First, consider CB. If we fix a k, each cluster has approximately the same
number of elements (cf. Fig. 2b). We get this number by dividing Mk (the
number of neighbor vectors for that degree) by the number of proportionality
clusters. To determine C, we simply aggregate this value over all degrees which
occur inside C. Next, consider CB. It denotes the number of points inside C
but next to one particular neighboring cluster. Luckily, our clustering has a nice
geometrical structure (namely a triangular one), which we exploit here. The size
of a face (surface area in one direction) of each cluster C ∈ C in n dimensions
for a fixed degree k is exactly the number of elements in a cluster for n − 1
dimensions for that k. For example, the face of a tetrahedron is a triangle, and
the size of the triangle can be determined by clustering three dimensions instead
of four.
We present two examples of approximative equation generation in Fig. 7.
First, we compare these with equations which are generated using the old ap-
proach (Fig. 7a). We find that their respective dynamics does not differ signifi-
cantly.
In addition, we test the approximative equation on a model for which the
traditional generation would introduce a significant overhead and where testing
the original AME is practically impossible (Fig. 7b). We choose a SIR model
(where nodes are trapped in state R) with kmax = 500, γ = 2.5, an infection rate
of 3.0 · m[I], and a recovery rate of 0.3. Again, we see that the lumped AME is
in excellent agreement with the numerical simulations.
(a)
(b)
Fig. 7: Dynamics of approximative equations. (a): The same SIR model as in
Fig. 3: Lumping of the AME (solid line) and approximation of the lumped equa-
tion (dashed line) corresponding to 10 proportionality clusters and 20 degree
clusters. (b): New SIR model with 50 degree clusters and 15 proportionality
clusters. Approximative equations (solid line) are compared with Monte-Carlo
simulations (diamonds). The lumped AME has 8583 clusters. In contrast to more
than 21 million clusters of the original model.
IRS
1.0
0.5
Fractions
0.0
0
1
2
3
Time
4
5
IRS
1.0
0.5
Fractions
0.0
0
1
2
3
Time
4
5
|
1705.02505 | 2 | 1705 | 2017-05-24T22:10:35 | HoloScope: Topology-and-Spike Aware Fraud Detection | [
"cs.SI"
] | As online fraudsters invest more resources, including purchasing large pools of fake user accounts and dedicated IPs, fraudulent attacks become less obvious and their detection becomes increasingly challenging. Existing approaches such as average degree maximization suffer from the bias of including more nodes than necessary, resulting in lower accuracy and increased need for manual verification. Hence, we propose HoloScope, which uses information from graph topology and temporal spikes to more accurately detect groups of fraudulent users. In terms of graph topology, we introduce "contrast suspiciousness," a dynamic weighting approach, which allows us to more accurately detect fraudulent blocks, particularly low-density blocks. In terms of temporal spikes, HoloScope takes into account the sudden bursts and drops of fraudsters' attacking patterns. In addition, we provide theoretical bounds for how much this increases the time cost needed for fraudsters to conduct adversarial attacks. Additionally, from the perspective of ratings, HoloScope incorporates the deviation of rating scores in order to catch fraudsters more accurately. Moreover, HoloScope has a concise framework and sub-quadratic time complexity, making the algorithm reproducible and scalable. Extensive experiments showed that HoloScope achieved significant accuracy improvements on synthetic and real data, compared with state-of-the-art fraud detection methods. | cs.SI | cs | HoloScope: Topology-and-Spike Aware Fraud Detection
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2∗
1CAS Key Laboratory of Network Data Science & Technology,
Institute of Computing Technology, Chinese Academy of Sciences
2Carnegie Mellon University
[email protected],{bhooi,christos}@cs.cmu.edu
7
1
0
2
y
a
M
4
2
]
I
S
.
s
c
[
2
v
5
0
5
2
0
.
5
0
7
1
:
v
i
X
r
a
ABSTRACT
As online fraudsters invest more resources, including purchasing
large pools of fake user accounts and dedicated IPs, fraudulent at-
tacks become less obvious and their detection becomes increasingly
challenging. Existing approaches such as average degree maximiza-
tion su(cid:130)er from the bias of including more nodes than necessary,
resulting in lower accuracy and increased need for manual veri-
(cid:128)cation. Hence, we propose HoloScope, which uses information
from graph topology and temporal spikes to more accurately de-
tect groups of fraudulent users. In terms of graph topology, we
introduce "contrast suspiciousness," a dynamic weighting approach,
which allows us to more accurately detect fraudulent blocks, partic-
ularly low-density blocks. In terms of temporal spikes, HoloScope
takes into account the sudden bursts and drops of fraudsters' at-
tacking pa(cid:138)erns. In addition, we provide theoretical bounds for
how much this increases the time cost needed for fraudsters to
conduct adversarial a(cid:138)acks. Additionally, from the perspective of
ratings, HoloScope incorporates the deviation of rating scores in
order to catch fraudsters more accurately. Moreover, HoloScope has
a concise framework and sub-quadratic time complexity, making
the algorithm reproducible and scalable. Extensive experiments
showed that HoloScope achieved signi(cid:128)cant accuracy improve-
ments on synthetic and real data, compared with state-of-the-art
fraud detection methods.
KEYWORDS
Graph Mining, Fraud Detection, Burst and Drop
ACM Reference format:
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2. 2016. HoloScope:
Topology-and-Spike Aware Fraud Detection. In Proceedings of ACM Confer-
ence, Washington, DC, USA, July 2017 (Conference'17), 9 pages.
DOI: 10.1145/nnnnnnn.nnnnnnn
1 INTRODUCTION
Online fraud has become an increasingly serious problem due to
the high pro(cid:128)t it o(cid:130)ers to fraudsters, which can be as much as
$5 million from 300 million fake "views" per day, according to a
report of Methbot [20] on Dec 2016. Meanwhile, to avoid detection,
fraudsters can manipulate their geolocation, internet providers, and
IP address, via large IP pools (852,992 dedicated IPs). Suppose a
fraudster has a accounts or IPs, and wants to rate or click b objects
(e.g., products) at least 200 times, as required by their customers.
∗(cid:140)e work was done when Shenghua Liu was a visiting researcher at Carnegie Mellon
University
Conference'17, Washington, DC, USA
2016. 978-x-xxxx-xxxx-x/YY/MM...$15.00
DOI: 10.1145/nnnnnnn.nnnnnnn
(cid:140)e edge density of the fraudulent block is then 200·b/(a·b) = 200/a.
We thus see that with enough user accounts or IPs, the fraudsters
can serve as many customers as they need while keeping density
low. (cid:140)is presents a di(cid:129)cult challenge for most existing fraud
detection methods.
Current dense block detection methods [4, 26, 27] maximize
the arithmetic or geometric average degree. We use "fraudulent
density" to indicate the edge density that fraudsters create for the
target objects. However, those methods have a bias of including
more nodes than necessary, especially as the fraudulent density
decreases, which we veri(cid:128)ed empirically. (cid:140)is bias results in low
precision, which then requires intensive manual work to verify each
user. Fraudar [11] proposed an edge weighting scheme based on
inverse logarithm of objects' degrees to reduce this bias, which was
inspired by IDF [23, 28]. However, their weighting scheme is (cid:128)xed
globally and a(cid:130)ects both suspicious and normal edges, lowering the
precision of Fraudar, which can be seen from results on semi-real
(with injected labels) and real data (see Fig. 1).
Accurately detecting fraudulent blocks of lower density requires
aggregating more sources of information. Consider the a(cid:138)ribute
of the creation time of edges: fraudulent a(cid:138)acks tend to be con-
centrated in time, e.g., fraudsters may surge to retweet a message,
creating one or more sudden bursts of activity, followed by sudden
drops a(cid:137)er the a(cid:138)ack is complete. Sudden bursts and drops have
not been directly considered together in previous work.
Tensor based methods [12, 26, 27] incorporate edge a(cid:138)ributes
into a multi-way tensor formulation, e.g., IPs, rating scores and time.
However, those methods rely on time-binning to incorporate tempo-
ral information, and then treat time bins independently, which loses
information about bursts and drops, which our approach captures.
(cid:140)erefore, we propose HoloScope, which combines suspicious
signals from graph topology, temporal bursts and drops, and rating
deviation. Our graph topology-based weighting scheme dynami-
cally reweights objects according to our beliefs about which users
are suspicious. Temporally, HoloScope detects suspicious spikes of
bursts and drops, which increases the time cost needed for fraud-
sters to conduct an a(cid:138)ack. In terms of rating deviation, our approach
takes into account how much di(cid:130)erence there is between an object's
ratings given by suspicious users and non-suspicious users.
In summary, our contributions are:
• Uni(cid:128)cation of signals: we make holistic use of several
signals, namely connectivity (i.e., topology), temporal bursts
and drops, and rating deviation in a systematic way.
• (cid:135)eoretical analysis of fraudsters' obstruction: we
show that if the fraudsters use less than an upper bound
of time for an a(cid:138)ack, they will cause a suspicious drop or
burst. In other words, HoloScope obstructs fraudsters by
increasing the time they need to perform an a(cid:138)ack.
Conference'17, July 2017, Washington, DC, USA
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2
(a) HS-α using topology information outper-
forms competitors
(b) HS using holistic attributes provides clear
improvement, and performs the best.
(c) HS achieves the best F-measure, on real
data from Sina Weibo
Figure 1: (a) and (b) show experimental results on BeerAdvocate dataset. HS-α and HS are both our methods, where the former
only uses topology information. We increase the # of injected fraudsters from 200 to 2000 for HS-α, and to 20000 for HS. (cid:135)e
corresponding density is shown on the horizontal axis. (c) shows accuracy results on Sina Weibo, with ground truth labels.
Table 1: Comparison between HoloScope and other fraud
detection algorithms.
(a) BeerAdvocate dataset
(b) Amazon Electronics dataset
Figure 2: HoloScope (HS) runs in near-linear time.
• E(cid:130)ectiveness: we achieved higher accuracy than the com-
petitors on semi-real and real datasets. In fact, HoloScope
using only topology information (HS-α) outperformed the
graph-based competitors (see Fig. 1a), while HoloScope
(HS) using all a(cid:138)ributes achieved further improvement,
and outperformed the tensor-based competitors (see Fig. 1b
and 1c).
• Scalability: HoloScope runs in subquadratic time in the
number of nodes, under reasonable assumptions. Fig. 2
shows that its running time increases near-linearly with
the number of edges.
In addition, in Microblog, Sina Weibo1 data, HoloScope achieved
higher F-measure than the competitors in detecting the ground
truth labels, with high precision and recall. (cid:140)e code of HoloScope
is open-sourced for reproducibility 2.
2 RELATED WORKS
Most existing works study anomaly detection based on the density
of blocks within adjacency matrices [13, 22], or multi-way ten-
sors [26, 27]. [8] and CoreScope [25] proposed to use Shingling and
K-core algorithms respectively to detect anomalous dense block in
1 (cid:140)e largest Microblog service in China, h(cid:138)p://www.weibo.com
2h(cid:138)ps://github.com/shenghua-liu/HoloScope
]
1
2
,
1
[
s
d
o
h
t
e
m
d
e
s
a
b
-
P
B
]
7
2
,
6
2
[
/
e
b
u
C
-
D
m
o
o
Z
-
M
e
p
o
c
S
o
l
o
H
]
1
1
[
r
a
d
u
a
r
F
]
2
2
[
n
E
k
o
p
S
]
3
[
h
c
t
a
C
y
p
o
C
]
2
1
[
t
o
p
S
s
s
o
r
C
scalability
camou(cid:131)age
hy-community
spike-aware
(cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88)
(cid:88) (cid:88)
(cid:88)
(cid:88)
(cid:88)
? (cid:88)
?
?
?
huge graphs. Taking into account the suspiciousness of each edge or
node in a real life graph potentially allows for more accurate detec-
tion. Fraudar [11] proposed to weight edges' suspiciousness by the
inverse logarithm of objects' indegrees, to discount popular objects.
[2] found that the degrees in a large community follow a power
law distribution, forming hyperbolic structures. (cid:140)is suggests pe-
nalizing high degree objects to avoid unnecessarily detecting the
dense core of hyperbolic community.
In addition to topological density, EdgeCentric [24] studied the
distribution of rating scores to (cid:128)nd the anomaly. In terms of tem-
poral a(cid:138)ribute, the identi(cid:128)cation of burst period has been studied
in [15]. A recent work, Sleep Beauty (SB) [14] more intuitively
de(cid:128)ned the awakening time for a paper's citation burst for burst
period. [31] clustered the temporal pa(cid:138)erns of text phrases and
hash tags in Twi(cid:138)er. Meanwhile, [9, 10] modeled the time stamped
rating scores with Bayesian model and autoregression model respec-
tively for anomalous behavior detection. Even though [6, 16, 30]
have used bursty pa(cid:138)erns to detect review spam, a sudden drop in
temporal spikes has not been considered yet.
However, aggregating suspiciousness signals from di(cid:130)erent at-
tributes is challenging. [5] proposed RRF (Reciprocal Rank Fusion)
scores for combining di(cid:130)erent rank lists in information retrieval.
However, RRF applies to ranks, not suspiciousness scores.
0.20.40.60.81density00.10.20.30.40.50.60.70.80.91F measureHS-α F of AFraudar F of AFraudar F of BSpokEn F of ASpokEn F of B00.050.10.20.30.40.50.6density00.20.40.60.81F measureHS F of AM-Zoom F of AM-Zoom F of BD-Cube F of AD-Cube F of BCrossSpot F of ACrossSpot F of B27.4%61.4%large result set withlow precision051015# of edges×10505001000Algorithm running Time (s)02468# of edges×1060200040006000800010000Algorithm running Time (s) HoloScope: Topology-and-Spike Aware Fraud Detection
Conference'17, July 2017, Washington, DC, USA
CrossSpot [12], a tensor-based algorithm, estimated the suspi-
ciousness of a block using a Poisson model. However, it did not take
into account the di(cid:130)erence between popular and unpopular objects.
Moreover, although CrossSpot, M-Zoom [26] and D-Cube [27] can
consider edge a(cid:138)ributes like rating time and scores via a multi-way
tensor approach, they require a time-binning approach. When time
is split into bins, a(cid:138)acks which create bursts and drops may not
stand out clearly a(cid:137)er time-binning. (cid:140)e problem of choosing bin
widths for histograms was studied by Sturges [29] assuming an ap-
proximately normal distribution, and Freedman-Diaconis [7] based
on statistical dispersion. However, the binning approaches were
proposed for the time series of a single object, which is not clear
for di(cid:130)erent kinds of objects in a real life graph, namely, popular
products and unpopular products should use di(cid:130)erent bin sizes.
Belief propagation [21] is another common approach for fraud
detection which can incorporate some speci(cid:128)c edge a(cid:138)ributes, like
sentiment [1]. However, its robustness against adversaries which
try to hide themselves is not well understood.
CopyCatch [3] detected lockstep behavior by maximizing the
number of edges in blocks constrained within time windows. How-
ever, this approach ignores the distribution of edge creation times
within the window, and does not capture bursts and drops directly.
Finally, we summarize our competitors compared to our Holo-
Scope in Table 1. We use "hy-community" to indicate whether
the method can avoid detecting the naturally-formed hyperbolic
topology which is unnecessary for fraud detection. HoloScope is
the only one which considers temporal spikes (sudden bursts and
drops and multiple bursts) and hyperbolic topology, in a uni(cid:128)ed
suspiciousness framework.
3 PROPOSED APPROACH
(cid:140)e de(cid:128)nition of our problem is as follows.
Problem 1 (informal definition). Given quadruplets (user,
object, timstamp, #stars), where timestamp is the time that a user
rates an object, and #stars is the categorical rating scores.
- Find a group of suspicious users, and suspicious objects or
its rank list with suspiciousness scores,
- to optimize the metric under the common knowledge of
suspiciousness from topology, rating time and scores.
To make the problem more general, timestamps and #stars are
optional. For example, in Twi(cid:138)er, we have (user , object, timestamp)
triples, where user retweets a message object at timestamp. In a
static following network, we have pairs (user , object), with user
following object.
As discussed in previous sections, our metric should capture the
following basic traits:
Trait 1 (Engagement). Fraudsters engage as much (cid:128)repower as
possible to boost customers' objects, i.e., suspicious objects.
Trait 2 (Less Involvement). Suspicious objects seldom a(cid:136)ract
non-fraudulent users to connect with them.
Trait 3 (Spikes: Bursts and Drops). Fraudulent a(cid:136)acks are
concentrated in time, sometimes over multiple waves of a(cid:136)acks, cre-
ating bursts of activity. Conversely, the end of an a(cid:136)ack corresponds
to sudden drops in activity.
Trait 4 (Rating Deviation). High-rating objects seldom a(cid:136)ract
extremely low ratings from normal users. Conversely, low-rating
objects seldom a(cid:136)ract extremely high ratings from normal users.
(cid:140)us we will show in the following sections, that our proposed
metric can make holistic use of several signals, namely topology,
temporal spikes, and rating deviation, to locate suspicious users
and objects satisfying the above traits. (cid:140)at is the reason we name
our method as HoloScope (HS).
3.1 HoloScope metric
To give a formal de(cid:128)nition of our metric, we describe the quadru-
plets (user, object, timstamp, #stars) as a bipartite and directed
graph G = {U , V , E}, which U is the source node set, V is the
sink node set, and connections E is the directed edges from U to
V . Generally, graph G is a multigraph, i.e., multiple edges can be
present between two nodes. Multiple edges mean that a user can
repeatedly comment or rate on the same product at a di(cid:130)erent time,
as common in practice. Users can also retweet message multiple
times in the Microblog Sina W eibo. Each edge can be associated
with rating scores (#stars), and timestamps, for which the data
structure is introduced in Subsection 3.1.2.
Our HoloScope metric detects fraud from three perspectives:
topology connection, timestamps, and rating score. To easily un-
derstand the framework, we (cid:128)rst introduce the HoloScope in a
perspective of topology connection. A(cid:137)erwards, we show how we
aggregate the other two perspectives into the HoloScope. We (cid:128)rst
view G as a weighted adjacency matrix M, with the number of
multiple edges (i.e., edge frequency) as matrix elements.
Our goal is to (cid:128)nd lockstep behavior of a group of suspicious
source nodes A ⊂ U who act on a group of sink nodes B ⊂ V . Based
on Trait 1, the total engagement of source nodes A to sink nodes B
can be basically measured via density measures. (cid:140)ere are many
density measures, such as arithmetic and geometric average degree.
Our HoloScope metric allows for any such measure. However, as
the average degree metrics have a bias toward including too many
nodes, we use a measure denoted by D(A, B) as the basis of the
HoloScope, de(cid:128)ned as:
D(A, B) =
vi ∈B fA(vi)
A + B
(1)
where fA(vi) is the total edge frequency from source nodes A to a
sink node vi. fA(vi) can also be viewed as an engagement from A
to vi, or A's lockstep on vi, which is de(cid:128)ned as
σji · eji
fA(vi) =
(2)
(uj,vi)∈E∧uj ∈A
where constant σ ji are the global suspiciousness on an edge, and
eji is the element of adjacency matrix M, i.e., the edge frequency
between a node pair (uj , vi). (cid:140)e edge frequency eji becomes a
binary in a simple graph. (cid:140)e global suspiciousness as a prior can
come from the degree, and the extra knowledge on fraudsters, such
as duplicated review sentences and unusual behaving time.
To maximize D(A, B), the suspicious source nodes A and the
suspicious sink nodes B are mutually dependent. (cid:140)erefore, we
introduce contrast suspiciousness in an informal de(cid:128)nition:
Conference'17, July 2017, Washington, DC, USA
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2
Figure 3: An intuitive view of our de(cid:128)nitions in the Holo-
Scope.
De(cid:128)nition 3.1 (contrast suspiciousness). (cid:140)e contrast suspicious-
ness denoted as P(vi ∈ BA) is de(cid:128)ned as the conditional likelihood
of a sink node vi that belongs to B (the suspicious object set), given
the suspicious source nodes A.
A visualization of the contrast suspiciousness is given in Fig. 3.
(cid:140)e intuitive idea behind contrast suspiciousness is that in the most
case, we need to judge the suspiciousness of objects by currently
chosen suspicious users A, e.g., an object is more suspicious if very
few users not in A are connected to it (see Trait 2); the sudden burst
of an object is mainly caused by A (see Trait 3); or the rating scores
from A to an object are quite di(cid:130)erent from other users (see Trait 4).
(cid:140)erefore, such suspiciousness makes use of the contrasts between
users in A and users not in A or the whole set.
Finally, instead of maximizing D(A, B), we maximize the follow-
ing expectation of suspiciousness D(A, B) over the probabilities
P(vi ∈ BA):
max
A
HS(A) := E[D(A, B)]
1
fA(vi)P(viA)
(3)
A +
k ∈V
=
i∈V
P(vk A)
where for simplicity we write P(viA) to mean P(vi ∈ BA). 1 −
P(viA) is the probability of vi being a normal sink node. We
dynamically calculate the contrast suspiciousness for all the objects,
a(cid:137)er every choice of source nodes A.
Using this overall framework for our proposed metric HS(A), we
next show how to satisfy the remaining Traits. To do this, we de(cid:128)ne
contrast suspiciousness P(viA) in a way that takes into account
various edge a(cid:138)ributes. (cid:140)is will allow greater accuracy particularly
for detecting low-density blocks.
3.1.1 HS-α: Less involvement from others. Based on Trait 2, a
sink node should be more suspicious if it only a(cid:138)racts connections
from the suspicious source nodes A, and less from other nodes.
Mathematically, we capture this by de(cid:128)ning
P(viA) ∝ q(αi), where αi =
(4)
where fU (vi) is the weighted indegree of sink node vi. Similar
to fA(vi), the edges are weighted by global suspiciousness. αi
measures the involvement ratio of A in the activity of sink node vi.
(cid:140)e scaling function q(·) is our belief about how this ratio relates to
suspiciousness, and we choose the exponential form q(x) = bx−1,
where base b > 1.
fA(vi)
fU (vi)
(a) HS-α (cid:128)nds the exact dense
block in the synthetic data
(b) Hyperbolic community in
BeerAdvocate data
Figure 4: (a) (cid:135)e synthetic data consists of hyperbolic and
rectangular blocks, with volume density around 0.84 and
0.60 respectively. (cid:135)e camou(cid:131)age is randomly biased to
columns of high degree. (b) A real data of naturally-formed
hyperbolic community, and injected dense block. (cid:135)e injec-
tion is 2000 × 200 with biased camou(cid:131)age.
As previous work showed, large communities form hyperbolic
structures, which is generated in our synthetic data (see the lower-
right block in Fig. 4a), and also exists in real BeerAdvocate data
(see Fig. 4b). For clarity, our HoloScope method are denoted as
HS-α when it is only applied on a connection graph. (cid:140)e results
of the synthetic data show that HS-α detected the exact dense
rectangular block (b = 128), while the other competitors included
a lot of non-suspicious nodes from the core part of hyperbolic
community resulting in low accuracy. In the beer review data from
the BeerAdvocate website, testing on di(cid:130)erent fraudulent density
(see Fig. 1a), our HS-α remained at high accuracy, while the other
methods' accuracy drops quickly when the density drops below
70%.
(cid:140)e main idea is that HS-α can do be(cid:138)er because it dynamically
adjusts the weights for sink nodes, penalizing those sink nodes
that also have many connections from other source nodes not in
A.
In contrast, although Fraudar proposed to penalize popular
sink nodes based their indegrees, these penalties also scaled down
the weights of suspicious edges. (cid:140)e Fraudar (green box) only
improved the unweighted "average degree" method (red box) by a
very limited amount. Moreover, with a heavier penalty, the "sqrt
weight" method (blue box) achieved be(cid:138)er accuracy on source
nodes but worse accuracy on sink nodes, since those methods used
globally (cid:128)xed weights, and the weights of suspicious were penalized
as well. Hence the hyperbolic structure pushes those methods to
include more nodes from its core part.
In summary, our HS-α using dynamic contrast suspiciousness
can improve the accuracy of fraud detection in 'noisy' graphs (con-
taining hyperbolic communities), even with low fraudulent density.
3.1.2 Temporal bursts and drops. Timestamps for edge creation
are commonly available in most real se(cid:138)ings. If two subgroups of
Microblog users have the same number of retweets to a message,
can we say they have the same suspiciousness? As an example
shown in Fig. 5a, the red line is the time series (histogram of time
bins) of the total retweets of a message in Microblog, Sina Weibo.
UVvi()Aifv()iPvAA050010001500200025003000050010001500200025003000avgdegree1172x1331Fraudar1120x1472sqrtweight1028x2057HS-α500x500injected fraud & camouflage hyperbolic block(naturally formed ) HoloScope: Topology-and-Spike Aware Fraud Detection
Conference'17, July 2017, Washington, DC, USA
(a) Group A1 is more suspicious than A2, due to
the sudden burst and drop
(b) Detection of temporal bursts and drops
(c) Proof the time cost obstruction
Figure 5: (a) and (b) are the time series (histogram) of a real message being retweeted in Microblog, Sina Weibo. (cid:135)e horizontal
axis is the seconds a(cid:133)er 2013 − 11 − 1. (c) illustrates our proof of time cost obstruction.
(cid:140)e blue do(cid:138)ed line and green dashed line are the retweeting time
series respectively from user groups A1 and A2. (cid:140)e two series have
the same area under the time series curves, i.e., the same number
of retweets. However, considering that fraudsters tend to surge to
retweet a message to reduce the time cost, the surge should create
one or more sudden bursts, along with sudden drops. (cid:140)erefore,
the suspiciousness of user groups A1 and A2 become quite di(cid:130)erent
even though they have the same number of retweets, which cannot
be detected solely based on connections in the graph. (cid:140)us we
include the temporal a(cid:138)ribute into our HoloScope framework for
de(cid:128)ning contrast suspiciousness.
Denote the list of timestamps of edges connected to a sink node
v as Tv. To simplify notation, we use T without subscript when
talking about a single given sink node v. Let T ={(t0, c0), (t1, c1), · · · ,
(te, ce)} as the time series of T , i.e., the histogram of T . (cid:140)e count ci
is the number of timestamps in the time bin [ti − ∆t/2, ti + ∆t/2),
with bin size ∆t. (cid:140)e bin size of histogram is calculated according to
the maximum of Sturges criteria and the robust Freedman-Diaconis'
criteria as mentioned in related works. It is worth noticing that the
HoloScope can tune di(cid:130)erent bin sizes for di(cid:130)erent sink nodes, e.g.,
popular objects need (cid:128)ne-grained bins to explore detailed pa(cid:138)erns.
Hence, the HoloScope is more (cid:131)exible than tensor based methods,
which use a globally (cid:128)xed bin size. Moreover, the HoloScope can
update the time series at a low cost when T is increasing.
To consider the burst and drop pa(cid:138)erns described in Trait 3,
we need to decide the start point of a burst and the end point of
a drop in time series T . Let the burst point be (tm, cm), having
the maximum value cm. According to the de(cid:128)nition in previous
work "Sleeping Beauty", we use an auxiliary straight line from the
beginning to the burst point to decide the start point, named the
awakening point of the burst. Fig. 5b shows the time series T (red
polygonal line) of a message from Sina Weibo, the auxiliary straight
line l (black do(cid:138)ed line) from the lower le(cid:137) point (t0, c0) to upper
right point (tm, cm), and the awakening point for the maximum
point (tm, cm), which is de(cid:128)ned as the point along the time series T
which maximizes the distance to l. As the do(cid:138)ed line perpendicular
to l suggests in this (cid:128)gure, the awakening point (ta, ca) satis(cid:128)es
(5)
(cm − c0)t − (tm − t0)c + tmc0 − cmt0
(cid:113)(cm − c0)2 + (tm − t0)2
ta = arg max
(c,t)∈T,t <tm
Finding the awakening point for one burst is not enough, as
multiple bursts may be present. (cid:140)us, sub-burst points and the as-
sociated awakening points should be considered. We then propose
a recursive algorithm MultiBurst in Alg. 1 for such a purpose.
Algorithm 1 MultiBurst algorithm.
Input Time series T of sink node v, beginning index i, end index j
Output A list of awakening-burst point pairs,
sam: slope of the line passing through each point pair,
∆c: altitude di(cid:130)erence of each point pair.
If j − i < 2 then return
(tm, cm) = point of maximum altitude between indices i and j.
(ta, ca) = the awakening point as Eq (5) between indices i and j.
∆cam = cm − ca, and sam = ∆cam/(tm−ta)
Append {(ta, ca), (tm, cm)}, sam, and ∆cam into the output.
MultiBurst(T, i, a − 1)
k = Find the (cid:128)rst local min position from indices m + 1 to j
MultiBurst(T, k, j)
∆cam ·sam
Φ(TU ), and Φ(T) =
A(cid:137)er (cid:128)nding awakening and burst points, the contrast suspi-
ciousness of burst awareness satis(cid:128)es P(viA) ∝ q(φi), where φi is
the involvement ratio of source nodes in A in multiple bursts. Let
the collection of timestamps from A to sink node vi be TA. (cid:140)en,
φi = Φ(TA)
1(t ∈ [ta, tm]) (6)
where sam is the slope from the output of MultiBurst algorithm.
Here sam is used as a weight based on how steep the current burst is.
(cid:140)is de(cid:128)nition of suspiciousness satis(cid:128)es Trait 3. It is worth noticing
that the MultiBurst algorithm only needs to be executed once.
With the preprocessed awakening and burst points, the contrast
suspiciousness of edges connected to v has O(dv) complexity, where
dv is the degree of sink node v. Hence the complexity for overall
sink nodes are O(E).
(ta,tm)
t ∈T
6789101112time (s)×105050100150200250300350400# of retweetsGroup 𝐴"Group 𝐴#involve more in sudden burstsudden dropawakening pointsburst pointsdying point𝑙(𝑡$,𝑐$)(𝑡(,𝑐()(𝑡),𝑐))(𝑡*,𝑐*)exceed 𝑆"exceed 𝑆#𝑆"𝑆#𝑐%𝑛"∆𝑡𝑛#∆𝑡 Conference'17, July 2017, Washington, DC, USA
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2
In fact, sudden drops are also a prominent pa(cid:138)ern of fraudulent
behavior as described in Trait 3, since a(cid:137)er creating the a(cid:138)ack is
complete, fraudsters usually stop their activity sharply. To make
use of the suspicious pa(cid:138)ern of a sudden drop, we de(cid:128)ne the dyin❕
point as the end of a drop. As Fig. 5b suggests, another auxiliary
straight line is drawn from the highest point (tm, cm) to the last
point (te, ce). (cid:140)e dying point (td, cd) can be found by maximizing
the distance to this straight line. (cid:140)us we can discover the "sudden
drop" by the absolute slope value sbd=(cm − cd)/(td − tm) between the
burst point and the dying point. Since there may be several drops in
a (cid:131)uctuated time series T , we choose the drop with the maximum
fall. To (cid:128)nd the maximum fall, we also need a recursive algorithm,
similar to Alg. 1:
1) Find a maximum point (tm, cm), and the corresponding dying
point (td , cd ) by de(cid:128)nition; 2) Calculate the current drop slope sbd ,
and the drop fall ∆cbd = cm − cd ; 3) Recursively (cid:128)nd drop slope and
drop fall for the le(cid:135) and right parts of T , i.e., t < tm and t ≥ td
respectively.
As a result, the algorithm returns the maximum drop fall ∆cbd,
and its drop slope sbd, which it has found recursively. Finally, we
use the weighted drop slope ∆cbd · sbd as a global suspiciousness
in equation (2), to measure the drop suspiciousness. Each edge con-
nected to the sink node v is assigned the same drop suspiciousness.
We use a logarithm scale for smoothing those edge weights.
With this approach to detect bursts and drops, we now show
that this provides a provable time obstruction for fraudsters.
(cid:113) 2N ∆t(S1+S1)
to create for an object.
Theorem 3.2. Let N be the number of edges that fraudsters want
If the fraudsters use time less than τ ≥
, then they will be tracked by a suspicious burst or
drop, where ∆t is the size of time bins, and S1 and S2 are the slopes of
normal rise and decline respectively.
S1·S1
Proof. (cid:140)e most e(cid:129)cient way to create N edges is to have one
burst and one drop, otherwise more time is needed. As shown
in Fig. 5c, in order to minimize the slope, every point in the time
series should in line with the two auxiliary straight lines to the
highest point cm, separately from the awakening and dying points.
Otherwise, the slopes will exceed the normal values S1 and S2.
Hence we only consider the triangle with the auxiliary lines as its
two edges. It is worth noticing that a trapezoid whose legs have
the same slopes as the triangle's edges cannot have a shorter time
cost. (cid:140)en
cm
n1∆t
(n1 + n2) · cm = 2N
Here n1 and n2 are the number of time bins before and a(cid:137)er the
burst. N (cid:48) is the total number of rating edges, and N (cid:48) ≥ N consider
some edges from normal users. (cid:140)us, solving the above equations,
we have
cm
n2∆t
= S2,
= S1,
(cid:48)
.
(cid:115)2N (cid:48)∆t(S1 + S2)
(cid:115)2N ∆t(S1 + S2)
≥
τ = (n1 + n2)∆t =
S1 · S2
S1 · S2
We also have the height of burst, cm ≥(cid:113) 2N ∆tS1S2
(cid:3)
. (cid:140)us, the
maximum height of time series T cannot be larger by far than that
of a normal sink node. (cid:140)at is the reason that we use the weighted
φi in equation (6) and weighted drop slope in equation (2).
S1+S2
3.1.3 Rating deviation and aggregation. We now consider edges
with categorical a(cid:138)ributes such as rating scores, text contents, etc.
For each sink node vi, we use the KL-divergence κi between the
distributions separately from the suspicious source nodes A and the
other nodes, i.e., U \ A. We use U \ A for KL-divergence instead of
the whole source nodes U , in order to avoid the trivial case where
most of the rating scores are from A. (cid:140)e rating deviation κi is
scaled into [0, 1] by the maximum value before being passed into
function q(·) to compute contrast suspiciousness. (cid:140)e neutral scores
can be ignored in the KL-divergence for the purpose of detecting
fraudulent boosting or defamation. Moreover, rating deviation is
meaningful when both A and U \A have the comparable numbers of
ratings. (cid:140)us, we weighted κi by a balance factor, min{ fA(vi)/fU \A(vi),
fU \A(vi)/fA(vi) }.
To make holistic use of di(cid:130)erent signals, i.e., topology, temporal
spikes, and rating deviation, we need a way to aggregate those
signals together. We have tried to use RRF (Reciprocal Rank Fusion)
scores from Information Retrieval, and wrapped the scores with
and without scaling function q(x). Compared to RRF score, we
found that a natural way of joint probability by multiplying those
signals together:
,
P(viA) = bαi +φi +κi−3
(7)
was the most e(cid:130)ective way to aggregate. In a joint probability,
we can consider the absolute suspicious value of each signal, as
opposite to the only use of ranking order. Moreover, being wrapped
with q(x), the signal values cannot be canceled out by multiplying
a very small value from other signals. A concrete example is that
a suspicious spike can still keep a high suspiciousness score by
multiplying a very small score from low fraudulent density.
Moreover, HoloScope dynamically updates the contrast suspi-
ciousness P(viA). (cid:140)us the sink nodes being added with camou-
(cid:131)age will have a very low contrast suspiciousness, with respect to
the suspicious source nodes A. (cid:140)is o(cid:130)ers HoloScope the resistance
to camou(cid:131)age.
3.2 Algorithm
Before designing the full algorithm for large scale datasets, we
(cid:128)rstly introduce the most important sub-procedure GreedyShavin❕
in Alg. 2.
At the beginning, this greedy shaving procedure starts with an
initial set A0 ⊂ U as input. It then greedily deletes source nodes
from A, according to users' scores S:
S(uj ∈ A) =
σ ji · eji · P(viA),
vi :(uj,vi)∈E
which can be interpreted as how many suspicious nodes that user
uj is involved in. So the user is less suspicious if he has a smaller
score, with respect to the current contrast suspiciousness P, where
we use P to denote a vector of contrast suspiciousness of all sink
nodes. We build a priority tree to help us e(cid:129)ciently (cid:128)nd the user
with minimum score. (cid:140)e priority tree updates the users' scores
and maintains the new minimum as the priorities change. With
removing source nodes A, the contrast suspiciousness P change,
in which we then update users' scores S. (cid:140)e algorithm keeps
reducing A until it is empty. (cid:140)e best A∗ maximizing objective HS
and P(vA∗) are returned at the end.
HoloScope: Topology-and-Spike Aware Fraud Detection
Conference'17, July 2017, Washington, DC, USA
Algorithm 2 GreedyShavin❕ Procedure.
Given bipartite multigraph G(U , V , E),
Initialize:
initial source nodes A0 ⊂ U .
A = A0
P= calculate contrast suspiciousness given A0
S = calculate suspiciousness scores of source nodes A.
MT = build priority tree of A with scores S.
while A is not empty do
u = pop the source node of the minimum score from MT .
A = A \ u, delete u from A.
Update P with respect to new source nodes A.
Update MT with respect to new P.
H S = estimate objective as Equation (3).
end while
return A∗ that maximizes objective H S(A∗) and P(v A∗), v ∈ V .
Since awakening and burst points have been already calculated
for each sink node as an initial step before the GreedyShavin❕
procedure, the calculation of the contrast suspicious P(vA) for a
sink node v only needs O(A) time. With source node j as the j-th
one removed from A0 by the GreedyShavin❕ procedure, A0 = m0,
and the out degree as di, the complexity is
O(dj · (j − 1) · log m0) = O(m0E0 log m0)
(8)
j =2,··· ,m0
where E0 is the set of edges connected to source nodes A0.
With the GreedyShavin❕ procedure, our scalable algorithm can
be designed so as to generate candidate suspicious source node sets.
In our implementation, we use singular vector decomposition (SVD)
for our algorithm. Each top singular vector gives a spectral view of
high connectivity communities. However, those singular vectors
are not associated with suspiciousness scores. (cid:140)us combined with
the top singular vectors, our fast greedy algorithm is given in Alg. 3.
Theorem 3.3 (Algorithm complexity). In the graph G(U , V , E),
given V = O(U ) and E = O(U 0), the complexity of FastGreedy
algorithm is subquadratic, i.e., o(U 2) in li(cid:136)le-o notation, if the size
of truncated user set U (k) ≤ U 1/ , where > max{1.5,
3−0 }.
2
Proof. (cid:140)e FastGreedy algorithm executes GreedyShavin❕ in a
constant iterations. A0 is assigned to U (k) in every GreedyShavin❕
procedure. (cid:140)en m0 = A0 = U (k). In the adjacency matrix M of
the graph, we consider the submatrix M0 with A0 as rows and V as
columns. If the fraudulent dense block is in submatrix M0, then we
assume that the block has at most O(m0) columns. Excluding the
dense block, the remaining part of M0 is assumed to have the same
density with the whole matrix M. (cid:140)erefore, the total number of
edges in M0 is
O(E0) = O(m
2
0 + m0 · E
U
) = O(U 2/ + U 1/−1E)
(cid:140)en based on equation (8), the algorithm complexity is
O(m0E0 log m0) = O((U 3/ + U 2/−1+0) log U )
(cid:140)erefore, if > max{1.5,
quadratic o(U 2).
3−0 }, then the complexity is sub-
2
(cid:3)
In real life graph, 0 ≤ 1.6, so if > 1.5 the complexity of
FastGreedy algorithm is subquadratic. (cid:140)erefore, without loss
of performance and e(cid:129)ciency, we can limit U (k) ≤ U 1/1.6 for
truncating an ordered U in the FastGreedy algorithm for a large
dataset.
In FastGreedy algorithm for HS-α, SVD on adjacency matrix M
is used to generate initial blocks for the GreedyShavin❕ procedure.
Although we can still use SVD on M for HS with holistic a(cid:138)ributes,
yet considering a(cid:138)ributes of timestamps and rating scores may
bring more bene(cid:128)ts. Observing that not every combination of # of
stars, timestamps and product ids has a value in a multi-way tenor
representation, we can only choose every existing triplets (object,
timestamp, #stars) as one column, and user as rows, to form a new
matrix. (cid:140)e above transformation is called the matricization of a
tensor, which outputs a new matrix. With proper time bins, e.g.,
one hour or day, and re-clustering of #stars, the (cid:131)a(cid:138)ening matrix
becomes more dense and contains more a(cid:138)ribute information. (cid:140)us
we use such a (cid:131)a(cid:138)ening matrix with each column weighted by the
sudden-drop suspiciousness for our FastGreedy algorithm.
Algorithm 3 FastGreedy Algorithm for Fraud detection.
Given bipartite multigraph G(U , V , E).
L = get (cid:128)rst several le(cid:137) singular vectors
for all L(k) ∈ L do
Rank source nodes U decreasingly on L(k)
(k)
u ≤ 1√U
U (k) = truncate u ∈ U when L
Gr eedyShavin❕ with initial U (k).
end for
return the best A∗ with maximized objective H S(A∗),
and the rank of v ∈ V by fA∗(v) · P(v A∗).
4 EXPERIMENTS
Table 2: Data Statistics
Data Name
BeerAdvocate [18]
Yelp
Amazon Phone & Acc [17]
Amazon Electronics [17]
Amazon Grocery [17]
Amazon mix category [19]
#nodes
26.5K x 50.8K
686K x 85.3K
2.26M x 329K
4.20M x 476K
763K x 165K
1.08M x 726K
time span
#edges
1.07M Jan 08 - Nov 11
2.68M Oct 04 - Jul 16
3.45M Jan 07 - Jul 14
7.82M Dec 98 - Jul 14
1.29M Jan 07 - Jul 14
2.72M Jan 04 - Jun 06
In the experiments, we only consider the signi(cid:128)cant multiple
bursts for (cid:131)uctuated time series of sink nodes. We keep those
awakening-burst point pairs with the altitude di(cid:130)erence ∆c at least
50% of the largest altitude di(cid:130)erence in the time series. Table 2
gives the statistics of our six datasets which are publicly available
for academic research 3. Our extensive experiments showed that
the performance was insensitive to scaling base b, and became
very stable when larger than 32. Hence we choose b = 32 in the
following experiments.
3Yelp dataset is from h(cid:138)ps://www.yelp.com/dataset challenge
Conference'17, July 2017, Washington, DC, USA
Shenghua Liu,1,2,∗ Bryan Hooi,2 Christos Faloutsos2
Table 3: Experimental results on real data with injected labels
Data Name
BeerAdvocate
metrics*
auc
F≥90%
auc
F≥90%
auc
F≥90%
auc
F≥90%
auc
F≥90%
auc
F≥90%
M-Zoom D-Cube
0.7280
0.7353
0.5000
0.5000
0.9137
0.9019
0.2000
0.2500
0.8042
0.9246
0.1667
0.5000
0.9117
0.9141
0.1250
0.2000
0.8428
0.8998
0.5000
0.1667
0.9001
0.8490
0.5000
0.2500
Yelp
Amazon
Phone & Acc
Amazon
Electronics
Amazon
Grocery
Amazon
mix category
* we use the two metrics: the area under the curve (abbrev as low-case "auc") of the accuracy curve as drawn in Fig. 1b, and the lowest
detection density that the method can detect in high accuracy(≥ 90%).
† one of the above fraudulent density was not detected in high accuracy.
0.5747
0.2000†
0.0009
0.0058
0.0515
0.0301
0.0200
–
–
–
0.0157
HS
0.9758
0.0333
0.9925
0.0143
0.9691
0.0200†
0.9250
0.1000
0.9250
0.1000
0.9922
0.0167
M-Zoom D-Cube
0.6221
0.6454
0.5000
0.5000
0.8863
0.9709
1.0000
0.0250
0.8810
0.9279
0.1429
0.1000
0.7868
0.9142
0.5000
0.1000
0.8241
0.8756
0.2500
0.1250
0.9937
0.9346
0.2000
0.0100
source nodes
CrossSpot
0.2259
–
0.9916
0.0200
0.0169
sink nodes
CrossSpot
0.1295
0.0415
HS
0.9945
0.0333
0.9950
0.0100
0.9823
0.0200†
0.9385
0.1250
0.9621
0.1000
0.9950
0.0100
–
–
–
–
–
–
4.1 Evaluation on di(cid:130)erent injection density
In the experiments, we mimic the fraudsters' behaviors and ran-
domly choose 200 objects that has no more than 100 indegree as
the fraudsters' customers, since less popular objects are more likely
to buy fake ratings. On the other hand, the fraudulent accounts
can come from the hijacked user accounts. (cid:140)us we can uniformly
sample out a number of users from the whole user set as fraud-
sters. To test on di(cid:130)erent fraudulent density, the number of sampled
fraudsters ranges from 200 to 20000. (cid:140)ose fraudsters as a whole
randomly rate each of the 200 products for 200 times, and also
create some biased camou(cid:131)age on other products. As a results, the
fraudulent density ranges from 1.0 to 0.01 for testing. (cid:140)e rating
time was generated for each fraudulent edge: (cid:128)rst randomly choos-
ing a start time between the earliest and the latest creation time
of the existing edges; and then plus a randomly and biased time
interval from the intervals of exiting creation time, to mimic the
surge of fraudsters' a(cid:138)acks. Besides, a high rating score, e.g. 4 or
4.5, is randomly chosen for each fraudulent edge4.
Fig. 1b shows the results of HoloScope HS on the BeerAdvocate
data. When the fraudulent density decreases from the right to the
le(cid:137) along the horizontal axis, HS can keep as high F-measure on
accuracy as more than 80% before reaching 0.025 in density, be(cid:138)er
by far than the competitors. Since HS returns suspiciousness scores
for sink nodes, we measure their accuracy using AUC (the area
under the ROC curve), where ROC stands for receiver operating
characteristic. For the detection of suspicious objects, HS achieved
more than 0.95 in AUC for all the testing injection density, with a
majority of tests reaching to 1.0.
In order to give a comparison on all six data sets with di(cid:130)erent
injection density, we propose to use the two metrics: a low-case
"auc" and the lowest detection density, described in the notes of
Table 3. (cid:140)e table reports the fraud detection results of our Holo-
Scope (HS) and competitors on the six datasets. Since the accuracy
curve stops at 0.01 (the minimum testing density); and we add zero
accuracy at zero density, the ideal value of auc is 0.995. (cid:140)e auc
on source and sink nodes are reported separately. As the table
4 the injection code is also open-sourced for reproducibility
suggests, our HS achieved the best auc among the competitors,
and even reached the ideal auc in two cases. Since the HS outputs
the suspiciousness scores for sink nodes, we used the area under
the upper-case AUC (similar to F-measure) accuracy curve along
all testing density. Although it is sometimes unfair to compare
F-measure with AUC, since the smallest value of AUC is around
0.5, yet the high auc values can indicate the high F-measure values
on our top suspicious list.
Furthermore, we compare the lowest detection density in Table 3.
(cid:140)e be(cid:138)er a method is, the lower density it should be able to detect
well. As we can see, HS has the smallest detection density in most
cases, which can be as small as 200/14000= 0.0143 on source nodes,
and reached the minimum testing density of 0.01 on sink nodes.
(cid:140)at means we can detect fraudsters in high accuracy even if they
use 14 thousand accounts to create 200 × 200 fake edges for 200
objects. (cid:140)e fraudulent objects can also be detected accurately, even
if 20 thousand fraudsters are hired to create 200 fake edges for each
object.
4.2 Evaluation on Sina Weibo with real labels
We also did experiments on a large real dataset from Sina Weibo,
which has 2.75 million users, 8.08 million messages, and 50.1 million
edges in Dec 2013. (cid:140)e user names and ids, and message ids are
from the online system. (cid:140)us we can check their existence status
in the system to evaluate the experiments. If the messages or the
users were deleted from the system, we treat them as the basis for
identifying suspicious users and messages. Since it is impossible to
check all of the users and messages, we (cid:128)rstly collected a candidate
set, which is the union of the output sets from the HS and the
baseline methods. (cid:140)e real labels are from the candidate set by
checking the status whether they still exists in Sina Weibo (checked
in Feb. 2017). We used a program on the API service of Sina Weibo
to check the candidate user and message id lists, (cid:128)nally resulting
in 3957 labeled users and 1615 labeled messages.
(cid:140)e experimental results in Fig. 1c show that HS achieved high
F-measure on accuracy, which detected 3781 labeled users higher
than M-Zoom's 1963 labeled users. (cid:140)e F-measure of HS improved
HoloScope: Topology-and-Spike Aware Fraud Detection
Conference'17, July 2017, Washington, DC, USA
about 30% and 60%, compared with M-Zoom and D-Cube respec-
tively. CrossSpot biased to include a large amount of users in their
detection results, which detected more than 100 labeled users but
with extremely low precision, i.e., less than 1%. (cid:140)at is the reason
CrossSpot got the lowest F-measure, which is less than 1.5%. For
labeled messages, the HS achieved around 0.8 in AUC, while M-
Zoom and D-Cube got lower recall, and CrossSpot still su(cid:130)ered
very low F-measure with higher recall. (cid:140)erefore, our HoloScope
outperformed the competitors in real-labeled data as well.
4.3 Scalability
To verify the complexity, we choose two representative datasets:
BeerAdvocate data which has the highest volume density, and Ama-
zon Electronics which has the most edges. We truncated the two
datasets according to di(cid:130)erent time ranges, i.e., from the past 3
months, 6 months, or several years to the last day, so that the gen-
erated data size increases. Our algorithm is implemented in Python.
As shown in Fig. 2, the running time of our algorithm increases
almost linearly with the number of the edges.
5 CONCLUSION
We proposed a fraud detection method, HoloScope, on a bipartite
graph which can have timestamps and rating scores. HoloScope
has the following advantages: 1) Uni(cid:128)cation of signals: we make
holistic use of several signals, namely topology, temporal spikes,
and rating deviation in our suspiciousness framework in a system-
atic way. 2) (cid:135)eoretical analysis of fraudsters' obstruction: we
showed that if the fraudsters use less than an upper bound of time
to rate an object, they will cause a suspicious drop or burst. In other
words, our HoloScope can obstruct fraudsters and increases their
time cost. 3) E(cid:130)ectiveness: we achieved higher accuracy on both
semi-real and real datasets than the competitors, achieving good
accuracy even when the fraudulent density is low. 4) Scalability:
while HoloScope needs to dynamically update the suspiciousness
of objects, the algorithm is sub-quadratic in the number of nodes,
under reasonable assumptions.
REFERENCES
[1] Leman Akoglu, Rishi Chandy, and Christos Faloutsos. 2013. Opinion Fraud
Detection in Online Reviews by Network E(cid:130)ects. ICWSM 13 (2013), 2–11.
[2] Miguel Araujo, Stephan Gunnemann, Gonzalo Mateos, and Christos Faloutsos.
2014. Beyond blocks: Hyperbolic community detection. In Joint European Con-
ference on Machine Learning and Knowledge Discovery in Databases. Springer,
50–65.
[3] Alex Beutel, Wanhong Xu, Venkatesan Guruswami, Christopher Palow, and
Christos Faloutsos. 2013. Copycatch: stopping group a(cid:138)acks by spo(cid:138)ing lockstep
behavior in social networks. In Proceedings of the 22nd international conference
on World Wide Web. ACM, 119–130.
[4] Moses Charikar. 2000. Greedy approximation algorithms for (cid:128)nding dense
components in a graph. In International Workshop on Approximation Algorithms
for Combinatorial Optimization. Springer, 84–95.
[5] Gordon V Cormack, Charles LA Clarke, and Stefan Bue(cid:138)cher. 2009. Reciprocal
rank fusion outperforms condorcet and individual rank learning methods. In
Proceedings of the 32nd international ACM SIGIR conference on Research and
development in information retrieval. ACM, 758–759.
[6] Geli Fei, Arjun Mukherjee, Bing Liu, Meichun Hsu, Malu Castellanos, and Rid-
dhiman Ghosh. 2013. Exploiting Burstiness in Reviews for Review Spammer
Detection. ICWSM 13 (2013), 175–184.
tor: L 2 theory. Probability theory and related (cid:128)elds 57, 4 (1981), 453–476.
[7] David Freedman and Persi Diaconis. 1981. On the histogram as a density estima-
[8] David Gibson, Ravi Kumar, and Andrew Tomkins. 2005. Discovering large dense
subgraphs in massive graphs. In Proceedings of the 31st international conference
on Very large data bases. VLDB Endowment, 721–732.
[9] Nikou Gunnemann, Stephan Gunnemann, and Christos Faloutsos. 2014. Robust
multivariate autoregression for anomaly detection in dynamic product ratings.
In Proceedings of the 23rd international conference on World wide web. ACM,
361–372.
[10] Stephan Gunnemann, Nikou Gunnemann, and Christos Faloutsos. 2014. Detect-
ing anomalies in dynamic rating data: A robust probabilistic model for rating
evolution. In Proceedings of the 20th ACM SIGKDD international conference on
Knowledge discovery and data mining. ACM, 841–850.
[11] Bryan Hooi, Hyun Ah Song, Alex Beutel, Neil Shah, Kijung Shin, and Christos
Faloutsos. 2016. Fraudar: bounding graph fraud in the face of camou(cid:131)age. In
Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. ACM, 895–904.
[12] Meng Jiang, Alex Beutel, Peng Cui, Bryan Hooi, Shiqiang Yang, and Christos
Faloutsos. 2015. A general suspiciousness metric for dense blocks in multimodal
data. In Data Mining (ICDM), 2015 IEEE International Conference on. IEEE, 781–
786.
[13] Meng Jiang, Peng Cui, Alex Beutel, Christos Faloutsos, and Shiqiang Yang.
2014. Inferring strange behavior from connectivity pa(cid:138)ern in social networks.
In Paci(cid:128)c-Asia Conference on Knowledge Discovery and Data Mining. Springer,
126–138.
[14] Qing Ke, Emilio Ferrara, Filippo Radicchi, and Alessandro Flammini. 2015. De(cid:128)n-
ing and identifying Sleeping Beauties in science. Proceedings of the National
Academy of Sciences 112, 24 (2015), 7426–7431.
[15] Ravi Kumar, Jasmine Novak, Prabhakar Raghavan, and Andrew Tomkins. 2005.
On the bursty evolution of blogspace. World Wide Web 8, 2 (2005), 159–178.
[16] Huayi Li, Geli Fei, Shuai Wang, Bing Liu, Weixiang Shao, Arjun Mukherjee,
and Jidong Shao. 2016. Modeling Review Spam Using Temporal Pa(cid:138)erns and
Co-bursting Behaviors. arXiv preprint arXiv:1611.06625 (2016).
Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics:
understanding rating dimensions with review text. In Proceedings of the 7th ACM
conference on Recommender systems. ACM, 165–172.
Julian John McAuley and Jure Leskovec. 2013. From amateurs to connoisseurs:
modeling the evolution of user expertise through online reviews. In Proceedings
of the 22nd international conference on World Wide Web. ACM, 897–908.
[19] Arjun Mukherjee, Bing Liu, and Natalie Glance. 2012. Spo(cid:138)ing fake reviewer
groups in consumer reviews. In Proceedings of the 21st international conference
on World Wide Web. ACM, 191–200.
h(cid:138)p://go.whiteops.com/rs/
179-SQE-823/images/WO Methbot Operation WP.pdf. (2016). [Online; accessed
Jan-7-2017].
[20] White Ops. 2016. (cid:140)e Methbot Operation.
[17]
[18]
[21] Shashank Pandit, Duen Horng Chau, Samuel Wang, and Christos Faloutsos.
2007. Netprobe: a fast and scalable system for fraud detection in online auction
networks. In Proceedings of the 16th international conference on World Wide Web.
ACM, 201–210.
[22] B Aditya Prakash, Ashwin Sridharan, Mukund Seshadri, Sridhar Machiraju,
and Christos Faloutsos. 2010. Eigenspokes: Surprising pa(cid:138)erns and scalable
community chipping in large graphs. In Paci(cid:128)c-Asia Conference on Knowledge
Discovery and Data Mining. Springer, 435–448.
[23] Stephen Robertson. 2004. Understanding inverse document frequency: on theo-
retical arguments for IDF. Journal of documentation 60, 5 (2004), 503–520.
[24] Neil Shah, Alex Beutel, Bryan Hooi, Leman Akoglu, Stephan Gunnemann, Disha
Makhija, Mohit Kumar, and Christos Faloutsos. 2015. EdgeCentric: Anomaly
Detection in Edge-A(cid:138)ributed Networks. arXiv preprint arXiv:1510.05544 (2015).
[25] Kijung Shin, Tina Eliassi-Rad, and Christos Faloutsos. 2016. CoreScope: Graph
Mining Using k-Core Analysis-Pa(cid:138)erns, Anomalies and Algorithms. ICDM.
[26] Kijung Shin, Bryan Hooi, and Christos Faloutsos. 2016. M-Zoom: Fast Dense-
Block Detection in Tensors with (cid:139)ality Guarantees. In Joint European Conference
on Machine Learning and Knowledge Discovery in Databases. 264–280.
[27] Kijung Shin, Bryan Hooi, Jisu Kim, and Christos Faloutsos. 2017. D-Cube:
Dense-Block Detection in Terabyte-Scale Tensors. In Proceedings of the 10th ACM
International Conference on Web Search and Data Mining (WSDM '17). ACM.
[28] Karen Sparck Jones. 1972. A statistical interpretation of term speci(cid:128)city and its
application in retrieval. Journal of documentation 28, 1 (1972), 11–21.
153 (1926), 65–66.
[29] Herbert A Sturges. 1926. (cid:140)e choice of a class interval. J. Amer. Statist. Assoc. 21,
[30] Sihong Xie, Guan Wang, Shuyang Lin, and Philip S Yu. 2012. Review spam
detection via temporal pa(cid:138)ern discovery. In Proceedings of the 18th ACM SIGKDD
international conference on Knowledge discovery and data mining. ACM, 823–831.
Jaewon Yang and Jure Leskovec. 2011. Pa(cid:138)erns of temporal variation in online
media. In Proceedings of the fourth ACM international conference on Web search
and data mining. ACM, 177–186.
[31]
|
1812.02712 | 1 | 1812 | 2018-12-06T18:46:32 | Analyzing the hate and counter speech accounts on Twitter | [
"cs.SI"
] | The online hate speech is proliferating with several organization and countries implementing laws to ban such harmful speech. While these restrictions might reduce the amount of such hateful content, it does so by restricting freedom of speech. Thus, an promising alternative supported by several organizations is to counter such hate speech with more speech. In this paper, We analyze hate speech and the corresponding counters (aka counterspeech) on Twitter. We perform several lexical, linguistic and psycholinguistic analysis on these user accounts and obverse that counter speakers employ several strategies depending on the target community. The hateful accounts express more negative sentiments and are more profane. We also find that the hate tweets by verified accounts have much more virality as compared to a tweet by a non-verified account. While the hate users seem to use words more about envy, hate, negative emotion, swearing terms, ugliness, the counter users use more words related to government, law, leader. We also build a supervised model for classifying the hateful and counterspeech accounts on Twitter and obtain an F-score of 0.77. We also make our dataset public to help advance the research on hate speech. | cs.SI | cs | Analyzing the hate and counter speech accounts on Twitter
Binny Mathew1 Navish Kumar1 Ravina2
Pawan Goyal1 Animesh Mukherjee1
1 Indian Institute of Technology, Kharagpur
2 National Institute of Technology, Patna
[email protected], [email protected], [email protected]
[email protected] , [email protected]
Abstract
The online hate speech is proliferating with several organi-
zation and countries implementing laws to ban such harm-
ful speech. While these restrictions might reduce the amount
of such hateful content, it does so by restricting freedom
of speech. Thus, an promising alternative supported by sev-
eral organizations is to counter such hate speech with more
speech. In this paper, We analyze hate speech and the cor-
responding counters (aka counterspeech) on Twitter. We per-
form several lexical, linguistic and psycholinguistic analysis
on these user accounts and obverse that counter speakers em-
ploy several strategies depending on the target community.
The hateful accounts express more negative sentiments and
are more profane. We also find that the hate tweets by verified
accounts have much more virality as compared to a tweet by a
non-verified account. While the hate users seem to use words
more about envy, hate, negative emotion, swearing terms, ug-
liness, the counter users use more words related to govern-
ment, law, leader. We also build a supervised model for clas-
sifying the hateful and counterspeech accounts on Twitter and
obtain an F-score of 0.77. We also make our dataset public to
help advance the research on hate speech.
1
Introduction
The online proliferation of hatespeech1 has caused several
countries and companies to implement laws against hate-
speech to enforce the citizens to restrain from such behavior.
Countries like Germany, United States of America, France
have laws banning hatespeech. Social media sites such as
Twitter, Facebook usually respond to hatespeech with the
suspension or deletion of the message or the user account
itself.
While these laws may reduce the amount of hatespeech in
online media, it does so at the cost of causing harm to the
freedom of speech. Another potential alternative to tackle
hatespeech is "counterspeech". The main idea behind coun-
terspeech is to add more speech to the conversation and try
to change the mindset of the hate speaker.
This requirement led countries and organizations to con-
sider counterspeech as an alternative to blocking (Gagliar-
done et al. 2015). The idea that 'more speech' is a remedy
for harmful speech has been familiar in liberal democratic
thought at least since the U.S. Supreme Court Justice Louis
1https://goo.gl/4rWGiF
Brandeis declared it in 1927. There are several initiatives
with the aim of using counterspeech to tackle hatespeech.
For example, the Council of Europe supports an initiative
called 'No Hate Speech Movement'2 with the aim to re-
duce the levels of acceptance of hatespeech and develop on-
line youth participation and citizenship, including in Internet
governance processes. UNESCO released a study (Gagliar-
done et al. 2015) titled 'Countering Online Hate Speech',
to help countries deal with this problem. Social platforms
like Facebook have started counterspeech programs to tackle
hatespeech3. Facebook has even publicly stated that it be-
lieves counterspeech is not only potentially more effective,
but also more likely to succeed in the long run (Bartlett and
Krasodomski-Jones 2015). Combating hatespeech in this
way has some advantages: it is faster, more flexible and re-
sponsive, capable of dealing with extremism from anywhere
and in any language and it does not form a barrier against
the principle of free and open public space for debate.
1.1 The present work
In this paper, we perform the first comparative study of
the interaction dynamics of the hatespeech and the corre-
sponding counterspeech replies. We choose Twitter as our
source of data and curate a dataset with 1290 hate tweet and
counterspeech reply pairs. After the annotation process, the
dataset consists of 558 unique hate tweets from 548 user and
1290 counterspeech replies from 1239 users. We found that
75.39% of these replies are counterspeech that oppose the
hatespeech spread by the user.
1.2 Contributions
The main contributions of our paper are as follows:
• We perform the first comparative study that looks into the
characteristics of the hateful and counter accounts.
• We provide a dataset of 1290 tweet-reply pair in which
the tweets are the hatespeech and their replies are coun-
terspeech. We plan to release the dataset upon acceptance.
2No Hate Speech Movement Campaign: http://www.
nohatespeechmovement.org/
3 Counterspeech Campaign
//counterspeech.fb.com/en/
by
Facebook: https:
• We develop a model which predicts if a given Twitter user
is a hateful or counterspeech account with an accuracy of
78%.
1.3 Observations
Our study results in several important observations.
• First, we find significant difference in the activity pat-
tern between these users. Hateful accounts tend to express
more negative sentiment and profanity in general. If the
hateful tweet is from a verified account, it seems to be
much more viral as compared to other hateful tweets.
• Another intriguing finding is that hateful users also act
as counterspeech users in some situations. In our dataset,
such users use hostile language as a counterspeech mea-
sure 55% of the times.
• In terms of personality traits, the counterspeakers seem to
have a higher quotient of 'agreeableness'. They are more
altruistic, modest and sympathetic. The hateful users, on
the other hand, seem to have a higher quotient of 'ex-
traversion' indicating that they are more energetic and
talkative in nature. They are more cheerful, excitement-
seeking, outgoing, and sociable in nature.
• Using the linguistic structure of the general tweets and
the account characteristics of the hateful and the counter
speakers, it is possible to distinguish them early with an
accuracy of 78%.
• We analyze the topical inclination of the users and ob-
serve that counterspeakers prefer topics related to 'poli-
tics', 'news', and 'journalism' while hate users prefer top-
ics like 'niggas'.
2 Preliminaries
• Hatespeech: We define hatespeech according to the Twit-
ter guidelines. Any tweet that 'promotes violence against
other people on the basis of race, ethnicity, national ori-
gin, sexual orientation, gender, gender identity, religious
affiliation, age, disability, or serious disease' is consid-
ered as a hatespeech4. A hate account (HA) is a Twitter
account which posts one or more such hateful tweets.
• Counterspeech: We follow the definition of counter-
speech used in (Mathew et al. 2018). We call a tweet as
a 'counterspeech' if the tweet is a direct reply to a hate-
ful tweet. A counter account (CA) is a Twitter account
which posts one or more such counterspeech in response
to a hateful post.
3 Dataset Collection
We use Twitter as the platform for this study. As we are in-
terested in the replies received by a particular hate tweet, we
could not use the Twitter API directly as it does not pro-
vide any service that helps in directly collecting the replies
of a particular tweet. To collect these replies we utilize the
4https://help.twitter.com/en/rules-and-policies/hateful-
conduct-policy
Figure 1: A flowchart showing the process by which we ob-
tained the dataset.
PHEME script5 which allows us to collect the set of tweets
replying to a specific tweet, forming a conversation. Other
works (Zubiaga et al. 2016) have used this technique as well.
Our overall methodology is a multi-step approach for creat-
ing the dataset which can be divided into the following three
steps:
• Hateful tweet collection.
• Filtration and annotation of the hateful tweets.
• Extraction and annotation of the reply tweets.
We explain these three steps in detail below.
3.1 Hatespeech collection
We rely on the hatespeech templates defined in (Silva et al.
2016) to collect the hateful tweets. These templates are of
the form:
I <intensity><userintent><hatetarget>
The subject "I" implies that the user is expressing her
own personal emotions. The verb, embodied by <userin-
tent> component is used to specify the user's intent which
in this case is the word 'hate' or its synonyms. The com-
ponent <intensity> is optional and acts as a qualifier which
some users use to amplify their emotions. Words such as 're-
ally','f**cking' are used to express the <intensity>. In our
work, we have used the words that are listed in (Mondal,
Silva, and Benevenuto 2017) for each of the component. The
component <hatetarget> is used to find the target commu-
nity on the receiving end of the hatred. Table 2 lists the key-
words that we have used to extract hateful tweets for each of
these communities. Some examples of the these templates
include: "I hate muslims", "I really despise white people"
etc.
Next, we utilize the Jefferson-Henrique's web scraping
script6 to collect the hateful posts using the templates de-
fined above. We run the script for all the communities de-
fined in Table 2 to collect around 578K tweets in total. We
removed all non-English tweets from the dataset.
3.2 Filtration and annotation of hateful tweets
As an initial filtering step to ensure that we get sufficient
replies to a hateful tweet, we filter out all the tweets col-
lected in the previous step that did not have at least two
replies. This reduced the number of hatespeech tweets to
5https://github.com/azubiaga/pheme-twitter-conversation-
collection
6https://github.com/Jefferson-Henrique/GetOldTweets-python
Counterspeech Type
Presentation of facts
Pointing our hypocrisy
Hatespeech
A tragedy that is a DIRECT result of allowing too many
f**king Muslims in their country. There should not be
any surprises here. #BANMUSLIMS
I hate Fat ppl bruhhhh ughhh
Warning of consequences
I hate lesbians so much omfg
Affiliation
Denouncing speech
Images
Humor
Positive tone
Hostile language
Miscellaneous
I'm christian so i hate muslims, muslim turkish people
and MUSLIM TURKISH BELIEBERS cause YOU ARE
THE TERRORIST!Turkish BELIEBERS Need Bieber
I hate gays
- - -
I hate gays
@user the problem with the world is you f**king mus-
lims! go choke to death on some bacon you child raper
pig! f**k you a**hole!
@user f**k the muslims!!!
F**k I hate women! All bit**es to a good guy
Total
Counterspeech
The guy was born in UK in 1995
"@user: I hate Fat ppl bruhhhh ughhh" you use to be fat
you hypocrite!
I report you to the LGBT community, this blatant dis-
crimination will not be tolerated
@user IM MUSLIM BUT DO NOT HATE CRISTIANS
"@user: I hate gays." Stop being homophobic
- - -
@user so uhhh I guess you hate yourself huh? HAHAH-
HAHAHAHAHA man I'm funny
@user, I'm sorry you feel that way. Islam is a beautiful
religion, so misunderstood. Not all Muslims are the same.
@user you are truly one stupid backwards thinking
mother**cker to believe negativity about Islam
@user Hey hey now.. ALL women are NOT the same.
Total
136
177
69
58
132
83
134
175
357
220
1481
Table 1: Example hatespeech and the corresponding counterspeech received for each type. Please note that the total is more
than the reported 1290 counterspeech replies because the users sometimes employ multiple strategies in a single tweet.
Gender
Men
Women
Female
Physical Trait
Fat
Ethnicity
Nigger
Nigga
White people
Black people
Sexual Orientation Nationality
Lesbian
Gay
LGBT
Transgender
American
Indian
Mexican
Arab
Asian
Religion
Jew
Islam
Muslim
Table 2: Keywords used during search to select hatespeech
for different communities.
13,321. Next, we manually tagged the tweets as containing
hatespeech or not. We follow the guidelines defined by Twit-
ter 4 to identify if a tweet contains hatespeech or not. Each
tweet was annotated by two users and in case of a disagree-
ment, we ask one more annotator to resolve it. The anno-
tators were instructed to take the context into consideration
when annotating a tweet as hateful. The annotations were
performed by a group of two undergraduate students with
Major in Computer Science and a PhD student in Social
Computing. The final dataset consisted of 558 tweets that
were tagged as hateful. Figure 1 illustrates the whole pro-
cess in a flowchart.
3.3 Extraction and annotation of the reply tweets
Once we selected the hatespeech tweets with at least two
replies, the next step involved using the PHEME script5 to
scrape the replies received to these hateful tweets. We ob-
serve that the 558 hate tweets received a total of 1711 replies
with a mean of 3.067 and a median of 2 replies per tweet.
We employ two annotators for the counterspeech tagging.
We follow the procedure used in (Mathew et al. 2018) to
annotate the counterspeech data. For each reply text to a
hatespeech tweet, the annotators were asked to perform two
tasks: (i) decide if the reply is a counterspeech or not, and
(ii) if the reply is indeed a counterspeech, then decide the
category of the counterspeech. We describe these categories
in the subsequent section.
Two independent annotators tagged each reply tweet as
counterspeech or not and got an accuracy of 81.35% and κ
score of 0.46, which is moderately good. For the second task
of deciding on the category of counterspeech, we use the
Exact Match Ratio and Accuracy defined in (Sorower 2010).
The two annotators obtain an Exact Match Ratio of 71.4%
and an Accuracy of 77.58%. We employ a third annotator to
resolve the conflicting cases and report the final distribution
in Table 1.
4 Taxonomy of Counterspeech
A successful counter to a hatespeech will require various
types of strategies. In (Benesch et al. 2016b), the authors de-
fine eight such strategies that are used by counterspeakers.
Similar to (Mathew et al. 2018), we divide the Tone cate-
gory into two parts: positive and negative Tone. We aggre-
gate all the counterspeech that would not fit in the above
categories into a 'Miscellaneous' category. Note that some-
times the users employ multiple counterspeech strategies in
a single tweet. We refer the readers to (Benesch et al. 2016b;
Mathew et al. 2018) for a detailed discussion on the taxon-
omy of counterspeech. Table 1 shows an example for each
type of counterspeech.
Hate Target
Presentation of facts
Pointing out hypocrisy
Warning of consequences
Affiliation
Denouncing speech
Images
Humor
Positive tone
Hostile language
Miscellaneous
Total counter
Total hate
Gender
1 (00.36%)
38 (13.77%)
3 (01.09%)
14 (05.07%)
15 (05.43%)
17 (06.16%)
32 (11.59%)
47 (17.03%)
50 (18.12%)
59 (21.38%)
276
120
Sexuality Nationality
5 (04.24%)
5 (02.54%)
19 (9.64%)
16 (13.56%)
4 (3.39%)
9 (4.57%)
9 (4.57%)
9 (7.63%)
12 (10.17%)
20 (10.15%)
10 (5.08%)
10 (8.47%)
30 (15.23%)
6 (5.08%)
13 (11.02%)
34 (17.26%)
32 (27.12%)
39 (19.80%)
11 (9.32%)
22 (11.17%)
118
197
110
43
Religion Physical Trait
0 (00.00%)
7 (4.86%)
2 (1.39%)
2 (1.39%)
3 (2.08%)
1 (0.69%)
12 (8.33%)
15 (10.42%)
65 (45.14%)
37 (25.69%)
144
91
125 (17.86%)
104 (14.86%)
35 (5.00%)
24 (3.43%)
53 (7.57%)
41 (5.86%)
51 (7.29%)
64 (9.14%)
124 (17.71%)
79 (11.29%)
700
143
Ethinicity
2 (00.96%)
7 (3.35%)
25 (11.96%)
4 (1.91%)
34 (16.27%)
10 (4.78%)
8 (3.83%)
13 (6.22%)
81 (38.76%)
25 (11.96%)
209
99
Total
138 (08.39%)
191 (11.62%)
78 (4.74%)
62 (3.77%)
137 (8.33%)
89 (5.41%)
139 (8.45%)
186 (11.31%)
391 (23.78%)
233 (14.17%)
1644
606
Table 3: Counterspeech strategies used by various target communities. The percentage given in the bracket are normalized by
the total number of counterspeech in each category. Please note that the total reported is more than the 1290 counterspeech
replies because of the presence of multiple target communities in a single tweet.
5 Target community analysis
Our final dataset consisted of 558 hatespeech tweets from
548 Hate Accounts (HAs) and 1290 direct replies that were
counterspeech from 1239 Counterspeech Accounts (CAs).
Of the total 1711 tweets that were replies to the 558 hate
tweets, 75.39% (1290) were counterspeech tweets. This is
almost double of what is reported in (Mathew et al. 2018).
We argue that the main reason behind this could be the pub-
lic nature of Twitter as opposed to the semi-anonymous na-
ture of Youtube. We also found that 82.09% of the coun-
terspeech employed only one kind of strategy. The p-values
reported are calculated using Mann-Whitney U test.
5.1 Strategies used by CAs
The CAs used different strategies to tackle the hatespeech.
We can observe from Table 3 that different target commu-
nities adopt different measures to respond to the hateful
tweets. The largest fraction (23.60%) of the hateful tweets
seems to be religious in nature. They also receive the high-
est number of counterspeech with an average of 4.9 counter-
speech for every religious hateful tweet as compared to only
2.7 counterspeech per hateful tweet for the entire dataset.
The religious CAs seems to be using two strategies more as
compared to other target communities: presentation of facts
and pointing out hypocrisy.
In case of the nationality target communities, CAs seem
to be relying more on affiliation and using images in their
counter speech. The CAs for the ethnicity use warning of the
consequences and denounce the hatespeech more as com-
pared to other target communities.
The CAs for the target communities associated with sex-
uality rely on humor and positive tone to counter the hateful
messages. In case of the hatespeech which target the phys-
ical traits, the CAs heavily use hostile language to counter
them.
Irrespective of the community, the counterspeakers seem
to be using hostile language a lot with the lowest being
for the religion (17.71%) and highest for the physical traits
(45.14%). We can also observe that 27.67% of the counter-
speech tweets used hostile language as a strategy. This is less
than the 35% reported in (Mathew et al. 2018) on YouTube.
One of the main reasons for this could be the more public
nature of Twitter.
5.2 Use of images/videos
We observe that only 12 (2.15%) of hate tweets contained
images/videos whereas 145 (11.24%) of the counter tweets
had images/videos. We also look into the replies of these
12 hate tweets which use images/videos to get a better un-
derstanding. Interestingly, we found that 69 (47.59%) out
of the 145 counter tweets which use images/videos were
in response to hateful tweets which themselves used im-
ages/videos. Another interesting observation was that the
counterspeech involving images/videos are liked more by
the Twitter community as compared to other strategies. The
counter images received the highest likes among all the
counterspeech strategies with a median of 2.5 likes (average
= 10).
6 User level analysis
In this section, we characterize the HAs and CAs based on
their activity and tweet history. We first collect for each user
their last 3200 tweets using the Twitter API7. This would
also give us other information about the users such as the
number of tweets, number of followers, friends etc.
6.1 User activity
We normalize the user-characteristics by dividing them by
the age of the account. Thus each of the user properties are
divided by the number of days since the start of the account.
In Figure 2, we can observe several striking differences in
the nature of the accounts. We observe that the hate users
are more "popular" in that they tweet more, have more fol-
lowers, and are part of more public lists as compared to the
counter users (p-value<0.001).
The counter users have more friends per day as compared
to hate users (p-value<0.005).
7Twitter API: https://developer.twitter.com/
en/docs/tweets/timelines/api-reference/
get-statuses-user_timeline.html
Figure 2: Average values for several user profile properties for Hateful and Counter user normalized by the account age. Error
bars represent 95% confidence interval.
6.2 Creation dates
We analyze the account creation dates of hate and counter
users as shown in Figure 3. Previous works on hate users
(Ribeiro et al. 2017; ElSherief et al. 2018b) have reported
that the Twitter accounts of hateful users have relatively less
age as compared to the normal users. Our results, in con-
trast, do not seem to support this observation. We find that
the hate accounts are older than the counter accounts (p-
value<0.01). We argue that the main reason for this is due
to the way we collected our dataset. Around 80% of the hate-
ful tweets are older than Dec 2017. Around this time Twit-
ter first started enforcing stricter rules on abusive content,
which means that they managed to bypass the platform's
strict guidelines. Thus, many of the hate accounts are much
older.
Figure 4: Boxplots for the distribution of Subjectivity, Neg-
ative Sentiment, Profanity, and Hatespeech for the Hateful
and Counter users
Figure 3: KDEs of the account creation dates of Hateful and
Counter users.
account or that the account was permanently deleted after
thirty days of deactivation. We found that 4.56% and 4.28%
of the hate and counter users, respectively, were deleted.
6.3 Suspended and deleted accounts
The Twitter API returns an error message when the user ac-
count is suspended or the user is not found. According to
Twitter8, the common reasons for account suspension in-
clude spam, risk to Twitter's security, or abusive tweet or
behavior. Twitter accounts are not found when the user does
not exist. This may be because the user deactivated her
8https://help.twitter.com/en/managing-your-
account/suspended-twitter-accounts
6.4 Verified accounts
Next, we look for the presence of verified accounts in our
dataset. We found that 3.28% of the hate users are verified; in
contrast, 0.56% of the counter users have verified accounts.
We found that the hateful tweets by the HAs generated much
more audience as compared to the CAs. On average (me-
dian) the verified HAs received 356.42 (57.5), 80.42 (13),
24.08 (16) likes, retweets, and replies, respectively. This is
much higher than the verified CAs who received 1.8 (2.0),
0.4 (0.0), 0.4 (0.0) likes, retweets, and replies, respectively.
(a) Hate users
(b) Counter users
Figure 5: Word cloud of the user hashtags. There are several hashtags such as 'maga', 'fakenews', 'worldcup' that are common
in both types of users. Notice the inclusion of hashtags such as 'WhiteGenocide' in the word cloud for the hate users. The
counter users seem to use the hashtags such as 'loveislove' to show their support for different communities.
6.5 Tweet analysis
To understand how the different set of users express their
emotions, we make use of the users' tweet history. We
apply VADER (Hutto and Gilbert 2014) to find the av-
erage negative sentiment expressed by the users. In Fig-
ure 4, we observe that the tweets from hateful users ex-
press more negative sentiments compared to counter users.
We use TextBlob9 to measure the subjectivity expressed in
the tweets. As observed from Figure 4, hate accounts use
more subjective tweets as compared to counters accounts
(p<0.001). In order to find the profanity expressed in the
tweets, we use Shutterstock's "List of Dirty, Naughty, Ob-
scene, and Otherwise Bad Words"10. We observe that hate
accounts use more profane words as compared to counter
accounts (p-value<0.05). We use the model provided by
(Davidson et al. 2017) to check for hatespeech and abusive
language in the tweet history of user. We found that the hate
users seem to use more hatespeech and abusive language as
compared to counter speakers.
6.6 Hashtags
Using the Twitter history of the users, we plot word clouds
for each type of users as shown in Figure 5. As we can ob-
serve, there are several hashtags such as '#maga', '#fake-
news', '#worldcup', '#newprofilepic' that are common for
both the types of users. The hate accounts used hashtags
such as '#whitegenocide' which is commonly used by white
nationalist/supremacists.
6.7 Lexical analysis
Empath (Fast, Chen, and Bernstein 2016) is a tool that can
be used to analyze text over 200 pre-built lexical categories.
We characterize the users over these categories and ob-
serve several interesting results as shown in Figure 6. The
9https://github.com/sloria/textblob
10https://github.com/LDNOOBW/List-of-Dirty-Naughty-
Obscene-and-Otherwise-Bad-Words
hate users seem to be using more words related to cate-
gories such as envy, hate, negative emotion, positive emo-
tion, ridicule, swearing terms, ugliness. These results are in
alignment with the sentiment values obtained for these users.
The counter users seem to be using words in the categories
such as government, law, leader, pride, religion, terrorism
indicating that the counter users are more civil in nature. All
the values reported have p-value<0.01.
6.8 Personality analysis
In order to get a better understanding of the user char-
acteristics, we perform personality analysis on the user
set. We make use of the IBM Watson Personality Insights
API (Gou, Zhou, and Yang 2014) 11 to understand the per-
sonality of the users. Previous research (Hu et al. 2016;
Liu et al. 2017) has used this tool to understand the person-
ality of the users. We provide all the tweets of a particular
user as input to the API and the service analyzes the linguis-
tic features to infer the Big-5 personality traits (Goldberg
1992). We observe distinctive difference between the hate
and counter users as shown in Figure 7.
CAs score higher in the personality traits such as agree-
ableness. They are more altruistic, modest, sympathetic.
This is intuitive since the counter speakers in many cases
have a simple motive to help the hate speakers. They also
seem to be more self-disciplined. The counter speakers score
higher than the hate speakers in all the sub-traits of the con-
scientiousness. The counter speakers are more driven, delib-
erate, dutiful, persistent, and self-assured.
The HAs seem to score higher in extraversion indicating
that they are more energetic and talkative in nature. They are
more cheerful, excitement-seeking, outgoing, and sociable in
nature.
11IBM Personality Insights API: https://www.ibm.com/
watson/services/personality-insights/
Figure 6: Lexical analysis using Empath for the hateful and counter users. We report the median values for the relative oc-
currence in several categories of Empath. Error bars represent 95% confidence intervals. That HAs score significantly higher
in topics like envy, hate, negative emotion, positive emotion, ridicule, swearing terms, ugliness while CAs score significantly
higher in topics like government, law, leader, pride, religion, terrorism
Openess
Conscientiousness
Extraversion
Agreeableness
Figure 7: Personality traits of the CAs and HAs. The CAs are more self-disciplined, driven, deliberate, dutiful, persistent, and
self-assured while the HAs are more energetic, talkative, cheerful, excitement-seeking, outgoing, and social in nature.
6.9 Topical inclination of users
We infer the topical interests of the users using the method
explained in (Bhattacharya et al. 2014) and utilize the 'Who
likes what?' web service12 for this task. In this method, for a
given Twitter user u (whose interests are to be inferred), the
method first checks which other users u is following, i.e.,
users from whom u is interested in receiving information. It
then identifies the topics of expertise of those users (whom u
is following) to infer u's interests, i.e., the topics on which u
is interested in receiving information. Expertise is defined by
the users bio or tweets via the lists feature in Twitter (Ghosh
et al. 2012).
12Topical
Interest: https://twitter-app.mpi-sws.
org/who-likes-what
In order to understand which topics are preferred by the
HAs and CAs, we analyze the topical interests of the users
as shown by the word cloud in Figure 8. We can observe
that hate and counter users have several topics of common
interest. Topics such as 'Fame', 'Music', 'Artist' seem to be
of interest to a large proportion of hate and counter users.
On a closer analysis, we found that topics such as 'politics',
'news', and 'journalism' are preferred more by the CAs as
compared to HAs (p-value<0.01). This is possibly indica-
tive of a constructive nature of the CAs. Also the HAs tend
to subscribe to topics like 'niggas' significantly more.
7 Classification
In this section, we leverage the unique characteristics of
the hate and counter user accounts to develop a predictive
(a) Hate users.
(b) Counter users.
Figure 8: Word cloud of the user topical interests. Several of the topics such as 'fame', 'music', 'artist' seems to be of interest
to both hate and counter users. However while CAs subscribe to topics like 'politics', 'journalism' and 'news' more, HAs
subscribe to topics like 'niggas' more.
model. We pose this as a binary classification problem - dis-
tinguishing hate users from counter users. This early auto-
matic distinction can help the platform to develop appropri-
ate incentive mechanisms to demote the hate user accounts
and promote the counter user accounts and also urge these
users to help in overall purging of the media. (Munger 2017)
showed that counterspeech using automated bots can reduce
instances of racist speech if the instigators are sanctioned by
a high-follower white male.
7.1 Features
We use the following feature set for our classifier.
• TF-IDF values: For each account, we calculate its tf-idf
• User profile properties: We use several user account
properties such as #favorites per day, #tweets per day,
#followers per day, #friends per day, #listed per day, and
whether the account is verified.
• Lexical properties: We use the vector of 200 pre-built
• Affect properties: For each user account we calculate the
average sentiment, profanity, and subjectivity expressed
in the tweet history.
topics as the feature set for each user.
vector using the users tweet history.
7.4 Results
Table 4 shows the results of the classification task. We can
observe that CatBoost (CB) performs the best with an accu-
racy of 78% followed by XGBoost (XGB) with an accuracy
of 74%. Both the classifiers perform much better than the
baseline classifier (LR + TF-IDF).
Model
LR + TFIDF
SVM
LR
ET
RF
XGB
CB
Precision Recall F-score Accuracy
0.68
0.64
0.66
0.72
0.72
0.74
0.83
0.68
0.63
0.66
0.70
0.72
0.74
0.78
0.68
0.63
0.66
0.70
0.72
0.74
0.78
0.68
0.62
0.66
0.69
0.72
0.74
0.77
Table 4: Evaluation results for various classifiers - Support
Vector Machines (SVM), Logistic Regression (LR), Ran-
dom Forest (RF), Extra-Tree (ET), XGBoost (XGB), and
CatBoost (CB) on the task of classification of an account
as hateful or counter. The evaluation metrics reported are
calculated by taking macro average.
7.2 Dataset
We first divide the dataset into training and test set in the
ratio 90:10 while keeping the test set balanced. In order to
find the hyperparameter of our models, we use 10% of the
training data as the validation set. We run randomized grid
search to find the optimal hyperparameter values.
7.3 Choice of classifier
We choose classifiers such as Support Vector Machines
(SVM), Logistic Regression (LR), Random Forest (RF),
Extra-Tree (ET), XGBoost (XGB), and CatBoost (CB) for
this task. We use TF-IDF features along with LR as the base-
line.
7.5 Feature ablation
We perform feature ablation to understand the importance of
the different feature types. Table 5 reports the feature abla-
tion study carried out by CatBoost (CB) classifier. From the
table, we can observe that TF-IDF feature is the most useful
followed by lexical features.
8 Related Work
In this section, we review some of the related literature. First,
we review the works done on analysis and detection of hate-
speech and harmful language. Next we summarize the litera-
ture available for countering these hateful or harmful speech.
Feature excluded Precision Recall F-score Accuracy
TF-IDF
User profile
Lexical
Affect
0.53
0.79
0.56
0.77
0.53
0.79
0.56
0.77
0.59
0.84
0.65
0.83
0.43
0.78
0.49
0.76
Table 5: Feature ablation study for the CatBoost (CB) classi-
fier. The evaluation metrics reported are calculated by taking
macro average.
8.1 Hateful or harmful speech
Hatespeech lies in a complex nexus with freedom of expres-
sion, group rights, as well as concepts of dignity, liberty, and
equality (Gagliardone et al. 2015). Owing to this, there can
be several issues in defining what constitutes as hatespeech
(Benesch 2014). The authors usually adopt a definition that
fits a general description of hatespeech. There is substantial
literature on the analysis of hatespeech. (Silva et al. 2016)
studies the targets of hatespeech on Twitter and Whisper
and observes that 'Blacks' are among the most frequent tar-
gets of hatespeech. (Mondal, Silva, and Benevenuto 2017)
studies the effects of anonymity on hatespeech. There are
works where the authors study the perceptions and ex-
perience of people regarding the hatespeech (Leets 2002;
Gelber and McNamara 2016). In (Van Spanje and De Vreese
2015), the authors study the impact of hatespeech prosecu-
tion of a politician on electoral support for his party and find
an immediate increase in the support for the party.
Another line of research corresponds to detection of
hatespeech in various social media platforms like Twit-
ter (Waseem and Hovy 2016; Chatzakou et al. 2017; David-
son et al. 2017; Badjatiya et al. 2017), Facebook (Del Vi-
gna et al. 2017), Yahoo! Finance and News (Warner and
Hirschberg 2012; Djuric et al. 2015; Nobata et al. 2016),
and Whisper (Mondal, Silva, and Benevenuto 2017). In an-
other online effort, a Canadian NGO, the Sentinel Project13,
launched a site in 2013 called HateBase14, which invites In-
ternet users to add to a list of slurs and insulting words in
many languages.
There are some works which have tried to characterize the
hateful users. In (Ribeiro et al. 2018; Ribeiro et al. 2017),
the authors study the user characteristics of hateful accounts
on twitter and found that the hateful user accounts differ sig-
nificantly from normal user accounts on the basis of activity,
network centrality, and the type of content they produce. In
(ElSherief et al. 2018b), the authors perform a comparative
study of the hate speech instigators and target users on twit-
ter. They found that the hate instigators target more popular
and high profile twitter users, which leads to greater online
visibility. In (ElSherief et al. 2018a), the authors focus on
studying the target of the hatespeech - directed and general-
ized. They look into linguistic and psycholinguistic proper-
ties of these two types of hatespeech and found that while
directed hate speech is more personal and directed, informal
and express anger, the generalized hate is more of religious
13https://thesentinelproject.org/
14https://www.hatebase.org/
type and uses lethal words such as 'murder', 'exterminate',
and 'kill'.
8.2 Countering hatespeech
The current solution adopted by several organization and
companies to tackle online hate speech involves block-
ing/suspending the account or the particular hateful post.
While these options are very powerful, they tend to violate
the free speech doctrine. Countespeech is considered to be a
promising solution in this direction as it can help in con-
trolling the hate speech problem and at the same time, it
supports free speech. Institutions can help in educating the
public about hatespeech and its implications, consequences
and how to respond. Programmes such as 'No Hate Speech'
movement 2 and Facebook's counterspeech program 3 help
in raising awareness, providing support and seeking creative
solutions (Citron and Norton 2011).
Silence in response to digital hate carries significant ex-
pressive costs as well. When powerful intermediaries rebut
demeaning stereotypes (like the Michelle Obama image) and
invidious falsehoods (such as holocaust denial), they send
a powerful message to readers. Because intermediaries of-
ten enjoy respect and a sense of legitimacy, using counter-
speech, they can demonstrate what it means to treat others
with respect and dignity (Citron and Norton 2011).
Online social media sites such as Twitter, YouTube, and
Facebook have been used to study counterspeech. (Wright
et al. 2017) study the conversations on Twitter, and find that
some arguments between strangers lead to favorable change
in discourse and even in attitudes. (Ernst et al. 2017) study
the comments in YouTube counterspeech videos related to
Islam and find that they are dominated by messages that deal
with devaluating prejudices and stereotypes corresponding
to Muslims and/or Islam. In (Schieb and Preuss 2016),
the authors study counterspeech on Facebook and through
simulation, find that the defining factors for the success of
counter speech are the proportion of the hatespeech and the
type of influence the counter speakers can exert on the un-
decided.
(Stroud and Cox 2018) perform case studies on
feminist counterspeech. They explore the range of feminist
counterspeech by engaging two representative case studies
from the current online environment. (Benesch et al. 2016a)
describes strategies that have favorable impact or are coun-
terproductive on users who tweet hateful or inflammatory
content. In (Mathew et al. 2018), the authors curate a coun-
terspeech dataset from YouTube and build a general classi-
fier which achieves an F1-score of 0.73 .
This work is the first to study the interaction dynamics
of the hate and counter speakers in a coupled fashion. We
would also release the dataset used in our experiment as
well as the model for classifying an account as hateful or
counter 15.
9 Discussion
Here we discuss some of the key insights that we obtain from
this study. We believe that many of these could be very help-
15Dataset and Model available here: https://goo.gl/
vqMPc7
ful for Twitter and, in general, other social media platforms
to meaningfully combat hate content.
9.1 Success of counterspeech
We found cases in which the counterspeech provided was
successful in changing the mindset of the user who had
posted the hateful tweet. In one case, the user who had
posted the hateful message on Twitter, later apologized to
everyone saying that she was really sorry for what she did.
Although, we rarely receive such direct evidence of a coun-
terspeech being successful, these cases prove that counter-
speech can actually help in changing the attitude of the
hate users without resorting to aggressive measures such as
blocking/suspension of the account.
9.2 New venues for hatespeech
The harsh nature of Twitter on hatespeech has led to several
user account suspension and deletion. This is true for other
social media sites as well. Due to this, several new social
media sites have sprung up in the recent years like Gab16,
Wrongthink17 which support free speech on their site and
allow users to posts contents that would not be allowed on
other sites. This has resulted in gab becoming an echo cham-
ber for right-leaning dissemination (de Lima et al. 2018).
9.3 Common users
We found some users which were common in HAs and CAs.
On examining these accounts we observe that these accounts
take the role of hate or counter user depending on the situa-
tion. For example, we observe that a user who had performed
a hate speech targeting the 'LGBTQ' community also per-
formed a counterspeech in response to a hatespeech target-
ing 'Fat people'. We found in cases where the HAs act as
counter speaker, 55% of the times they use the hostile lan-
guage strategy.
9.4 Hate supporters
While annotating the dataset, we also found several tweets
that were in support of the hate tweet. Specifically, we an-
notated 223 direct replies to the hate tweets that were in
support. We found several interesting properties of these
hate support accounts (HSAs). To begin with, we found that
HSAs had several similarities with the HAs. They use more
profane and subjective words, their accounts are relatively
newer as compared to HAs and CAs, and hashtags such as
'whitegenocide' were also used by them. As per lexical anal-
ysis, the HSAs use more words in the categories such as
anger, independence, suffering. The HSAs seem to be more
open in nature as compared to CAs and HAs. They are more
imaginative, philosophical and authority-challenging.
10 Conclusion
In this paper, we perform the first characteristic study com-
paring the hateful and counterspeech accounts in Twitter. We
provide a dataset of 1290 tweet-reply pairs of hatespeech
16https://gab.ai/home
17https://wrongthink.net/
and the corresponding counterspeech tweets. We observe
that the counter speakers of different communities adopt dif-
ferent techniques to counter the hateful tweets. We perform
several interesting analysis on these accounts and find that
hateful accounts express more negative sentiments and are
more profane. The hateful users in our dataset seem to be
more popular as they tweet more and have more followers.
We also find that the hate tweets by verified accounts have
much more virality as compared to a tweet by a non-verified
account. While the hate users seem to use words more about
envy, hate, negative emotion, swearing terms, ugliness, the
counter users use more words related to government, law,
leader. We also build a supervised model for classifying the
hateful and counterspeech accounts on Twitter and obtain an
F-score of 0.77. Our work should be useful in appropriately
designing incentive mechanisms to make Twitter-like plat-
forms free of hate content.
References
2015.
J.,
for
2016a.
successful counterspeech.
[Badjatiya et al. 2017] Badjatiya, P.; Gupta, S.; Gupta, M.; and
Varma, V. 2017. Deep learning for hate speech detection in tweets.
WWW, 759 -- 760.
and
[Bartlett and Krasodomski-Jones 2015] Bartlett,
Krasodomski-Jones, A.
Counter-speech examining
content that challenges extremism online. Demos. Available at:
http://www.demos.co.uk/wp-content/uploads/2015/10/Counter-
speech.pdf.
[Benesch et al. 2016a] Benesch, S.; Ruths, D.; Dillon, K. P.;
Saleem, H. M.; and Wright, L.
Consider-
Dangerous Speech
ations
Project. Available at: https://dangerousspeech.org/considerations-
for-successful-counterspeech/.
[Benesch et al. 2016b] Benesch, S.; Ruths, D.; Dillon, K. P.;
Saleem, H. M.; and Wright, L.
2016b. Counterspeech on
twitter: A field study. Dangerous Speech Project. Available
at: https://dangerousspeech.org/counterspeech-on-twitter-a-field-
study/.
[Benesch 2014] Benesch, S. 2014. Defining and diminishing hate
speech. Freedom from hate: State of the worlds minorities and
indigenous peoples 2014 18 -- 25.
[Bhattacharya et al. 2014] Bhattacharya, P.; Zafar, M. B.; Ganguly,
N.; Ghosh, S.; and Gummadi, K. P. 2014.
Inferring user inter-
ests in the twitter social network. In Proceedings of the 8th ACM
Conference on Recommender systems, 357 -- 360. ACM.
[Chatzakou et al. 2017] Chatzakou, D.; Kourtellis, N.; Blackburn,
J.; De Cristofaro, E.; Stringhini, G.; and Vakali, A. 2017. Measur-
ing #gamergate: A tale of hate, sexism, and bullying. In WWW '17
Companion, 1285 -- 1290.
[Citron and Norton 2011] Citron, D. K., and Norton, H. 2011. In-
termediaries and hate speech: Fostering digital citizenship for our
information age. BUL Rev. 91:1435.
[Davidson et al. 2017] Davidson, T.; Warmsley, D.; Macy, M.; and
Weber, I. 2017. Automated hate speech detection and the problem
of offensive language.
[de Lima et al. 2018] de Lima, L. R. P.; Reis, J. C. S.; Melo, P. F.;
Murai, F.; Silva, L. A.; Vikatos, P.; and Benevenuto, F. 2018. Inside
the right-leaning echo chambers: Characterizing gab, an unmoder-
ated social system.
[Del Vigna et al. 2017] Del Vigna, F.; Cimino, A.; DellOrletta, F.;
[Nobata et al. 2016] Nobata, C.; Tetreault, J.; Thomas, A.; Mehdad,
Y.; and Chang, Y. 2016. Abusive language detection in online user
content. In WWW '16, 145 -- 153.
[Ribeiro et al. 2017] Ribeiro, M. H.; Calais, P. H.; Santos, Y. A.;
Almeida, V. A.; and Meira Jr, W. 2017.
" like sheep among
wolves": Characterizing hateful users on twitter. In WSDM work-
shop on Misinformation and Misbehavior Mining on the Web
(MIS2).
[Ribeiro et al. 2018] Ribeiro, M.; Calais, P.; Santos, Y.; Almeida,
V.; and Jr., W. M. 2018. Characterizing and detecting hateful users
on twitter.
[Schieb and Preuss 2016] Schieb, C., and Preuss, M. 2016. Gov-
erning hate speech by means of counterspeech on facebook.
In
66th ICA Annual Conference, At Fukuoka, Japan, 1 -- 23.
[Silva et al. 2016] Silva, L. A.; Mondal, M.; Correa, D.; Ben-
evenuto, F.; and Weber, I. 2016. Analyzing the targets of hate
in online social media. In ICWSM, 687 -- 690.
[Sorower 2010] Sorower, M. S. 2010. A literature survey on algo-
rithms for multi-label learning. Oregon State University, Corvallis
18.
[Stroud and Cox 2018] Stroud, S. R., and Cox, W. 2018. The va-
rieties of feminist counterspeech in the misogynistic online world.
In Mediating Misogyny. Springer. 293 -- 310.
[Van Spanje and De Vreese 2015] Van Spanje, J., and De Vreese,
C. 2015. The good, the bad and the voter: The impact of hate
speech prosecution of a politician on electoral support for his party.
Party Politics 21(1):115 -- 130.
[Warner and Hirschberg 2012] Warner, W., and Hirschberg,
J.
2012. Detecting hate speech on the world wide web. In Proceed-
ings of the Second Workshop on Language in Social Media, LSM
'12, 19 -- 26.
[Waseem and Hovy 2016] Waseem, Z., and Hovy, D. 2016. Hate-
ful symbols or hateful people? predictive features for hate speech
detection on twitter. In Proceedings of the NAACL student research
workshop, 88 -- 93.
[Wright et al. 2017] Wright, L.; Ruths, D.; Dillon, K. P.; Saleem,
H. M.; and Benesch, S. 2017. Vectors for counterspeech on twitter.
In Proceedings of the First Workshop on Abusive Language Online,
57 -- 62.
[Zubiaga et al. 2016] Zubiaga, A.; Liakata, M.; Procter, R.; Wong
Sak Hoi, G.; and Tolmie, P. 2016. Analysing how people orient
to and spread rumours in social media by looking at conversational
threads. PLOS ONE 11:1 -- 29.
Petrocchi, M.; and Tesconi, M. 2017. Hate me, hate me not: Hate
speech detection on facebook.
[Djuric et al. 2015] Djuric, N.; Zhou, J.; Morris, R.; Grbovic, M.;
Radosavljevic, V.; and Bhamidipati, N. 2015. Hate speech detec-
tion with comment embeddings. In WWW '15 Companion, 29 -- 30.
[ElSherief et al. 2018a] ElSherief, M.; Kulkarni, V.; Nguyen, D.;
Wang, W. Y.; and Belding, E. 2018a. Hate lingo: A target-based
linguistic analysis of hate speech in social media. ICWSM '18.
[ElSherief et al. 2018b] ElSherief, M.; Nilizadeh, S.; Nguyen, D.;
Vigna, G.; and Belding, E. 2018b. Peer to peer hate: Hate speech
instigators and their targets.
[Ernst et al. 2017] Ernst, J.; Schmitt, J. B.; Rieger, D.; Beier, A. K.;
Vorderer, P.; Bente, G.; and Roth, H.-J. 2017. Hate beneath the
counter speech? a qualitative content analysis of user comments on
youtube related to counter speech videos. Journal for Deradical-
ization 1 -- 49.
[Fast, Chen, and Bernstein 2016] Fast, E.; Chen, B.; and Bernstein,
M. S. 2016. Empath: Understanding topic signals in large-scale
text. In Proceedings of the 2016 CHI Conference on Human Fac-
tors in Computing Systems, 4647 -- 4657. ACM.
[Gagliardone et al. 2015] Gagliardone, I.; Gal, D.; Alves, T.; and
Martinez, G. 2015. Countering online hate speech. UNESCO
Publishing.
[Gelber and McNamara 2016] Gelber, K., and McNamara, L. 2016.
Evidencing the harms of hate speech. Social Identities 22(3):324 --
341.
[Ghosh et al. 2012] Ghosh, S.; Sharma, N.; Benevenuto, F.; Gan-
guly, N.; and Gummadi, K. 2012. Cognos: crowdsourcing search
for topic experts in microblogs. In Proceedings of the 35th inter-
national ACM SIGIR conference on Research and development in
information retrieval, 575 -- 590. ACM.
[Goldberg 1992] Goldberg, L. R. 1992. The development of mark-
ers for the big-five factor structure. Psychological assessment
4(1):26.
[Gou, Zhou, and Yang 2014] Gou, L.; Zhou, M. X.; and Yang, H.
2014. Knowme and shareme: Understanding automatically dis-
covered personality traits from social media and user sharing pref-
erences. CHI '14, 955 -- 964. ACM.
[Hu et al. 2016] Hu, T.; Xiao, H.; Luo, J.; and Nguyen, T.-v. T.
2016. What the language you tweet says about your occupation.
In Tenth International AAAI Conference on Web and Social Media.
[Hutto and Gilbert 2014] Hutto, C. J., and Gilbert, E. 2014. Vader:
A parsimonious rule-based model for sentiment analysis of social
media text. In ICWSM.
[Leets 2002] Leets, L. 2002. Experiencing hate speech: Percep-
tions and responses to anti-semitism and antigay speech. Journal
of social issues 58(2):341 -- 361.
[Liu et al. 2017] Liu, Z.; Xu, A.; Wang, Y.; Schoudt, J.; Mahmud,
J.; and Akkiraju, R. 2017. Does personality matter?: A study of
personality and situational effects on consumer behavior. HT '17,
185 -- 193. ACM.
[Mathew et al. 2018] Mathew, B.; Tharad, H.; Rajgaria, S.; Singha-
nia, P.; Maity, S. K.; Goyal, P.; and Mukherje, A. 2018. Thou
arXiv preprint
shalt not hate: Countering online hate speech.
arXiv:1808.04409.
[Mondal, Silva, and Benevenuto 2017] Mondal, M.; Silva, L. A.;
and Benevenuto, F. 2017. A measurement study of hate speech
in social media. In HT.
Tweetment effects on the
[Munger 2017] Munger, K.
tweeted: Experimentally reducing racist harassment. Political Be-
havior 39(3):629 -- 649.
2017.
|
1710.10642 | 1 | 1710 | 2017-10-29T16:26:32 | Homophily of Music Listening in Online Social Networks | [
"cs.SI",
"physics.soc-ph"
] | Homophily, ranging from demographics to sentiments, breeds connections in social networks, either offline or online. However, with the prosperous growth of music streaming service, whether homophily exists in online music listening remains unclear. In this study, two online social networks of a same group of active users are established respectively in Netease Music and Weibo. Through presented multiple similarity measures, it is evidently demonstrated that homophily does exist in music listening of both online social networks. The unexpected music similarity in Weibo also implies that knowledge from generic social networks can be confidently transfered to domain-oriented networks for context enrichment and algorithm enhancement. Comprehensive factors that might function in formation of homophily are further probed and many interesting patterns are profoundly revealed. It is found that female friends are more homogeneous in music listening and positive and energetic songs significantly pull users close. Our methodology and findings would shed lights on realistic applications in online music services. | cs.SI | cs |
Homophily of Music Listening in Online Social Networks
Zhenkun Zhou, Ke Xu
State Key Lab of Software Development Environment, Beihang University, Beijing, China
Jichang Zhao∗
School of Economics and Management, Beihang University, Beijing, China
∗Corresponding author: [email protected]
Abstract
Homophily, ranging from demographics to sentiments, breeds connections in
social networks, either offline or online. However, with the prosperous growth
of music streaming service, whether homophily exists in online music listening
remains unclear. In this study, two online social networks of a same group of
active users are established respectively in Netease Music and Weibo. Through
presented multiple similarity measures, it is evidently demonstrated that ho-
mophily does exist in music listening of both online social networks. The un-
expected music similarity in Weibo also implies that knowledge from generic
social networks can be confidently transfered to domain-oriented networks for
context enrichment and algorithm enhancement. Comprehensive factors that
might function in formation of homophily are further probed and many inter-
esting patterns are profoundly revealed. It is found that female friends are more
homogeneous in music listening and positive and energetic songs significantly
pull users close. Our methodology and findings would shed lights on realistic
applications in online music services.
Keywords: Online Social Networks, Homophily, Music Listening, Music
Genres
Preprint submitted to Elsevier
October 31, 2017
1. Introduction
One of the best established findings in social networks is that people who
are friends exhibit plenty of similarities in human behaviors [1, 2]. Friendship
relations, either offline or online, in which individuals socially interact, involve a
need for shared mutual understandings. Tremendous efforts have been devoted
on homophily of social networks from many aspects, ranging from demograph-
ics [3] to mental states [4, 5]. Even it is revealed from the recent study that
personality similarity exists among close relationships [6]. Particularly, with
the booming of online social media, previous studies have extensively inves-
tigated the homophily of behaviors in online social networks [7, 8, 9] and the
friend similarity that exists in face-to-face offline networks is identically revealed.
However, as a prominent element of daily life that possesses the cultural uni-
versality [10, 11, 12], music listening is rarely explored in the context of social
networks and little knowledge is established about the behavior referring to
homophily, especially in the circumstance of online social networks.
In fact, individuals that embedded in social networks come across music of
varying categories, including vast kinds of genres, languages and moods, and
continually judge whether or not like the music [13]. In addition, music is al-
ways shared with families, friends and other folks socially around us. Before
times of the Internet, CDs and cassettes were the main media to record the
music and the music communication was thus limited and awkward [14]. Yet
in last twenty years, with the prosperous development of the Internet, portable
music players have exploded in popularity, promoting the music communication
essentially [15, 16]. Friends are willing to exchange the iPods from each other.
And now, the music streaming platforms offer low-latency access to the large-
scale database, such as Spotify [17], Last.fm [18], QQ Music [19] and Netease
Music [20]. Since then, people exchange the music with each other online and
share the amazing music they respectively like, and music even creates inter-
personal bonds between different individuals in turn [21]. Though evidence of
music similarity in offline friendships has been demonstrated [22], the relation-
2
ship between listening similarity and online social networks has not been com-
prehensively explored and understood yet. Specifically, questions like whether
we enjoy the music at which the online friends are enchanted, in other words,
whether the homophily of music listening exists in online social networks, still
deserve a systemic investigation.
Until recently, empirical research willing to answer the questions about mu-
sic listening had to depend on interviews and surveys in controlled environ-
ments [23, 24] with inevitable limitations in both data scale and granularity.
While in fact, the technological and societal evolutions that sustain the emer-
gence of online music listening indeed provide unparalleled opportunities for
human behavior understanding. Detailed footprints, including where, when
and how massive individuals listen music can be regarded as a big-data window
and through which the homophily in music listening can be collectively or in-
dividually probed and studied thoroughly. For instance, Netease Music, one of
the most popular online music providers in China, contributes the high-quality
music streaming service to millions of users and accordingly accumulate the
detailed behavior records of these users continuously. According to the official
report of Netease Music, its active users is over 200 million. With music playlists
creation being the core listening pattern, each day users establish around 420
thousand playlists and user-generated playlists total 800 million.
In the half
of 2016, users play songs 1.82 billion times and the duration amounts to 7.2
billion minutes, implying the impressive vitality of users in online music listen-
ing. Even more inspiring, Netease Music develops one extraordinary trait of
socializing its users. Specifically, it firstly provides a domain-specialized social
network through which users engage in sharing interests of music. Like generic
online social networks, users can be networked through following others, not
only ordinary users but also artists. Indeed, the social network sourced in mu-
sic listening profoundly facilitates the acquiring and sharing of music interests,
implying an ideal entanglement between social ties and music listening for the
present study. Thus anonymous digital traces of massive users are collected to
quantitatively support the investigation of homophily in music listening.
3
However, domain-oriented social networks like the one established by Netease
Music can not be a typical representation of online social networks that are gen-
erally resulted by comprehensive factors and music might be just one of them.
Specifically, Netease music social network relies predominately on the musical
interest and its digital traces are insufficient to describe other individual traits.
In the meantime, evidence of musical preferences being linked to individual
traits like personalities [24], cognitive styles [25] and even socioeconomic sta-
tuses [26] has been extensively demonstrated, implying the consideration more
generic online social networks. We argue that aiming at a systemic understand-
ing, it is necessary to study listening practices based on other more typical social
networks in which users are jointed sophisticatedly but realistically by psycho-
logical traits, extensive interests or other individual characteristics. Considering
the booming of social media in recent decades, prosperous networks like Twitter
or its variant Weibo of China that aggressively replicate offline social networks
to online counterparts, can be ideal targets. These online social networks are
natural, long-term and diverse and thus the objective footprints of massive in-
dividuals can be promising proxies for the present study. Nonetheless, it is
still extremely difficult to correctly match each individual of the music social
network to the identical one in networks like Weibo, which is the essence of em-
bedding music listening into a generic online social network. Very fortunately,
users can log into Netease Music through their Weibo accounts and along this
line we can obtain the digital traces of Weibo for these users. Therefore, it is
possible to further study the similarity of music listening for friends that linked
in Weibo. In addition, demographics and tweets in Weibo are excellent supple-
ments for enriching individual characteristics. We can even explore which key
factors influence the similarity of music listening.
Starting from the above motivations and assumptions, in this study, dig-
ital footprints of over seventy thousand individuals from both Netease Music
and Weibo are thoroughly collected through a novel crawler. Then 25, 953 ac-
tive users with plenty of times for listening and tweeting are sampled as the
subjectives for further explorations. Two online social networks, the Netease
4
network and Weibo network are respectively constructed through user follow-
ings in Netease Music and Weibo. To examine whether the music listening is
homogeneous for users jointed by online social networks, similarities from six
perspectives are defined and measured. In order to investigate the crucial fac-
tors influencing the music listening homophily, subjectives are clustered into
different categories from multiple perspectives like social attributes, musical
preferences or etc. Our results demonstrate that friends linked by online social
networks indeed appreciate the identical songs and possess the similar music
preferences. As for gender, patterns for music listening between female friends
are closer than male friends, implying that women are more sensitive for emo-
tional expression through music [27, 28]. Listening practices of the friends with
common music preferences in the current culture are similar especially for music
languages (Chinese) and genres (pop and folk). With regard to musical mood,
the users who enjoy exciting, wild and happy music share more similarities in
music listening.
It is also difficult for users owning high musical diversity to
find friends with similar musical tastes. The present study confirms evidently
the existing of homophily in music listening of online social networks and elab-
orately clarify the roles of human demographics and music traits in influencing
the homophily. Our findings would shed insightful lights on music recommen-
dation and friend suggestion in online applications. We merge different social
circles of an individual and surprisingly reveal that generic social networks like
Weibo still demonstrate significantly the homophily that intuitively only exists
in domain-oriented networks. It indeed implies that rich information in general
social media can be confidently introduced into the study of specialized social
networks [29].
2. Methods
2.1. Netease Music Dataset
Digital service providers began to amass large user bases, offering the pri-
mary sources of digital music streaming via the Internet increasingly. New
5
innovations, including digitalization and the Internet, transform the existing
landscape over the past decade and attracted new artists and listeners into the
fields, and digital music streaming services has also profoundly reshaped the
user behaviors. China follows this global trend and has become a leading dig-
ital country in terms of music services. The Netease Music is one of the most
trending music streaming providers in China, whose special trait is the music so-
cial network. Users are allowed to follow others in the platform, such as friends
and artists. Therefore, this music platform has become a new and booming
social network of domain-oriented.
The web site of Netease Music provides abundant online information about
users, playlists and music. We develop an excellent crawler aiming at Netease
Music platform first to perform the data collection. Our crawler adapts massive
agent servers and each agent will simulate a real user to visit pages and click
links from the browser. By traversing the following links, from August in 2016
to May in 2017, we obtain a dataset of over 200 thousand users, 1.5 million
playlists and 3.2 million songs.
The collected data referring to users include the following lists, historical
listening and like records. For each user, the following list contains the complete
ID numbers of users (UID) that she/he followed. The top 100 songs listened
by one user are recorded in her/his historical listening. The platform provides
the 'like' button to label their favorite playlists or songs which are restored in
like records. According to the historical listening and like records, we obtain
the corresponding detailed playlists and attributes of songs. In each playlist,
ID numbers of songs (SID) it contains and style tags labeled by its creator are
both collected. The attributes of a song include the album, artists and several
acoustic features.
2.2. Weibo Dataset
Tremendous social relationships are forged in popular online services like
Facebook, Twitter and Weibo with the explosive development of online social
networks. Twitter, as the most popular social networking and micro-blog ser-
6
vice, enables registered users to read and post short messages, so-called tweets.
At the beginning of 2016, Twitter had reached 310 monthly active users (MAU).
As of the third quarter of 2017, it averaged at 330 million MAU. However, more
people are using Weibo, the Chinese variant of Twitter now. According to the
Chinese company's first quarter reports, it has 340 million MAU, 30% up on
the previous year, implying that social relationships and online behaviors of
massive individuals from Weibo can be sensed and profiled with unparalleled
richness and granularity. Moreover, replicating offline social ties to online and
replacing face-to-face communication with interactions of cyberspace have es-
sentially digitized the daily life and reshaped the social networks, suggesting
that the online social network is of fundamental significance for explorations in
both social theories and applications.
In the fore part mentioning Netease Music Dataset, 74,056 users connect
their Weibo accounts to Netease Music, offering us an opportunity to establish
a perfect match between these two different platforms. Employing the crawler
agents, their corresponding ID numbers in Weibo are obtained. With the help
of Weibo's open APIs, for each ID, the public profile, following relationships
and historical tweets can be accordingly collected. And in order to speed up the
the collection, cloud servers are also extensively utilized. The profiles of Weibo
users include demographics (e.g. gender), social attributes (e.g. the number of
followers) and other individual information. For each Weibo user, the following
relations contains the Weibo ID numbers (WID) of users it follows. Unfortu-
nately, due to the official limit on viewing following list, we only obtain the
latest 1,000 following relations. Nevertheless, according to theory of Dunbar's
number, the maximum size of ego-network averages 150 [30, 31, 32], implying
that the latest one thousand followings could sufficiently reflect recently active
social connections the user possesses. Similarly, due the limit on tweets viewing,
for each Weibo user only its latest 1000 (or less if not enough) tweets can be
obtained.
7
2.3. Two Social Networks
We further refine the subjects of the present study by selecting active users
from both datasets. Specifically, active users are defined as those have listened
complete songs at least 1000 times and posted over 100 tweets. The complemen-
tary cumulative distribution function (CCDF) of the number of users in music
listening and tweeting is depicted in Fig 1. As can be seen in Fig 1a, for the
majority of users, the number of music listening is lower than 104. The percent
of users that listen music over 1000 times is lower than 55%. As shown in Fig 1b,
about 70% of users tweet 100 times at least. Note that owing to the limit of
Weibo, the maximum of number of tweets is 1000. After the refinement, totally
25,953 active users are filtered out to be subjectives of further explorations.
Regarding to the same group of active users, two social networks are then
respectively established from the following relationships in Netease Music and
Weibo. The first one, named Netease network, contains 89,988 links. The sec-
ond one, named Weibo network, contains 112,753 links. And these two networks
share the 25,953 identical nodes, implying that we construct an ideal scenario to
reveal the homophily of music listening in both domain-oriented social network
and generic social network. It is worthy noting that here we regard social net-
works as undirected graphs. The degree distributions of Netease network and
Weibo network are demonstrated in Fig 2 and the power-law trending implies
the consistence to existing understanding of typical social networks [33]. As for
Netease network (γ = 2.12), the distribution evidently shows a long tail, indi-
cating there are few users with large amount of relationships in social network
of music-oriented. As for Weibo network (γ = 2.07), a power-law like distribu-
tion is observed as degree lower than 500, however, a exponential cutoff then
emerges and the maximum degree is 100, both due to the API limit of Weibo.
2.4. Similarities of music listening
Based on the personal historical listening and like records, we define six
types of similarities with respect to music listening for any pair of users A and
8
(a)
(b)
Figure 1: CCDF of the number of music listening and tweeting.
9
100101102103104Number of music listening0.00.20.40.60.81.0P(X>x)100101102103Number of tweets0.20.40.60.81.0P(X>x) (a)
(b)
Figure 2: Probability distribution of the degree of Weibo network and Netease Music network.
The degree exponent γ of networks are respectively 2.12 and 2.07 (2 < γ < 3).
10
101102103104Degree of Netease Music user104103102101Probability=2.12101102103Degree of Weibo user104103102101Probability=2.07 B in both social networks. And higher similarities suggest more significant
homophily in social networks.
The historical listening records for A and B are denoted as HA and HB,
respectively and in which top 100 songs of high-frequency are included. The
historical similarity simsong, indicating the number of co-occurrence songs in
top records can be thus defined as
simsong = HA ∩ HB
.
(1)
For presenting measures of similarity from aspects of music traits, we first
develop ways to label songs and users in terms of traits vectors. Given the
missing of tags of songs in Netease dataset, the tags of playlists can be fused
to infer traits vectors of songs. Specifically, Netease officially provides 6 tags of
languages, 24 tags of genres, 13 tags of moods and 12 tags of scenes and among
these 55 tags at most three are selected to label a playlist by its creator. Assum-
ing that a song can be labeled confidently by tags of the playlists it belongs to,
a 55-dimension traits vector (initialized to be zero) can be accordingly derived
for each song by examining 1.5 million playlists, i.e., a tag occurrence will result
in addition of 1 to the corresponding dimension. However, considering the fact
that values of the traits vector of a song can be significantly influenced by the
number of playlists it appears, we split the long vector into four sub-vectors
including Slanguage, Sgenre, Smood and Sscene (S refers to songs) and then do
normalization on each sub-vector to avoid bias. To be specific, we calculate
the proportion of each tag respectively in four vectors and the sum of values in
each sub-vector should be 1. In fact, Slanguage, Sgenre, Smood and Sscene can be
features to appropriately represent the category distribution of songs. Following
the same idea, supposing that music preferences of a user can be well reflected
by the top songs listened, the traits vectors of users can be inferred through
100(cid:88)
Uc =
Sc,i,
(2)
in which c stands for traits and can be language, genre, mood or scene. Finally,
based on the feature vectors of 3.2 million songs, traits vectors for 25,953 active
i=1
11
users are derived and normalized to z-cores. Then the similarity of user pair A
and B can be intuitively measured through cosine distance between their traits
vectors, which is defined as
simc =
Uc,A · Uc,B
Uc,A Uc,B .
(3)
Note that −1 (cid:54) simc (cid:54) 1 and values closer to 1 indicate more similar music
preferences. And as c varies over different music traits, we accordingly obtain
four music preference similarities from perspectives of languages, genres, moods
and scenes.
In addition, the similarity of favorite songs can also reflect the strength of
homophily in music preference. Different from historical listening records, the
number of songs in like record is not limited. Therefore we define the Jaccard
distance as the similarity for like records between users A and B as
simlike =
LA ∩ LB
LA ∪ LB
(4)
in which L stands for the set of one's favorite songs.
To sum up, from the above we obtain six similarities referring to music
listening and higher similarities indicate more significant homophily in music
listening of online social networks. These measures pave the way for quantitative
investigations of homophily in this study.
2.5. User classifications
Many factors in social networks might influence the homophily between
friends. In order to reveal the crucial factors, demographics, social attributes
and listening behaviors of active users are further probed for detailed under-
standing of their roles in homophily of music listening. We argue that regarding
these factors as features, active users can be thus clustered into groups from
multiple perspectives and discriminations of inter-groups offer windows of ho-
mophily investigation and factor weighing.
Gender is one of the most significant demographics in understanding human
behaviors. We extract gender from Weibo profiles and split active users into
12
10,388 female ones and 15,565 male ones. Whether a user is officially verified
by Weibo can be an indicator of influence and through the 'verified' labels
active users are split into 24,368 non-verified ones and 1,585 verified ones. In
addition, social attributes that reflect users' ranks in the social network can be
well modeled through the number of the followers (N ER) as well as RF F that
defined as the rate of f ollowing numbers to f ollowed numbers (log( N ER+1
N EE+1 ),
in which N EE refers to the number of the users' followees). Given the fact
that both N ER and RF F are continuous variables, the discretization based on
K-means clustering is employed to categorize active users into groups of low
rank and high rank respectively based on these two attributes.
Patterns of online music consumption could be sufficiently reflected in terms
of users' music preferences that result in behavioral differences directly. Hence
active users can also be clustered into groups from the perspective of music
preference. The traits vector Uc can be features representing users in clustering
and the preference feature matrix of c is accordingly constructed for all users.
The matrix is de-correlated by principal component analysis (PCA) with full
covariance and then the approach of K-means is employed to cluster active users
into groups. Clusterings based on all traits are denoted separately as Cmood,
Clanguage and Cgenre, indicating that for each trait active users are grouped into
three clusters with divergent patterns of music preferences. In order to better
interpret the clusterings, for groups of each clustering, the group feature defined
as the mean user feature within the group is calculated. Then the semantic of
each group is explained by the top 3 attributes of the group feature. Table 1
reports the clusterings and group interpretations. As can be seen, groups of
each clustering could be well explained by the top attributes, suggesting that
our methods indeed capture the music preferences and behavioral patterns in
music listening can be effectively detected.
In the meantime, as exploited in previous efforts [34, 26], diversity of mu-
sical preferences is also an indicator of great significance in reflecting listening
behaviors. We draw on the concept of entropy [35] to define users' diversity in
13
Table 1: User classifications from perspectives about music preferences.
Traits
Groups
Interpretations
language
genre
mood
0
1
2
0
1
2
0
1
2
Chinese (including Cantonese)
Japanese, Korean
English, minority
rap, dance, alternative
pop, folk
light, New age, classic,
exciting, wild, happy
sad, missing, lonely
fresh, sanative, easy
music preferences as
3(cid:88)
n(cid:88)
c=1
i=1
div = − 1
3
p(uc,i)log(p(uc,i)),
(5)
in which p(uc,i) refers to the proportion of i-th attribute in user vector of trait c.
Higher diversity implies the user's more appreciations for variety of experience
in music listening. We then employ K-means to cluster active users into high-
diversity and low-diversity groups respectively.
The above user classifications can be finally utilized to define the homophily
of relationships in online social networks. According to eight groups of all the
clusterings, these are totally 19 factors1. As the dummy variable, the value of
factor is 1 when the two linked users are classified to the same group, otherwise
the value is 0. For example, Factors f emale is 1 and male is 0 when user
A (female) is linked with user B (female).
If f emale and male are both 0,
1The factors range from demographics to music traits and include f emale, male, V 0, V 1,
N ER0, N ER1, RF F 0, RF F 1, lang0, lang1, lang2, genre0, genre1, genre2, mood0, mood1,
mood2, div0 and div1.
14
Figure 3: Distribution of similarities of listening historical records. Through the two-sample
Kolmogorov-Smirnov test, except for the two random networks (p = 0.99), the distributions
of other pairs have significant differences (p < 0.001,∗∗∗).
indicating that A and B possess different genders.
3. Results
3.1. Existence of homophily
Being direct indicators of homophily, similarities from different perspectives
are first investigated in the two social networks constructed from Netease and
Weibo. Meanwhile, in order to testify the significance of the similarity distribu-
tion, for each realistic network, in terms of shuffling nodes a random counterpart
is also built for base lines.
The CCDF of simsong, as can be seen in Fig. 3, demonstrates that similarities
of actual networks are evidently higher than that of random networks, especially
when simsong > 5, indicating that friends in online social network inclines
to listen same songs and the homophily in music listening significantly exists.
Particularly, as compared to Weibo network, friendships in Netease network
15
0.02.55.07.510.012.515.017.520.0simsong104103102101100CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork (a)
(b)
(c)
(d)
Figure 4: Distribution of similarities of musical preferences. Through the two-sample
Kolmogorov-Smirnov test, except for the two random networks (p > 0.05), the distributions
of other pairs have significant differences (p < 0.001,∗∗∗).
are slightly closer, implying that from the perspective of listening same songs,
homophily is enhanced in the music-oriented social network.
CCDFs of similarities in perspectives of music preferences are further in-
vestigated, as can be seen in Fig 4. Similar to observations of simsong, it is
consistently demonstrated that the friend similarity in music preferences is sig-
nificant, suggesting the existing of homophily in music listening of online social
network. Even more interesting, homophily in Netease network is more evident
than that of Weibo network, implying again that friends are more homogeneous
in domain-oriented social networks than the generic counterparts. However, as
for the measure of simlike that shown in Fig 5, there is no evidence for the ex-
isting of homophily because the similarity of friends in 'like' behavior of online
networks are close to or even lower than that in random networks.
16
1.000.750.500.250.000.250.500.751.00simlanguage0.00.20.40.60.81.0CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork1.000.750.500.250.000.250.500.751.00simgenre0.00.20.40.60.81.0CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork1.000.750.500.250.000.250.500.751.00simmood0.00.20.40.60.81.0CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork1.000.750.500.250.000.250.500.751.00simscene0.00.20.40.60.81.0CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork Figure 5: Distribution of similarities of favorite songs between users. Through the two-sample
Kolmogorov-Smirnov test, except for the two random networks (p = 0.163), the distributions
of other pairs have significant differences (p < 0.001,∗∗∗).
To sum up, in terms of similarity metrics except the one regarding to 'like'
behavior, homophily of music listening in online social networks, especially the
one of music-oriented, is significantly demonstrated and confirmed. Friends con-
nected in online social networks indeed enjoy identical music and share similar
music preferences. The absence of homophily in 'like' behavior suggests that
even in homogeneous online social networks, individual differences still exist
across friends. Although the homophily in specialized network is slightly more
significant, we indeed find the same behavior pattern in generic networks like
Weibo where the relationships are bonded with more common interests. It fur-
ther implies that social media like Weibo can be generic but typical instances
of online social networks. We further explore the important factors which affect
the listening similarity in the Weibo network.
3.2. Critical factors for homophily
Given the rich backgrounds of active users carried by Weibo, we further
investigate how characteristics of social ties, i.e., factors influence the homophily
17
0.000.010.020.030.040.05simlike102101100CCDFWeiboNetworkNeteaseNetworkRandomWeiboNetworkRandomNeteaseNetwork Table 2: Results of multiple regression to predict simsong based on user classifications. The
column of "Coef." (coefficient) shows the influence of factors on homophily. Note: ∗∗p <
.01,∗∗∗ p < .001.
Traits
Factors
Coef.
Std. Error
t-Value P r(> t )
gender
verified
NER
RFF
female
male
V0
V1
NER0
NER1
RFF0
RFF1
lang0
language
lang1
lang2
genre0
genre1
genre2
mood0
mood1
mood2
div0
div1
genre
mood
diversity
0.369
0.254
0.119
0.014
0.444
-0.149
0.517
0.036
1.047
0.569
0.228
-0.023
1.043
0.061
0.668
0.359
0.126
0.770
0.208
0.021
0.015
0.017
0.035
0.042
0.018
0.031
0.026
0.026
0.036
0.020
0.022
0.020
0.041
0.024
0.027
0.023
0.015
0.023
17.25
16.91
6.91
0.41
10.58
-8.49
16.72
1.35
40.43
15.79
11.28
-1.05
50.90
1.51
27.71
13.32
5.35
50.59
8.89
Intercept
0.396
0.019
20.50
Observations
112,753
R2
0.1701
Adjusted R2
0.1699
F-Statistic
1100
18
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
disclosed in music listening. Except for simsong, other similarities referring to
music preference are measured with Uc (traits vectors of users), however, the
vectors are also used to cluster the individuals. It could take a mistake when
the dependent variable and some independent variables are calculated from the
same source. Therefore, we only consider simsong as the dependent variable to
predict in regression analysis. As can be seen, Table 2 presents the multiple
regression results of nineteen factors function on the homophily.
As for gender, the coefficients of male are higher than that of female, indi-
cating that relationships in online social network between female users are more
homogeneous in music listening. Yet the correlations between listening similar-
ities and the verification statuses are not significant due to trivial coefficients.
For social attributes N ER0 and RF F 0 that reflecting individual extraversion
and openness [36, 37], coefficients are significantly higher than that of N ER1
and RF F 1, implying that users of low extroversion and openness demonstrate
more homophily in music listening.
For factors from perspective of music preferences, the positive coefficients
shown in Table 2 indicate that users within the same group of clusterings re-
ferring to preferences of language and mood are similar. As for the language
preference, users in the group of Chinese are more similar than those in other
groups. Pop music and folk music are known as the mainstream genres in China.
As for genres, it can be found that the users following pop and folk genre are
significantly more similar. According to the results for languages and genres,
we suggest that the friends in online social networks, whose music preferences
accord with the mass taste, are always listening to the similar music. From the
perspective of moods, the coefficients of the factor 'exciting, wild and happy'
is the highest for simsong. It demonstrates that people who like the positive
and energetic songs are more homogeneous in music listening. In addition, we
investigate whether the diversity (div0 and div1) of music consumption is an
influential factor for listening similarity. The results show that users with low
musical diversity are more similar. However, users that enjoys multiple types of
music, always own the uncommon and unique preferences and it is difficult for
19
them to find friends sharing alike musical preferences.
4. Discussion
The booming of social media greatly facilitates exploitations of social net-
works by offering an unprecedented big-data window.
In the present study,
aiming at understanding homophily of music listening in online social networks,
two networks with a same group of active users from Netease Music and Weibo
are separately built for probing six measures of homophily. It is confirmed sig-
nificantly that homophily exists in music listening of online social networks,
even for the generic one from Weibo. Factors of social ties ranging from demo-
graphics to music preferences are further investigated in influencing homophily
and many interesting patterns are revealed. To our best knowledge, for the first
time a systematic and comprehensive study of music homophily is performed in
a manner of data-driven solution.
In the meantime, we are also the first to construct a big music data through
matching active users in both music stream platform and Weibo. The unex-
pected existing of homophily in Weibo suggests that the knowledge from generic
social networks can be confidently introduced and transfered to the study [38]
and application [29] of domain-oriented social networks, which will essentially
enrich the context of subjectives. Previous investigation shows that human av-
eragely spend 17% of life in music listening [39]. Nevertheless, massive behavior
data is still missing in existing studies. By combing behavioral and demograph-
ical data from different online social networks, our study demonstrates that
explorations of music listening can be profoundly boosted.
In fact, music steaming platforms have been experiencing a fast growing
stage with increasing influence through establishment of musical environments
tailored to individual preferences. Plenty of music service providers use the
relationships between online friends to estimate users' musical tastes and then
recommend music [40]. In our study, we provide a solid evidence for the ho-
mophily of music listening both in specialized online social network and the
20
generic network, suggesting that no matter specialized or generic social net-
works can serve as high quality sources of music recommendation [41, 42]. We
argue that the homophily should be exploited in building and enhancing music
services. Specifically, the influential factors that function on listening similarities
can be directly introduced into the design of music recommendation algorithms
and coefficients we obtained can be principal weights of those factors. Tak-
ing gender as an example, music platforms could put more emphasis on female
friends when recommend music to female users.
Acknowledgments
This work was supported by NSFC (Grant No. 71501005) and the fund of the
State Key Lab of Software Development Environment (Grant Nos. SKLSDE-
2015ZX-05 and SKLSDE-2015ZX-28).
References
References
[1] M. Mcpherson, L. Smithlovin, J. M. Cook, Birds of a feather: Homophily
in social networks, Annual Review of Sociology 27 (1) (2001) 415–444.
[2] M. De Klepper, E. Sleebos, G. Van de Bunt, F. Agneessens, Similarity in
friendship networks: Selection or influence? the effect of constraining con-
texts and non-visible individual attributes, Social Networks 32 (1) (2010)
82–90.
[3] A. Chmiel, P. Sobkowicz, J. Sienkiewicz, G. Paltoglou, K. Buckley, M. Thel-
wall, J. A. Ho(cid:32)lyst, Negative emotions boost user activity at bbc forum,
Physica A: statistical mechanics and its applications 390 (16) (2011) 2936–
2944.
[4] J. Bollen, B. Gon¸calves, G. Ruan, H. Mao, Happiness is assortative in
online social networks, Artificial life 17 (3) (2011) 237–251.
21
[5] R. Fan, J. Zhao, Y. Chen, K. Xu, Anger is more influential than joy:
Sentiment correlation in weibo, PloS one 9 (10) (2014) e110184.
[6] W. Youyou, D. Stillwell, H. A. Schwartz, M. Kosinski, Birds of a feather do
flock together: Behavior-based personality-assessment method reveals per-
sonality similarity among couples and friends, Psychological science 28 (3)
(2017) 276–284.
[7] J. Brown, A. J. Broderick, N. Lee, Word of mouth communication within
online communities: Conceptualizing the online social network, Journal of
interactive marketing 21 (3) (2007) 2–20.
[8] A. Mislove, B. Viswanath, K. P. Gummadi, P. Druschel, You are who you
know: inferring user profiles in online social networks, in: Proceedings of
the third ACM international conference on Web search and data mining,
ACM, 2010, pp. 251–260.
[9] D. Centola, The spread of behavior in an online social network experiment,
science 329 (5996) (2010) 1194–1197.
[10] J. Blacking, Music, culture, and experience: Selected papers of John Black-
ing, University of Chicago Press, 1995.
[11] A. C. North, Uses of music in everyday life, Music Perception An Interdis-
ciplinary Journal 22 (1) (2004) 41–77.
[12] C. Wright, Listening to music, Nelson Education, 2013.
[13] S. Frith, Music and everyday life, Critical Quarterly 44 (1) (2002) 35–48.
[14] T. Bostrom, S. Eliasson, P. Lindtorp, F. Moioli, M. Nystrom, Mobile audio
distribution, Personal Technologies 3 (4) (1999) 166–172.
[15] L. E. Holmquist, Ubiquitous music, Interactions 12 (4) (2005) 71.
[16] D. B. Rondeau, For mobile applications, branding is experience, Commu-
nications of the Acm 48 (7) (2005) 61–66.
22
[17] G. Kreitz, F. Niemela, Spotify–large scale,
low latency, p2p music-on-
demand streaming, in: Peer-to-Peer Computing (P2P), 2010 IEEE Tenth
International Conference on, IEEE, 2010, pp. 1–10.
[18] V. Henning, J. Reichelt, Mendeley-a last. fm for research?, in: eScience,
2008. eScience'08. IEEE Fourth International Conference on, IEEE, 2008,
pp. 327–328.
[19] E. Priest, The future of music and film piracy in china, Berkeley Technology
Law Journal (2006) 795–871.
[20] A. Y. Fung, The emerging (national) popular music culture in china, Inter-
Asia Cultural Studies 8 (3) (2007) 425–437.
[21] D. Boer, R. Fischer, M. Strack, M. H. Bond, E. Lo, J. Lam, How shared
preferences in music create bonds between people: Values as the missing
link, Personality and Social Psychology Bulletin 37 (9) (2011) 1159–1171.
[22] M. H. Selfhout, S. J. Branje, T. F. ter Bogt, W. H. Meeus, The role of
music preferences in early adolescents' friendship formation and stability.,
Journal of Adolescence 32 (1) (2009) 95–107.
[23] J. A. Sloboda, Everyday uses of music listening: A preliminary study,
Music, mind and science (1999) 354–369.
[24] D. M. Greenberg, M. Kosinski, D. J. Stillwell, B. L. Monteiro, D. J. Levitin,
P. J. Rentfrow, The song is you: Preferences for musical attribute dimen-
sions reflect personality, Social Psychological and Personality Science 7 (6)
(2016) 597–605.
[25] D. M. Greenberg, S. Baron-Cohen, D. J. Stillwell, M. Kosinski, P. J. Rent-
frow, Musical preferences are linked to cognitive styles, PloS one 10 (7)
(2015) e0131151.
[26] M. Park, I. Weber, M. Naaman, S. Vieweg, Understanding musical diversity
via online social media., in: ICWSM, 2015, pp. 308–317.
23
[27] A. Wells, E. A. Hakanen, The emotional use of popular music by adoles-
cents, Journalism Quarterly 68 (3) (1991) 445–454.
[28] C. Robazza, C. Macaluso, V. D'Urso, Emotional reactions to music by
gender, age, and expertise, Perceptual and Motor skills 79 (2) (1994) 939–
944.
[29] F. Carmagnola, F. Cena, User identification for cross-system personalisa-
tion, Information Sciences 179 (1) (2009) 16–32.
[30] R. Dunbar, Neocortex size as a constraint on group size in primates, Journal
of Human Evolution 22 (6) (1992) 469 – 493.
[31] R. A. Hill, R. I. Dunbar, Social network size in humans, Human nature
14 (1) (2003) 53–72.
[32] J. Zhao, J. Wu, G. Liu, D. Tao, K. Xu, C. Liu, Being rational or aggressive?
a revisit to dunbars number in online social networks, Neurocomputing
142 (Supplement C) (2014) 343 – 353.
[33] M. Newman, Networks: an introduction, Oxford university press, 2010.
[34] K. Van Eijck, Social differentiation in musical taste patterns, Social forces
79 (3) (2001) 1163–1185.
[35] P. J. Alexander, Entropy and popular culture: product diversity in the
popular music recording industry, American Sociological Review 61 (1)
(1996) 171–174.
[36] Y. Amichai-Hamburger, G. Vinitzky, Social network use and personality,
Computers in Human Behavior 26 (6) (2010) 1289–1295.
[37] Y. Bachrach, M. Kosinski, T. Graepel, P. Kohli, D. Stillwell, Personality
and patterns of facebook usage, in: Acm Web Science Conference, 2012,
pp. 24–32.
24
[38] A. Malhotra, L. Totti, W. Meira Jr, P. Kumaraguru, V. Almeida, Study-
ing user footprints in different online social networks, in: Advances in So-
cial Networks Analysis and Mining (ASONAM), 2012 IEEE/ACM Inter-
national Conference on, IEEE, 2012, pp. 1065–1070.
[39] P. J. Rentfrow, The role of music in everyday life: Current directions in
the social psychology of music, Social and personality psychology compass
6 (5) (2012) 402–416.
[40] C. Zhong, N. Sastry, Systems applications of social networks, ACM Com-
put. Surv. 50 (5) (2017) 63:1–63:42. doi:10.1145/3092742.
[41] O. Simek, D. Jensen, Navigating networks by using homophily and degree,
Proceedings of the National Academy of Sciences of the United States of
America 105 (35) (2008) 12758.
[42] J. Bu, S. Tan, C. Chen, C. Wang, H. Wu, L. Zhang, X. He, Music recom-
mendation by unified hypergraph:combining social media information and
music content, in: ACM International Conference on Multimedia, 2010,
pp. 391–400.
25
|
1905.08067 | 1 | 1905 | 2019-05-15T17:14:34 | Understanding the Radical Mind: Identifying Signals to Detect Extremist Content on Twitter | [
"cs.SI",
"cs.CL",
"cs.CY",
"cs.LG",
"stat.ML"
] | The Internet and, in particular, Online Social Networks have changed the way that terrorist and extremist groups can influence and radicalise individuals. Recent reports show that the mode of operation of these groups starts by exposing a wide audience to extremist material online, before migrating them to less open online platforms for further radicalization. Thus, identifying radical content online is crucial to limit the reach and spread of the extremist narrative. In this paper, our aim is to identify measures to automatically detect radical content in social media. We identify several signals, including textual, psychological and behavioural, that together allow for the classification of radical messages. Our contribution is three-fold: (1) we analyze propaganda material published by extremist groups and create a contextual text-based model of radical content, (2) we build a model of psychological properties inferred from these material, and (3) we evaluate these models on Twitter to determine the extent to which it is possible to automatically identify online radical tweets. Our results show that radical users do exhibit distinguishable textual, psychological, and behavioural properties. We find that the psychological properties are among the most distinguishing features. Additionally, our results show that textual models using vector embedding features significantly improves the detection over TF-IDF features. We validate our approach on two experiments achieving high accuracy. Our findings can be utilized as signals for detecting online radicalization activities. | cs.SI | cs | Understanding the Radical Mind: Identifying
Signals to Detect Extremist Content on Twitter
Mariam Nouh∗, Jason R.C. Nurse†, and Michael Goldsmith∗
∗Department of Computer Science, University of Oxford, UK
Email: {mariam.nouh, michael.goldsmith}@cs.ox.ac.uk
†School of Computing, University of Kent, UK
Email: [email protected]
9
1
0
2
y
a
M
5
1
]
I
S
.
s
c
[
1
v
7
6
0
8
0
.
5
0
9
1
:
v
i
X
r
a
Abstract -- The Internet and, in particular, Online Social Net-
works have changed the way that terrorist and extremist groups
can influence and radicalise individuals. Recent reports show
that the mode of operation of these groups starts by exposing
a wide audience to extremist material online, before migrating
them to less open online platforms for further radicalization.
Thus, identifying radical content online is crucial to limit the
reach and spread of the extremist narrative. In this paper,
our aim is to identify measures to automatically detect radical
content in social media. We identify several signals, including
textual, psychological and behavioural, that together allow for
the classification of radical messages. Our contribution is three-
fold: (1) we analyze propaganda material published by extremist
groups and create a contextual text-based model of radical
content, (2) we build a model of psychological properties inferred
from these material, and (3) we evaluate these models on Twitter
to determine the extent to which it is possible to automatically
identify online radical tweets. Our results show that radical users
do exhibit distinguishable textual, psychological, and behavioural
properties. We find that the psychological properties are among
the most distinguishing features. Additionally, our results show
that textual models using vector embedding features significantly
improves the detection over TF-IDF features. We validate our
approach on two experiments achieving high accuracy. Our find-
ings can be utilized as signals for detecting online radicalization
activities.
Index Terms -- radicalization, extremism, data mining, social
media, machine learning, Twitter
I. INTRODUCTION
The rise of Online Social Networks (OSN) has facilitated
a wide application of its data as sensors for information to
solve different problems. For example, Twitter data has been
used for predicting election results, detecting the spread of
flu epidemics, and a source for finding eye-witnesses during
criminal incidents and crises [1], [2]. This phenomenon is
possible due to the great overlap between our online and
offline worlds. Such seamless shift between both worlds has
also affected the modus operandi of cyber-criminals and ex-
tremist groups [3]. They have benefited tremendously from
the Internet and OSN platforms as it provides them with
opportunities to spread their propaganda, widen their reach
for victims, and facilitate potential recruitment opportunities.
For instance, recent studies show that the Internet and social
media played an important role in the increased amount of
©2019 IEEE. Personal use of this material is permitted. Permission from
IEEE must be obtained for all other uses, in any current or future media,
including reprinting/republishing this material for advertising or promotional
purposes,creating new collective works, for resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works.
violent, right-wing extremism [4]. Similarly, radical groups
such as Al-Qaeda and ISIS have used social media to spread
their propaganda and promoted their digital magazine, which
inspired the Boston Marathon bombers in 2010 [5].
To limit the reach of cyber-terrorists, several private and
governmental organizations are policing online content and
utilising big data technologies to minimize the damage and
counter the spread of such information. For example, the UK
launched a Counter Terrorism Internet Referral Unit in 2010
aiming to remove unlawful Internet content and it supports
the police in investigating terrorist and radicalizing activities
online. The Unit reports that among the most frequently
referred links were those coming from several OSNs, such
as Facebook and Twitter [3]. Similarly, several OSNs are
constantly working on detecting and removing users promoting
extremist content. In 2018, Twitter announced that over 1.2
million accounts were suspended for terrorist content [6].
Realizing the danger of violent extremism and radicaliza-
tion and how it is becoming a major challenge to societies
worldwide, many researchers have attempted to study the
behaviour of pro-extremist users online. Looking at existing
literature, we find that a number of existing studies incorporate
methods to identify distinguishing properties that can aid in
automatic detection of these users [7], [8]. However, many of
them depend on performing a keyword-based textual analysis
which, if used alone, may have several shortcomings, such as
producing a large number of false positives and having a high
dependency on the data being studied. In addition, it can be
evaded using automated tools to adjust the writing style.
Another angle for analyzing written text is by looking at the
psychological properties that can be inferred regarding their
authors. This is typically called psycholinguistics, where one
examines how the use of the language can be indicative of
different psychological states. Examples of such psychological
properties include introversion, extroversion, sensitivity, and
emotions. One of the tools that automates the process of
extracting psychological meaning from text is the Linguistic
Inquiry and Word Count (LIWC) [9] tool. This approach
has been used in the literature to study the behaviour of
different groups and to predict
their psychological states,
such as predicting depression [10]. More recently, it has also
been applied to uncover different psychological properties of
extremist groups and understand their intentions behind the
recruitment campaigns [11].
Building on the findings of previous research efforts, this
paper aims to study the effects of using new textual and
psycholinguistic signals to detect extremist content online.
These signals are developed based on insights gathered from
analyzing propaganda material published by known extremist
groups. In this study, we focus mainly on the ISIS group
as they are one of the leading terrorist groups that utilise
social media to share their propaganda and recruit individuals.
We analyze the propaganda material they publish in their
online English magazine called Dabiq, and use data-mining
techniques to computationally uncover contextual
text and
psychological properties associated with these groups. From
our analysis of these texts, we are able to extract a set of
signals that provide some insight into the mindset of the radical
group. This allows us to create a general radical profile that we
apply as a signal to detect pro-ISIS supporters on Twitter. Our
results show that these identified signals are indeed critical to
help improve existing efforts to detect online radicalization.
II. RELATED WORK
In recent years, there has been an increase in online accounts
advocating and supporting terrorist groups such as ISIS [6].
This phenomenon has attracted researchers to study their
online existence, and research ways to automatically detect
these accounts and limit
their spread. Ashcroft et al. [7]
make an attempt to automatically detect Jihadist messages
on Twitter. They adopt a machine-learning method to classify
tweets as ISIS supporters or not. In the article, the authors
focus on English tweets that contain a reference to a set of
predefined English hashtags related to ISIS. Three different
classes of features are used, including stylometric features,
temporal features and sentiment features. However, one of the
main limitations of their approach is that it is highly dependent
on the data. Rowe and Saif [8] focused on studying Europe-
based Twitter accounts in order to understand what happens
before, during, and after they exhibit pro-ISIS behaviour.
They define such behaviour as sharing of pro-ISIS content
and/or using pro-ISIS terms. To achieve this, they use a term-
based approach such that a user is considered to exhibit a
radicalization behaviour if he/she uses more pro-ISIS terms
than anti-ISIS terms. While such an approach seems effective
in distinguishing radicalised users, it is unable to properly
deal with lexical ambiguity (i.e., polysemy). Furthermore,
in [12] the authors focused on detecting Twitter users who
are involved with "Media Mujahideen", a Jihadist group who
distribute propaganda content online. They used a machine
learning approach using a combination of data-dependent and
data-independent features. Similar to [8] they used textual
features as well as temporal features to classify tweets and
accounts. The experiment was based on a limited set of Twitter
accounts, which makes it difficult to generalize the results for
a more complex and realistic scenario.
Radicalization literature also looked at psychological factors
involved with adopting such behaviour. Torok [13] used a
grounded theory approach to develop an explanatory model
for the radicalization process utilizing concepts of psychiatric
power. Their findings show that the process typically starts
with the social isolation of individuals. This isolation seems
to be self-imposed as individuals tend to spend a long time
engaging with radical content. This leads to the concept
of homophily, the tendency to interact and associate with
similar others. Through constant interaction with like-minded
people, an individual gradually strengthens their mindset and
progresses to more extreme levels. Similarly, they start to feel
as being part of a group with a strong group identity which
leads to group polarization. In psychology, group polarization
occurs when discussion leads the group to adopt actions that
are more extreme than the initial actions of the individual
group members [14]. Moreover, the National Police Service
Agency of the Netherlands developed a model to describe the
phases a Jihadist may pass through before committing an act of
terrorism [15]. These sequential phases of radicalism include
strong links between the person's psychological and emotional
state (e.g., social alienation, depression, lack of confidence in
authority) and their susceptibility to radicalization.
III. METHODOLOGY
tasks. In the first
As illustrated in Fig. 1, our approach consists of two
main phases: Phase 1:Radical Properties Extraction, where
articles from Dabiq extremist magazines are input into this
step to perform two parallel
task, we
build a language model using (i) Term-Frequency Inverse-
Document-Frequency (TF-IDF) scores of uni-, bi-, and tri-
grams, and (ii) Word embeddings generated from a word2vec
model [16]. The output of this task is a radical corpus of
top k-grams, and a word embedding model giving a vector
representation for each word in the corpus. The second task
seeks to create a psychological profile based on the language
used in the extremist propaganda articles, consisting of a set
of emotional and topical categories using LIWC dictionary-
based tool. Phase 2: Tweet classification involves the use of
the models generated from Phase 1 to engineer features related
to radical activities. We identify three groups of features and
then train a binary classifier to detect radical tweets.
A. Feature Engineering
Feature engineering is the process of exploring large spaces
of heterogeneous features with the aim of discovering mean-
ingful features that may aid in modeling the problem at hand.
We explore three categories of information to identify relevant
features to detect radical content. Some features are user-based
while others are message-based. The three categories are: 1)
Radical language (Textual features FT ); 2) Psychological sig-
nals (Psychological features FP ); and 3) Behavioural features
(FB). In the following, we detail each of these categories.
1) Radical Language: In order to understand how radical
messages are constructed and used, as mentioned earlier, we
analyze content of ISIS propaganda material published in
Dabiq magazine. Dabiq is an online magazine published by
ISIS terrorist groups with the purpose of recruiting people
and promoting their propaganda and ideology. Using this data
Fig. 1: Approach overview
topics,
source, we investigate what
textual properties, and
linguistic cues exist in these magazines. Our intuition is that
utilising these linguistic cues from the extremist propaganda
would allow us to detect supporters of ISIS group who are
influenced by their propaganda.
We use two methods to extract the radical language from
the propaganda corpus. First we calculate tf-idf scores for
each gram in the propaganda corpus. We use uni-grams,
bi-grams, and tri-grams to capture phrases and context in
which words are being used. We then select the top scoring
grams to be used as features for the language model. N-
grams and words frequency have been used in the literature to
classify similar problems, such as hate-speech and extremist
text and have proven successful [17]. The second method
we use is word embeddings to capture semantic meanings.
Research in NLP has compared the effectiveness of word
embedding methods for encoding semantic meaning and found
that semantic relationships between words are best captured by
word vectors within word embedding models [18]. Therefore,
we train word2vec model on our propaganda corpus to build
the lexical semantic aspects of the text using vector space
models. We learn word embeddings using skip-gram word2vec
model implemented in the gensim package1 with vector size
of 100 and window size of 5. This word embedding model
is used to obtain the vector representation for each word.
We aggregate the vectors for each word in the tweet, and
concatenate the maximum and average for each word vector
dimension, such that any given tweet is represented in 200
dimension sized vector. This approach of aggregating vectors
was used successfully in previous research [19]. Moreover,
since ISIS supporters typically advocate for violent behaviour
and tend to use offensive curse words, we use dictionaries of
violent words2 and curse words3 to record the ratio of such
words in the tweet. We also count the frequency of words
with all capital letters as they are traditionally used to convey
yelling behaviour.
2) Psychological Signals: Research in fields such as lin-
guistics, social science, and psychology suggest that the use
of language and the word choices we make in our daily
1https://radimrehurek.com/gensim/models/word2vec.html
2https://myvocabulary.com/word-list/terrorism-vocabulary
3https://www.cs.cmu.edu/∼biglou/resources/bad-words.txt
communication, can act as a powerful signal to detect our
emotional and psychological states [9]. Several psychological
properties are unintentionally transmitted when we commu-
nicate. Additionally, literature from the fields of terrorism
and psychology suggests that terrorists may differ from non-
terrorists in their psychological profiles [20]. A number of
studies looked at the motivating factors surrounding terrorism,
radicalization, and recruitment tactics, and found that terrorist
groups tend to target vulnerable individuals who have feelings
of desperation and displaced aggression. In particular research
into the recruiting tactics of ISIS groups, it was found that
they focus on harnessing the individual's need for signifi-
cance. They seek out vulnerable people and provide them
with constant attention [21]. Similarly, these groups create a
dichotomy and promote the mentality of dividing the world
into "us" versus "them" [22]. Inspired by previous research,
we extract psychological properties from the radical corpus in
order to understand the personality, emotions, and the different
psychological properties conveyed in these articles.
We utilise LIWC dictionaries to assign a score to a set of
psychological, personality, and emotional categories. Mainly,
we look at the following properties: (1) Summary variables:
Analytically thinking which reflects formal, logical, and hi-
erarchical thinking (high value), versus informal, personal,
and narrative thinking (low value). Clout which reflects high
expertise and confidence levels (high value), versus tentative,
humble, and anxious levels (low value). Tone which reflects
positive emotions (high value) versus more negative emotions
such as anxiety, sadness, or anger (low value). Authentic
which reflects whether the text
is conveying honesty and
disclosing (high value) versus more guarded, and distanced
(low value). (2) Big five: Measures the five psychological
properties (OCEAN), namely Openness, Conscientiousness,
Extraversion, Agreeableness, and Neuroticism. (3) Emotional
Analysis: Measures the positive emotions conveyed in the text,
and the negative emotions (including anger, sadness, anxiety).
(4) Personal Drives: Focuses on five personal drives, namely
power, reward, risk, achievement, and affiliation. (5) Personal
Pronouns: Counts the number of 1st, 2nd, and 3rd personal
pronouns used. For each Twitter user, we calculate their
psychological profiles across these categories. Additionally,
using Minkowski distance measure, we calculate the distance
between each of these profiles and the average values of the
psychological properties created from the ISIS magazines.
3) Behaviour Signals: This category consists of measuring
behavioural features to capture different properties related to
the user and their behaviour. This includes how active the user
is (frequency of tweets posted) and the followers/following
ratio. Additionally, we use features to capture users' interac-
tions with others through using hashtags, and engagement in
discussions using mention action. To capture this, we construct
the mention interaction graph (GM ) from our dataset, such
that GM = (U, E), where U represents the user nodes and E
represents the set of edges. The graph GM is a directed graph,
where an edge e exists between two user nodes A and B, if
user A mentions user B. After constructing the graph, we mea-
sure the degree of influence each user has over their network
using different centrality measures, such as degree centrality,
betweenness centrality, and HITS-Hub. Such properties have
been adopted in the research literature to study properties of
cyber-criminal networks and their behaviour [23], [24].
IV. EXPERIMENTS
A. Dataset
We acquired a publicly available dataset of tweets posted
by known pro-ISIS Twitter accounts that was published during
the 2015 Paris attacks by Kaggle data science community4.
The dataset consists of around 17, 000 tweets posted by more
than 100 users. These tweets were labelled as being pro-ISIS
by looking at specific indicators, such as a set of keywords
used (in the user's name, description, tweet text), their network
of follower/following of other known radical accounts, and
sharing of images of the ISIS flag or some radical leaders. To
validate that these accounts are indeed malicious, we checked
the current status of the users' accounts in the dataset and
found that most of them had been suspended by Twitter. This
suggests that they did, in fact, possess a malicious behaviour
that opposes the Twitter platform terms of use which caused
them to be suspended. We filter out any tweets posted by
existing active users and label this dataset as known-bad.
To model the normal behaviour, we collected a random
sample of tweets from ten-trending topics in Twitter using
the Twitter streaming API. These topics were related to news
events and on-going social events (e.g., sports, music). We
filter out any topics and keywords that may be connected to
extremist views. This second dataset consists of around 8, 000
tweets published by around 1, 000 users. A random sample of
200 tweets was manually verified to ascertain it did not contain
radical views. We label this dataset as our random-good data.
A third dataset is used which was acquired from Kaggle
community5. This dataset is created to be a counterpoise to
the pro-ISIS dataset (our known-bad) as it consists of tweets
talking about topics concerning ISIS without being radical. It
contains 122, 000 tweets from around 95, 000 users collected
on two separate days. We verify that this dataset is indeed non
4https://www.kaggle.com/fifthtribe/how-isis-uses-twitter/data
5https://www.kaggle.com/activegalaxy/isis-related-tweets/home
radical by checking the status of users in Twitter and found
that a subset (24, 000 users) was suspended. We remove those
from the dataset and only keep users that are still active on
Twitter. This dataset is labelled as counterpoise data.
We performed a series of preprocessing steps to clean the
complete dataset and prepare it for feature extraction. These
steps are: (1) We remove any duplicates and re-tweets from
the dataset in order to reduce noise. (2) We remove tweets
that have been authored by verified users accounts, as they
are typically accounts associated with known public figures.
(3) All stop words (e.g., and, or, the) and punctuation marks
are removed from the text of the tweet. (4) If the tweet text
contains a URL, we record the existence of the URL in a new
attribute, hasURL, and then remove it from the tweet text. (5)
If the tweet text contains emojis (e.g., :-), :), :P), we record
the existence of the emoji in a new attribute, hasEmj, and then
remove it from the tweet text. (6) If the tweet text contains
any words with all capital characters, we record its existence
in a new attribute, allCaps, and then normalize the text to
lower-case and filter out any non-alphabetic characters. (7) We
tokenize the cleansed tweet text into words, then we perform
lemmatization, the process of reducing inflected words to their
roots (lemma), and store the result in a vector.
B. Experimental Set-up
We conducted two experiments using the datasets described
in Section IV-A. Our hypothesis is that supporters of groups
such as ISIS may exhibit similar textual and psychological
properties when communicating in social media to the proper-
ties seen in the propaganda magazines. A tweet is considered
radical if it promotes violence, racism, or supports violent
behaviour. In Exp 1 we use the first two datasets, i.e., the
known-bad and the random-good datasets to classify tweets
to radical and normal classes. For Exp 2 we examine if
our classifier can also distinguish between tweets that are
discussing similar topics (ISIS related) by using the known-
bad and the counterpoise datasets.
The classification task is binomial (binary) classification
where the output of the model predicts whether the input
tweet
is considered radical or normal. In order to handle
the imbalanced class problem in the dataset, there are mul-
tiple techniques suggested in the literature Oversampling or
undersampling of the minority/majority classes are common
techniques. Another technique that
is more related to the
classification algorithm is cost sensitive learning, which pe-
nalizes the classification model for making a mistake on the
minority class. This is achieved by applying a weighted cost
on misclassifying of the minority class [25]. We will use the
last approach to avoid downsampling of our dataset.
Previous research investigating similar problems reported
better performances for Random Forest (RF) classifiers [26].
RF usually performs very well as it is scalable and is robust
to outliers. RF typically outperforms decision trees as it has
a hierarchical structure and is based on multiple trees. This
allows RF to be able to model non-linear decision boundaries.
Moreover, Neural Networks (NN) also produced good results
TABLE I: Exp 1: Evaluation metrics across all feature groups
Features
FT (tf − idf )
FT (w2v)
FT
FB
FP
FALL
AC
0.52
0.81
0.84
0.94
1.0
1.0
Precision
0.76
0.82
0.84
0.95
1.0
1.0
Recall
0.52
0.81
0.84
0.94
1.0
1.0
F-measure
0.37
0.81
0.84
0.94
1.0
1.0
TABLE II: Exp 2: Evaluation metrics across all feature groups
Features
FT (tf − idf )
FT (w2v)
FT
FB
FP
FALL
AC
0.56
0.73
0.80
0.91
1.0
1.0
Precision
0.69
0.73
0.80
0.92
1.0
1.0
Recall
0.56
0.73
0.80
0.91
1.0
1.0
F-measure
0.48
0.73
0.80
0.91
1.0
1.0
when applied to problems related to image recognition, text
and natural language processing [27]. However, they usually
tend to require very large amounts of data to train. For
the purpose of this study, we experimented with multiple
classification algorithms, including RF, NN, SVM, and KNN
and found that RF and NN produced the best performance.
Due to space limitation, we only report results obtained using
RF model. We configured the model to use 100 estimators trees
with a maximum depth of 50, and we selected gini impurity
for the split criteria. We used the out-of-bag samples (oob)
score to estimate the generalization accuracy of the model.
Additionally, since RF tends to be biased towards the majority
class, we apply the cost sensitive learning method described
earlier to make RF more suitable for imbalanced data [25].
We divided the dataset to training set (80%) and testing set
(20%), where the testing set is held out for validation. We
reported validation results using different combinations of the
features categories (i.e., FT , FB, FP ) and different evaluation
metrics: accuracy, recall, precision, f-measure, and area under
the ROC curve. Recall measures how many radical tweets we
are able to detect, while precision measures how many radical
tweets we can detect without falsely accusing anyone. For
instance, if we identify every single tweet as radical, we will
expose all radical tweets and thus obtain high recall, but at the
same time, we will call everyone in the population a radical
and thus obtain low precision. F-measure is the average of
both precision and recall.
Fig. 2: ROC curve for Exp1 (top), Exp2 (bottom).
Exp2: In this experiment, we tested the performance of
our classifier in distinguishing between radical and normal
tweets that discusses ISIS-related topics. Although this task
is more challenging given the similarity of the topic discussed
in the two classes, we find that the model still achieves high
performance. Table II shows the different metrics obtained
from each feature category. The FT feature group obtains
80% accuracy, and 91%, 100% for FB and FP feature groups,
respectively. The results are consistent with the ones obtained
from the first experiment with the features from FP group
contributing to the high accuracy of the model. The area under
the Receiver Operator Characteristic (ROC) curve, which
measures accuracy based on TP, and FP rates, is shown in
Fig. 2 for each classification model.
C. Results
D. Features Significance
Exp 1: The classification results using the known-bad and
random-good datasets are reported in Table I. The table shows
the average accuracy, precision, recall and f-measure scores
obtained from each feature category (FT , FP , FB) and their
combination (FAll). We also compared the two textual models,
and find that results obtained from using word embedding out-
performs the use of n-grams tf-idf scores. This confirms that
contextual information is important in detecting radicalization
activities. Furthermore, our model performed best using the
FP features across all metrics. This means that the model is
able to distinguish between both radical and non-radical with
high confidence using only FP .
We investigated which features contribute most to the clas-
sification task to distinguish between radical and non-radical
tweets. We used the mean decrease impurity method of random
forests [28] to identify the most important features in each
feature category. The ten most important features are shown
in Table III. We found that the most important feature for dis-
tinguishing radical tweets is the psychological feature distance
measure. This measures how similar the Twitter user is to the
average psychological profile calculated from the propaganda
magazine articles. Following this is the Us-them dichotomy
which looks at the total number of pronouns used (I,they,
we, you). This finding is in line with the tactics reported in
TABLE III: Features Importance
Top 10
1
2
3
4
5
6
7
8
9
10
Features
Radical psych-profile distance
Us-Them dichotomy
# of mentions a user make
User rank (hub and authority)
Sad emotion
Risk driver
All-caps count
URL count
Violent-word ratio
Hash count
Category
FP
FP
FB
FB
FP
FP
FT
FT
FT
FT
the radicalization literature with regards to emphasizing the
separation between the radical group and the world.
Moreover, among the top contributing features are be-
havioural features related to the number of mentions a single
user makes, and their HITS hub and authority rank among
their interaction network. This relates to how active the user
is in interacting with other users and how much attention
they receive from their community. This links to the ob-
jectives of those radical users in spreading their ideologies
and reaching out to potential like-minded people. As for the
FT category, we find that the use of word2vec embedding
improves the performance in comparison with using the tf-idf
features. Additionally, all bi-grams and tri-grams features did
not contribute much to the classification; only uni-grams did.
This can be related to the differences in the writing styles when
constructing sentences and phrases in articles and in the social
media context (especially given the limitation of the number
of words allowed by the Twitter platform). Additionally, the
violent word ratio, longWords, and allCaps features are among
the top contributing features from this category. This finding
agrees to a large extent with observations from the literature
regarding dealing with similar problems, where the use of
dictionaries of violent words aids with the prediction of violent
extremist narrative.
V. CONCLUSION AND FUTURE WORK
In this paper, we identified different signals that can be
utilized to detect evidence of online radicalization. We de-
rived linguistic and psychological properties from propaganda
published by ISIS for recruitment purposes. We utilize these
properties to detect pro-ISIS tweets that are influenced by their
ideology. Unlike previous efforts, these properties do not only
focus on lexical keyword analysis of the messages, but also
add a contextual and psychological dimension. We validated
our approach in different experiments and the results show
that this method is robust across multiple datasets. This system
can aid law enforcement and OSN companies to better address
such threats and help solve a challenging real-world problem.
In future work, we aim to investigate if the model is resilient
to different evasion techniques that users may adopt. We will
also expand the analysis to other languages.
REFERENCES
[1] F. Morstatter, N. Lubold, H. Pon-Barry, J. Pfeffer, and H. Liu, "Finding
eyewitness tweets during crises," arXiv preprint arXiv:1403.1773, 2014.
[2] A. Tumasjan, T. Sprenger, P. Sandner, and I. Welpe, "Predicting elections
with twitter: What 140 characters reveal about political sentiment," in
Int. Conference on Web and Social Media, 2010.
[3] C. Edwards and L. Gribbon, "Pathways to violent extremism in the
digital era," The RUSI Journal, vol. 158, no. 5, pp. 40 -- 47, 2013.
[4] The Editorial Board. The New York Times, "The New Radicalization of
the Internet," www.nytimes.com/2018/11/24/opinion/sunday/facebook-
twitter-terrorism-extremism.html, 2018, [Online; accessed 09-04-2019].
[5] A. Ward, "ISIS's Use of Social Media Still Poses a Threat to Stability
in the Middle East and Africa," Georgetown Security Studies Review,
Tech. Rep., 2018.
[6] Twitter Public Policy, "Expanding and building #TwitterTransparency,"
https://blog.twitter.com/en us/topics/company/2018/twitter-
transparency-report-12.html, 2018, [Online; accessed 10-April-2019].
[7] M. Ashcroft, A. Fisher, L. Kaati, E. Omer, and N. Prucha, "Detecting
Jihadist Messages on Twitter," in Proceedings of the Intelligence and
Security Informatics Conference, Sept 2015, pp. 161 -- 164.
[8] M. Rowe and H. Saif, "Mining pro-isis radicalisation signals from social
media users." in ICWSM, 2016, pp. 329 -- 338.
[9] J. Pennebaker, R. Boyd, K. Jordan, and K. Blackburn, "The development
and psychometric properties of LIWC2015," University of Texas at
Austin, Tech. Rep., 2015.
[10] M. De Choudhury, M. Gamon, S. Counts, and E. Horvitz, "Predicting
depression via social media." ICWSM, vol. 13, pp. 1 -- 10, 2013.
[11] M. Vergani and A.-M. Bliuc, "The language of new terrorism: dif-
ferences in psychological dimensions of communication in Dabiq and
inspire," Journal of Language and Social Psychology, vol. 37, no. 5, pp.
523 -- 540, 2018.
[12] L. Kaati, E. Omer, N. Prucha, and A. Shrestha, "Detecting Multipliers
of Jihadism on Twitter," in IEEE ICDMW, 2015, pp. 954 -- 960.
[13] R. Torok, "Developing an explanatory model for the process of online
radicalisation and terrorism," Sec. Informatics, vol. 2, no. 1, p. 6, 2013.
[14] D. G. Myers and H. Lamm, "The group polarization phenomenon."
Psychological Bulletin, vol. 83, no. 4, p. 602, 1976.
[15] P. Elzinga, J. Poelmans, S. Viaene, G. Dedene, and S. Morsing, "Terrorist
threat assessment with formal concept analysis," in IEEE ISI, 2010, pp.
77 -- 82.
[16] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean,
"Distributed representations of words and phrases and their composi-
tionality," in Advances in neural information processing systems, 2013,
pp. 3111 -- 3119.
[17] D. Devyatkin, I. Smirnov, M. Ananyeva, M. Kobozeva, A. Chepovskiy,
and F. Solovyev, "Exploring linguistic features for extremist
texts
detection (on the material of russian-speaking illegal texts)," in IEEE
ISI, 2017, pp. 188 -- 190.
[18] W. L. Hamilton, J. Leskovec, and D. Jurafsky, "Diachronic word embed-
dings reveal statistical laws of semantic change," in Proceedings of the
54th Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers). Association for Computational Linguistics,
2016, pp. 1489 -- 1501.
[19] P. Liu, J. Guberman, L. Hemphill, and A. Culotta, "Forecasting the
presence and intensity of hostility on Instagram using linguistic and
social features," in ICWSM, 2018.
[20] A. W. Kruglanski and S. Fishman, "The psychology of terrorism: "Syn-
drome" versus "Tool" perspectives," Terrorism and Political Violence,
vol. 18, no. 2, pp. 193 -- 215, 2006.
[21] V. Pierson, "Western radicalization: rethinking the psychology of terror-
ism," Naval Postgraduate School Monterey, US, Tech. Rep., 2017.
[22] M. Lopez, E. Grace, R. Matteson, S. Henry, M. Link, and D. Axler,
"A Forensic Psychological Assessment of Terrorists: An Anti-Terrorism
Approach for Radicalized Westerners," Journal of Aggression, Maltreat-
ment & Trauma, vol. 9, no. 1-2, pp. 33 -- 43, 2015.
[23] M. Nouh and J. R. C. Nurse, "Identifying key-players in online activist
groups on the Facebook social network," in IEEE International Confer-
ence on Data Mining Workshop (ICDMW), 2015, pp. 969 -- 978.
[24] P. J. Carrington, "Crime and social network analysis," The SAGE
handbook of social network analysis, pp. 236 -- 255, 2011.
[25] C. Chen, A. Liaw, and L. Breiman, "Using random forest to learn
imbalanced data," Dept. Statistics, University of California, Berkeley,
Tech. Rep., 2004.
[26] M. Alfifi and J. Caverlee, Badly Evolved? Exploring Long-Surviving
Suspicious Users on Twitter. Cham: Springer International Publishing,
2017, pp. 218 -- 233.
[27] R. Collobert and J. Weston, "A unified architecture for natural language
processing: Deep neural networks with multitask learning," in Proceed-
ings of ICML. ACM, 2008, pp. 160 -- 167.
[28] G. Louppe, L. Wehenkel, A. Sutera, and P. Geurts, "Understanding
variable importances in forests of randomized trees," in Advances in
neural information processing systems, 2013, pp. 431 -- 439.
|
1804.05705 | 2 | 1804 | 2018-04-23T09:07:36 | And Now for Something Completely Different: Visual Novelty in an Online Network of Designers | [
"cs.SI"
] | Novelty is a key ingredient of innovation but quantifying it is difficult. This is especially true for visual work like graphic design. Using designs shared on an online social network of professional digital designers, we measure visual novelty using statistical learning methods to compare an images features with those of images that have been created before. We then relate social network position to the novelty of the designers images. We find that on this professional platform, users with dense local networks tend to produce more novel but generally less successful images, with important exceptions. Namely, users making novel images while embedded in cohesive local networks are more successful. | cs.SI | cs | And Now for Something Completely Different:
Visual Novelty in an Online Network of Designers
Johannes Wachs
Anikó Hannák
Bálint Daróczy
Institute for Computer Science and
Control, Hungarian Academy of
Sciences (MTA SZTAKI)
Budapest, Hungary
[email protected]
Central European University
Budapest, Hungary
[email protected]
Central European University
Budapest, Hungary
[email protected]
8
1
0
2
r
p
A
3
2
]
I
S
.
s
c
[
2
v
5
0
7
5
0
.
4
0
8
1
:
v
i
X
r
a
Katinka Páll
Institute for Computer Science and
Control, Hungarian Academy of
Sciences (MTA SZTAKI)
Budapest, Hungary
[email protected]
Christoph Riedl
Northeastern University
Boston, MA
Harvard University
Cambridge, MA
[email protected]
ABSTRACT
Novelty is a key ingredient of innovation but quantifying it is dif-
ficult. This is especially true for visual work like graphic design.
Using designs shared on an online social network of professional
digital designers, we measure visual novelty using statistical learn-
ing methods to compare an image's features with those of images
that have been created before. We then relate social network po-
sition to the novelty of the designer's images. We find that on
this professional platform, users with dense local networks tend
to produce more novel but generally less successful images, with
important exceptions. Namely, users making novel images while
embedded in cohesive local networks are more successful.
KEYWORDS
Novelty; image analysis; neural networks; Fisher information; social
networks
ACM Reference Format:
Johannes Wachs, Bálint Daróczy, Anikó Hannák, Katinka Páll, and Christoph
Riedl. 2018. And Now for Something Completely Different: Visual Novelty
in an Online Network of Designers. In WebSci '18: 10th ACM Conference on
Web Science, May 27–30, 2018, Amsterdam, Netherlands. ACM, New York,
NY, USA, 11 pages. https://doi.org/10.1145/3201064.3201088
1 INTRODUCTION
High-quality creative design work can create tremendous value for
organizations. It helps technical products gain acceptance [26] and
it often serves as the basis for competition in cultural markets [59].
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].
WebSci '18, May 27–30, 2018, Amsterdam, Netherlands
© 2018 Association for Computing Machinery.
ACM ISBN 978-1-4503-5563-6/18/05...$15.00
https://doi.org/10.1145/3201064.3201088
Consequently, there has been mounting interest in the use of design-
ers by organizations as a source of value creation [43, 46, 47]. One
important ingredient to successful designs is novelty: the degree
to which a design is new, original, or unusual relative to what has
come before. One reason for this is that derivative work is frowned
up in creative fields [6]. Indeed novelty is the prime ingredient
of innovation and the production of new things [19]. Economists
have long known that innovation is the driving influence behind
economic growth and development [50], and recent studies suggest
that successful companies make 80% of their revenue with products
younger than five years [34].
Despite its importance novelty is difficult to measure, especially
in the context of creative design. In this paper we investigate three
research questions related to novelty in design: (1) how can we
measure novelty in digital design? (2) who produces novel work?
and (3) what is the relationship between novelty and success? We
develop and compare different mathematically-grounded measures
of novelty or distinctiveness of digital images to better understand
its antecedents and subsequent effect on success in a community
of professional designers.
To investigate these questions we collect roughly 40,000 images
posted by over four thousand professional designers on an online
community over a period of about four years. We propose and eval-
uate a measure of novelty for digital design at the image level using
two feature sets: one capturing content and structure defined using
an Inception neural network, the other capturing visual aesthetics
using classical compositional features. We visualize the distribu-
tions of images in low dimensional projections of these feature
spaces to better understand what these features capture and how
they may capture novelty of an image.
We calculate novelty by comparing an image with prior images
in terms of these derived features using information theoretic meth-
ods. This focus on temporal order distinguishes novelty from more
"timeless" notions like beauty or appeal [18]. Calculating novelty
using the compositional features yields a measure of aesthetic or
style novelty based on colors, spatial arrangement, and symmetry,
while using Inception features results in a measure of content nov-
elty. We validate our measures by showing that the earliest images
annotated with emerging labels or "tags" for new kinds of designs
are indeed more content-novel.
With these measures of novelty for digital design in hand, we
ask two questions: who produces novel images? How does novelty
relate to success? The social networks literature makes two sugges-
tions. Individuals with open, diverse social networks have access to
diverse sources of information, which they may synthesize in novel
ways [10, 24]. But individuals in cohesive, closed networks have
greater access to trust and social support, allowing them to more
easily take the risk inherent in the creation of novelties [12, 35].
The literature suggests that when the domain is quickly changing
and when the space of possible novelties is large, it is rather cohe-
sive networks that facilitate novelty [3]. We argue that our topic of
study is such a domain: design evolves quickly and new trends can
be drastically different, and so we hypothesize that cohesive local
networks do more to facilitate novelty in this domain than diverse
ones.
Using a regression framework to analyze our panel data, we find
a positive relationship between the local cohesion of a user's net-
work on the site and the novelty of her images. Users in the global
center of the network make less novel images. We suggest one
possible explanation: that standing out is a form of risk-taking and
that local network density facilitates this behavior. Furthermore,
we find that novel images are on average less successful, but can
be successful when originating from the right network position.
Finally, we demonstrate that our novelty measures add explanatory
power to a machine learning model predicting success, above and
beyond a user's network position. This suggests that network po-
sition does not entirely mediate the relationship between novelty
and success.
Our paper makes three contributions to the literature. First, we
qualitatively compare the data encoded in different feature sets that
can be derived from images. Second, we define a statistically-sound
measure of the novelty of images, applicable to either set of features.
Third, we provide empirical evidence for relationships between
novelty, network position, and success, showing that novelty and
network position together can predict success.
2 RELATED WORK
In this section we first survey research on the quantification of
novelty. Next, we overview literature on the relationship between
novelty, social network position, and success, and finally, introduce
studies that have looked at design in an online setting.
2.1 Quantifying Novelty
As novelty is a complex construct with various dimensions, many
different measures of it have been proposed [9, 21, 44]. One key
notion underlying the measurement of novelty is the concept of
recombination: that novelty is the result of reconfiguration of old
ideas [58]. Novelty is distinguished from aesthetic quality or beauty
because it carries an intrinsic temporal property. For example, it is
difficult to judge in retrospect how novel a product was at the time
of its release. Previous studies on the beauty of images utilize the
fact that crowdsourced judgments of beauty are relatively stable
over time [49].
Recent models of novelty frame it in terms of the "actual" and
the "possible". Models consider what it means for something to
be new in terms of a path of discovery in an evolving complex
space [37]. When something new is done for the first time, the
space of the "adjacent possible" grows, making new things possible.
In this framework, novelties are discrete, binary events.
Previous work from the data mining community on novelty of
images has mostly been concerned with the detection of outliers or
anomalies within images, rather than across images [8, 52]. Most ap-
plications concern the detection of verifiable facts about an image:
the presence of specific objects in satellite images, detecting biolog-
ical abnormalities like cancer, etc. One commonality across these
efforts, and indeed our own, is that features need to be extracted
from an image to make computational analysis tractable.
Several recent data-driven studies quantify novelty in creative
fields. In a study of popular music, Askin and Mauskapf compare
songs with their predecessors using cosine similarity of a set of
derived features like danceability and tempo [4]. Past work has
quantified the creativity of visual art as a combination of both nov-
elty and influence using visual features [18]. Redi et al. quantify the
novelty of short video clips using a similar approach to ours [44],
while Khosla et al. use image features to predict engagement on so-
cial media [32]. Natural language processing has also been applied
to measure the novelty of textual content including scientific article
abstracts [9, 20] and equity crowdfunding campaigns [29]. We do
not define novelty of a thing in terms of success [51] or surprise [5].
Novelty of a thing as we consider it says nothing intrinsically about
its impact, influence, or outcomes. At the same time, we acknowl-
edge that any attempt to measure novelty or distinctiveness can
only capture a small facet of the phenomenon.
2.2 Network Position, Novelty, Success
Psychological research emphasizes that creativity is a demanding
enterprise, requiring focus and concentration [13]. Given the appar-
ent difficulty of creative endeavors, it is perhaps no surprise that
social network structure plays a significant role in both facilitating
novelty and shaping its reception. In fact, recent studies of creativity
emphasize that novel products, even nominally created by a single
author, can sometimes be understood as "products of a momentary
collective process" [27]. How the networks that synthesize creative
products fit together have strong predictive power of their eventual
success [16].
So what kind of network position facilitates novelty? Creators
embedded in a cohesive social network can hope to benefit from
high amounts of social capital and support [12]. Strong ties rep-
resent avenues of trust, which greatly facilitates the kind of risk-
taking inherent in making a novel product in a professional, cre-
ative environment [35]. One study indicates that central actors in a
network of research scientists produce more creative outputs, indi-
cating that established actors can feel the freedom to experiment
more broadly [42].
It is also true that diversity of social connections has been shown
to foster creativity. Weak ties in social networks tend to bridge
groups and provide an actor with access to novel information [24].
Indeed the same study of research scientists cited above shows
that creativity increases with the number of weak ties [42]. This
line of thought is built on the idea that bridging actors occupying
"structural holes" can create their own social capital by leveraging
their unique access to diverse information [10]. Whether open or
closed networks better support novelty creation in our context is
therefore an empirical question.
Besides the relationship of network position and novelty, the
perception of novelty is also of interest to the research community.
What ratio of traditional and novel maximizes success? Work across
many disciplines find an inverse-U shaped relationship between
novelty and success [4, 9]. One prolific strand of the literature
models novelty as the recombination of known ideas in new ways,
and that the key to successful novelty is the combination of many
conventional ingredients with relatively few new ones [56].
2.3 Online Design Communities
Closest to our work empirically are studies on online design commu-
nities, like Dribbble, Behance, or Threadless. These studies generally
focus on the question of how users or products become successful,
and how different groups of users fare [17, 45]. For instance several
studies find significant differences in the behavior and success of
men and women on these sites [33, 57].
Dribbble has received attention from researchers because of its
importance to the professional design community and its exclusive,
invitation-only nature. In an interview-based study researchers
found that users leverage the site and its social network to gather
inspiration, learn skills by reverse engineering examples, anticipate
trends in the marketplace, and to gather feedback [40]. The study
also found that users invested significant effort in developing a
professional identity through the site. As in many other online
communities, users reported the status importance of having many
followers and collecting likes.
More recently, machine vision researchers have taken an interest
in learning from image data taken from online design communi-
ties, as they offer substantively different opportunities to develop
machine vision than, say, photographs [60]. Similarly, the dual func-
tions of online digital communities as places to post and places to
be inspired offer interesting opportunities for bespoke recommen-
dation systems [48].
3 DATA
In this section we describe the Dribbble platform, our data collection
method, and outline the extracted features at the image, user, and
network levels.
3.1 Dribbble
Dribbble, founded in 2009, is an online community where designers
share their work by posting images. It is a highly-visited site, with
an Alexa rank of 1104, the second most popular website for design
sharing after Behance. Unlike most content-sharing platforms, the
site operates on an invitation-only basis: though the site can be
viewed by anyone, only invited users can post images. Active users
are occasionally given invitations which they can use to invite other
designers. Moreover, the number of images a user can post in a
given time frame is capped. All together, this leads to high-quality
Figure 1: Shot (Image) and User Pages on Dribbble.
content and the feeling of belonging to an "elite" community among
users.
The stakes on Dribbble are high. Interviews with users on the
site reveal that individuals use the site to develop their professional
identities [40]. Indeed most users use their real names, post pho-
tographs of themselves for their account image, and link to their
accounts on other online platforms including Linkedin and Twit-
ter. Users build their portfolio of designs over many years. They
accumulate reputation by gathering views and likes (engagement)
on their images, called shots on the site, and followers on their
account. The social network aspect of the site facilitates continued
interactions as users see more and more of each others' work. Suc-
cess on Dribbble has impact outside the site itself, as it can bring
significant employment opportunities and influence. The platform
has recently added a job board and special recruiter accounts.
3.2 Data Collection
Our data sample consists of all Dribbble users who were members of
a team at the time of the data collection. Typically companies form
teams on Dribbble as paid umbrella accounts that users can join.
We select this sample in order to gather a comparable set of users
who are both active and committed members of the site. We then
crawled the profiles of 6,215 users identified as team members. Next,
we crawled all 60,406 images made by these users. In subsequent
analysis, we discard users making fewer than five images1. We
also discard images posted by the team account with identifiable
individual author. We share examples of an image and a user page
in Figure 1. Data collection took place between September and
November 2016 and observed listed rate-limits on the Dribbble API.
3.3 Extracted User features
At the shot level we first record the image itself, the date it was
made, and the identity of the author. We also note the tags the
author annotated the shot with. Tags are free-form key words that
say something about the image. Others can search for images listing
specific tags. Tags therefore serve a dual purpose: to describe what
the author is doing, and to help others find the image. Each shot
has a count of the likes that it received, which can be thought of as
the main success measure in the community.
At the user level we collect the name of the author, whether the
author has a "pro-badge", and the author's tenure on the platform
1Our results are robust to including these users.
(in days). A pro-badge is a sign that the user has paid for a premium
account, which facilitates job search features on the site and lifts the
cap on the number of shots a user can make in a given amount of
time. We consider pro-badges as a proxy for buy-in on the platform.
At the shot level we calculate how many shots a user has made
before to quantify their experience. Finally, we also estimate the
gender of each user. Since the profiles do not directly list gender,
we infer them from the users' first names using the US baby name
data set [1]. For any user with a name not in the database or an
ambiguous gender score (i.e. greater than 10% and less than 90%)
we manually check their self-portrait on Dribbble and on linked
social media accounts.
3.4 Network Features
Like many other online communities, Dribbble is built on top of a
social network. When a user follows another user, the second user's
future shots are included in the default newsfeed of the first user
and so following a user has bandwidth costs. We collect a list of
all following relationships amongst our users and when they were
created. These timestamped edges allow us to recreate the social
network of our users at the time when an image was submitted .
For each image we calculate several network measures quantifying
the position of the user at the time of creation.
• In-degree: How many followers the user has.
• Out-degree: How many other users the user follows.
• Closeness centrality: One over the average distance of the
user from all other nodes [7]. This measures how close the
user is to the center of network.
• Constraint: Burt's measure of the extent to which a user's
• Density: The ratio of observed ties to possible ties among the
outgoing connections are redundant [10].
users the user follows.
In- and out-degree quantify the simple connectivity of a user.
Closeness centrality is a global network measure which increases
as the user is closer to the center of the network. Constraint and
density of the user measure the cohesiveness of his local social
network.
4 EXTRACTING IMAGE FEATURES
In this section we describe two sets of images features upon which
we calculate an image's novelty. First we calculate compositional
features. Then we use a neural network framework to extract a
set of unsupervised features. We compare the two feature spaces
by projecting them to a low-dimensional space in which similar
images are placed closer to one another. We examine what kind of
images are similar according to the two feature sets, finding that
the compositional features capture color and style while the neural
network features capture content.
4.1 Compositional Features
Imitating precisely previous work on the qualitative features of
images [49], we define 47 compositional features for each image.
These features are derived from aesthetic considerations and have
proven to have significant predictive power of the beauty or at-
tractiveness of images. Previous work groups the features into the
following categories: colors, spatial arrangements, and texture.
Color features include contrast (defined in terms of luminance)
and the averages of hue, saturation, and brightness across both the
whole image and a subset in its center [15]. We also include three
"emotional" features which are linear combinations of saturation
and brightness: pleasure, arousal, and dominance [39]. Binning hue,
saturation, and brightness yield Itten Color Histograms and taking
their standard deviations yields Itten Color Contrasts after a careful
segmentation. Spatial features include symmetry and salience [30],
the distribution of which describes how attention-grabbing different
regions of the image are. Finally, Haralick's texture features quantify
image complexity: entropy, energy, homogeneity, and contrast [25].
4.2 Neural Network Features
Feedforward-based neural networks have made tremendous strides
in object-in-image classification tasks in recent years. Many such
networks have penultimate layers which reduce images input for
classification into a feature space for the classification layer. It is
possible to extract these features from pre-trained neural networks.
We harness one such network: the Inception v3 [54], originally
constructed to optimally classify a large dataset of images into 1000
categories. We acknowledge here that there are many alternative
specifications to generate similar sets of features. Passing our im-
ages through the network we generate 2048 features that encode
highly discriminating facets of the data.
4.3 Visualizing Image Features
Before proceeding, we pause to visualize and inspect our data in the
two visual feature spaces. We reduce the 47 and 2048 dimensional
spaces to two-dimensions using t-SNE, a popular dimensionality
reduction method that uses information theoretic methods to mini-
mize distances between data points in the projection as a function
of their similarity [38]. In Figure 2, we visualize the 2-D t-SNE
projections of a random sample of 200 images a year from 2012 to
2016 using the Inception and compositional features, respectively.
In both projections we observe the clustering of images into
groups. The qualitative attributes that define the clustering, how-
ever, are quite different. As highlighted in Figure 2, clustering on
compositional features is based on color and aesthetic style, as ex-
pected. In the projection based on Inception features, however, we
observe that images cluster based on their content. In other words,
images with highly similar Inception features are likely to represent
similar concepts, be they logos, mobile phone interfaces, icons, wire-
frames, etc. This is perhaps not surprising given Inception's origin
as an object-in-image classification tool. This characterization of
the two features sets as describing style and content is important
for understanding their novelty.
5 NOVELTY MEASURES
In this section we define a reference novelty based on user annota-
tions or tags of an image by defining the relative surprise of seeing
a set of tags on image, compared with the tags that came before.
We then define a measure of novelty for our visual feature spaces
using Gaussian mixtures and Fisher information.
Figure 2: Visualizing sample images using t-SNE dimensionality reduction of Inception and compositional features. We high-
light three example groups of images. Images from the gold group are close together in the compositional feature space but
spread out in the Inception feature space. The teal group is close in both feature spaces, with one exception in the composi-
tional space. Images from the purple group are close in Inception space but scattered in compositional space. The gold group
consists of a logo, a collection of icons, a web page design, and an email flier: they are likely clustered in compositional space
because of their color. The members of the purple group are all mobile phone screens. Members of the teal group are likely
clustered in both spaces because they share both structural and compositional qualities.
5.1 Tag Novelty
Before calculating novelty using visual features, we create a novelty
measure using the tags an author gives an image. Following [53],
we calculate the "surprise" of each tag of an image. That is, given
all the images and their tags posted before the image, we define the
probability of observing a tag t as P(t), the proportion of previous
images listing that tag. The log of P(t) is our measure of the surprise
of a tag. As we are especially interested in completely new tags,
we also include the focal image and its tag when we calculate P(t),
to avoid taking the log of 0. We then define tag novelty Ni of an
image i with tags t1, t2, . . . tn ∈ Ti as the aggregate the surprise of
an image's tags:
t ∈Ti
Ni = − 1
Ti
log P(t)
In order to make our measure robust to the order of the images,
we scale each image's tag novelty by the maximum possible novelty.
Namely, if I is the number of images made before image i, we
normalize the equation above by − log(I).
5.2 Visual Novelty via Fisher Information
To study the visual novelty of images, we define a parametric model
for images in terms of their position in a given feature space. Given
a new image, we consider the distribution of previous images in a
feature space and approximate them using Gaussian mixture mod-
els. We calculate the likelihood of the focal image relative to these
(GMM) with N Gaussians, the pdf is p(xθ) =N
distributions using its Fisher information, an information theoretic
measure which we prefer to alternatives such as the Akaike in-
formation criterion because of its reparametrization-invariance.
Specifically we define novelty as one minus the norm of the Fisher
vector of an image over the Gaussian mixture models. This approach
is similar in style to a recent method to calculate novelty using a
data point's distance to the centroids of a k-means clustering [44].
Formally, let be x ∈ Rd a finite d-dimensional real representation
of an image and a parametric model p(xθ) where θ is the parameter
of the density function. If the model is a Gaussian mixture model
i ωiдi(x) where
the дi(x) is the density function of the i-th Gaussian. The continu-
ously evolving model changes the parameters of the probabilistic
model with the emergence of new images in time. We consider two
different likelihood measures to apply to the probabilistic model:
• Akaike's information criterion (AIC) [2]: we measure the
AIC per image according the actual state of our generative
model.
• Fisher information: after calculating the Fisher score [31]
of for each image according to the shape of the model we
can measure the similarity of images x and y with the Fisher
kernel, as
Kθ(x, y) = ∇θ log p(xθ)T F
(1)
where Fθ is the Fisher information matrix. The gradient of
the likelihood indicates how the model may change to fit the
actual point, in our case an image. Our choice was driven by
−1
θ ∇θ log p(yθ)
the unique invariance properties (e.g. reparametrization in-
variance) of the Fisher information matrix and the Fisher ker-
nel [11, 36, 55]. Applying Cholesky decomposition, the ker-
nel can be defined as a simple scalar product, as Kθ(x, y) =
−1/2
Gθ(x)T Gθ(y) where Gθ(x) = ∇θ log p(xθ)F
is the nor-
θ
malized Fisher score or the Fisher vector of image x. We note
that the Fisher vector has dimension O(dθ).
On account of its reparametrization invariance we choose to
continue with the Fisher information as our measure of likelihood.
Although estimation of the Fisher information matrix is difficult,
there are known closed form approximations for both Gaussian mix-
ture models [41] and special classes of Markov random fields [14].
We suggest two potential definitions of novelty measures based on
the Fisher information:
• Norm of the Fisher Vector over Gaussian Mixture (FVGMM):
as the Fisher score highlights how the model parameters
should change to best fit the focal image, our first novelty
measures the norm of the Fisher vector for each image as
−1/2
θ
.
NFV (x) = Gθ(x) = ∇θ log p(xθ)F
In case of Gaussian Mixtures the pdf is p(xθ) =N
(2)
i ωiдi(x)
where θ consists of the mixture weights, mean, and covari-
ance parameters of the Gaussian mixture. In practice we
observe that the Fisher score for both compositional and
Inception features is very sparse because of the "peakness"
property of the membership probability, defined as the prob-
ability that a point is generated from one of the Gaussians.
In comparison with [4] this method puts the most weight on
the most similar images that came before the focal image.
• Similarity graph over the Gaussian Mixture (FVMRF): one
approach to overcoming the "peakness" property while still
capturing the temporal distribution is to define a Markov
random field following [14] with the mean of the Gaussian
mixture as the sample set. The main idea is to define an
undirected random field, which is a graph with N nodes
consisting of random variables and sample points, connected
to our image as a separate random variable in a star. The
probability density function of the new distribution can be
factorized over the maximal cliques in the resulting graph.
In our case the edges and therefore the pdf are:
e−
∫
x ∈X e−
i αi x−µi
i αi x−µi dx
p(xα, θ) =
(3)
where µi is the mean vector of the i-th Gaussian, α is the
relative importance of the cliques. The Fisher vector can be
approximated in this context with a simple formula [14]:
NFV MRF(x) = { di(x) − E[di(x)]
− 1
2 (di(x))
where di(x) = x − µi and i ∈ 1, ..., N .
V ar
}
Given the relatively high complexity of the random field ap-
proach, we define novelty using the norm of the Fisher vector2. As
2In applications where the aforementioned peakness issue is more pronounced, we
recommend using the random field approach
Figure 3: Kernel density estimated distributions of tag, com-
positional, and Inception novelty.
the method returns a similarity score, we subtract one to define vi-
sual novelty. For the rest of the paper we refer to this novelty score
as Inception novelty when it is calculated using Inception features,
and compositional novelty when it is calculated using compositional
features.
5.3 Comparison of Novelty Scores and
Validation
We visualize the distribution of tag, Inception, and compositional
novelty scores in Figure 3. We correlate the two novelties with
tag novelty and several user-level features in Table 1. We find that
both visual novelties are weakly correlated with tag novelty. The
correlation is roughly twice as strong for Inception novelty than
compositional novelty. This suggests that tags are used to describe
images in a conceptional rather than stylistic manner. The two
visual novelties are significantly correlated, and, together with tag
novelty, are negatively correlated with engagement. We note that
the platform's design may explain the trade-off between engage-
ment and tag novelty: users can search for images by tags.
5.3.1 Validation of Visual Novelty. As discussed, novelty is an
ephemeral quality of a cultural product and its measurement im-
plicitly requires comparison, more so than, for example, its beauty.
We cannot, for instance, ask someone to evaluate the novelty of a
four-year-old mobile phone application layout. In this case success
and perceptions of novelty are likely anti-correlated: success breeds
familiarity.
One approach to validate our measures of visual novelty, be-
sides the correlations with tag novelty noted above, is to identify a
population of images which are likely to be covering a new kind
of product that emerges in the middle of our dataset. We identify
emerging product types by finding tags which are used only after
2013, yet still are among the 200 most used tags. We find two such
(1)
(2)
(3)
(4)
Tag Novelty (1)
Inception Novelty (2) 0.123
Compositional Novelty (3) 0.067
0.274
Likes (Log) (4) -0.138 -0.082 -0.014
Views (Log) (5) -0.138 -0.114 -0.058 0.927
Table 1: Correlation matrix of novelty and success features.
6 NOVELTY, NETWORKS, AND SUCCESS
In this section we investigate which users are more likely to create
novel images and whether novel images are more or less likely to be
successful. We consider both Inception and compositional novelty.
First we use hierarchical linear regression [23] on data at the
image level with user random-effectsand controls to predict nov-
elty. Our aim is understand who makes novel images. Then we
predict success using novelty and network position. In both cases
we control for gender, the (log) number of shots made previously,
the (log) number of days the user has been active on the site at the
time of the shot, and whether the user has a paid account. In other
words we control for gender, productivity/experience, tenure, and
investment into the site.
6.1 Who makes novel shots?
We find several significant predictors of Inception novelty, both
among our control variables and network variables. Interestingly,
the network features we consider do not impact compositional
novelty. We summarize these findings in Table 2.
For both compositional and Inception-based measures we find
that pro-users are less likely to make novel images. One inter-
pretation is that users who take the site more seriously are more
risk-averse and less likely to experiment. Users making more shots
in the past make slightly more novel shots. There is mixed evidence
that users active for a longer period of time make less novel shots.
We detect no gender disparity.
The two novelty measures diverge when we consider the impact
of network features. The Inception-based measure of novelty is
significantly lower for users closer to the core of the network, and
higher for users with cohesive local networks defined by density
and constraint. This supports our hypothesis that cohesion facili-
tates novelty. We find no significant relationship between network
position and compositional novelty.
6.2 When are novel shots successful?
We now turn to the question of predicting engagement, measured
by likes, using novelty. We find that novel shots are generally less
successful. We summarize our findings in Table 3. Pro users are
more successful, as are those who have many followers. We find
that constrained users are less successful. Finally, novel images are
in general less successful.
We find an interesting interaction between constraint and In-
ception novelty. Namely, users embedded in highly constrained
networks making novel images do better than those in uncon-
strained networks making novel images. To better interpret this
finding we visualize this relationship in Figure 5. In other words,
the least constrained users have a penalty for novelty while the
most constrained users have a bonus for novelty. We also find a
significant interaction between inception novelty and closeness
centrality: novelty has an increasingly negative relationship with
success as a user is more central in the network, but no relationship
between local density and either novelty measure.
Finally, using a machine learning framework, we check how
well our features can predict success binned into three separate
class labels: less than ten likes, between ten and one hundred likes,
and more than one hundred likes. As an initialization we used the
Figure 4: Comparison of visual novelty scores of images with
the tags "material" and "principle". We consider those im-
ages in the first 10% and most recent 10% of all images cre-
ated using the tags. We find that Inception novelty is signif-
icantly higher for images listing these "emerging" tags.
tags3 which we can verify as representing truly emerging novelties:
"material" and "principle".
Material design4 is a design language or vocabulary created by
Google, announced to the public in June 2014. Like other design
languages, it has guidelines and principles that shape the design pro-
cess, resulting in a consistent look with certain qualities. Material
design was created especially for use in digital and technological
areas. It emphasizes the use of print design best practices together
with motion. Material or "material design" appears as a tag in 748
images in our dataset.
Principle5 is a new software design tool for creating interactive
and dynamic user interfaces. Released in August 2015, it is a popular
tool for designers to prototype UIs. 243 images in our dataset include
a "principle" tag.
For both tags we compare the distributions of novelty for the first
10% of images using the tag, with the most recent 10% of images
using the tag. In figure 4 we plot the resulting distributions. We
find that Inception novelty is significantly higher for the earliest
images tagged with "material" (Mann-Whitney U = 1897, p<.01)
and "princple" (Mann-Whitney U = 190, p<.01) compared with the
most recent ones. Though the average compositional novelty is
higher for the earliest images in both cases, the differences are not
statistically significant (resp. U = 2465, p .26; U = 288, p .32).
3Other examples of tag fitting our quantitative criteria are tags used by groups of
designers to indicate group membership. Though these tags certainly merit further
study, they do not capture the emergence of a new design approach or method
4https://material.io/
5http://principleformac.com/
Inception Novelty
Composition Novelty
Dependent variable:
(1)
−0.015∗∗∗ (0.002)
0.020∗∗∗ (0.007)
−0.002
(0.013)
−0.036∗∗∗ (0.010)
0.004
(0.004)
−0.042∗∗∗ (0.008)
0.002
(0.025)
37,799
−25,740.880
51,576.620
Days Active (log)
nShots Previous
Male
Pro
In-Degree (log)
Closeness
Constraint
Density
Constant
Observations
Log Likelihood
Bayesian Inf. Crit.
User random effects
(2)
−0.016∗∗∗ (0.002)
0.022∗∗∗ (0.007)
0.001
(0.013)
−0.034∗∗∗ (0.010)
0.002
(0.004)
0.060∗∗ (0.025)
−0.009
(0.026)
37,799
−25,749.400
51,593.660
(3)
−0.017∗∗∗ (0.002)
0.022∗∗∗ (0.007)
0.001
(0.013)
−0.034∗∗∗ (0.011)
(0.004)
0.001
0.046∗∗ (0.023)
−0.002
(0.026)
37,799
−25,750.350
51,595.570
(4)
−0.000
(0.002)
0.013∗
(0.007)
−0.000
(0.013)
−0.034∗∗∗ (0.011)
(0.004)
0.004
0.001
(0.008)
−0.048∗
(0.026)
37,799
−25,731.900
51,558.660
Table 2: Predicting novelty with network position.
(5)
0.000
(0.002)
0.013∗
(0.007)
−0.000
(0.013)
−0.034∗∗∗ (0.011)
(0.004)
0.004
0.018
(0.026)
−0.053∗∗ (0.026)
37,799
−25,730.540
51,555.950
(6)
−0.000
(0.002)
0.013∗
(0.007)
−0.000
(0.013)
−0.034∗∗∗ (0.011)
(0.004)
0.004
−0.005
−0.047∗
(0.024)
(0.026)
37,799
−25,730.860
51,556.580
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
Dependent variable:
Log Likes
(1)
−0.006∗∗ (0.003)
0.086∗∗∗ (0.023)
−0.046
(0.044)
0.196∗∗∗ (0.037)
0.371∗∗∗ (0.009)
−0.046∗∗∗ (0.013)
−0.234∗∗∗ (0.059)
−0.108∗∗∗ (0.009)
0.084∗∗ (0.039)
2.930∗∗∗ (0.088)
37,799
−36,353.290
72,833.060
(2)
−0.004∗
(0.003)
0.090∗∗∗ (0.023)
−0.047
(0.045)
0.196∗∗∗ (0.037)
0.373∗∗∗ (0.009)
−0.046∗∗∗ (0.013)
−0.233∗∗∗ (0.059)
−0.025∗∗∗ (0.010)
0.017
(0.040)
2.908∗∗∗ (0.089)
37,799
−36,450.650
73,027.780
Days Active (log)
nShots Previous
Male
Pro
In-Degree (log)
Out-Degree (log)
Constraint
Incep. Nov.
Incep. Nov. × Constraint
Comp. Nov.
Comp. Nov. × Constraint
Constant
Observations
Log Likelihood
Bayesian Inf. Crit.
∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01
User random effects
Table 3: Predicting success with novelty and network posi-
tion.
first year as the first training period and for every consecutive
quarter thereafter we consider the previous year. We found that
the random field approach to calculating the Fisher vector (FVMRF)
was most effective in predicting engagement. Using the area under
the receiver operating characteristic curve (AUC), we find that a
gradient boosted trees model [22] on network and content features
has significant predictive power. As we can see in Figure 6, even
Figure 5: Relationship between success and novelty as con-
straint varies. Low constraint users have less success when
making novel shots. High constraint users have more suc-
cess with novel shots.
though the network features are the best indicators of success, the
content and novelty of the images, encoded using the Inception-
based Fisher vectors offer additional predictive power. This suggests
that it is possible to use image features to predict success on the
site. It is likely possible to do better if features are extracted with
the aim of predicting success.
7 CONCLUSIONS
In this paper we developed, evaluated, and compared measures of
novelty of images using data from an online community of digital
designers. We first compared different feature sets of images, noting
that compositional features like entropy, contrast, and brightness
participation, the presence of significant players and companies
in the field, and the potential for economic opportunities, it seems
reasonable that established designers may have reason to make
more conventional images. That Dribbble is an online community
only compounds the potential costs of creating unsuccessful nov-
elty: though a designer's support system and network of strong ties
cannot vastly grow, her audience can scale drastically. The underes-
timated permanence of online identities makes this asymmetry all
the more important when we consider what it means for a designer
to take a risk with a distinctive image.
Indeed professional online communities present a dilemma for
users in general. Though the feelings of anonymity and distance
may facilitate bold experimentation, members of online communi-
ties who wish to leverage their investment of time and effort into
professional advancement must credibly link their online identi-
ties to their real ones. Even users who want to stay anonymous
often have a hard time doing so [28]. Once this identification has
occurred, the individual must consider that anything they share
online is widely broadcast and more consistently recorded and pre-
served than what they may say or share offline. We claim that as the
labor market becomes increasingly digital, online social networks
merit closer study.
Turning to the relationship between novelty and success, we find
that novelty is related to worse outcomes. We also find that users in
highly constrained positions are less successful. On the other hand,
the interaction between constraint and novelty is positive: users
with cohesive local networks of strong ties making novel images
find more success. We argue that these relationships merit further
study. Are these embedded designers better positioned to take
risks? Can we interpret images with high novelty score, according
to our definition, as being risky? The negative relationship between
novelty and network centrality raises even more questions.
Our study has several limitations. Given the transient nature of
novelty, we have only limited tests of validity for our measures.
Given the ubiquity of digital technology, a highly novel digital de-
sign from five years ago likely looks highly outdated now. Moreover,
the networking behavior of designers on this platform is highly
tailored to the situation. For example, users adopting a strategy of
aggressive following anticipating reciprocity, may end up in highly
dense networks. All at once, Dribbble serves as a social network,
professional portfolio, information network, and status hierarchy
for the field. Any attempt to infer causal relations between social
network structure and the creation of new ideas on this platform
must disentangle the complicated layers driving interactions. We
also concede that novelty is multi-faceted: no single measure can
totally capture such a broad concept. In future work we aim to
better understand influence and spreading of novelty.
8 ACKNOWLEDGEMENTS
The authors wish to thank Zsófia Czémán, Anna May, and anony-
mous referees for their helpful suggestions. This research has been
funded in part by NSF grant IIS-1514283. D.B was supported by the
Momentum Grant of the Hungarian Academy of Sciences (LP2012-
19/2012).
Figure 6: Average AUC of quarter to quarter success pre-
diction. We predict success of the images using gradient
boosted trees on the visual image features, novelty scores,
and network position of the users. We find that novelty
scores extracted image features increase the predictive
power of the model including the network features, This
suggests that network position does not entirely mediate the
relationship between novelty and success.
capture qualitatively different facets of an image and features de-
rived from an Inception neural network learning framework capture
qualitatively different facets of an image. Specifically, compositional
features seem to capture stylistic aspects and while Inception fea-
tures capture content, in line with their origins.
Next, we created a mathematical framework to compare images
with all images that came before in terms of either set of image
features. To calculate the novelty of an image, we estimate the
distribution of previous shots in the given feature space using a
Gaussian mixture model. We then calculate the likelihood of the
the image - in other words we quantify how statistically similar the
image is to those that came before. We define novelty of an image
as one minus this similarity score.
We find that both novelties calculated from the Inception fea-
tures and compositional features are significantly correlated with a
measure of novelty based on author text annotations or "tags" of
their images. We also found that Inception novelty was significantly
higher for images created in the early stages of an emerging tag
compared with images using the same tag later.
Attempting to understand the profile of a user who makes more
novel shots, we turned to the site's social network. Using temporal
following data, we related social network position at the time of
the creation of an image to its novelty. We found that users with
cohesive local networks (quantified by density or Burt's constraint
measure) tend to post images with higher Inception novelty.
We also find that users close to the center of the network, in a
global sense, make less novel shots. Users with a "pro-badge" (paid
account) likewise make less novel images. Given the professional
atmosphere of the site, including for example its invitation-only
Baby Names from Social Security Card Applications-National
https://catalog.data.gov/dataset/
Level Data.
baby-names-from-social-security-card-applications-national-level-data.
data.gov.
(2016).
[2] Hirotugu Akaike. 1981. Likelihood of a model and information criteria. Journal
of econometrics 16, 1 (1981), 3–14.
[3] Sinan Aral and Marshall Van Alstyne. 2011. The diversity-bandwidth trade-off.
Amer. J. Sociology 117, 1 (2011), 90–171.
[4] Noah Askin and Michael Mauskapf. 2017. What Makes Popular Culture Popular?
Product Features and Optimal Differentiation in Music. American Sociological
Review 82, 5 (2017), 910–944.
[5] Andrew Barto, Marco Mirolli, and Gianluca Baldassarre. 2013. Novelty or sur-
prise? Frontiers in psychology 4 (2013).
[6] Julia Bauer, Nikolaus Franke, and Philipp Tuertscher. 2016. Intellectual property
norms in online communities: How user-organized intellectual property regula-
tion supports innovation. Information Systems Research 27, 4 (2016), 724–750.
[7] Alex Bavelas. 1950. Communication patterns in task-oriented groups. The Journal
of the Acoustical Society of America 22, 6 (1950), 725–730.
[8] Giacomo Boracchi, Diego Carrera, and Brendt Wohlberg. 2014. Novelty detection
in images by sparse representations. In Intelligent Embedded Systems (IES), 2014
IEEE Symposium on. IEEE, 47–54.
[9] Kevin J Boudreau, Eva C Guinan, Karim R Lakhani, and Christoph Riedl. 2016.
Looking across and looking beyond the knowledge frontier: Intellectual distance,
novelty, and resource allocation in science. Management Science 62, 10 (2016),
2765–2783.
[10] Ronald S Burt. 2004. Structural holes and good ideas. American journal of sociology
110, 2 (2004), 349–399.
[11] LL Campbell. 1986. An extended Čencov characterization of the information
metric. Proc. Amer. Math. Soc. 98, 1 (1986), 135–141.
[12] James S Coleman. 1988. Social capital in the creation of human capital. American
journal of sociology 94 (1988), S95–S120.
[13] Mihaly Csikszentmihalyi. 1996. Flow and the psychology of discovery and invention.
New York: Harper Collins.
[14] Balint Daroczy, David Siklois, Robert Palovics, and Andras A Benczur. 2015. Text
Classification Kernels for Quality Prediction over the C3 Data Set. In Proceedings
of the 24th International Conference on World Wide Web. ACM, 1441–1446.
[15] Ritendra Datta, Dhiraj Joshi, Jia Li, and James Z Wang. 2006. Studying aesthetics
in photographic images using a computational approach. In European Conference
on Computer Vision. Springer, 288–301.
[16] Mathijs De Vaan, David Stark, and Balazs Vedres. 2015. Game changer: The
topology of creativity. Amer. J. Sociology 120, 4 (2015), 1144–1194.
[17] Biplab Deka, Haizi Yu, Devin Ho, Zifeng Huang, Jerry O Talton, and Ranjitha
Kumar. 2015. Ranking designs and users in online social networks. In Proceedings
of the 33rd Annual ACM Conference Extended Abstracts on Human Factors in
Computing Systems. ACM, 1887–1892.
[18] Ahmed Elgammal and Babak Saleh. 2015. Quantifying Creativity in Art Networks.
In Proceedings of the Sixth International Conference on Computational Creativity
June. 39.
[19] Maria-Isabel Encinar and Felix-Fernando Munoz. 2006. On novelty and economics:
Schumpeter's paradox. Journal of Evolutionary Economics 16, 3 (2006), 255–277.
[20] James A Evans and Jacob G Foster. 2011. Metaknowledge. Science 331, 6018
(2011), 721–725.
Science 47, 1 (2001), 117–132.
[21] L. Fleming. 2001. Recombinant uncertainty in technological search. Management
[22] Jerome H Friedman. 2001. Greedy function approximation: a gradient boosting
machine. Annals of statistics (2001), 1189–1232.
[23] Andrew Gelman and Jennifer Hill. 2006. Data analysis using regression and
multilevel/hierarchical models. Cambridge university press.
[24] Mark S Granovetter. 1973. The strength of weak ties. American journal of sociology
REFERENCES
[1] 2016.
78, 6 (1973), 1360–1380.
IEEE 67, 5 (1979), 786–804.
[25] Robert M Haralick. 1979. Statistical and structural approaches to texture. Proc.
[26] A. Hargadon and R.I. Sutton. 1997. Technology brokering and innovation in a
product development firm. Administration Science Quarterly 42, 4 (1997), 716–749.
[27] Andrew B Hargadon and Beth A Bechky. 2006. When collections of creatives be-
come creative collectives: A field study of problem solving at work. Organization
Science 17, 4 (2006), 484–500.
[28] Emöke-Ágnes Horvát, Michael Hanselmann, Fred A Hamprecht, and Katharina A
Zweig. 2012. One plus one makes three (for social networks). PloS one 7, 4 (2012),
e34740.
[29] Emöke-Ágnes Horvát, Johannes Wachs, Aniko Hannak, and Rong Wang. 2018.
The Role of Novelty in Securing Investors for Equity Crowdfunding Campaigns.
In The 6th AAAI Conference on Human Computation and Crowdsourcing (HCOMP).
AAAI.
[30] Xiaodi Hou and Liqing Zhang. 2007. Saliency detection: A spectral residual
approach. In Computer Vision and Pattern Recognition, 2007. CVPR'07. IEEE Con-
ference on. IEEE, 1–8.
[31] Tommi S Jaakkola, David Haussler, et al. 1999. Exploiting generative models
in discriminative classifiers. Advances in neural information processing systems
(1999), 487–493.
[32] Aditya Khosla, Atish Das Sarma, and Raffay Hamid. 2014. What makes an image
popular?. In Proceedings of the 23rd international conference on World wide web.
ACM, 867–876.
[33] Nam Wook Kim. 2017. Creative Community Demystified: A Statistical Overview
of Behance. arXiv preprint arXiv:1703.00800 (2017).
[34] W Chan Kim and Renée Mauborgne. 1997. Value innovation: The strategic logic
of high growth. Harvard Business School Pub.
[35] David Krackhardt. 2003. The strength of strong ties. Networks in the knowledge
economy (2003), 82.
[36] Guy Lebanon. 2004. An extended Čencov-Campbell characterization of condi-
tional information geometry. In Proceedings of the 20th conference on Uncertainty
in artificial intelligence. AUAI Press, 341–348.
[37] Vittorio Loreto, Vito DP Servedio, Steven H Strogatz, and Francesca Tria. 2016.
Dynamics on expanding spaces: modeling the emergence of novelties. In Cre-
ativity and Universality in Language. Springer, 59–83.
[38] Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE.
Journal of Machine Learning Research 9, Nov (2008), 2579–2605.
[39] Jana Machajdik and Allan Hanbury. 2010. Affective image classification using
features inspired by psychology and art theory. In Proceedings of the 18th ACM
international conference on Multimedia. ACM, 83–92.
[40] Jennifer Marlow and Laura Dabbish. 2014. From rookie to all-star: professional
development in a graphic design social networking site. In Proceedings of the
17th ACM conference on Computer supported cooperative work & social computing.
ACM, 922–933.
[41] Florent Perronnin and Christopher Dance. 2007. Fisher kernels on visual vocabu-
laries for image categorization. In Computer Vision and Pattern Recognition, 2007.
CVPR'07. IEEE Conference on. IEEE, 1–8.
[42] Jill E Perry-Smith. 2006. Social yet creative: The role of social relationships in
facilitating individual creativity. Academy of Management journal 49, 1 (2006),
85–101.
[43] D. Ravasi and G. Lojacono. 2005. Managing design and designers for strategic
renewal. Long Range Planning 38 (2005), 51–77.
[44] Miriam Redi, Neil OHare, Rossano Schifanella, Michele Trevisiol, and Alejandro
Jaimes. 2014. 6 seconds of sound and vision: Creativity in micro-videos. In
Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE,
4272–4279.
[45] Christoph Riedl and V Seidel. 2018. Learning from Mixed Signals: Evidence from
a Contest-based Online Innovation Community. Organization Science (2018).
[46] V. Rindova, E. Dalpiaz, and D. Ravasi. 2011. A cultural quest: A study of organiza-
tional use of new cultural resources in strategy formation. Organization Science
22, 2 (2011), 413–431.
[47] V.P. Rindova and A.P. Petkova. 2007. When is a new thing a good thing? Tech-
nological change, product form design, and perceptions of value for product
innovations. Organization Science 18, 2 (2007), 217–232.
[48] Maja R Rudolph, Matthew Hoffman, and Aaron Hertzmann. 2016. A joint model
for who-to-follow and what-to-view recommendations on behance. In Proceedings
of the 25th International Conference Companion on World Wide Web. International
World Wide Web Conferences Steering Committee, 581–584.
[49] Rossano Schifanella, Miriam Redi, and Luca Maria Aiello. 2015. An Image Is
Worth More than a Thousand Favorites: Surfacing the Hidden Beauty of Flickr
Pictures. In Ninth International AAAI Conference on Web and Social Media.
[50] Joseph Schumpeter and Ursula Backhaus. 2003. The theory of economic develop-
ment. Joseph Alois Schumpeter (2003), 61–116.
[51] Roberta Sinatra, Dashun Wang, Pierre Deville, Chaoming Song, and Albert-László
Barabási. 2016. Quantifying the evolution of individual scientific impact. Science
354, 6312 (2016), aaf5239.
[52] Alex Smola, Le Song, and Choon Hui Teo. 2009. Relative novelty detection. In
Artificial Intelligence and Statistics. 536–543.
[53] Sameet Sreenivasan. 2013. Quantitative analysis of the evolution of novelty in
cinema through crowdsourced keywords. Scientific reports 3 (2013).
[54] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir
Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. 2015.
Going deeper with convolutions. In Proceedings of the IEEE conference on computer
vision and pattern recognition. 1–9.
[55] N. N. Cencov. 1982. Statistical decision rules and optimal inference. American
Mathematical Society 53 (1982).
[56] Brian Uzzi, Satyam Mukherjee, Michael Stringer, and Ben Jones. 2013. Atypical
combinations and scientific impact. Science 342, 6157 (2013), 468–472.
[57] Johannes Wachs, Anikó Hannák, András Vörös, and Bálint Daróczy. 2017. Why
Do Men Get More Attention? Exploring Factors Behind Success in an Online
Design Community. In Eleventh International AAAI Conference on Web and Social
Media.
[58] M.L. Weitzman. 1998. Recombinant growth. Quarterly Journal of Economics 113,
2 (1998), 331–360.
[59] N.M. Wijnberg and G. Gemser. 2000. Adding Value to Innovation: Impressionism
and the Transformation of the Selection System in Visual Arts. Organization
Science 11, 3 (2000), 323–329.
[60] Michael J Wilber, Chen Fang, Hailin Jin, Aaron Hertzmann, John Collomosse, and
Serge Belongie. 2017. BAM! The Behance Artistic Media Dataset for Recognition
Beyond Photography. arXiv preprint arXiv:1704.08614 (2017).
|
1711.10251 | 1 | 1711 | 2017-11-28T12:29:40 | Joint Non-negative Matrix Factorization for Learning Ideological Leaning on Twitter | [
"cs.SI"
] | People are shifting from traditional news sources to online news at an incredibly fast rate. However, the technology behind online news consumption promotes content that confirms the users' existing point of view. This phenomenon has led to polarization of opinions and intolerance towards opposing views. Thus, a key problem is to model information filter bubbles on social media and design methods to eliminate them. In this paper, we use a machine-learning approach to learn a liberal-conservative ideology space on Twitter, and show how we can use the learned latent space to tackle the filter bubble problem.
We model the problem of learning the liberal-conservative ideology space of social media users and media sources as a constrained non-negative matrix-factorization problem. Our model incorporates the social-network structure and content-consumption information in a joint factorization problem with shared latent factors. We validate our model and solution on a real-world Twitter dataset consisting of controversial topics, and show that we are able to separate users by ideology with over 90% purity. When applied to media sources, our approach estimates ideology scores that are highly correlated (Pearson correlation 0.9) with ground-truth ideology scores. Finally, we demonstrate the utility of our model in real-world scenarios, by illustrating how the learned ideology latent space can be used to develop exploratory and interactive interfaces that can help users in diffusing their information filter bubble. | cs.SI | cs | Joint Non-negative Matrix Factorization for
Learning Ideological Leaning on Twi(cid:29)er
Preethi Lahoti∗
Max Planck Institute for Informatics
Saarbruecken, Germany
[email protected]
Kiran Garimella
Aalto University
Helsinki, Finland
kiran.garimella@aalto.(cid:27)
Aristides Gionis
Aalto University
Helsinki, Finland
aristides.gionis@aalto.(cid:27)
7
1
0
2
v
o
N
8
2
]
I
S
.
s
c
[
1
v
1
5
2
0
1
.
1
1
7
1
:
v
i
X
r
a
ABSTRACT
People are shifting from traditional news sources to online news at
an incredibly fast rate. However, the technology behind online news
consumption promotes content that con(cid:27)rms the users' existing
point of view. This phenomenon has led to polarization of opinions
and intolerance towards opposing views. Thus, a key problem is
to model information (cid:27)lter bubbles on social media and design
methods to eliminate them. In this paper, we use a machine-learning
approach to learn a liberal-conservative ideology space on Twitter,
and show how we can use the learned latent space to tackle the
(cid:27)lter bubble problem.
We model the problem of learning the liberal-conservative ideol-
ogy space of social media users and media sources as a constrained
non-negative matrix-factorization problem. Our model incorporates
the social-network structure and content-consumption informa-
tion in a joint factorization problem with shared latent factors. We
validate our model and solution on a real-world Twitter dataset
consisting of controversial topics, and show that we are able to
separate users by ideology with over 90 % purity. When applied
to media sources, our approach estimates ideology scores that are
highly correlated (Pearson correlation 0.9) with ground-truth ide-
ology scores. Finally, we demonstrate the utility of our model in
real-world scenarios, by illustrating how the learned ideology latent
space can be used to develop exploratory and interactive interfaces
that can help users in di(cid:29)using their information (cid:27)lter bubble.
1 INTRODUCTION
Social media and the web have provided a foundation where users
can easily access diverse information from around the world. How-
ever, over the years, various factors, such as user homophily (social
network structure), and algorithmic (cid:27)ltering (e.g., news feeds and
recommendations) have narrowed the content that a user consumes.
As an example, imagine two users of opposite ideological stances
(liberal and conservative). Though the two users may be looking
at the same topic (e.g., a presidential debate), they might be seeing
∗Author was at Aalto University at the time of this work.
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].
WSDM 2018, Marina Del Rey, CA, USA
© 2018 Copyright held by the owner/author(s). Publication rights licensed to ACM.
978-1-4503-5581-0/18/02...$15.00
DOI: 10.1145/3159652.3159669
completely di(cid:29)erent viewpoints due to the diverse network sur-
rounding and di(cid:29)erent content sources they get their information
from. Consequently, users on di(cid:29)erent ends of the ideological spec-
trum live in their own information bubbles [31], oblivious to the
views on the other side and creating their own world-view of truth.
This phenomenon has led to the polarization of viewpoints, intol-
erance towards opposing views, and ideological segregation [36].
Many studies suggest that increasingly users live in their echo
chambers [35] and polarization of the public has intensi(cid:27)ed [11].
In this paper, we propose a principled approach to infer the
ideological stances (also known as ideology or polarity or leaning)
of both the users in a social network, and the media sources that
provide news content in the network. Our approach is based on a
non-negative matrix-factorization model, which jointly decomposes
the social network of users and the content they consume in a
shared latent space. Learning the ideological stances of social media
users and content sources is an important step in building useful
tools that can make users aware of their informational bias and
consequently, help in reducing those biases, thus mitigating the
increasing polarization in the society.
Existing approaches to identifying the ideological leaning of
users either (i) require large amounts of manually annotated data [10,
32]; or (ii) consider only the structure of social ties [4] or user
interactions [16, 37]; or (iii) analyze only the content shared by
users [33]. Each of these families of approaches has its own limita-
tions. First, obtaining manual annotations is both expensive and
time consuming. Second, when using only the structure of social
ties or interactions, it is assumed that users with the same ideology
are more likely to interact with or follow each other. However,
social networks are extremely sparse and noisy, and users with
the same ideological leaning may never interact with or follow
each other, while users with di(cid:29)erent ideological leaning may still
interact due to di(cid:29)erent reasons. Last, inferring ideological leaning
by using only content is a challenging task and prone to errors due
to the inherent complexity of natural-language understanding.
To overcome these problems, we propose an unsupervised ap-
proach that uses simultaneously the network structure and the
information about content shared by users. We are motivated by
the observation that a user's ideological stance on a topic depends
on both the surrounding network structure as well as the content
sources that users consume their information from. Thus, we ex-
ploit the inherent connection between the two data types: Users are
not only more likely to interact with or follow like-minded users,
but also to share content of aligned ideological leaning. Taking
into account the full available information results in more accurate
estimation of ideological leaning both for users and content sources.
In addition, we formulate the task of learning ideological leaning
as a joint matrix factorization problem in which users and sources
are represented in a shared latent space. This formulation allows us
to identify the relationship between data points of the two types,
which is particularly useful for visualizing social-media users and
sources in a common space, and building applications for exploring
the ideological landscape in one's media neighborhood, or making
recommendations to escape the (cid:27)lter bubble. In fact, we demon-
strate concretely how to use the learned ideology latent space to
develop exploration and recommendation tools (Section 6).
Experiments comparing our approach to the state of the art show
the bene(cid:27)t of such a principled joint approach. Our method is able
to separate users into ideological clusters with over 90% purity.
When applied to media sources, our approach estimates ideology
scores that are highly correlated with ground-truth ideology scores.
All our analysis is done on Twitter, though the methods gener-
alize to any other social network. In the rest of the paper, for the
sake of clarity we use Twitter-speci(cid:27)c nomenclature (e.g., retweets,
follow, etc.).
In summary, our contributions are as follows:
• We present a principled approach to jointly compute the ide-
ology scores of both users and content sources by formulating
the problem as a joint constrained matrix-factorization task.
This formulation allows us to jointly cluster the two data types,
"user" and "source". We apply our learned ideological latent
space to the problems of "who" and "what" to recommend to
users so as to reduce polarization and di(cid:29)use their information
(cid:27)lter bubble.
• To the best of our knowledge, this is the (cid:27)rst work to compute
ideological stances for both users and content sources, in a
common latent space, based on both the network structure and
the users' interaction with sources.
• We provide an extensive experimental evaluation, presenting
both qualitative and quantitative results on real-world Twitter
data. Our method shows promise when compared to existing
state of the art approaches.
2 RELATED WORK
In this paper, we propose an approach that can identify the ideolog-
ical leaning of users on Twitter. We de(cid:27)ne ideology based on the
policy dimension that articulates a user's political preference. Our
de(cid:27)nition is inspired by work in political-science literature, such
as work by Bafumi et al. [2], where ideology is de(cid:27)ned as "a line
whose left end is understood to re(cid:30)ect an extremely liberal position
and whose right end corresponds to extreme conservatism."
Estimation of user ideology. Traditionally, the most common
sources for estimating ideology comprised of behavioral data gener-
ated from roll call votes [34], co-sponsorship records [1], or political
contributions [7]. These datasets were often only available for the
political elite, like members of congress, and hence getting such
estimates for a large population of ordinary citizens was di(cid:28)cult, if
not impossible.
With the proliferation of social media platforms, behavioral data
started being available at an individual level and researchers have
tried to use such data for identifying political ideology for social
media users at scale. Initial work started with supervised meth-
ods [10, 32] for predicting a (binary) political alignment of users
on Twitter. Though these works report accuracies over 90%, Cohen
at al. [9] warns about the limitations of such approaches and their
dependence on politically active users.
Unsupervised approaches have also been proposed, mainly based
on the structure of user interests [23], social connections [4], and
interactions [6, 16, 37]. The main idea behind these methods is
that users typically either surround themselves (follow/friend) with
other users who are similar in their ideology (homophily), or inter-
act with others (retweet/like) similar to them.
Perhaps the closest approach to this paper is the work by Lu et
al. [28] (biaswatch), who seek to identify the bias of a user on a
topic by combining their retweet and content networks, where a
content network is obtained based on the similarity of users tweets.
biaswatch, however, assumes the presence of a set of labeled
bias anchors (seed hashtags), making it not completely unsupervised.
Second, fusing the content and retweet networks is somewhat
arbitrary, since there is no common underlying principle that holds
the two networks together and hence a graph that results from
such a merger contains di(cid:29)erent types of edges (multigraph) simply
merged together. We compared our approach with biaswatch in
Section 5 and show that our method outperforms their bias scores.
Estimation of source ideology. Polarization and bias of media
outlets has existed long before the time of internet, however, data
availability makes it easier to study and quantify nowadays. Mitchell
et al. [29] study the media habits of American public using a large
scale survey and show how fractured the media production and con-
sumption have become in the recent years. Groseclosea et al. [21]
propose a method to estimate the ideology of various media sources
by comparing the number of citations to think tanks and policy
groups to those of Congress members. For a complete survey of
methods on measuring ideology of media and media bias, please
refer to Groeling et al. [20].
Reducing polarization. The problem of reducing polarization
and di(cid:29)using information bubbles on social media has been tack-
led before. Most papers approach the problem by recommending
content outside a user's (cid:27)lter bubble. Studies have looked at what
to recommend [19], how to recommend [25, 26] and to whom to
recommend such content [17].
Based on the papers discussed above, we make the following
observations:
a. Most existing approaches only consider one dimension (con-
tent or network) for computing user ideology. We propose a princi-
pled approach that can compute a user's ideology by taking into
account both content and network.
b. We compute the ideology of users and sources simultaneously.
No existing approach, to the best of our knowledge, proposes such
a method to do it simultaneously.
c. Most approaches for reducing polarization do not take a user's
choice into account. Our methods (in Section 6) can help users
di(cid:29)use their information bubbles based on their own choices.
3 PROBLEM FORMULATION
In this section we discuss the problem setting, introduce the nota-
tion, and formally de(cid:27)ne the problem we consider.
3.1 Data model
The social graph. We consider a social graph of Twitter users
G = (U , E, w), where U represents the set of users, E is the set
of edges representing social interactions between the users, and
w : E → R is a weighting function that assigns real-valued weights
to each edge in G.
We also represent the graph G by its adjacency matrix A ∈ Rn×n.
In particular, it is A(u, v) = wi for each edge ei = (u, v, wi) ∈ E,
and A(u, v) = 0 if there is no edge between u and v. Since the social
graph G may not represent a symmetric relation, the matrix A is
not necessarily symmetric.
To de(cid:27)ne the edges of the graph G one can consider social links
between users based on their retweet and follow networks. Two
commonly-used options are to consider an edge ei = (u, v, wi) if (i)
users u and v follow a common set of users, and wi is the number of
common users; or if (ii) user u retweets user v, and wi is the number
of retweets. In the (cid:27)rst case the social graph is symmetric, while in
the second case it is not.
Content sources. We model the presence of content in the network
by considering a set of content sources S. The set S represents a set
of items that are shared in the network. The number of sources is
denoted by S = m. We associate with each user u the subset of
sources in S that the user shares (i.e., posts) in the network. Thus
the overall user activity in the network is denoted by a matrix
C ∈ Rn×m, where C(u, s) denotes the number of times that user
u ∈ U shares source s ∈ S.
We derive the content features (i.e., items of the set S) based
on users' tweets. Speci(cid:27)cally, we experiment with two variants of
feature sets: (i) extracted urls from the tweets; and (ii) hostnames
of urls extracted from the tweets. In other words, in the latter case,
the content is aggregated by their source of authorship (i.e., various
news media channels).
For the presentation of our method, we also need the following
de(cid:27)nitions.
A(cid:28)nity matrix. Given a matrix X, we de(cid:27)ne the a(cid:28)nity matrix of
the rows of X to be XrnXT
rn, where Xrn is row normalized X. That is,
the a(cid:28)nity matrix of rows of X is formed by the cosine similarities
of all pairs of rows of X. Similarly, we de(cid:27)ne the a(cid:28)nity matrix of
the columns of X to be XT
cnXcn, where Xcn is column normalized X.
Laplacian matrix. Given a square and symmetric matrix X, we
consider its Laplacian to be the matrix L = D − X, where D is a
diagonal matrix whose (i, i) entry is the i-th row-sum of X.
3.2 Motivation of the approach
Our goal in this paper is to learn the ideological leaning of users
and content sources on Twitter. The underlying motivation is that
learning the ideological leaning of Twitter users and sources is an
important step in building useful tools that can help users perceive
their informational bias and consequently improve their news diet.
A simple approach to identify ideological leaning of users is to
consider only the social graph and apply one of the many community-
detection algorithms. The intuition here is that users with the same
ideological leaning are more likely to interact with each other, and
thus, to form graph communities. The drawback of this simple
approach is that community detection on real-world social graphs
is an extremely challenging task, due to sparsity and overlapping
communities. For instance, two users u and v can have the same
ideological leaning even though they do not have any social inter-
action, or they have di(cid:29)erent topical interests.
Similarly, one simple approach to attempt identifying the ideo-
logical leaning of sources is to cluster them using deep-NLP or
semantic-analysis techniques. However, this approach is again
prone to errors due to the inherent complexity of the text-analysis
task. Furthermore, it ignores the rich user information about how
content is shared in the social network.
In contrast to these simple techniques, which rely on one-sided
clustering of either users or content, with no association between
them, the proposed approach seeks to jointly learn the ideological
leaning of users and content sources.
Combining di(cid:29)erent data types in a uni(cid:27)ed learning framework
has several advantages. First, we exploit the inherent connection
between the two data types: users are not only more likely to
interact with, or follow, like-minded users, but also to share content
of aligned ideological leaning. Consequently, taking into account
the full information on social structure and content will result in
better clustering performance. Furthermore, considering both data
types simultaneously allows us to learn the ideological leanings of
users and content sources in a shared latent space. This means that
not only do we separate users and sources into ideological clusters,
but we also identify the relationship between the clusters of the
two data types. The applications we present in the Section 6 rely
heavily on the ability to represent users and content sources in a
shared latent space.
3.3 Problem formulation
To learn the latent space of the input data (users and sources) we use
non-negative matrix factorization (nmf) techniques. In particular,
we propose a joint matrix factorization formulation, which exploits
the duality between user and source clustering.
First, assume that the ideological leaning is represented by k
factors (dimensions). Our model is described by two components.
The (cid:27)rst component, represented by a n × k matrix U, captures user
information. In particular, the entry (u, i) of the matrix U represents
the degree to which user u aligns with ideology factor i.
The second component, represented by a m×k matrix V, captures
source information: the entry (s, i) of the matrix V represents the
degree to which source item s aligns with ideology factor i.
To determine the user and source ideology clusters, we decom-
pose the two input matrices A and C, using the components U
and V as latent factors. For decomposing A and C so as to capture
ideology clusters we require the following constraints.
Partitioning constraints:
1. Users in a user-cluster interact with each other more often than
with users outside the cluster.
2. Users in the same user-cluster post content from the same
3. Content in the same content-cluster is posted by users in the
content-cluster.
same user-cluster.
Co-partitioning constraints:
4. Users in a user-cluster share more articles from their corre-
sponding content-cluster than from other content-clusters.
5. Content in a content-cluster is shared by more users from their
corresponding user-cluster than from other user-clusters.
Non-negative matrix factorization for co-clustering: For a
given input data matrix X, the bi-orthogonal non-negative 3-factor
decomposition (onmtf) [13], formulated as X ≈ WHZT , provides
a good framework for simultaneously clustering the rows and the
columns of X. The left factor W provides a clustering of the rows
of X, while the right matrix Z provides a clustering of the columns
of X. The middle factor H provides association between the clusters
and additional degrees of freedom.
Problem 1. (onmtf [13]) Given an n×m matrix X, and integer k,
with k << n, m, (cid:27)nd non-negative matrices W, H, and Z, of respective
dimensions n × k, k × k, and k × m, so as to
minimize (cid:107)X − WHZT (cid:107)2
F ,
subject to W ≥ 0, H ≥ 0, Z ≥ 0,
and WT W = ZT Z = I.
subject to orthogonality constraints for U and V.
As shown by Ding et al. [13], the bi-orthogonality constraints
provide an interpretation of the onmtf problem as a simultaneous
clustering of the rows and columns of the input data matrix.
In our approach we use onmtf to decompose the input matrices
(social-graph matrix A and user–source matrix C) using the latent
factors U and V.
First, the user–source matrix is decomposed by C ≈ UHs VT ,
For the social-graph matrix A, as both the rows and columns
represent users, we equate the left and right factors and require
A ≈ UHu UT , subject to orthogonality constraints for U.
Note that as Hu is not necessarily symmetric, the decomposition
UHu UT can produce a non-symmetric matrix. Furthermore, the
formulation can capture link transitivity [39]: consider a path ui →
uk → uj. A non-symmetric factorization A ≈ XZT represents
the values of A as links from set of users to a set of objects, say,
O = {oi}. Hence, it would split the path (ui → uk → uj) into
two parts ui → ok and uk → oj, which is a misinterpretation
of the original path. Whereas, A ≈ UHu UT considers the links
to be amongst the same set of objects. Hence, the transitive link
ui → uk → uj is correctly captured by the latent factors in U.
Combining link and content information is achieved by using
a common latent factor U and formulating a joint factorization
problem asking to minimize
(cid:107)A − UHu UT (cid:107)2
F
+ (cid:107)C − UHs VT (cid:107)2
F ,
(1)
subject to non-negativity of U, Hu, V, Hs, and orthogonality of
U and V. This approach is inspired by the formulation of Zhu et
al. [39] for classifying web-pages by exploiting both content and
link information.
Graph regularization. From a geometric perspective, a dataset
can be viewed as a set of data points on a continuous manifold. The
task of clustering is to (cid:27)nd these intrinsic manifolds in the data.
However, the clustering formulation of onmtf fails to consider
this geometric structure in the data. To address this problem, Cai et
al. [8] introduced a graph-regularized nmf based on the manifold
assumption that if two data points xi, xj are close in the input data
matrix X, their projections ui and uj in the new basis U are also
close. This is formulated by seeking to minimize
(cid:107)ui − uj (cid:107)2Wij = tr(UT LU),
1
2
i, j
where W is the a(cid:28)nity matrix of rows of X (see Section 3.1), and L
is the laplacian of W.
Motivated by the duality between row and column manifolds,
Gu et al. [22] proposed a dual-manifold regularized co-clustering
method as a decomposition of an input matrix X, by asking to
minimize
(cid:107)X − WHZT (cid:107)2
F
+ α · tr(WT Lw W) + β · tr(ZT LzZ),
subject to non-negativity constraints of the factor matrices. Here
Lw and Lz are the Laplacians on the a(cid:28)nity matrices of rows and
columns of X, respectively. As before, the matrix H captures the
association between row and column clusters.
We extend our joint factorization model (Equation 1) by including
dual-graph regularization constraints on users and sources. We
refer to this problem by ifd, for ideology factor decomposition. The
formal problem de(cid:27)nition is the following.
Problem 2 (ifd). Given a user–user social matrix A of dimension
n × n, a user–source matrix C of dimension n × m, an integer k, with
k << n, m, and regularization parameters α and β, (cid:27)nd factors U,
Hu , V, Hs , of dimensions n × k, k × k, m × k, n × k, respectively, so
as to
minimize (cid:107)A − UHu UT (cid:107)2
F
+ (cid:107)C − UHs VT (cid:107)2
F
+ α · tr(UT Lu U) + β · tr(VT Ls V),
subject to U ≥ 0, Hu ≥ 0, V ≥ 0, Hs ≥ 0,
(2)
and UT U = VT V = I,
where Lu and Ls are the Laplacians of the a(cid:28)nity matrices of rows
(users) and columns (sources) of C, respectively.
In the ifd problem we factorize A and C jointly based on the dual
manifold assumption i.e., both users and content share the same
latent space and the cluster labels of users are smooth with respect
to the content manifold, while the cluster labels of content are
smooth with respect to the user manifold. To apply these manifold
constraints, we consider a(cid:28)nity matrices for users and content
sources. While there are many ways to construct such an a(cid:28)nity
matrices, in our experiments we consider the a(cid:28)nity matrices of
the rows (for users) and columns (for sources) of the C matrix.
Finally, it is instructive to review how ifd addresses the par-
titioning constraints 1–3 and co-partitioning constraints 4–5: the
bi-orthogonality constraints on the tri-factorization problem pro-
vide a clustering interpretation for the rows and columns of the
input matrices, while the manifold constraints provide a geometric
interpretation to the discovered latent space. The correspondance
between the user clusters and source clusters is achieved by using
a shared latent factor.
3.4 Solving the optimization problem
Following the standard theory of constrained optimization, we solve
Problem 2 by introducing Lagrangian multipliers Λ (a symmetric
matrix of size k × k) and minimizing the Lagrangian function
L = (cid:107)A − UHu UT (cid:107)2
F
+ (cid:107)C − UHs VT (cid:107)2
F
+α · tr(UT Lu U) + β · tr(VT Ls V)
tr(Λ(UT U − I)) + tr(Λ(VT V − I)).
(3)
We can compute the gradient of L with respect to U, V, Hu,
and Hs. A locally-optimal solution for Problem 2 can be found
using an iterative-update algorithm, similar to the one proposed by
Ding et al. [13]. The multiplicative update rules are as follows:
U ← U
V ← V
(cid:118)(cid:116)
(cid:115)
(cid:115)
(cid:115)
H1 ← H1
H3 ← H3
U T AU
U T U H1U T U
U T CV
U T U H3V T V
,
,
AU HT1 + CV HT3 + αSuU
U H1U T U HT1 + U H3V T V HT3 + αDuU + U λu
CT U H3 + βSdV
βDdV + V H3U T U H3 + V λv
,
,
(4)
(5)
(6)
(7)
where
λu = U T AU HT1 + U T CV HT3 − αU T LuU − H1U T U HT1 − H3V T V HT3
and λv = V T CT U H3 − βV T LdV − H3U T U H3.
4 ESTIMATING IDEOLOGICAL LEANING
In this section we discuss how to use the latent factors U and V com-
puted using ifd in order to estimate ideological leaning scores for
Twitter users and media channels (sources). Our approach utilizes
the probabilistic model of nmf factorizations, discussed in Yoo et
al. [38], to derive a probabilistic interpretation of latent factors. We
also discuss how to derive a hard cross-ideological separation (hard
clustering) from the latent factors.
As proposed in the literature [24, 38], the latent factors U and V
have the following probabilistic interpretation:
– the entry (i, (cid:96)) of matrix U indicates the degree to which user i
– the entry (j, (cid:96)) of matrix V indicates the degree to which media
belongs to the user-cluster (cid:96); and
source j belongs to the content-cluster (cid:96).
In the context of our problem setting, we are interested in identify-
ing two main ideologies, liberal and conservative, and thus, we set
the number of latent dimensions equal to 2 (k = 2). It follows that
each user and each media source are represented by a 2-dimensional
vector (x, y) in the latent space - a row of U for users, and a row of
V for media sources. Thus, user and source ideology hard clusters
are derived as arg max Uij and arg max Vij, respectively. To esti-
mate a single score for users and media sources, we compute the
angle of line de(cid:27)ned by the center of origin and the latent vector
(x, y), normalized to be between 0 and 1, i.e.,1
arctan(y/x)
i(x, y) = θ
π/2
1Recall that arctan(0) = 0 and limz→∞ arctan(z) = π/2.
π/2 =
.
(8)
(a) original latent space
(b) transformed coordinate space
(cid:113)
Figure 1: Projection of a subset data points in the learned ide-
ology latent space and the transformed ideology-popularity
coordinate space.
We also compute the magnitude of the latent vector (x, y)
2
,
x
ρ(x, y) =
2 + y
(9)
which, as can be shown easily, represents the popularity of the cor-
responding user or source - in particular, it is correlated with the
number of re-tweets and follows of a given user, and the number of
tweets containing a given source. Figure 1 visualizes a subset of real
data points projected in the original latent space and their transfor-
mation to the corresponding ideology/popularity co-ordinate space
according to the aforementioned computations.
In summary, given a user or media source represented by a latent
vector (x, y), we can estimate a single ideological leaning score by
Equation (8), as well as its popularity by Equation (9). When using
more than 2 dimensions for the latent ideology space (k > 2) it is
not possible to estimate ideology scores with a single number, but
we can still handle the user and source representations by standard
vector operations.
5 EXPERIMENTS
5.1 Dataset
Our dataset is collected using Twitter's streaming API from 2011 to
2016, by (cid:27)ltering for keywords related to three popular controver-
sial topics: gun control, abortion and obamacare. We use the list of
keywords proposed by Lu et al. [28] to (cid:27)lter the tweets related to
these topics. We only consider users who tweeted about all three
topics at least once, obtaining a set of n = 6 391 users, and collect
all their tweets, which gave us 19 million tweets.
As discussed in Section 3, we consider two variants of matrix
A (retweet and follow) and two variants of matrix C (urls and host-
names). We observe that the follow user-user matrix A along with
hostname user-content matrix C gives signi(cid:27)cantly superior results
compared to all other variants, thus, all subsequent results use that
variant. We omit results with the other variants due to lack of space.
5.2 Ground truth
Ideology scores for sources: We collect ground truth for news-
media channels from multiple studies in the literature: (i) 500 most
shared news domains on Facebook [3] (ii) 100 most visited domains
in Bing toolbar [15] and (iii) 27 domains from an o(cid:31)ine survey and
latentdimension1(x)latentdimension2(y)011hx,yiρθideologyscoreipopularityscoreρ011(i,ρ) webpage visit data [18]. Each of these scores roughly measures the
fraction of views/shares/clicks by a conservative user. We map all
scores in the [0, 1] range, 1 being conservative. For the domains
listed in multiple lists we compute the ideology by averaging the
scores. We also remove domains that are not necessarily news
sources (e.g., wikipedia.org, reddit.com, etc.). In total, we collect 559
news domains with ground-truth ideology scores. We refer to this
dataset as content ground truth.
Ideology scores for Twitter users: We use two di(cid:29)erent ground-
truth scores for users: (i) barbera: ideology score estimated by
Barberá et al. [5], which applies Bayesian ideal point estimate on
nearly 12 million Twitter users, and (ii) avg_content: average
ground-truth ideology scores of the sources tweeted by the user.
Popularity scores for sources: We use the aggregated number
of tweets about each news media channel in the collected data set
as a proxy for the popularity of the source.
Popularity scores for users: Since the collection of users is a
random set of people on user, we do not have any ground truth for
popularity of Twitter users.
5.3 Baseline algorithms
We compare our method with three types of methods for ideol-
ogy detection: network-only, content-only, and a combination of
network and content.
Network-only: We consider two types of network-only methods:
(i) nmf-based methods that can provide a continuous ideology
score for a user between 0 and 1; and (ii) other methods that only
produce binary labels for ideology (a user is either liberal or conser-
vative). We use symmetric nmf (nmf-symm) [13], a 3-factor nmf
shown to be equivalent to normalized-cut spectral clustering [12],
retweet a method based on partitioning the retweet graph [16]
and follow a graph partitioning approach on the follow network.
In order to construct a source-source relationship matrix, we use
CT C. It is noteworthy that network-only methods perform only
one-side clustering - one data type at a time. Hence, we need to ap-
ply the methods separately for users and content sources. As such,
network-only methods do not provide any information about the
correspondence between the two clusterings. Further, retweet
and follow return only binary labels, hence we do not use this
baseline for comparing ideology scores.
Content-only: We use orthogonalnmf tri-factorization (onmtf),
a co-clustering approach [13], and dual manifold co-clustering
(dmcc) [22]. In these methods the bipartite content matrix C is used
to co-cluster the rows (users) and columns (sources) of the matrix
simultaneously using bi-orthogonality and graph-regularization
constraints.
Network and content: We compute ideology scores of Twitter
users estimated by kulshrestha et al. [23] (kulshrestha) and
Lu et al. [28] (biaswatch).
Proposed methods: We use the proposed method ifd, and a vari-
ant of ifd without graph-regularization constraints (ifd-ngr).
We initialized U and V randomly from a uniform distribution in
[0,1] and Hu and Hs as identity matrices of size k. Parameters α
and β are chosen using grid search. Additional details on various
approaches tried for parameter initialization, parameter tuning,
Figure 2: Popular media outlets and their ideology leaning
scores computed by our method.
and stability of the algorithms with respect to the parameters are
omitted due to lack of space, and will be provided in the full version
of the paper.
5.4 Experimental setup
Evaluation measures. We perform two types of qualitative evalu-
ation tasks: (i) quality of ideological cluster separation (into liberal
or conservative clusters) and (ii) correlation between the computed
ideology scores and ground-truth scores. In order to evaluate clus-
ter separation, we measure purity, adjusted Rand index (ari),
adjusted mutual information (ami), and normalized mutual infor-
mation (nmi) between the clusters detected by the algorithm and
the set of ground-truth communities derived by separating users
at ideology score threshold at 0.5. In order to measure correlation
between the computed ideology scores and the ground-truth scores,
we use Pearson mutual correlation coe(cid:28)cient (corr).
5.5 Results
Ideology estimates for users and sources. At a (cid:27)rst look, the
user ideology scores seem intuitive with the top liberal users being
@barackobama (score: 0.0), @berniesanders (0.0), @thedemocrats
(0.0) and top conservative users @tedcruz (score: 0.99), @sean-
hannity (0.99), and @davidlimbaugh (0.9). Figure 2 shows popular
news-media outlets and their ideology leaning scores computed by
our method. We observe that the position of the news sources is as
expected: Liberal-leaning news outlets (e.g., nytimes, washington
post, the guardian) are on the left, and conservative news outlets
(e.g., fox news, breitbart, rushlimbaugh) on the right. This is also
consistent with the survey-based results found by [29]. While it
is easy to identify the extreme left and right, it is more di(cid:28)cult to
identify the neutral users and sources (like yahoo, mediaite, white-
house.gov, etc), which, in fact, is the most important subset of users
and sources to tackle the information (cid:27)lter-bubble issue.
Evaluation of clustering and ideology scores. We compare the
proposed methods with the baselines on (i) quality of ideological
cluster separation (purity, ari, ami and nmi) and (ii) correla-
tion (corr) between computed ideology score and ground-truth
00.51nytimeshuffingtonposttheguardiandailykosmsnbcwashingtonpostcnnpoliticobbcthehillusatodayreutersforbesbloombergchicagotribuneyahoomediaitewhitehouse.govwsjtelegraphnationalreviewdailymail.co.ukfreebeaconbreitbartfoxnewsdailycallerwashingtonexaminerthegatewaypundittheblazerushlimbaughliberalconservative scores. The results of both evaluation tasks are listed in Table 1.
Following are some noteworthy observations:
(i) ifd has the best performance among all the methods for
both user as well as source clustering. We observe that combining
network and content information gives consistently best results.
For example, purity of clustering for combined methods is 20%
higher than content-only methods for users and 27% higher for
source clustering.
(ii) When comparing ideology scores, ifd performs better than
state of the art baselines kulshrestha and biaswatch. Note
that biaswatch almost has no correlation with ground truth
and performs poorly. This could be due to reasons mentioned in
Section 2. Though kulshrestha works slightly better on the
clustering task, it doesn't do so well with the correlation. This indi-
cates that our method is able to identify the (cid:27)ne grained ideology
scores better, where as kulshrestha can do better at separating
users into binary clusters.
(iii) nmf-symm, a network-only method, performs quite well
for user clustering, whereas the results are not satisfactory for
source clustering. Perhaps this can be attributed to the noise in
the input matrix CT C, caused by, e.g., sparsity and topical diver-
sity of user interests, which is mitigated in ifd because of the
joint factorization. Similarly, our method performs much better
than retweet and follow, which do not use both content and
network information.
Audience of news sources. Figure 3 shows a kernel density es-
timate (kde) of ideology scores computed using the proposed
method (solid line) and ground truth of ideology scores (dashed
line) for all 6 391 users for a selection of 5 representative news
channels. The (cid:27)ndings from our experiments are strikingly similar
to the results computed using extensive user surveys [29]. We can
clearly observe that there is a non-trivial association between news
sources and polarization of users.
6 DIFFUSING THE INFORMATION BUBBLE
In the recent years, a few papers [17] have aimed to solve the
problem of echo chambers and (cid:27)lter bubbles on social media by
connecting users with others outside of their bubble. The problem
with most of these approaches is that they typically suggest tech-
nical solutions that do not take into account user biases, such as,
cognitive dissonance [14] and biased assimilation [27]. In addition,
in many cases, users themselves are not aware of being present in a
bubble, due to the non-transparent nature of algorithmic (cid:27)ltering.2
To handle both these issues, we propose that a better solution to
take users out of their bubble is to give them information and choice.
We do this in two steps: (i) make a user aware of their information
bubble, and (ii) provide content recommendations that can help
users di(cid:29)use their bubble. In this section, we demonstrate how to
use the latent space representations learned in this paper to perform
both of these steps.
2E.g., the (cid:27)lter bubble is claimed to be a reason why many people did not predict
correctly the results of Brexit or US elections
http://(cid:30)amingogroup.com/trump-brexit-and-why-we-didnt-see-them-coming
Table 1: Comparison of the proposed method with baseline
methods on (i) quality of ideological cluster separation (pu-
rity, ari, ami and nmi) and (ii) correlation (corr) between
computed ideology score and ground truth. corr for ideol-
ogy is represented as corri and for popularity as corrρ.
Since, we do not have ground truth for popularity of users
we do not compute corrρ for users. Best results for each
measure are marked in bold.
(a) Evaluation for users using barbera ground truth (upper value
in each row) and avg_content ground truth (lower value in
each row)
Type
Network Only
Content Only
Network + Content
Method
nmf-symm
retweet
follow
onmtf
dmcc
ifd
ifd-ngr
kulshrestha
biaswatch
purity ari
0.928
0.733
0.522
0.861
0.476
0.844
0.461
0.839
0.538
0.867
0.476
0.845
0.743
0.234
0.247
0.749
0.229
0.74
0.241
0.746
0.722
0.925
0.528
0.863
0.722
0.925
0.863
0.528
0.744
0.931
0.547
0.869
0.000
0.541
0.543
0.000
ami
0.628
0.418
0.385
0.395
0.454
0.382
0.233
0.223
0.23
0.218
0.62
0.441
0.62
0.441
0.637
0.448
0.000
0.000
nmi
0.629
0.418
0.385
0.399
0.456
0.382
0.266
0.246
0.263
0.242
0.621
0.442
0.621
0.442
0.638
0.449
0.000
0.000
corri
0.912
0.744
-
-
-
-
0.756
0.715
0.755
0.715
0.904
0.772
0.904
0.772
0.875
0.744
0.002
0.018
corrρ
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
(b) Evaluation for sources using content ground truth
0.752
-
0.746
0.746
0.827
0.813
nmf-symm 0.597
follow
0.819
onmtf
0.606
dmcc
0.606
ifd
0.826
ifd-ngr
0.822
0.031
0.405
0.039
0.039
0.424
0.415
0.135
0.318
0.145
0.145
0.346
0.339
0.171
0.32
0.181
0.181
0.348
0.341
Network Only
Content Only
Network + Content
0.597
-
0.593
0.592
0.929
0.93
6.1 Visualizing the information bubble
Recent studies have shown that making users aware of their im-
balance in media consumption can encourage them to make small
yet signi(cid:27)cant improvements in increasing the diversity of their
reading [30]. Inspired by these studies, we suggest providing to the
users a visual way to explore their information (cid:27)lter bubble.
The way we approach this is by visualizing the user and content
in the same space to allow users to explore their content consump-
tion. To this end, we use the learned latent space to map the users
and content in the same space. Using their estimated ideological po-
sitions (from Section 4), we project users as well as sources in a two
dimensional ideology-popularity coordinate space with computed
ideology score on x-axis and popularity score on y-axis. Since we
do not have popularity score for users, we determine the position of
a user on y-axis using the average popularity score of the content
that the user engages with. Since all user scores are on the same
scale and relative to each other, we would observe that users and
sources with similar ideology lie close to each other. Finally, we
connect users to the sources that they consume by drawing a link
between them. The size of a source node is proportionate to the
Figure 3: Polarization of the audience of news sources. Values on the x-axis represent the ideology score of users and values
on the y-axis represent the kernel density estimate of the number of users at each point.
(a) Democratic Party
(b) Republican Party
Figure 4: Ideological position of @thedemocrats and @gop
(black dots) and their content engagement. Points in the grey
are the sources that the user never interacted with.
number of times a user has consumed content from the said source.
In order to increase the ease of visual interpretation, we color the
content according to the ideological learning (blue: liberal, green:
neutral and red: conservative). Content not consumed by the user
is colored gray.
Figure 4 presents a prototype for two popular Twitter accounts
from the two ends of the political spectrum: the Republican Party
(@gop) and the Democratic party (@thedemocrats).3 From this
(cid:27)gure, one can visually observe their own ideological positioning
as well as the ideology of the content that they engage with. For
instance, @thedemocrats is heavily liberal in their ideology (ide-
ology score 0.0). The content consumed by @thedemocrats is also
heavily biased on the liberal side. As expected, a large fraction of the
content they engage with is from the left (mainly liberal media like
nytimes.com and washingtonpost.com), and negligible amount from
the opposite point of view, whereas the opposite is true for @gop.
It is interesting to observe that the Republican party account has a
higher engagement with diverse view points than the Democrats.
6.2 Making ideologically diverse content
recommendations
Garimella et al. [17] proposed an approach to di(cid:29)use a user's (cid:27)lter
bubble by connecting him to a user outside his bubble from the
opposing viewpoint. Their approach is mainly based on identify-
ing users from opposing sides and optimizing a global function.
Here, we build on top of that idea and use our computed ideology
to di(cid:29)use a user's bubble by recommending him content from an
3An interactive web version of these plots can be accessed at
http://resources.mpi-inf.mpg.de/d5/(cid:27)lterbubble.
(a) original latent space
(b) transformed latent space
Figure 5: Logical diagram of user content recommendation
by sampling from the Gaussian over "ideology" and "popu-
larity" positioning.
opposing viewpoint, along with an option to choose how willing
the user is to explore the other side. Recommending ideologically
diverse content to a user can be controlled by the user using two
parameters: ideology tolerance threshold θ and popularity thresh-
old δ. Intuitively, a user is more likely to accept content within the
region +θ and −θ on either side of the user's ideological positioning,
and +δ and −δ on either side of his popularity position. Figure 5a
visualizes a hypothetical user in the original ideology latent space
and the transformed ideology-popularity coordinate space (detailed
in Section 4). Consider that we build two Gaussian distributions
around the user box (see Figure 5b) with their means centered at
user's ideology and popularity score respectively, and variance as
a function of the tolerance threshold given as input by the user.
We can now sample content from these Gaussian distributions and
use it for recommending content to the user. As desired, in such a
sampling, the content close to the user's own ideology and popu-
larity score has a higher probability of being selected. As we move
closer to the thresholds, the probability of an article being selected
gradually decreases. This "box" gives the space of exploration for a
user and depending on the user's willingness to explore (based on
parameters θ, δ), they can see content outside their bubble.
7 CONCLUSIONS
We considered the problem of identifying ideological leaning of
users and news sources (content) on Twitter. The paper tackles
two main challenges: (i) learning the ideological latent factors of
users and content in a joint model that explores simultaneously
user-to-user and user-to-content relations; and (ii) embedding the
discovered factors in a common latent space so as to support visu-
alization and exploration of the results. Our approach distinguishes
/044
/0389: 00/
/044
047%208
/044
,$970094:73,
/044
4 08
/044
%0,0.425:90/8.47074:3/97:9
/0448.470
545:,798.470,7.3943,54 89-442-07-:100/:1139435489549.439 28,83943548939208;490/024.7,98
/0448.470
14308890549.4,839435489,839430,2307549 .4-442-0739208;490 4545 !latentdimension1(x)latentdimension2(y)011userθθδδideologyscoreipopularityscoreρ011userδδθπ/2θπ/2 itself from most existing work in the area in three major ways. First,
our model aims to learn ideology on a continuous scale rather than
a binary liberal-conservative opinion, which is a much simpler and
well studied problem. Second, our model de(cid:27)nes polarization as a
multidimensional problem and allows for learning any number of
dimensions (ideology, popularity, etc) rather than just one (ideol-
ogy). Third, our model can identify the ideology of both the user
and content simultaneously, thus making use of the interdependent
structure between content and network. We also demonstrate how
to use the learned latent space in an application of providing tools
to visualize a users information bubble and for ideologically di-
verse content recommendation with the purpose of di(cid:29)using their
information (cid:27)lter bubble.
Future work. One of the main focuses of this work was to create
a strong technical foundation to understand the problem of online
polarization. We believe that the methods presented in this paper
provide several avenues for future work in multiple emerging in-
terdisciplinary research areas, for instance, humanly interpretable
and explanatory machine learning, transparent recommendations
and a new era of social media platforms that encourage discussion
and debates between users of diverse view points, thus helping to
reduce the ideological segregation of users instead of reinforcing it.
Acknowledgments. This work was supported by the Academy of
Finland projects "Nestor" (286211) and "Agra" (313927), and the EC
H2020 RIA project "SoBigData" (654024).
8 REFERENCES
[1] E. Alemán, E. Calvo, M. P. Jones, and N. Kaplan. Comparing
cosponsorship and roll-call ideal points. Legislative Studies Quarterly,
34(1):87–116, 2009.
[2] J. Bafumi, A. Gelman, D. K. Park, and N. Kaplan. Practical issues in
implementing and understanding bayesian ideal point estimation.
Political Analysis, 13(2):171–187, 2005.
[3] E. Bakshy, S. Messing, and L. A. Adamic. Exposure to ideologically
diverse news and opinion on facebook. Science, 348(6239):1130–1132,
2015.
[4] P. Barberá. Birds of the same feather tweet together: Bayesian ideal
point estimation using twitter data. Political Analysis, 23(1):76–91,
2015.
[5] P. Barberá, J. T. Jost, J. Nagler, J. A. Tucker, and R. Bonneau. Tweeting
from left to right: Is online political communication more than an
echo chamber? Psychological science, 26(10):1531–1542, 2015.
[6] R. Bond and S. Messing. Quantifying social media's political space:
Estimating ideology from publicly revealed preferences on facebook.
American Political Science Review, 109(1):62–78, 2015.
[7] A. Bonica. Ideology and interests in the political marketplace.
American Journal of Political Science, 57, 2013.
[8] D. Cai, X. He, X. Wu, and J. Han. Non-negative matrix factorization
on manifold. In ICDM, pages 63–72. IEEE, 2008.
[9] R. Cohen and D. Ruths. Classifying political orientation on twitter:
It's not easy! In ICWSM, 2013.
[10] M. D. Conover, B. Gonçalves, J. Ratkiewicz, A. Flammini, and
F. Menczer. Predicting the political alignment of twitter users. In
SocialCom, pages 192–199. IEEE, 2011.
[11] M. Dimock, C. Doherty, J. Kiley, and R. Oates. Political polarization in
the american public. Pew Research Center, 2014.
[12] C. Ding, X. He, and H. D. Simon. On the equivalence of nonnegative
matrix factorization and spectral clustering. In SDM, pages 606–610.
SIAM, 2005.
[13] C. Ding, T. Li, W. Peng, and H. Park. Orthogonal nonnegative matrix
tri-factorizations for clustering. In KDD, pages 126–135. ACM, 2006.
[14] L. Festinger. A theory of cognitive dissonance, volume 2. Stanford
university press, 1962.
[15] S. Flaxman, S. Goel, and J. M. Rao. Filter bubbles, echo chambers, and
online news consumption. Public Opinion Quarterly, 80(S1):298–320,
2016.
[16] K. Garimella, G. De Francisci Morales, A. Gionis, and
M. Mathioudakis. Quantifying controversy in social media. In WSDM,
pages 33–42. ACM, 2016.
[17] K. Garimella, G. De Francisci Morales, A. Gionis, and
M. Mathioudakis. Reducing controversy by connecting opposing
views. In WSDM, pages 81–90. ACM, 2017.
[18] M. Gentzkow and J. M. Shapiro. Ideological segregation online and
o(cid:31)ine. The Quarterly Journal of Economics, 126(4):1799–1839, 2011.
[19] E. Graells-Garrido, M. Lalmas, and D. Quercia. People of opposing
views can share common interests. In WWW Companion, pages
281–282. ACM, 2014.
[20] T. Groeling. Media bias by the numbers: Challenges and
opportunities in the empirical study of partisan news. Annual Review
of Political Science, 16, 2013.
[21] T. Groseclose and J. Milyo. A measure of media bias. The Quarterly
Journal of Economics, 120(4):1191–1237, 2005.
[22] Q. Gu and J. Zhou. Co-clustering on manifolds. In KDD, pages
359–368. ACM, 2009.
[23] J. Kulshrestha, M. Eslami, J. Messias, M. B. Zafar, S. Ghosh, K. P.
Gummadi, and K. Karahalios. Quantifying search bias: Investigating
sources of bias for political searches in social media. In CSCW, pages
417–432. ACM, 2017.
[24] T. Li and C. Ding. The relationships among various nonnegative
matrix factorization methods for clustering. In ICDM, pages 362–371.
IEEE, 2006.
[25] Q. V. Liao and W.-T. Fu. Can you hear me now?: mitigating the echo
chamber e(cid:29)ect by source position indicators. In CSCW, pages
184–196. ACM, 2014.
[26] Q. V. Liao and W.-T. Fu. Expert voices in echo chambers: e(cid:29)ects of
source expertise indicators on exposure to diverse opinions. In CHI,
pages 2745–2754. ACM, 2014.
[27] C. G. Lord, L. Ross, and M. R. Lepper. Biased assimilation and attitude
polarization: The e(cid:29)ects of prior theories on subsequently considered
evidence. Journal of personality and social psychology, 37(11):2098,
1979.
[28] H. Lu, J. Caverlee, and W. Niu. Biaswatch: A lightweight system for
discovering and tracking topic-sensitive opinion bias in social media.
In CIKM, pages 213–222. ACM, 2015.
[29] A. Mitchell, J. Gottfried, J. Kiley, and K. E. Matsa. Political
polarization & media habits. Pew Research Center, 2014.
[30] S. A. Munson, S. Y. Lee, and P. Resnick. Encouraging reading of
diverse political viewpoints with a browser widget. In ICWSM, 2013.
[31] E. Pariser. The (cid:27)lter bubble: What the Internet is hiding from you.
Penguin UK, 2011.
[32] M. Pennacchiotti and A.-M. Popescu. Democrats, republicans and
starbucks a(cid:28)cionados: user classi(cid:27)cation in twitter. In KDD, pages
430–438. ACM, 2011.
[33] F. Pla and L. F. Hurtado. Political tendency identi(cid:27)cation in twitter
using sentiment analysis techniques. In COLING, pages 183–192,
2014.
[34] K. T. Poole and H. L. Rosenthal. Ideology and congress, volume 1.
Transaction Publishers, 1997.
[35] M. Prior. Post-broadcast democracy: How media choice increases
inequality in political involvement and polarizes elections. Cambridge
University Press, 2007.
[36] C. R. Sunstein. Republic. com 2.0. Princeton University Press, 2009.
[37] F. M. F. Wong, C. W. Tan, S. Sen, and M. Chiang. Quantifying political
leaning from tweets, retweets, and retweeters. TKDE,
28(8):2158–2172, 2016.
[38] J. Yoo and S. Choi. Orthogonal nonnegative matrix tri-factorization
for co-clustering: Multiplicative updates on stiefel manifolds.
Information processing & management, 46(5):559–570, 2010.
[39] S. Zhu, K. Yu, Y. Chi, and Y. Gong. Combining content and link for
classi(cid:27)cation using matrix factorization. In SIGIR, pages 487–494.
ACM, 2007.
|
1704.06920 | 1 | 1704 | 2017-04-23T13:22:41 | Maximizing the overall profit of a word-of-mouth marketing campaign: A modeling study | [
"cs.SI"
] | As compared to the traditional advertising, the word-of-mouth (WOM) communications have striking advantages such as significantly lower cost and rapid delivery; this is especially the case with the popularity of online social networks. This paper addresses the issue of maximizing the overall profit of a WOM marketing campaign. A marketing process with both positive and negative WOM is modeled as a dynamical model knwn as the SIPNS model, and the profit maximization problem is modeled as a constrained optimization problem. The influence of different factors on the dynamics of the SIPNS model is revealed experimentally. Also, the impact of different factors on the expected overall profit of a WOM marketing campaign is uncovered experimentally. On this basis, some promotion strategies are suggested. To our knowledge, this is the first time a WOM marketing campaign is treated this way. | cs.SI | cs | Maximizing the overall profit of a word-of-mouth marketing campaign:
A modeling study
Pengdeng Lia, Xiaofan Yanga, Lu-Xing Yanga,b,∗, Qingyu Xionga, Yingbo Wua, Yuan Yan Tangc
bFaculty of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology, Delft, GA 2600, The Netherlands
cDepartment of Computer and Information Science, The University of Macau, Macau
aSchool of Software Engineering, Chongqing University, Chongqing, 400044, China
7
1
0
2
r
p
A
3
2
]
I
S
.
s
c
[
1
v
0
2
9
6
0
.
4
0
7
1
:
v
i
X
r
a
Abstract
As compared to the traditional advertising, the word-of-mouth (WOM) communications have striking advantages such
as significantly lower cost and rapid delivery; this is especially the case with the popularity of online social networks.
This paper addresses the issue of maximizing the overall profit of a WOM marketing campaign. A marketing process
with both positive and negative WOM is modeled as a dynamical model knwn as the SIPNS model, and the profit
maximization problem is modeled as a constrained optimization problem. The influence of different factors on the
dynamics of the SIPNS model is revealed experimentally. Also, the impact of different factors on the expected overall
profit of a WOM marketing campaign is uncovered experimentally. On this basis, some promotion strategies are
suggested. To our knowledge, this is the first time a WOM marketing campaign is treated this way.
Keywords: word-of-mouth, marketing campaign, overall profit, dynamical model, constrained optimization
1. Introduction
Promotion is a common form of product sales. The third-party advertising on mass media such as TV and news-
paper has long been taken as the major means of promoting products. However, this promotion strategy suffers from
high cost [1, 2]. Furthermore, it has been found that, beyond the early stage of product promotion, the efficacy of
advertising diminishes [3].
Word-of-mouth (WOM) communications are a pervasive and intriguing phenomenon. It has been found that both
satisfied and dissatisfied consumers tend to spread positive and negative WOM, respectively, regarding the products
they have purchased and used [4, 5]. As compared to positive WOM, negative WOM is more emotional and is more
likely to influence the receiver's opinion. By contrast, positive WOM is more cognitive and more considered [6 -- 9].
The significant role of WOM in product sales is supported by broad agreement among practitioners and academics.
Both positive and negative WOM will affect consumers' purchase decisions. Due to significantly lower cost and
fast propagation, WOM-based promotion campaign outperforms their advertising counterparts [10, 11]. With the
increasing popularity of online social networks such as Facebook, Myspace, and Twitter, WOM has come to be the
main form of product marketing [12].
Currently, the major concern on WOM-based viral marketing focuses on the influence maximization problem: find
a set of seeds such that the expected number of individuals activated from this seed set is maximized [13]. Toward
this direction, large number of seeding algorithms have been proposed [14 -- 23]. Additionally, a number of epidemic
models capturing the WOM spreading process have been suggested [24 -- 34]. However, all of these works build on the
premise that a single product or a few competing products are to be promoted.
Typically, a large-scale marketing campaign often involves a consistent collection of products, that is, a customer
may purchase more than one product out of the collection. The ultimate goal of such a marketing campaign is to
∗Corresponding author
Email addresses: [email protected] (Pengdeng Li), [email protected] (Xiaofan Yang), [email protected] (Lu-Xing
Yang), [email protected] (Qingyu Xiong), [email protected] (Yingbo Wu), [email protected] (Yuan Yan Tang)
Preprint submitted to Physica A
October 8, 2018
maximize the overall profit. To achieve the goal, it is crucial to determine those factors that have significant influence
on the profit. To our knowledge, so far there is no literature in this aspect.
This paper addresses the issue of maximizing the overall profit of a WOM marketing campaign with rich products.
A marketing process with both positive and negative WOM is modeled as a dynamical model known as the SIPNS
model, and the profit maximization is modeled as an optimization problem. The influence of different factors on
the dynamics of the SIPNS model is revealed experimentally. Also, the impact of different factors on the expected
overall profit of a WOM marketing campaign is uncovered experimentally. On this basis, some promotion strategies
are recommended. To our knowledge, this is the first time a WOM marketing campaign is treated this way.
The subsequent materials are organized as follows. Section 2 models a WOM marketing process as the SIPNS
model, and models the profit maximization as an optimization problem. Section 3 experimentally reveals the influence
of different factors on the dynamics of the SIPNS model. Section 4 experimentally uncovers the influence of different
factors on the expected overall profit of a WOM marketing campaign. Finally, Section 5 summarizes this work and
points out some future topics.
2. The problem and its modeling
Suppose a marketer will launch a WOM promotion campaign for promoting a batch of products, with the goal
of achieving the maximum possible profit. To achieve the goal, let us modeling the profit maximization problem as
follows.
Let [0, T] denote the time interval for the campaign, Mt the target market at time t for the campaign, M(t) = Mt
the number of individuals in Mt. Here, Mt is classified as four categories, which are listed below.
(a) Susceptible individuals, i.e., those who haven't recently bought any product but are inclined to buy one.
(b) Infected individuals, i.e., those who have recently bought a product but haven't made comment on it.
(c) Positive individuals, i.e., those who have recently bought a product and have made a positive comment on it.
(d) Negative individuals, i.e., those who have recently bought a product and have made a negative comment on it.
Let S (t), I(t), P(t), and N(t) denote the expected number at time t of susceptible, infected, positive, and negative
individuals in Mt, respectively. Then, S (t) + I(t) + P(t) + N(t) = M(t).
Now, let us impose a set of statistical assumptions as follows.
(H1) At any time, the individuals outside of the target market enter the market at constant rate µ > 0, where µ is
referred to as the entrance rate.
(H2) At any time, a positive individual spontaneously exits from the market at constant rate δP, where δP is referred
to as the P-exit rate.
(H3) At any time, an infected individual spontaneously exits from the market at constant rate δI, where δI is referred
to as the I-exit rate.
(H4) At any time, a negative individual spontaneously exits from the market at constant rate δN, where δN is referred
to as the N-exit rate.
(H5) Encouraged by positive comments, at time t a susceptible individual buys a product (and hence becomes in-
fected) at rate βPP(t), where the constant βP > 0 is referred to as the P-infection force.
(H6) Discouraged by negative comments, at time t a susceptible individual exits from the market at rate βNN(t),
where the constant βN > 0 is referred to as the N-infection force.
(H7) At any time, an infected individual makes a positive comment on the product he/she has recently bought (and
hence becomes positive) at constant rate αP > 0, where αP > 0 is referred to as the P-comment rate.
(H8) At any time, an infected individual makes a negative comment on the product he/she has recently bought (and
hence becomes negative) at constant rate αN > 0, where αN > 0 is referred to as the N-comment rate.
(H9) At any time, a positive individual is inclined to buy a new product (and hence becomes susceptible) at constant
rate γP > 0, where γP > 0 is referred to as the P-viscosity rate.
(H10) At any time, an infected individual is inclined to buy a new product (and hence becomes susceptible) at constant
rate γI > 0, where γI > 0 is referred to as the I-viscosity rate.
2
dS (t)
dt
dI(t)
dt
dP(t)
dt
dN(t)
dt
= µ − βPP(t)S (t) − βNN(t)S (t) + γPP(t) + γII(t),
= βPP(t)S (t) − αPI(t) − αNI(t) − γII(t) − δII(t),
= αPI(t) − γPP(t) − δPP(t),
= αNI(t) − δNN(t).
(1)
(cid:90) T
βP
0
Fig. 1 shows these assumptions schematically.
Figure 1. Diagram of assumptions (H1)-(H10).
This collection of assumptions can be formulated as the dynamical model
Here, 0 ≤ t ≤ T. We refer to the model as the Susceptible-Infected-Positive-Negative-Susceptible (SIPNS) model.
Suppose the profit per item is one unit. Then the expected profit during the infinitesimal time interval [t, t + dt) is
βPP(t)S (t)dt. As a result, the expected overall profit during the campaign is
P(t)S (t)dt.
(2)
The marketer's goal is to maximize the expected overall profit subject to the SIPNS model.
3. The dynamics of the SIPNS model
The SIPNS model plays a key role in the overall profit of a WOM marketing campaign. However, it is difficult
to theoretically study the dynamics of the four-dimensional SIPNS model. This section is intended to experimentally
reveal the influence of different factors on the dynamics of the SIPNS model.
3.1. The entrance rate
Comprehensive simulation experiments show that, typically, the influence of the entrance rate on the dynamics of
the SIPNS model is as shown in Fig. 2. In general, the following conclusions are drawn.
(a) For any entrance rate, the expected number of susceptible, infected, positive and negative individuals levels off,
respectively.
(b) The steady expected number of susceptible individuals is not dependent upon the entrance rate.
(c) With the rise of the entrance rate, the steady expected number of infected, positive, and negative individuals
goes up, respectively.
3
()St()It()Pt()Nt()PPtPNPIPNI()NNt Figure 2. The time plots of S (t), I(t), P(t), and N(t) for different entrance rates.
3.2. The I-exit rate
Comprehensive simulation experiments show that, typically, the influence of the I-exit rate on the dynamics of the
SIPNS model is as shown in Fig. 3. In general, the following conclusions are drawn.
(a) For any I-exit rate, the expected number of susceptible, infected, positive, and negative individuals levels off,
respectively.
(b) With the rise of the I-exit rate, the steady expected number of susceptible individuals goes up.
(c) With the rise of the I-exit rate, the steady expected number of infected, positive, and negative individuals goes
down, respectively.
3.3. The P-exit rate
Comprehensive simulation experiments show that, typically, the influence of the P-exit rate on the dynamics of
the SIPNS model is as shown in Fig. 4. In general, the following conclusions are drawn.
(a) For any P-exit rate, the expected number of susceptible, infected, positive, and negative individuals levels off,
respectively.
(b) With the rise of the P-exit rate, the steady expected number of susceptible individuals goes up.
(c) With the rise of the P-exit rate, the steady expected number of infected, positive, and negative individuals goes
down, respectively.
3.4. The N-exit rate
Comprehensive simulation experiments show that, typically, the influence of the N-exit rate on the dynamics of
the SIPNS model is as shown in Fig. 5. In general, the following conclusions are drawn.
(a) For any N-exit rate, the expected number of susceptible, infected, positive, and negative individuals levels off,
respectively.
4
543210Number of susceptible individuals200150100500 Time (a)543210Number of infected individuals4003002001000 Time (b)1.00.80.60.40.20.0Number of positive individuals4003002001000 Time (c)100x10-3806040200Number of negative individuals4003002001000 Time (d) Figure 3. The time plots of S (t), I(t), P(t), and N(t) for different I-exit rates.
(b) The steady expected number of susceptible individuals is not dependent upon the N-exit rate.
(c) With the rise of the N-exit rate, the steady expected number of infected and positive individuals goes up,
respectively.
(d) With the rise of the N-exit rate, the steady expected number of infected and positive individuals goes down.
3.5. The P-comment rate
Comprehensive simulation experiments show that, typically, the influence of the P-comment rate on the dynamics
of the SIPNS model is as shown in Fig. 6. In general, the following conclusions are drawn.
(a) For any P-comment rate, the expected number of susceptible, infected, positive, and negative individuals levels
off, respectively.
(b) With the rise of the P-comment rate, the steady expected number of susceptible individuals goes down.
(c) With the rise of the P-comment rate, the steady expected number of infected, positive, and negative individuals
goes up, respectively.
3.6. The N-comment rate
Comprehensive simulation experiments show that, typically, the influence of the N-comment rate on the dynamics
of the SIPNS model is as shown in Fig. 7. In general, the following conclusions are drawn.
(a) For any N-comment rate, the expected number of susceptible, infected, positive, and negative individuals levels
off, respectively.
(b) With the rise of the N-comment rate, the steady expected number of susceptible and negative individuals goes
up, respectively.
(c) With the rise of the N-comment rate, the steady expected number of infected and positive individuals goes
down, respectively.
5
14121086420Number of susceptible individuals200150100500 Time (a)543210Number of infected individuals200150100500 Time(b) 0.350.300.250.200.150.100.050.00Number of positive individuals200150100500 Time (c)0.300.250.200.150.100.050.00Number of negative individuals200150100500 Time(d) Figure 4. The time plots of S (t), I(t), P(t), and N(t) for different P-exit rates.
3.7. The P-infection force
Comprehensive simulation experiments show that, typically, the influence of the P-infection force on the dynamics
of the SIPNS model is as shown in Fig. 8. In general, the following conclusions are drawn.
(a) For any P-infection force, the expected number of susceptible, infected, positive, and negative individuals levels
(b) With the rise of the P-infection force, the steady expected number of susceptible individuals goes down.
(c) With the rise of the P-infection force, the steady expected number of infected, positive, and negative individuals
off, respectively.
goes up, respectively.
3.8. The N-infection force
Comprehensive simulation experiments show that, typically, the influence of the N-infection force on the dynamics
of the SIPNS model is as shown in Fig. 9. In general, the following conclusions are drawn.
(a) For any N-infection force, the expected number of susceptible, infected, positive, and negative individuals levels
(b) The steady expected number of susceptible individuals is not dependent upon the N-infection force.
(c) With the rise of the N-infection force, the steady expected number of infected, positive, and negative individuals
off, respectively.
goes down, respectively.
3.9. The I-viscosity rate
Comprehensive simulation experiments show that, typically, the influence of the I-viscosity rate on the dynamics
of the SIPNS model is as shown in Fig. 10. In general, the following conclusions are drawn.
(a) For any I-viscosity rate, the expected number of susceptible, infected, positive, and negative individuals levels
(b) With the rise of I-viscosity rate, the steady expected number of susceptible individuals goes up.
(c) With the rise of the I-viscosity rate, the steady expected number of infected, positive, and negative individuals
off, respectively.
goes down, respectively.
6
1612840Number of susceptible individuals200150100500 Time (a)543210Number of infected individuals200150100500 Time (b)0.300.250.200.150.100.050.00Number of positive individuals200150100500 Time (c)0.40.30.20.10.0Number of negative individuals200150100500 Time (d) Figure 5. The time plots of S (t), I(t), P(t), and N(t) for different N-exit rates.
3.10. The P-viscosity rate
Comprehensive simulation experiments show that, typically, the influence of the P-viscosity rate on the dynamics
of the SIPNS model is as shown in Fig. 11. In general, the following conclusions are drawn.
(a) For any P-viscosity rate, the expected number of susceptible, infected, positive, and negative individuals levels
off, respectively.
(b) With the rise of P-viscosity rate, the steady expected number of susceptible individuals goes up.
(c) With the rise of the P-viscosity rate, the steady expected number of infected, positive, and negative individuals
goes down, respectively.
4. The overall profit of a WOM marketing campaign
This section aims to experimentally uncover the influence of different factors on the expected overall profit of a
WOM marketing campaign.
4.1. The entrance rate and the exit rates
Comprehensive simulation experiments show that, typically, the influence of the entrance rate and the three exit
rates on the expected overall profit is as shown in Fig. 12. In general, the following conclusions are drawn.
(a) With the rise of the entrance rate, the expected overall profit goes up.
(b) With the rise of the I-exit rate entrance rate, the expected overall profit goes down.
(c) With the rise of the P-exit rate entrance rate, the expected overall profit goes down.
(d) With the rise of the N-exit rate entrance rate, the expected overall profit goes up.
7
3.53.02.52.01.51.00.50.0Number of susceptible individuals100806040200 Time (a)543210Number of infected individuals100806040200 Time (b)1.21.00.80.60.40.20.0Number of positive individuals100806040200 Time (c)1.21.00.80.60.40.20.0Number of negative individuals100806040200 Time (d) Figure 6. The time plots of S (t), I(t), P(t), and N(t) for different P-comment rates.
The entrance rate can be enhanced by launching a viral marketing (VM) campaign, that is, marketers develop a
marketing message and encourage customers to forward this message to their contacts. There are quite a number of
successful VM cases: Hotmail generated 12 million subscribers in just 18 months with a marketing budget of only
$50,000, and Unilevers Dove Evolution campaign generated over 2.3 million views in its first 10 days. These VM
campaigns were successful in part because the marketers effectively utilized VM's unique potential to reach large
numbers of potential customers in a short period of time at a lower cost.
The I-exit rate or the P-exit rate can be reduced by enhancing customers' experiences with the purchased products
and taking promotional measures such as discounting and distributing coupons.
Typically, the N-exit rate is uncontrollable.
4.2. The comment rates
Comprehensive simulation experiments show that, typically, the influence of the two comment rates on the ex-
pected overall profit is as shown in Fig. 13. In general, the following conclusions are drawn.
(a) With the rise of the P-comment rate, the overall profit goes up.
(b) With the rise of the N-comment rate rate, the overall profit goes down.
The P-comment rate can be enhanced by enhancing customers' experiences. Also, the N-comment rate can be
enhanced by enhancing customers' experiences.
4.3. The infection forces
Comprehensive simulation experiments show that, typically, the influence of the two infection forces on the ex-
pected overall profit is as shown in Fig. 14. In general, the following conclusions are drawn.
(a) With the rise of the P-infection force, the overall profit goes up.
(b) With the rise of the N-infection force, the overall profit goes down.
The P-infection force can be enhanced by enhancing customers' experiences and, hence, earning positive WOM.
Also, the N-infection force can be reduced by enhancing customers' experiences.
8
121086420Number of susceptible individuals200150100500 Time (a)543210Number of infected individuals200150100500 Time (b)1.41.21.00.80.60.40.20.0Number of positive individuals200150100500 Time (c)0.300.250.200.150.100.050.00Number of negative individuals200150100500 Time (d) Figure 7. The time plots of S (t), I(t), P(t) and N(t) for different N-comment rates.
4.4. The viscosity rates
Comprehensive simulation experiments show that, typically, the influence of the two viscosity rates on the ex-
pected overall profit is as shown in Fig. 15. In general, the following conclusions are drawn.
(a) With the rise of the I-viscosity rate, the overall profit goes up.
(b) There is a threshold θ such that (1) when the P-viscosity is lower than θ, the expected overall profit goes up with
the rise of the P-viscosity, and (2) when the P-viscosity exceeds θ, the expected overall profit goes down with
the rise of the P-viscosity.
The I-viscosity rate can be enhanced by enhancing customers' experiences or by taking promotional measures.
5. Conclusions and remarks
The issue of maximizing the overall profit of a WOM-based promotion campaign has been modeled as a con-
strained optimization problem. Through comprehensive computer simulations, the influence of different factors on
the profit has been revealed. On this basis, some promotion strategies have been suggested.
Toward this direction, great efforts have yet to be made. It is well known that the structure of the underlying WOM
network has significant influence on the performance of a viral marketing [35, 36]. The proposed SIPNS model is
a population-level model and hence does not allow the analysis of this influence. To reveal the impact of the WOM
network on the marketing profit, a network-level epidemic model [37 -- 40] or an individual-level epidemic model
[41 -- 45] is an appropriate choice. The profit model presented in this paper builds on the uniform-profit assumption.
However, in everyday life different products may have seperate profits. Hence, it is of practical importance to construct
a non-uniform profit model. Also, a customer may purchase more than one items a time, and the corresponding model
is yet to be developed. Typically, a viral marketing campaign is subject to a limited budget. The dynamic optimal
control strategy against malicious epidemics [46 -- 50] may be borrowed to the analysis of viral marketing so as to
achieve the maximum possible net profit.
9
6543210Number of susceptible individuals200150100500 Time (a)543210Number of infected individuals200150100500 Time (b)1.21.00.80.60.40.20.0Number of positive individuals200150100500 Time (c)1.21.00.80.60.40.20.0Number of negative individuals200150100500 Time (d) Figure 8. The time plots of S (t), I(t), P(t) and N(t) under different P-infection forces.
Acknowledgments
This work is supported by Natural Science Foundation of China (Grant Nos. 61572006, 71301177), National Sci-
Tech Support Program of China (Grant No. 2015BAF05B03), Basic and Advanced Research Program of Chongqing
(Grant No. cstc2013jcyjA1658), and Fundamental Research Funds for the Central Universities (Grant No. 106112014CD-
JZR008823).
References
[1] G. Armstrong, P. Kotler, Marketing: An introduction (11th Edition), Prentice Hall, 2012.
[2] D. Grewal, M. Levy, Marketing (5th Edition), McGraw-Hill Education, 2016.
[3] J. Goldenberg, B. Libai, E. Muller, Talk of the network: A complex systems look at the underlying process of word-of-mouth, Marketing
[4] V. Mahajan, E. Muller, R.A. Kerin, Introduction strategy for new products with positive and negative word-of-mouth, Management Science,
Letters, 2001, 12(3): 211-223.
1984, 30(12): 1389-1404.
[5] E. Anderson, Customer satisfaction and word-of-mouth, Journal of Service Research, 1998, 1(1): 5-17.
[6] P.M. Herr, F.R. Kardes, J. Kim, Effects of word-of-mouth and product attribute information on persuasion: An accessibility-diagnosticity
perspective, Journal of Consumer Research, 1991, 17, 454-462.
[7] D. Charlett, R. Garland, N. Marr, How damaging is negative word of mouth? Marketing Bulletin, 1995, 6: 42-50.
[8] J.C. Sweeney, G.N. Soutar, T. Mazzarol, The differences between positive and negative word-of-mouth: Emotion as a differentiator? S.
Gopalan, N. Taher (eds), Viral Marketing: Concepts and Cases, Icfai University Press, 2007, pp. 156-168.
[9] N. Ahmad, J. Vveinhardt, R.R. Ahmed, Impact of word of mouth on consumer buying decision, European Journal of Business and Manage-
ment, 2014, 6(31): 394-403.
[10] I.R. Misner, The World's Best Known Marketing Secret: Building Your Business with Word-of-Mouth Marketing, 2nd ed. Bard Press, 1999.
[11] J. Chevalier, M. Dina, The effect of word of mouth on sales: online book reviews, Journal of Marketing Research, 2006, 43: 345-354.
[12] M. Trusov, R.E. Bucklin, K. Pauwels, Effects of word-of-mouth versus traditional marketing: Findings from an Internet social networking
[13] R. Peres, E. Muller, V. Mahajan, Innovation diffusion and new product growth models: A critical review and research directions, International
sites, Journal of Marketing, 2009, 73(5): 90-102.
Journal of Research in Marketing, 2010, 27: 91-106.
[14] D. Kempe, J. Kleinberg, E, Tardos, Influential nodes in a diffusion model for social networks, Proceedings of the 32nd International Confer-
ence on Automata, Languages and Programming, 2005, 1127-1138.
[15] W. Chen, Y. Wang, S. Yang, Efficient influence maximization in social networks, Proceedings of the 15th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, 2009, pp. 199-208.
10
2520151050Number of susceptible individuals4003002001000 Time (a)543210Number of infected individuals4003002001000 Time (b)1.41.21.00.80.60.40.20.0Number of positive individuals4003002001000 Time (c)0.250.200.150.100.050.00Number of negative individuals4003002001000 Time (d) Figure 9. The time plots of S (t), I(t), P(t) and N(t) under different N-infection forces.
[16] O. Hinz, B. Skiera, C. Barrot, J.U. Becker, Seeding strategies for viral marketing: An empirical comparison, Journal of Marketing, 2011, 75:
[17] T.N. Dinh, H. Zhang, D.T. Nguyen, M.T. Thai, Cost-effective viral marketing for tme-critical campaigns in large-scale social networks,
IEEE/ACM Transactions on Networking, 2014, 22(6): 2001-2011.
[18] A. Mochalova, A. Nanopoulos, A targeted approach to viral marketing, Electronic Commerce Research and Applications, 2014, 13: 283-294.
[19] D. Kempe, J. Kleinberg, E, Tardos, Maximizing the spread of influence through a social network, THEORY OF COMPUTING, 2015, 11(4):
[20] M. Samadi, A. Nikolaev, R. Nagi, A subjective evidence model for influence maximization in social networks, Omega, 2016, 59(B): 263-278.
[21] H. Zhang, D.T. Nguyen, S. Das, H. Zhang, M.T. Thai, Least cost influence maximization across multiple social networks, IEEE/ACM
Transactions on Networking, 2016, 24(2): 929-939.
[22] S. Bharathi, D. Kempe, M. Salek, Competitive influence maximization in social networks, X. Deng, F. C. Graham (eds) Internet and Network
Economics, WINE2007, Lecture Notes in Computer Science, vol. 4858, 2007.
[23] H. Zhang, H. Zhang, A. Kuhnle, M.T. Thai, Profit maximization for multiple products in online social networks, Proceedings of the IEEE
International Conference on Computer Communications (INFOCOM), 2016.
[24] F.M. Bass, A new product growth for model consumer durables, Management Science, 1969, 15(5): 215-227.
[25] J.T. Gardner, K. Sohn, J.Y. Seo, J.L. Weaver, Analysis of an epidemiological model of viral marketing: when viral marketing efforts fall flat,
Journal of Marketing Development and Competitiveness, 2013, 7(4): 25-46.
[26] K. Sohn, J. Gardner, J. Weaver, Viral marketing -- more than a buzzword, Journal of Applied Business and Economics, 2013, 14(1): 21-42.
[27] S. Li, Z. Jin, Global dynamics analysis of homogeneous new products diffusion model, Discrete Dynamics in Nature and Society, 2013, 2013:
55-71.
105-147.
158901.
940623.
[28] S. Li, Z. Jin, Modeling and analysis of new products diffusion on heterogeneous networks, Journal of Applied Mathematics, 2014, 2014:
[29] H.S. Rodrigues, M. Fonseca, Viral marketing as epidemiological model, Proceedings of the 15th International Conference on Computational
and Mathematical Methods in Science and Engineering, 2015, pp. 946-955.
[30] H.S. Rodrigues, M. Fonseca, Can information be spread as a virus? Viral marketing as epidemiological model, Mathematical Methods in the
[31] P. Jiang, X. Yan, L. Wang, A viral product diffusion model to forecast the market performance of products, Discrete Dynamics in Nature and
[32] W. Wang, P. Fergola, C. Tenneriello, Innovation diffusion model in patch environment, Applied Mathematics and Computation, 2003, 134:
[33] Y. Yu, W. Wang, Y. Zhang, An innovation diffusion model for three competitive products, Computers and Mathematics with Applications,
[34] X. Wei, N. Valler, B.A. Prakash, I. Neamtiu, M. Faloutsos, C. Faloutsos, Competing memes propagation on networks: A network science
perspective, IEEE Journal on Selected Areas in Communications, 2013, 31(6): 1049-1060.
Applied Sciences, 2016, 39: 4780-4786.
Society, 2017, 2017: 9121032.
51-67.
2003, 46: 1473-1481.
11
2520151050Number of susceptible individuals3002001000 Time (a)543210Number of infected individuals3002001000 Time (b)2.52.01.51.00.50.0Number of positive individuals3002001000 Time (c)0.40.30.20.10.0Number of negative individuals3002001000 Time (d) Figure 10. The time plots of S (t), I(t), P(t), and N(t) for different I-viscosity rates.
[35] M. Bampo, M. Ewing, D. Mather, D. Stewart, M. Wallace, The effects of the social structure of digital networks on viral marketing perfor-
mance, Information Systems Research, 2008, 19(3): 273-290.
[36] A. Siri, T. Thaiupathump, Measuring the performance of viral marketing based on the dynamic behavior of social Networks, Proceedings of
the IEEE International Conference on Industrial Engineering and Engineering Management, 2013, 432.
[37] P. Pastor-Satorras, A. Vespignani, Epidemic spreading in scale-free networks, Physical Review Letters, 2001, 86(14): 3200.
[38] P. Pastor-Satorras, A. Vespignani, Epidemic dynamics in finite size scale-free networks, Physical Review E, 2002, 65(3): 035108.
[39] L.X. Yang, X. Yang, The spread of computer viruses over a reduced scale-free network, Physica A, 2014, 396(15): 173-184.
[40] L.X. Yang, X. Yang, The effect of network topology on the spread of computer viruses: a modelling study, International Journal of Computer
Mathematics, DOI: 10.1080/00207160.2016.1226499.
[41] P. Van Mieghem, J. Omic, R. Kooij, Virus spread in networks, IEEE/ACM Transactions on Networking, 2009, 17(1): 1-14.
[42] F.D. Sahneh, C. Scoglio, P. Van Mieghem, Generalized epidemic mean-field model for spreading processes over multilayer complex networks,
IEEE/ACM Transactions on Networking, 2013, 21(5): 1609-1620.
[43] L.X. Yang, M. Draief, X. Yang, The impact of the network topology on the viral prevalence: a node-based approach, Plos One, 2015, 10:
[44] L.X. Yang, M. Draief, X. Yang, Heterogeneous virus propagation in networks: a theoretical study. Mathematical Methods in the Applied
[45] L.X. Yang, X. Yang, Y. Wu, The impact of patch forwarding on the prevalence of computer virus: A theoretical assessment approach, Applied
e0134507.
Sciences, 2017, 40(5): 1396-1413.
Mathematical Modelling, 2017, 43: 110-125.
[46] V.M. Preciado, M. Zargham, C. Enyioha, A. Jadbabaie, G. Pappas, Optimal resource allocation for network protection against spreading
processes, IEEE Transactions on Control of Network Systems, 2014, 1(1): 99-108.
[47] H. Shakeri, F.D. Sahneh, C. Scoglio, Optimal information dissemination strategy to promote preventive behaviours in multilayer networks.
Mathematical Bioscience and Engineering, 2015, 12(3): 609-623.
[48] L.X. Yang, M. Draief, X. Yang, The optimal dynamic immunization under a controlled heterogeneous node-based SIRS model, Physic A,
[49] C. Nowzari, V.M. Preciado, G.J. Pappas, Analysis and control of epidemics: A survey of spreading processes on complex networks, IEEE
[50] T.R. Zhang, L.X. Yang, X. Yang, Y. Wu, Y.Y. Tang, The dynamic malware containment under a node-level epidemic model with the alert
[51] J. Bi, X. Yang, Y. Wu, Q. Xiong, J. Wen, Y.Y. Tang, On the optimal dynamic control strategy of disruptive computer virus, Discrete Dynamics
2016, 450: 403-415.
Control Systems, 2016, 36(1): 26-46.
compartment, Physica A, 2017, 470: 249-260.
in Nature and Society, 2017, 2017: 8390784.
12
20151050Number of susceptible individuals3002001000 Time (a)543210Number of infected individuals3002001000 Time (b)0.250.200.150.100.050.00Number of positive individuals3002001000 Time (c)0.300.250.200.150.100.050.00Number of negative individuals3002001000 Time (d) Figure 11. The time plots of S (t), I(t), P(t), and N(t) for different P-viscosity rates.
Figure 12. The expected overall profit versus (a) the entrance rate, (b) the I-exit rate, (c) the P-exit rate, and (d) the N-exit rate.
13
1612840Number of susceptible individuals3002001000 Time (a)543210Number of infected individuals3002001000 Time (b)0.300.250.200.150.100.050.00Number of positive individuals3002001000 Time (c)0.250.200.150.100.050.00Number of negative individuals3002001000 Time (d)3500300025002000150010005000Overall profit1.00.80.60.40.20.0(a)450400350300250200150Overall profit0.80.60.40.20.0(b)280260240220200180160140Overall profit0.400.300.200.100.00(c)240236232228224Overall profit1.00.90.80.70.60.5(d) Figure 13. The expected overall profit versus (a) the P-comment rate, and (b) the N-comment rate.
Figure 14. The expected overall profit versus (a) the P-infection force, and (b) the N-infection force.
Figure 15. The overall profit versus (a) the P-viscosity rate, and (b) the I-viscosity rate.
14
250200150100500Overall profit1.00.80.60.40.20.0(a)300250200150100500Overall profit1.00.80.60.40.20.0(b)200150100500Overall profit1.00.80.60.40.20.0(a)200180160140120100Overall profit1.00.80.60.40.20.0(b)750700650600550500Overall profit1.00.80.60.40.2(a)500450400350300Overall profit0.80.60.40.20.0(b) |
1608.06196 | 5 | 1608 | 2019-12-11T16:38:45 | A Framework for the Construction of Generative Models for Mesoscale Structure in Multilayer Networks | [
"cs.SI",
"cond-mat.stat-mech",
"nlin.AO",
"physics.soc-ph",
"stat.ME"
] | Multilayer networks allow one to represent diverse and coupled connectivity patterns --- e.g., time-dependence, multiple subsystems, or both --- that arise in many applications and which are difficult or awkward to incorporate into standard network representations. In the study of multilayer networks, it is important to investigate mesoscale (i.e., intermediate-scale) structures, such as dense sets of nodes known as communities, to discover network features that are not apparent at the microscale or the macroscale. The ill-defined nature of mesoscale structure and its ubiquity in empirical networks make it crucial to develop generative models that can produce the features that one encounters in empirical networks. Key purposes of such generative models include generating synthetic networks with empirical properties of interest, benchmarking mesoscale-detection methods and algorithms, and inferring structure in empirical multilayer networks. In this paper, we introduce a framework for the construction of generative models for mesoscale structures in multilayer networks. Our framework provides a standardized set of generative models, together with an associated set of principles from which they are derived, for studies of mesoscale structures in multilayer networks. It unifies and generalizes many existing models for mesoscale structures in fully-ordered (e.g., temporal) and unordered (e.g., multiplex) multilayer networks. One can also use it to construct generative models for mesoscale structures in partially-ordered multilayer networks (e.g., networks that are both temporal and multiplex). Our framework has the ability to produce many features of empirical multilayer networks, and it explicitly incorporates a user-specified dependency structure between layers. | cs.SI | cs |
A framework for the construction of generative models
for mesoscale structure in multilayer networks
Marya Bazzi,1, 2, 3, ∗ Lucas G. S. Jeub,1, 4, 5, ∗ Alex Arenas,6 Sam D. Howison,1 and Mason A. Porter1, 7, 8
1Oxford Centre for Industrial and Applied Mathematics, Mathematical Institute,
University of Oxford, Oxford OX2 6GG, United Kingdom
2The Alan Turing Institute, London NW1 2DB, United Kingdom
3Warwick Mathematics Institute, University of Warwick, Coventry CV4 7AL, United Kingdom
4Center for Complex Networks and Systems Research, School of Informatics and Computing,
Indiana University, Bloomington, Indiana 47408, USA
5ISI Foundation, Turin, Italy
6Departament d'Enginyeria Inform`atica i Matem`atiques,
Universitat Rovira i Virgili, 43007 Tarragona, Spain
7CABDyN Complexity Centre, University of Oxford, Oxford OX1 1HP, United Kingdom
8Department of Mathematics, University of California,
Los Angeles, Los Angeles, California 90095, USA
Multilayer networks allow one to represent diverse and coupled connectivity patterns -- e.g.,
time-dependence, multiple subsystems, or both -- that arise in many applications and which are
difficult or awkward to incorporate into standard network representations. In the study of multi-
layer networks, it is important to investigate mesoscale (i.e., intermediate-scale) structures, such
as dense sets of nodes known as communities, to discover network features that are not apparent
at the microscale or the macroscale. The ill-defined nature of mesoscale structure and its ubiquity
in empirical networks make it crucial to develop generative models that can produce the features
that one encounters in empirical networks. Key purposes of such generative models include gen-
erating synthetic networks with empirical properties of interest, benchmarking mesoscale-detection
methods and algorithms, and inferring structure in empirical multilayer networks. In this paper, we
introduce a framework for the construction of generative models for mesoscale structures in mul-
tilayer networks. Our framework provides a standardized set of generative models, together with
an associated set of principles from which they are derived, for studies of mesoscale structures in
multilayer networks.
It unifies and generalizes many existing models for mesoscale structures in
fully-ordered (e.g., temporal) and unordered (e.g., multiplex) multilayer networks. One can also use
it to construct generative models for mesoscale structures in partially-ordered multilayer networks
(e.g., networks that are both temporal and multiplex). Our framework has the ability to produce
many features of empirical multilayer networks, and it explicitly incorporates a user-specified de-
pendency structure between layers. We discuss the parameters and properties of our framework,
and we illustrate examples of its use with benchmark models for community-detection methods and
algorithms in multilayer networks.
I.
INTRODUCTION
One can model many physical, technological, biologi-
cal, financial, and social systems as networks. The sim-
plest type of network is a graph [1], which consists of
a set of nodes (which represent entities) and a set of
edges between pairs of nodes that encode interactions be-
tween those entities. One can consider either unweighted
graphs or weighted graphs, in which each edge has a
weight that quantifies the strength of the associated in-
teraction. Edges can also incorporate directions to rep-
resent asymmetric interactions or signs to differentiate
between positive and negative interactions.
However, this relatively simple structure cannot cap-
ture many of the possible intricacies of connectivity pat-
terns between entities. For example, in temporal net-
works [2, 3], nodes and/or edges change in time; and
∗ Both authors contributed equally.
in multiplex networks [4], multiple types of interactions
can occur between the same pairs of nodes. To better
account for the complexity, diversity, and dependencies
in real-world interactions, one can represent such con-
nectivity patterns using "multilayer networks" (see Sec-
tion II A) [4 -- 8]. We use the term single-layer network
(which is also called a "monolayer network" or a "graph")
for a multilayer network with a single layer (i.e., an 'ordi-
nary' network), and we use the term "network" to refer
to both single-layer and multilayer networks.
By using a single multilayer network instead of several
independent single-layer networks, one can account for
the fact that connectivity patterns in different layers of-
ten "depend" on each other. (We use the term dependent
in a probabilistic sense throughout the paper: an object
A depends on B if and only if P(AB) (cid:54)= P(A).) For ex-
ample, the connectivity patterns in somebody's Facebook
friendship network today may depend both on the con-
nectivity patterns in that person's Facebook friendship
network last year (temporal) and on the connectivity pat-
terns in that person's Twitter followership network today
(multiplex). Data sets that have multilayer structures are
increasingly available (e.g., see Table 2 of [4]). A natu-
ral type of multilayer network consists of a sequence of
dependent single-layer networks, where layers may corre-
spond to different temporal snapshots, different types of
related interactions that occur during a given time inter-
val, and so on. Following existing terminology [8, 9], we
refer to an instance of a node in one "layer" as a "state
node" (see Section II A).
There is a diverse set of models for multilayer networks.
(We overview them in Section II C.) Many of these mod-
els take a specific type of dependency (e.g., temporal) as
their starting point. In the present paper, we introduce
a framework for the construction of generative models
for multilayer networks that incorporate a wide variety
of structures and dependencies.
It is broad enough to
unify many existing, more restrictive interlayer specifica-
tions, but it is also easy to customize to yield multilayer
network models for many specific cases of interest. Key
purposes of such generative models include (1) generat-
ing synthetic networks with empirical features of interest,
(2) benchmarking methods and algorithms for detecting
mesoscale structures, and (3) inferring structure in em-
pirical multilayer networks.
A unifying framework
A key feature of multilayer networks is their flexibil-
ity, which allows one to incorporate many different types
of data as part of a single structure. In this spirit, our
goal is to provide a general, unifying framework that en-
ables users to construct generative models of multilayer
networks with a large variety of features of interest in em-
pirical multilayer networks by appropriately constraining
the parameter space. We accomplish this in two consec-
utive steps. First, we partition the set of state nodes of
a multilayer network. Second, we allocate edges, given a
multilayer partition. We focus on modeling dependency
at the level of partitions (as was done in [10]), rather than
with respect to edges. Additionally, we treat the process
of generating a multilayer partition separately from that
of generating edges for a given multilayer partition. This
modular approach yields random structures that can cap-
ture a wide variety of interlayer-dependency structures,
including temporal and/or multiplex networks, appear-
ance and/or disappearance of entities, uniform or nonuni-
form dependencies between state nodes from different
layers, and others. For a specified interlayer-dependency
structure, one can then use any (single-layer or multi-
layer) network model with a planted partition (i.e., a
planted-partition network model ) to generate a wide va-
riety of network features, including weighted edges, di-
rected edges, and spatially-embedded layers.
The flexibility of our framework to generate multilayer
networks with a specified dependency structure between
different layers makes it possible to (1) gain insight into
whether, when, and how to build interlayer dependen-
2
cies into methods for studying many different types of
multilayer networks and (2) generate tunable "bench-
mark models" that allow a principled comparison for
community-detection (and, more generally, "meso-set-
detection") tools for multilayer networks, including for
complicated situations that arise in many applications
(such as networks that are both temporal and multiplex)
but thus far have seldom or never been studied. In many
benchmark models, one plants a partition of a network
into well-separated "meso-sets" (e.g., communities), and
one thereby imposes a so-called "ground truth" (should
one wish to use such a notion) [11] that a properly de-
ployed meso-set-detection method ought to be able to
find. Benchmark networks with known structural prop-
erties can be important for (1) analyzing and compar-
ing the performance of different meso-set-detection tools;
(2) better understanding the inner workings of meso-set-
detection tools; and (3) determining which tool(s) may
be most appropriate in a given situation.
One can also use our framework to generate synthetic
networks with desired empirical properties, to generate
null networks, and to explore "detectability limits" of
mesoscale structures. (See, for example, the study of de-
tectability thresholds in [10] for a model that is a special
case of our framework.) With some further work, it is
also possible to use our framework to develop models for
statistical inference.
In our concluding discussion (see
Section VII), we suggest directions for how to apply our
framework to the task of statistical inference. Our inten-
tion in designing such a flexible framework is to ensure
that the generative models that it provides remain useful
as researchers consider progressively more general multi-
layer networks in the coming years.
Paper outline
Our paper proceeds as follows. In Section II, we give an
introduction to multilayer networks, overview mesoscale
structure in networks, and review related work on gener-
ative models for mesoscale structure. In Section III, we
formally introduce the notation that we use throughout
the paper.
In Section IV, we explain how we generate
a multilayer partition with a specified dependency struc-
ture between layers. We also give examples of how to
constrain the parameter space of our framework to gen-
erate several different types of multilayer networks and
discuss what we expect to be common use cases (includ-
ing temporal structure [10, 12, 13] and multiplex struc-
ture [9, 13]) in detail. In Section V, we describe how we
generate edges that are consistent with the planted parti-
tion. In Section VI, we illustrate the use of our framework
as a way to construct benchmark models for multilayer
community detection. In Section VII, we summarize our
main results and briefly discuss possible extensions of
our work to enable statistical inference on empirical mul-
tilayer networks with our framework.
Along with this paper, we include code [14] that users
3
multilayer network, a node can be present in a variety
of different states, where each state is also characterized
by a variety of different aspects. In this setting, edges
connect state nodes, each of which is the instantiation of
a given node in a particular state. The set of all state
nodes of a given entity corresponds to a single physi-
cal node (which represents the entity), and the set of all
state nodes in a given state is in one layer of a multi-
layer network. In the remainder of this paper, we use the
terms "physical node" and "node" interchangeably and
the terms "layer" and "state" interchangeably. One can
think of the aspects as features that one needs to specify
to identify the state of a node. In other words, a state is
a collection of exactly one element from each aspect. For
convenience, we introduce a mapping that assigns an in-
teger label to each element of an aspect. That is, we map
the la elements of the ath aspect to the elements of a set
{1, . . . , la} of integers. Aspects can be unordered (e.g.,
social-media platform) or ordered (e.g., time). Most em-
pirical investigations of multilayer networks focus on a
single aspect (e.g., temporal [16 -- 18] or multiplex [19 --
21]). However, many real-world situations include more
than one aspect (e.g., a multiplex network that changes
over time). For an ordered aspect, we require that the
mapping respects the order of the aspect (e.g., ti → i for
time, where t1 ≤ . . . ≤ tla is a set of discrete time points).
A multilayer network can include an arbitrary number of
ordered aspects and an arbitrary number of unordered
aspects, and one can generalize these ideas further (e.g.,
by introducing a time horizon) [4].
To illustrate the above ideas, consider a hypotheti-
cal social network with connections on multiple social-
media platforms ('Facebook', 'Twitter', and 'LinkedIn')
between the same set of people at different points in time.
In this example network, there are two aspects: social-
media platform and time.
(One can consider 'type of
connection' (e.g., 'friendship' and 'following') as a third
aspect, but we restrict our example to two aspects for
simplicity.) The first aspect is ordered, and the num-
ber of its elements is equal to the number of time points
or time intervals. (For simplicity, we often refer simply
to "time points" in our discussions.) The second aspect
is unordered and consists of three elements: 'Facebook'
(which we label with the integer "1"), 'Twitter' (which
we label with "2"), and 'LinkedIn' (which we label with
"3"). If we assume that the time resolution is daily and
spans the year 2010, an example of a state is the tuple
('01-Jan-2010', 'Twitter'), which is (1, 2) in our short-
hand notation. We show a multilayer representation of
this example social network in Fig. 1.
Edges in a multilayer network can occur either between
nodes in the same state (i.e., intralayer edges) or between
nodes in different states (i.e., interlayer edges). An ex-
ample of an intralayer edge in Fig. 1 is (1, (1, 1)) ↔ (2,
(1, 1)), indicating that entity 1 is friends with entity 2
on Facebook at time 1. All interlayer edges in Fig. 1 are
diagonal (because they connect state nodes that corre-
spond to the same physical node). Interlayer edges be-
Figure 1. Toy example of a multilayer network with three
physical nodes and two aspects. We represent undirected in-
tralayer edges using solid black lines, directed interlayer edges
using dotted red lines with an arrowhead, and undirected in-
terlayer edges using dashed blue arcs. The first aspect is
ordered and corresponds to time.
It has two time points
(which we label as "1" and "2"). Therefore, L1 = {1, 2}.
The second aspect is unordered and represents a social-media
platform.
It has three elements: Facebook (labeled "1"),
Twitter (labeled "2"), and LinkedIn (labeled "3"). There-
fore, L2 = {1, 2, 3}.
In this example, a state node takes
the form (i, (α1, α2)), with physical node i ∈ {1, 2, 3}, as-
pect α1 ∈ {1, 2}, and aspect α2 ∈ {1, 2, 3}. The total number
of layers is l = L1 × L2 = 6.
can modify to readily incorporate different types of "null
distributions" (see Section IV A), "interlayer-dependency
structures" (see Section IV B), and planted-partition net-
work models (see Section V). The model instantiations
that one needs for generating the figures in Section VI
are available at https://dx.doi.org/10.5281/zenodo.
3304059.
II. BACKGROUND AND RELATED WORK
A. Multilayer networks
The simplest type of network is a graph G = (V,E),
where V = {1, . . . , n} is a set of nodes (which correspond
to entities) and E ⊆ V×V is a set of edges. Using a graph,
one can encode the presence or absence of connections
(the edges) between entities (the nodes). However, in
many situations, it is desirable to include more detailed
information about connections between entities. A com-
mon extension is to allow each edge to have a weight,
which one can use to represent the strength of a connec-
tion. We assign a weight to each edge using a weight
function w : E → R.
As we mentioned in Section I, one can also general-
ize the notion of a graph to encode different aspects of
connections between entities, such as connections at dif-
ferent points in time or multiple types of relationships.
We adopt the framework of multilayer networks [4, 6 --
8, 15] to encode such connections in a network.
In a
(1,(1,1))(2,(1,1))timeFacebook(1,(2,1))(2,(2,1))(1,(1,2))(2,(1,2))(3,(1,2))Twitter(1,(2,2))(2,(2,2))(1,(1,3))(2,(1,3))(3,(1,3))LinkedIn(1,(2,3))(2,(2,3))(3,(2,3)) tween layers at different times for a given social-media
platform are ordinal (because they connect state nodes
with successive time labels [22]) and directed. Interlayer
edges between concurrent layers that correspond to dif-
ferent social-media platforms are categorical and undi-
rected. (In this example, such edges occur between state
nodes from all pairs of concurrent layers, although one
can envision situations in which interlayer edges occur
only between state nodes from a subset of such layers.)
An example of an ordinal intralayer edge in Fig. 1 is (1,
(1, 1)) → (1, (2, 1)), indicating a connection from entity
1 in Facebook at time 1 to entity 1 in Facebook at time
2. An example of a categorical intralayer edge in Fig. 1 is
(1, (1, 1)) ↔ (1, (1, 2)), indicating a connection between
node 1 in Facebook at time 1 and node 1 in Twitter at
time 1. One can also imagine other types of interlayer
edges in an example like the one in Fig. 1, as there may be
edges between state nodes at successive times and differ-
ent social-media platforms (e.g., (1, (1, 1)) → (1, (2, 2))).
B. Mesoscale structures in networks
Given a (single-layer or multilayer) network represen-
tation of a system, it is often useful to apply a coarse-
graining technique to investigate features that lie be-
tween those at the microscale (e.g., nodes, pairwise in-
teractions between nodes, or local properties of nodes)
and those at the macroscale (e.g., total edge weight, de-
gree distribution, or mean clustering coefficient) [1]. One
thereby studies mesoscale features such as community
structure [23 -- 26], core -- periphery structure [27, 28], role
structure [29], and others.
Our framework can produce multilayer networks with
any of the above mesoscale structures. Notwithstand-
ing this flexibility, an important situation is multilayer
networks with "community structure", which are the
most commonly studied type of mesoscale structure [1].
Community detection is part of the standard toolkit for
studying single-layer networks [23, 24, 30], and efforts
at community detection in multilayer networks have led
to insights in applications such as brain and behavioral
networks in neuroscience [31], financial correlation net-
works [16], committee and voting networks in political
science [32, 33], networks of interactions between bacte-
rial species [21], disease-spreading networks [12], social
networks in Indian villages [34], and much more.
Loosely speaking, a community in a network is a set
of nodes that are "more densely" connected to each
other than they are to nodes in the rest of the net-
work [23, 24, 30, 35, 36]. Typically, a 'good commu-
nity' should be a set of nodes that are 'surprisingly
well-connected' in some sense, but what one means by
'surprising' and 'well-connected'
is often application-
dependent and subjective. In many cases, a precise defi-
nition of "community" depends on the method that one
uses to detect communities. In particular, many popu-
lar community-detection approaches in single-layer and
4
multilayer networks define communities as sets in a par-
tition of a network that optimizes a quality function such
as modularity [1, 24, 37]; stability [38 -- 40]; Infomap and
its variants [9, 41]; likelihood functions that are derived
from stochastic block models (SBMs), which are models
for partitioning a network into sets of nodes with sta-
tistically homogeneous connectivity patterns [21, 42 -- 45];
and others [4, 46, 47]. In this paper, we refer to a par-
tition of the set of nodes of a single-layer network as
a single-layer partition and to a partition of the set of
state nodes of a multilayer network as a multilayer parti-
tion. The two primary differences between a community
in a multilayer partition and a community in a single-
layer partition are that (1) the former can include state
nodes from different layers; and (2) "induced communi-
ties" (see Section III) in one layer of a multilayer network
may depend on connectivity patterns in other layers. For
many notions of (single-layer or multilayer) community
structure -- including the most prominent methods --
one cannot exactly solve a community-assignment prob-
lem in polynomial time (unless P = NP) [30, 48 -- 50]; and
popular scalable heuristics currently have few or no the-
oretical guarantees on how closely an identified partition
resembles an optimal partition. These issues apply more
generally to the field of cluster analysis, such as in graph
partitioning [51], and many of the problems that plague
community detection also apply to detecting other kinds
of mesoscale structures.
Throughout our paper, to make it clear which results
and observations apply to community structure in par-
ticular and which apply to mesoscale structure more gen-
erally, we use the term "community" when referring to
a set in a partition of the set of nodes (or the set of
state nodes) of a network that corresponds to commu-
nity structure and the term "meso-set" (see Section III)
when referring to a set in a partition that corresponds to
any type of mesoscale structure. (In particular, a com-
munity is a type of meso-set.)
C. Generative models for mesoscale structure
The ubiquity and diversity of mesoscale structures in
empirical networks make it crucial to develop genera-
tive models of mesoscale structure that can produce fea-
tures that one encounters in empirical networks. Broadly
speaking, the goal of such generative models is to con-
struct synthetic networks that resemble real-world net-
works when one appropriately constrains and/or cali-
brates the parameters of a model using information about
a scenario or application. Generative models of mesoscale
structure can serve a variety of purposes, such as (1) gen-
erating benchmark network models for comparing meso-
set-detection methods and algorithms [10, 52 -- 56]; (2)
undertaking statistical inference on empirical networks
[10, 43, 57]; (3) generating synthetic networks with a de-
sired set of properties [20, 58]; (4) generating null models
to take into account available information about an em-
pirical network [59]; and (5) investigating "detectability
limits" for mesoscale structure, as one can plant par-
titions that, under suitable conditions, cannot subse-
quently be detected algorithmically, despite the fact that
they exist by construction [10, 60, 61].
One of the main challenges in constructing a realis-
tic generative model (even for single-layer networks) is
the breadth of possible empirical features in networks.
The available generative models for mesoscale structure
in single-layer networks usually focus on replicating a
few empirical features at a time (rather than all of
them at once): heterogeneous degree distributions and
community-size distributions [43, 53, 62], edge-weight
distribution [52, 57, 63], spatial embeddedness [12, 64],
and so on. Multilayer networks inherit all of the empiri-
cal features of single-layer networks, and they also have a
key additional one: dependencies between layers. These
interlayer dependencies can be ordered (as in most mod-
els of temporal networks), unordered (as in multiplex net-
works), combinations of these, or something more compli-
cated. However, despite this variety, existing generative
models for mesoscale structure in multilayer networks al-
low only a restrictive set of interlayer dependencies. They
assume either a temporal structure [10, 12, 54, 65 -- 67], a
simplified multiplex structure with the same planted par-
titions across all layers [20, 44, 68 -- 70], or independent
groups of layers such that layers in the same group have
identical planted partitions [9, 21]. Using an alternative
approach, a very recent model is able to generate multi-
layer partitions that satisfy the constraint that nonempty
induced meso-sets in different layers that correspond to
the same meso-set in the multilayer partition are iden-
tical [71]. Recent work [13] on the link between multi-
layer modularity maximization and maximum-likelihood
estimation of multilayer SBMs allows either temporal
or multiplex interlayer dependencies with induced parti-
tions that can vary across layers, but it makes restrictive
assumptions on interlayer dependencies (e.g., all layers
have the same set of nodes, interdependencies are "diag-
onal" and "layer-coupled", and so on). Importantly, in
all aforementioned generative models of mesoscale struc-
ture in multilayer networks, interlayer dependencies are
either (1) not explicitly specifiable or (2) special cases of
the framework that we discuss in the present paper.
III. NOTATION
We now present a comprehensive set of notation for
multilayer networks. Different subsets of notation are
useful for different situations.
We consider a multilayer network M = (VM ,EM ,V,L)
with n = V nodes (i.e., physical nodes) and l = L
layers (i.e., states). We use d to denote the number of
aspects and La = {1, . . . , la} to denote the labels of the
ath aspect (where a ∈ {1, . . . , d}). We use O to denote
the set of ordered aspects and U to denote the set of
unordered aspects. For each ordered aspect a ∈ O, we
5
assume that the labels La reflect the order of the aspect.
That is, for all α, β ∈ La, we require that α < β if and
only if α precedes β. We say that a multilayer network is
fully-ordered if U = ∅, unordered if O = ∅, and partially-
ordered otherwise. The set L = L1 × . . . × Ld of states
is the Cartesian product of the aspects, where a state
α ∈ L is an integer-valued vector of length d and each
entry specifies an element of the corresponding aspect.
Note that l = L =(cid:81)d
a=1 la.
We use (i, α) ∈ VM ⊆ V × L to denote the state node
(i.e., "node-layer tuple" [4]) of physical node i ∈ V in
state α ∈ L. We include a state node in VM if and only
if the corresponding node exists in that state. The edges
EM ⊆ VM × VM in a multilayer network connect state
nodes to each other. We use ((i, α), (j, β)) to denote a
directed edge from (i, α) to (j, β). For two state nodes,
(i, α) and (j, β), that are connected to each other via
a directed edge ((i, α), (j, β)) ∈ EM , we say that (i, α)
is an in-neighbor of (j, β) and (j, β) is an out-neighbor
of (i, α). We categorize the edges into intralayer edges
EL, which have the form ((i, α), (j, α)) and link entities i
and j in the same state α, and interlayer (i.e., coupling)
edges EC, which have the form ((i, α), (j, β)) for α (cid:54)= β.
We thereby decompose the edge set as EM = EL ∪ EC.
We define a weighted multilayer network by intro-
ducing a weight function w : EM → R (analogous to
the weight function for weighted single-layer networks),
which encodes the edge weights within and between lay-
ers. For an unweighted multilayer network, w(e) = 1 for
all e ∈ EM . We encode the connectivity pattern of a
multilayer network using an adjacency tensor A, which
is analogous to the adjacency matrix for single-layer net-
works, with entries
i,α =(cid:40)w(((i, α), (j, β))) ,
0 ,
Aj,β
((i, α), (j, β)) ∈ EM ,
otherwise .
(1)
Note that GM = (VM ,EM ) is a graph on the state
nodes of the multilayer network M . We refer to GM
as the flattened network of M . The adjacency matrix
of the flattened network is the "supra-adjacency ma-
trix" [4, 15, 72, 73] of the multilayer network. One ob-
tains a supra-adjacency matrix by flattening [74] an ad-
jacency tensor (see Eq. (1)) of a multilayer network. The
multilayer network and the corresponding flattened net-
work encode the same information [4, 75], provided one
keeps track of the correspondence between the nodes in
the flattened network and the physical nodes and lay-
ers of the multilayer network. In a similar vein, aspects
provide a convenient way to keep track of the correspon-
dence between state nodes and, for example, temporal
and multiplex relationships.
{S1, . . . ,Snset}, where(cid:83)nset
We denote a multilayer partition with nset sets by S =
s=1 Ss = VM and Ss∩Sr = ∅ for
s (cid:54)= r. We represent a partition S using a partition tensor
S with entries Si,α, where Si,α = s if and only if the state
node (i, α) is in the set Ss. We refer to a partition of a
temporal network (a common type of multilayer network)
as a temporal partition and a partition of a multiplex
network (another common type of multilayer network)
as a multiplex partition. A multilayer partition induces a
partition Sα = {S1α, . . . ,Snsetα} on each layer, where
Ssα = {i ∈ V : (i, α) ∈ Ss}. We refer to a set Ss
of a partition S as a meso-set, to Sα as the induced
partition of S on layer α, and to Ssα as the induced
meso-set of Ss on layer α. A community is a set Ss in a
partition that corresponds to community structure. We
call s ∈ {1, . . . , nset} the label of meso-set Ss. One way to
examine overlapping meso-sets is by identifying multiple
state nodes from a single layer with the same physical
node.
IV. GENERATING MULTILAYER PARTITIONS
The systematic analysis of dependencies between layers
is a key motivation for analyzing a single multilayer net-
work, instead of examining several single-layer networks
independently. The goal of our partition-generation pro-
cess is to model interlayer dependency in a way that can
incorporate diverse types of dependencies. We now mo-
tivate our partition-modeling approach, and we describe
it in detail in Sections IV A -- D.
The complexity of dependencies between layers can
make it difficult to explicitly specify a joint probabil-
ity distribution for meso-set assignments, especially for
unordered or partially-ordered multilayer networks. To
address this issue, we require only the specification of
conditional probabilities for a state node's meso-set as-
signment, given the assignments of all other state nodes.
The idea of conditional models is old and follows nat-
urally from Markov-chain theory [76, 77]. Specifying
conditional models (which capture different dependency
features separately) rather than joint models (which try
to capture many dependency features at once) is con-
venient for numerous situations. For example, condi-
tionally specified distributions have been applied to areas
such as spatial data modeling [78], imputation of missing
data [77], secure disclosure of information [79], depen-
dence networks for combining databases from different
sources [80], and Gibbs sampling [81, 82].
A problem with conditional models is that the differ-
ent conditional probability distributions are not neces-
sarily compatible, in the sense that there may not ex-
ist a joint distribution that realizes all conditional dis-
tributions [80, 82 -- 85]. Although several methods have
been developed in recent years for checking the compat-
ibility of discrete conditional distributions, these either
make restrictive assumptions on the conditional distri-
butions or have scalability constraints that significantly
limit practical use [77, 86]. Nevertheless, the employment
of conditional models (even if potentially incompatible) is
common [87], and they have been useful in many applica-
tions, provided that one cautiously handles any potential
incompatibility [77]. For our use case, potential incom-
patibility arises for unordered or partially-ordered mul-
6
tilayer networks. An issue that can result from it is the
non-uniqueness of a joint distribution. We carefully de-
sign our partition-generation process such that it is well-
defined irrespective of whether conditional distributions
are compatible. In particular, we show that convergence
is guaranteed and we address non-uniqueness by appro-
priately sampling initial conditions (see Appendix B). We
also suggest empirical checks to ensure that a generated
partition reflects planted interlayer dependencies.
In our framework, we define conditional probabilities
in two parts: we separately specify (1) independent layer-
specific random components and (2) interlayer dependen-
cies. For a choice of interlayer dependencies, there are
several features that one may want to allow in a multi-
layer partition. These include variation in the numbers
and sizes of meso-sets across layers (e.g., meso-sets can
gain state nodes, lose state nodes, appear, and disap-
pear) and the possibility for these meso-set variations
to incorporate features of the application at hand. (For
example, in a temporal network, one may not want a
meso-set to reappear after it disappears.) We ensure that
such features are possible for any choice of interlayer de-
pendency via independent layer-specific null distributions
that specify the set of possible meso-set assignments for
each layer and determine the expected size and expected
number of meso-sets in the absence of interlayer depen-
dencies.
Interlayer dependencies should reflect the type of mul-
tilayer network that one is investigating. For example,
in a temporal network, dependencies tend to be stronger
between adjacent layers and weaker for pairs of layers
that are farther apart. It is common to assume that in-
terlayer dependencies are uniform across state nodes for
a given pair of layers (i.e., interlayer dependencies are
"layer-coupled" [4]) and occur only between state nodes
that correspond to the same physical node (i.e., inter-
layer dependencies are "diagonal" [4]) [12, 16, 21, 33, 61,
72, 73, 88]. However, these assumptions are too restric-
tive in many cases (e.g., situations in which dependencies
are state-node-dependent or in which dependencies exist
between state nodes that do not correspond to the same
physical node) [9, 46, 70, 89 -- 91]. To ensure that one
can relax these assumptions, we allow dependencies to
be specified at the level of state nodes (or at the level of
layers, when a user assumes that dependencies are layer-
coupled). We encode these interlayer dependencies in
a user-specified interlayer-dependency tensor that deter-
mines the extent to which a state node's meso-set assign-
ment in one layer "depends directly" on the assignment of
state nodes in other layers (see Sections IV B and C). Our
independence assumption on the null distributions allows
us to encode all of the interlayer dependencies in a single
object (an interlayer-dependency tensor). The entries of
an interlayer-dependency tensor specify an "interlayer-
dependency network" and correspond to the causal links
for the flow of information (in the information-theoretic
sense) between different layers of a multilayer network.
Therefore, they should reflect any constraints that one
wishes to impose on the direct flow of information be-
tween layers. Longer paths in an interlayer dependency
network yield indirect dependencies between structures
in different layers.
After specifying the interlayer-dependency structure,
we define the conditional meso-set assignment probabil-
ities such that we either sample a state node's meso-
set assignment from the corresponding null distribution
or obtain it by copying the assignment of another state
node (based on the interlayer-dependency tensor). Using
these conditional probabilities, we define an iterative up-
date process on the meso-set assignments of state nodes
to generate multilayer partitions with dependencies be-
tween induced partitions in different layers. When updat-
ing the meso-set assignments of state nodes, we respect
the order of an aspect (e.g., temporal ordering). For a
fully-ordered multilayer network, our update process re-
duces to sequentially sampling an induced partition for
each layer based on the induced partitions of previous
layers. For an unordered multilayer network, our update
process defines a Markov chain on the space of multilayer
partitions. We sample partitions from a stationary dis-
tribution of this Markov chain. This sampling strategy is
known as (pseudo-)Gibbs sampling [80 -- 82, 87]. (We use
the word "pseudo" because the conditional probabilities
that we use to define the Markov chain are not necessarily
compatible.) For a partially-ordered multilayer network,
our update process combines these two sampling strate-
gies.
In Section IV A, we describe possible choices for the
independent layer-specific null distributions.
In Sec-
tion IV B, we explain our framework for generating a mul-
tilayer partition with general interlayer dependencies. In
Section IV C, we focus on the specific situation in which
interlayer dependencies are layer-coupled and diagonal,
and we also assume that a physical node is present in
all layers (i.e., the network is "fully interconnected" [4]).
In Section IV D, we illustrate the properties of example
temporal and multiplex partitions that are generated by
models that we construct from our framework. Addition-
ally, we take advantage of the tractability of the special
case of temporal networks to highlight some of its prop-
erties analytically.
A. Null distribution
We denote the null distribution of layer α by Pα
the set of all null distributions by P0 = {Pα
0 and
0 , α ∈ L}.
A simple choice for the null distributions is a categorical
distribution, where for each layer α and each meso-set
label s, we fix the probability pα
s that an arbitrary state
node in layer α is assigned to a meso-set s in the absence
of interlayer dependencies. That is,
7
s=1 pα
tilayer partition and (cid:80)nset
where nset is the total number of meso-sets in the mul-
s = 1 for all α ∈ L. The
set {1, . . . , nset} is the set of meso-set labels, and the
support of a null distribution is the set of labels that
have nonzero probability.
In the absence of interlayer
dependencies, a categorical null distribution corresponds
to specifying independent multinomial distributions for
the sizes of induced meso-sets on each layer and fixing
the expected size npα
s of each induced meso-set. There-
fore, by choosing the probabilities pα
s , one has some con-
trol over the expected number and sizes of meso-sets in
a sampled multilayer partition. A natural choice for pα
is to sample it from a Dirichlet distribution, which is the
conjugate prior for the categorical distribution [92, 93].
One can think of the Dirichlet distribution, which is the
multivariate form of the beta distribution, as a proba-
bility distribution over the space of all possible categori-
cal distributions with a given number of categories. Any
other (probabilistic or deterministic) choice for pα is also
possible. We give further detail about the Dirichlet dis-
tribution in Appendix A, where we discuss how one can
vary its parameters to control the expected number and
sizes of meso-sets in the absence of interlayer dependen-
cies. This can allow a user to generate, for example, a null
distribution with equally-sized meso-sets or a null distri-
bution with a few large meso-sets and many small meso-
sets. Furthermore, irrespective of the particular choice
of categorical distribution, it may be desirable to have
meso-sets that have a nonzero probability of obtaining
state nodes from the null distribution in some, but not
all, layers. In Appendix A, we give examples of how one
can sample the support of the null distributions before
the probability vectors to incorporate this property when
modeling the birth and/or death of meso-sets in temporal
networks and the presence and/or absence of meso-sets
in multiplex networks.
In general, the choice of the null distributions can have
a large effect on the set of sampled multilayer parti-
tions and ought to be guided by one's use case. For
our numerical examples in Section VI, we fix a value
of nset ∈ {1, . . . , nl}, where n is the number of physi-
cal nodes in each layer and l the number of layers (see
Section III); and we use a symmetric Dirichlet distri-
bution with parameters q = nset and θ = 1 (see Ap-
pendix A) to sample probability vectors pα of length
nset. This produces multilayer partitions in which the
expected meso-set labels are the same across layers (and
given by {1, . . . , nset}) in the absence of interlayer de-
pendencies and for which the expected induced meso-set
sizes (which are npα
s in the absence of interlayer depen-
dencies) differ across layers.
B. General interlayer dependencies
Pα
0 [s] =(cid:40)pα
s ,
0 ,
s ∈ {1, . . . , nset} ,
otherwise ,
(2)
We denote the user-specified interlayer-dependency
tensor by P , where P j,β
i,α is the probability that state
node (j, β) copies its meso-set assignment from state
node (i, α), for any two state nodes (i, α), (j, β) ∈ VM ,
and where P is fixed throughout the copying process.
The probability that state node (j, β) copies its meso-set
assignment from an arbitrary state node when state node
(j, β)'s meso-set assignment is updated is
pj,β = (cid:88)(i,α)∈VM
P j,β
i,α ,
(3)
where we require that pj,β ≤ 1 for all state nodes
(j, β) ∈ VM . We also require that all intralayer proba-
bilities are 0; that is, P j,α
i,α = 0 for all i, j ∈ V and α ∈ L.
We say that a state node (j, β) depends directly on a
state node (i, α) if and only if P j,β
i,α is nonzero. By ex-
tension, we say that a layer β depends directly on a
layer α if there exists at least one state node in layer
β that depends directly on a state node in layer α.
The interlayer-dependency tensor induces the interlayer-
dependency network, whose edges are all interlayer, di-
rected, and pointing in the direction of information flow
between layers. The in-neighbors (see Section III) of a
state node (j, β) in this network consists of the state
nodes from which (j, β) can copy a meso-set assignment
with nonzero probability. The support of the null distri-
bution Pβ
0 corresponds to the only possible meso-set as-
signments for a state node (j, β) whenever the state node
does not copy its assignment from one of its in-neighbors
in the interlayer-dependency network.
For a given null distribution P0 and interlayer-
dependency tensor P , a multilayer partition that results
from our sampling process depends on four choices: (1)
the way in which we update a state node assignment at a
given step; (2) the order in which we update state node
assignments; (3) the initial multilayer partition; and (4)
the criterion for convergence of the iterative update pro-
cess. We discuss points (1), (2), and (3) in the remainder
of this subsection. We describe the sampling process in
more detail and discuss convergence in Appendix B.
Update equation
A single meso-set-assignment update depends only on
the choice of state node to update and on the current
multilayer partition. Let τ be an arbitrary update step
of the copying process. Suppose that we are updating
the meso-set assignment of state node (j, β) at step τ
and that the current multilayer partition is S(τ ) (with
partition tensor S(τ )). We update the meso-set assign-
ment of state node (j, β) either by copying the meso-set
assignment in S(τ ) from one of its in-neighbors in the
interlayer-dependency network or by obtaining a new,
random meso-set assignment from the null distribution
Pβ
0 for layer β.
In particular, with probability pj,β, a
state node (j, β) copies its meso-set assignment from one
of its in-neighbors in the interlayer-dependency network;
and with probability 1 − pj,β, it obtains its meso-set as-
signment from the null distribution Pβ
0 . This yields the
following update equation at step τ of our copying pro-
cess:
8
P[Sj,β(τ + 1) = sS(τ )]
=(cid:88)(i,α)∈VM
P j,β
i,α δ(Si,α(τ ), s)
(4)
+ (1 − pj,β) Pβ
0 [Sj,β = s] .
The update equation (4) is at the heart of our frame-
work. It is clear from Eq. (4) that the set of null distribu-
tions P0 is responsible for the specification of meso-set as-
signments in the absence of interlayer dependencies (i.e.,
if P j,β
i,α = 0 for all (i, α), (j, β)). In general, pj,β deter-
mines the relative importance of interlayer dependencies
and the null distribution on the meso-set assignment of
state node (j, β). Specifically, when pj,β = 0, the meso-
set assignment of (j, β) depends only on the null distri-
bution; and when pj,β = 1, the meso-set assignment of
(j, β) depends only on the meso-set assignments of its
in-neighbors in the interlayer-dependency network.
Update order
The order in which we update meso-set assignments of
state nodes via Eq. (4) can influence a generated mul-
tilayer partition. As we mentioned in Section II A, an
aspect of a multilayer network can be either ordered or
unordered, and the update order is of particular impor-
tance when generating a multilayer partition with at least
one ordered aspect.
In particular, the structure of the
interlayer-dependency tensor should reflect the causality
that arises from the order of the aspect's elements. For an
ordered aspect, structure in a given layer should depend
directly only on structure in previous layers. Formally,
for each ordered aspect a ∈ O of a multilayer network,
we require that
P j,β
i,α > 0 ⇒ αa ≤ βa ,
(5)
where αa denotes the element of state α corresponding
to aspect a and where (as we stated in Section III) we
require that the labels of an aspect's elements reflect the
ordering of those elements [94]. For example, if we think
of the interlayer edges in Fig. 1 as interlayer edges in the
interlayer-dependency network, then Eq. (5) states that
all edges with nonzero edge weights must respect the ar-
row of time. That is, it is impossible to have edges of the
form ((j, (2, α2)), (i, (1, α2))), with nodes i, j ∈ {1, 2, 3}
and aspect α2 ∈ {1, 2, 3}.
The update order for the state nodes also needs to
reflect the causality from the ordered aspects. In partic-
ular, if a ∈ O is an ordered aspect and we update a state
node in a layer β, then the final meso-set assignment for
state nodes in layers α with αa < βa should already be
fixed. For example, in Fig. 1, an edge direction should re-
spect the arrow of time; and it is also necessary that the
assignment of all state nodes in the first time point are
fixed before one updates the assignments of state nodes
in the second time point. To achieve the latter, we di-
vide the layers into sets of "order-equivalent" layers, such
that two layers α and β are order-equivalent if and only
if αa = βa for all a ∈ O. We also say that layer α pre-
cedes layer β if and only if αa < βa for all a ∈ O. (For
example, the three layers in Fig. 1 that correspond to
the first time point are order-equivalent and precede the
three layers that correspond to the second time point.)
Based on this equivalence relation, we obtain an ordered
set of equivalence classes by inheriting the order from the
ordered aspects. If we sort the classes of order-equivalent
layers based on "lexicographic" order [95] of the ordered
aspects, then (as a consequence of Eq. (5)) the meso-set
assignment of state nodes in a given layer depends only on
the assignment of state nodes in order-equivalent or pre-
ceding layers. This ensures that our partition-generation
process reflects notions of causality that arise in multi-
layer networks with at least one ordered aspect.
9
same distribution.) For an unordered multilayer network,
our sampling algorithm reduces to (pseudo-)Gibbs sam-
pling, because all layers are order-equivalent and there
is thus only one class of order-equivalent layers. For a
partially-ordered multilayer network, there are multiple
non-singleton classes of order-equivalent layers. We use
(pseudo-)Gibbs sampling on each class, conditional on
the meso-set assignment of state nodes in preceding lay-
ers.
In Appendix B, we explain our sampling process in
more detail. We describe our (pseudo-)Gibbs sampling
procedure for each class of order-equivalent layers, and
we show pseudo-code that one can use to sample parti-
tions in unordered, partially-ordered, and fully-ordered
multilayer networks in Algorithm 1. We also discuss
the convergence properties of our sampling procedure for
unordered and partially-ordered multilayer networks, in-
cluding the effect of the potential incompatibility of the
distributions that are defined by Eq. (4).
Sampling process
C. Layer-coupled and diagonal interlayer
dependencies
The general idea of our sampling process is to simul-
taneously sample the meso-set assignment of state nodes
within each class of order-equivalent layers and to se-
quentially sample the meso-set assignment of state nodes
in non-order-equivalent layers, conditional on the fixed
assignment of state nodes in preceding layers.
from the null distribution (i.e., Si,α(0) ∼ Pα
More specifically, our sampling algorithm proceeds as
follows. First, we sample an initial multilayer partition
0 ). Sec-
ond, we sample a partition for the first class of order-
equivalent layers using (pseudo-)Gibbs sampling. In par-
ticular, we iteratively sample a layer uniformly at ran-
dom from the first class of order-equivalent layers and
update the meso-set assignment of all state nodes in that
layer [96] based on Eq. (4). This defines a Markov chain
on a subspace of multilayer partitions. We repeat the
update process for sufficiently many iterations such that
we approximately sample the meso-set assignments of
the first class of order-equivalent layers from a station-
ary distribution of this Markov chain. Third, we sample
state node assignments for subsequent classes of order-
equivalent layers in the same way, based on a fixed state-
node assignment from preceding layers. In particular, at
each update step τ in Eq. (4), a state node can either
copy a fixed meso-set assignment from an in-neighbor in
a preceding layer, copy a current meso-set assignment
from an in-neighbor in an order-equivalent layer, or ob-
tain a meso-set assignment from the null distribution.
For a fully-ordered multilayer network (i.e., a multi-
layer network with U = ∅), each class of order-equivalent
layers consists only of a single layer. Consequently,
one needs only a single update for each state node
for the sampling process to converge. (Subsequent up-
dates would constitute independent samples from the
As we mentioned at the beginning of Section IV, a par-
ticularly useful restriction of the interlayer-dependency
tensor that still allows us to represent many situations of
interest is to assume that it is layer-coupled and diago-
nal. For simplicity, we also assume that the multilayer
network is fully interconnected. Under these assump-
tions, we can express the interlayer-dependency tensor
P using a layer-coupled interlayer-dependency tensor (cid:101)P
with elements P j,β
α . The probability that
state node (j, β) copies its meso-set assignment from an
arbitrary state node when (j, β)'s meso-set assignment is
updated is given by
i,α = δ(i, j)(cid:101)P β
pβ = (cid:88)α∈L(cid:101)P β
α .
(6)
(7)
α = 0 for all
α ∈ L. The update equation (Eq. (4)) then simplifies to
As before, we require that pβ ≤ 1 and (cid:101)P α
P[Sj,β(τ + 1) = sS(τ )]
=(cid:88)α∈L(cid:101)P β
+ (1 − pβ) Pβ
α δ(Sj,α(τ ), s)
0 [Sj,β = s] ,
(which is now independent of state nodes) and the null
α quantifies the extent to
which an induced partition in layer β depends directly on
an induced partition in layer α. By considering different
which depends on the interlayer-dependency tensor (cid:101)P
distributions P0. Each term (cid:101)P β
(cid:101)P , one can generate multilayer networks that correspond
to several important scenarios, including temporal net-
works, multiplex networks, and multilayer networks with
10
(a) Layer-coupled temporal
(b) Layer-coupled multiplex
dependencies
dependencies
Figure 2.
Layer-coupled interlayer dependency tensors
(which, in this case, are matrices) for different types of multi-
layer networks with a single aspect. (a) In a typical temporal
network, an induced partition in a layer depends directly only
on the induced partition in the previous layer. Therefore, the
only nonzero elements of the layer-coupled interlayer depen-
dency tensor occur in the first superdiagonal. (b) In a typical
multiplex network, an induced partition in a layer depends di-
rectly on the induced partitions in all other layers. We show
a layer-coupled example. The copying probability in Eq. (6)
is pβ for the temporal layer-coupled interlayer-dependency
matrix in panel (a) and p for the multiplex layer-coupled
interlayer-dependency matrix in panel (b). We suppress the
subscript β in panel (b), as the copying probability is the
same for all layers.
more than one aspect (e.g., with combinations of tempo-
ral and multiplex features). That is, the above restriction
reduces the dimension of the interlayer-dependency ten-
sor significantly, while still allowing one to analyze several
very important situations.
In Fig. 2, we show layer-coupled interlayer-dependency
tensors for two types of single-aspect multilayer networks:
a temporal network and a multiplex network. As we men-
tioned in Section IV, it is useful to think of interlayer
dependencies as causal links for the flow of information
between layers. In a temporal network, it is typical to
assume that an induced partition in a layer depends di-
rectly only on induced partitions in the previous layer.
There are thus l − 1 copying probabilities (one for each
pair of consecutive layers), which we are free to choose.
Common examples include choosing the same probability
for each pair of consecutive layers [10, 12] or making some
of the probabilities significantly smaller than the others
to introduce change points [44, 97]. In a multiplex net-
work, an induced partition in any layer can in principle
depend directly on induced partitions in all other layers.
This yields l(l − 1) copying probabilities to choose. In
Fig. 2b, we illustrate the simplest case, in which each
layer depends equally on every other layer. The layer-
coupled interlayer-dependency tensors in Fig. 2 are ma-
trices, so we sometimes refer to them and other examples
of layer-coupled interlayer-dependency tensors with a sin-
gle aspect as layer-coupled interlayer-dependency matri-
Figure 3. Block-matrix representation of a layer-coupled
interlayer-dependency tensor for a multilayer network that
is both temporal and multiplex. This example has more
than one aspect, and it combines the features of the exam-
ples in Fig. 2.
It has two classes of order-equivalent layers
(where each corresponds to one time point), as indicated by
the presence of two diagonal blocks. An induced partition
in a layer depends directly on the induced partitions of its
order-equivalent layers and on the induced partition of the
corresponding layer in the preceding class of order-equivalent
layers. If we think of the interlayer edges in Fig. 1 as edges in
an interlayer-dependency network, this matrix with diagonal
blocks of size 3 × 3 would be the corresponding layer-coupled
interlayer-dependency tensor.
ces.
We can also generate multilayer networks with more
than one aspect and can thereby combine temporal and
multiplex features. In Fig. 3, we illustrate how to con-
struct a layer-coupled interlayer-dependency tensor to
generate such a multilayer network on a simple exam-
ple with two aspects, one of which is multiplex and the
other of which is temporal.
D. Temporal and multiplex partitions
In Fig. 4 and Fig. 5, we show example multilayer parti-
tions that we obtain with the interlayer-dependency ten-
sors of Fig. 2. The examples in Fig. 4 are temporal,
and the examples in Fig. 5 are multiplex. For simplicity,
we assume in Fig. 2a that dependencies between con-
tiguous layers are uniform (i.e., pβ = p ∈ [0, 1] for all
β ∈ {2, . . . , l}). To illustrate the effect of the interlayer
dependencies, we also show a heat map of the normalized
mutual information (NMI) [98] between induced parti-
tions in different layers. NMI is a measure of similarity
between partitions, so we expect the values of NMI to
0p2000pl00eP=ePβα=δ(α+1,β)pβ,pβ=pβ≤10pppppp0eP=ePβα=(1−δ(α,β))p,p=(l−1)p≤10p1,1p1,1p1,200p1,10p1,10p1,1p1,1000p1,2000p2,2p2,2p2,2p2,200p2,2p2,20α2=1α2=2β2=1β2=2eP=ePβ1,β2α1,α2=(1−δ(α1,β1))δ(α2,β2)pβ1,β2+δ(α1,β1)δ(α2+1,β2)pβ1,β2 11
(a) p = 0
(b) p = 0.5
(c) p = 0.85
(a) p = 0
(b) p = 0.5
(c) p = 0.85
(d) p = 0.95
(e) p = 0.99
(f) p = 1
(d) p = 0.95
(e) p = 0.99
(f) p = 1
Figure 4. Example temporal partitions for (n, l) = (150, 100).
(Recall that n is the number of physical nodes and l is the
number of layers.) We use the interlayer-dependency ten-
sor from Fig. 2a with uniform probabilities pβ = p for all
β ∈ {1, . . . , l} and a Dirichlet null distribution with q = 1,
θ = 1, and nset = 5 (see Appendix A). For (a) p = 0, (b)
p = 0.5, (c) p = 0.85, (d) p = 0.95, (e) p = 0.99, and (f) p = 1,
, top) for a
we show color-coded meso-set assignments (
single example output partition and NMI values [98] (0
1,
bottom) between induced partitions in different layers aver-
aged over a sample of 10 output partitions. The parameter
values match those in the numerical examples in Section VI
(with the exception of p = 0, which we include for complete-
ness). We choose a node ordering for each visualization that
(whenever possible) emphasizes "persistent" mesoscale struc-
ture [16]. We show only the first 15 layers of each multilayer
partition, because (as one can see in the NMI heat maps) sim-
ilarities between induced partitions for p < 1 decay steeply
with the number of layers when dependencies exist only be-
tween contiguous layers.
reflect the planted dependencies. As expected, for the
temporal examples, partitions are most similar for adja-
cent layers. Similarities decay with the distance between
layers and increase with the value of p. For the multiplex
examples, we obtain approximately uniform similarities
between pairs of layers (for all pairs of layers), where
the similarities increase with the value of p. We use the
examples of interlayer dependency of Figs. 4 and 5, as
well as non-uniform and multi-aspect examples, in our
Figure 5. Example multiplex partitions for (n, l) = (1000, 15).
We use the interlayer-dependency tensor in Fig. 2b and a
Dirichlet null distribution with q = 1, θ = 1, and nset = 10
(see Appendix A). We perform 200 updating iterations (see
Appendix B). For (a) p = 0, (b) p = 0.5, (c) p = 0.85, (d)
p = 0.95, (e) p = 0.99, and (f) p = 1, where p = (l− 1)p is the
probability that a state node copies its assignment from an-
other state node, we show color-coded meso-set assignments
, top) for a single example output partition and NMI
(
values [98] (0
1, bottom) between induced partitions in
different layers averaged over a sample of 10 output parti-
tions.
(For the temporal case in Fig. 4, note that p = p.)
The parameter values match those in the numerical examples
in Section VI (with the exception of p = 0, which we include
for completeness). We choose a node ordering that (whenever
possible) emphasizes "persistent" mesoscale structure [16].
numerical experiments of Section VI.
To provide a detailed illustration of the steps that re-
sult in a multilayer partition, we focus on the temporal
examples in Fig. 4. For the important special case of
temporal interlayer dependencies, the generative model
for multilayer partitions simplifies significantly. In par-
ticular, there is a single ordered aspect, so the layer index
α ∈ N is a scalar and the order of the layers corresponds
to temporal ordering. Furthermore, as we mentioned in
Section IV B, for a fully-ordered multilayer network, we
require that the order of the meso-set-assignment update
process in Eq. (4) respects the order of the layers. The
update order of state nodes (1, α) . . . (n, α) in any given
151015150100150layernodePartition150100150100layerlayerNMI151015150100150layernodePartition150100150100layerlayerNMI151015150100150layernodePartition150100150100layerlayerNMI151015150100150layernodePartition150100150100layerlayerNMI151015150100150layernodePartition150100150100layerlayerNMI151015150100150layernodePartition150100150100layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI15101515001000layernodePartition151015151015layerlayerNMI 0
layer α can be arbitrary (so we can update them simulta-
neously), but each update is conditional on the meso-set
assignment of state nodes in layer α − 1.
The update process that we described in Section IV B
reduces to three steps:
(1) initialize (in an arbitrary
order) the meso-set assignments in layer α = 1; (2)
take the meso-set assignment of (i, α + 1) to be that of
(i, α) with probability p, and sample the meso-set as-
signment of (i, α + 1) from Pα+1
with complementary
probability 1 − p; and (3) repeat steps (1) and (2) until
α + 1 = l. As we mentioned in Section IV B, conver-
gence is not an issue for this case (or, more generally,
for any fully-ordered multilayer network), as we need
only one iteration through the layers. This three-step
generative model for temporal partitions was also sug-
gested by Ghasemian et al [50]. They used it to derive
a detectability threshold for the case in which the null
distributions are uniform across meso-sets (i.e., θ → ∞
in Section IV A) and intralayer edges are generated in-
dependently using the standard SBM. In other words,
one replaces the degree-corrected SBM in Section V with
the non-degree-corrected SBM of [42]. In Appendix C,
we highlight properties of a sampled temporal partition
that illustrate how the interplay between p and the null
distributions affects the evolution of meso-sets across lay-
ers (e.g., growth, birth, and death of induced meso-sets).
The properties that we highlight hold for any choice of
null distribution. In the same appendix, we subsequently
illustrate that the particular choice of null distributions
can greatly influence resulting partitions. For example, a
non-empty overlap between the supports of null distribu-
tions that correspond to contiguous layers is a necessary
condition for meso-sets to gain new state nodes over time.
The properties and observations in Appendix C are in-
dependent of one's choice of planted-partition network
model.
V. GENERATING NETWORK EDGES
There are diverse different types of multilayer networks
[4, 8]. One common type of empirical multilayer net-
work to study is one with only intralayer edges. There
are also many empirical multilayer networks with both
intralayer and interlayer edges (e.g., multimodal trans-
portation networks [99]), as well as ones with only in-
terlayer edges (e.g., temporal networks with edge delays,
such as departures and arrivals of flights between air-
ports [5, 100]). One can use our framework to generate
all three types of examples, provided the underlying edge
generation model is a planted-partition network model.
Having generated a multilayer partition S with de-
pendencies between induced partitions in different layers,
the simplest way to generate edges is to use any single-
layer planted-partition network model (e.g., SBMs [42,
43, 49, 57, 63, 101], models for spatially-embedded net-
works [12, 64], and so on) and to generate edges for each
layer independently. This yields a multilayer network
12
with only intralayer edges, such that any dependencies
between different layers result only from dependencies
between the partitions that are induced on the differ-
ent layers. For our numerical experiments in Section VI,
we use a single-layer network model that is a slight
variant (avoiding the creation of self-edges and multi-
edges) of the degree-corrected SBM (DCSBM) bench-
mark from [43], where "DCSBM benchmark" designates
the specific type of DCSBM that was used in the nu-
merical experiments of [43]. We include pseudocode for
our implementation of the DCSBM benchmark in Algo-
rithm 3 of Appendix D.
One can also include dependencies between layers
other than those that are induced by planted mesoscale
structures. For example, one can introduce dependen-
cies between the parameters of a single-layer planted-
partition network model by (1) sampling them from a
common probability distribution (e.g., interlayer degree
correlations [4, 102] in a DCSBM) or by (2) introduc-
ing interlayer edge correlations, given a single-layer par-
tition on each layer [103]. For temporal networks, one
can also incorporate "burstiness" [2, 3, 104] in the inter-
event-time distribution of edges. In such a scenario, the
probability for an edge to exist in a given layer depends
not only on the induced partition on that layer but also
on the existence of the edge in previous layers. For ex-
ample, one can use a Hawkes process to specify the time
points at which edges are active [105, 106].
In Appendix D, we describe a generalization of the
DCSBM in [43] to multilayer networks that one can use
to generate intralayer edges and/or interlayer edges. Our
generalization constitutes a framework within which we
formulate the parameters of a single-layer DCSBM in a
multilayer setting. One can use our multilayer DCSBM
(M-DCSBM) framework to incorporate some of the fea-
tures that we described in the previous paragraph (e.g.,
degree correlations) in a multilayer network with in-
tralayer and/or interlayer edges.
VI. NUMERICAL EXAMPLES
In this section, we use our framework to construct
benchmark models for multilayer community-detection
methods and algorithms. We use the examples of
interlayer-dependency tensors from Figs. 2 and 3 to gen-
erate benchmark models. We also consider a variant of
Fig. 2b in which we split the layers into groups, use uni-
form dependencies between layers in the same groups,
and treat layers in different groups as independent of each
other. These examples cover commonly studied temporal
and multiplex dependencies, and they illustrate how one
can generate benchmark multilayer networks with more
than one aspect. We focus on a couple of popular qual-
ity functions for multilayer community detection in our
numerical examples, rather than investigating any given
method or algorithm in detail.
We compare the behavior of several variants of [107],
which are similar to the the locally greedy Louvain com-
putational heuristic [108], to optimize a multilayer mod-
ularity objective function [16, 33] using the standard
Newman -- Girvan null model (which is a variant of a "con-
figuration model" [109]). Modularity is an objective func-
tion that is often used to partition sets of nodes into
communities that have a larger total internal edge weight
than the expected total internal edge weight in the same
sets in a "null network" [16], which is generated from
some null model. Modularity maximization consists of
finding a partition that maximizes this difference. For
our numerical experiments, we use the generalization of
modularity to multilayer networks of [33]. For multilayer
modularity, the strength of the interactions between dif-
ferent layers are governed by an interlayer-coupling ten-
sor that controls the incentive for state nodes in different
layers to be assigned to the same community. We use
multilayer modularity with uniform interlayer coupling,
so the strength of the interactions between different lay-
ers of the network depends on a layer-independent and
node-independent interlayer coupling weight ω ≥ 0. We
use diagonal and categorical (i.e., between all pairs of lay-
ers) interlayer coupling with weight ω for the multiplex
examples in Section VI A. We use diagonal and ordinal
(i.e., between contiguous layers) interlayer coupling with
weight ω for the temporal examples in Section VI B. In
Appendix E, we describe the Louvain algorithm and the
variants of it that we use in this section.
We compare the results of multilayer modularity max-
imization with those of multilayer InfoMap [110] [9],
which uses an objective function called the "map equa-
tion" (which is not an equation), based on a discrete-time
random walk and ideas from coding theory, to coarse-
grain sets of nodes into communities [111].
In multi-
layer InfoMap, one uses a probability r ∈ [0, 1] called
the "relaxation rate" to control the relative frequency
with which a random walker remains in the same layer
or moves to other layers.
(A random walker cannot
change layers when r = 0.) The relaxation rate thus con-
trols the interactions between different layers of a mul-
tilayer network. We allow the random walker to move
to all other layers when r (cid:54)= 0 for the multiplex exam-
ples in Section VI A, and we allow the random walker
to move only to adjacent layers for the temporal exam-
ples in Section VI B. In contrast to multilayer modularity
(where ω = 0 yields single-layer modularity), single-layer
InfoMap is not equivalent to choosing r = 0 in mul-
tilayer InfoMap because placing state nodes that cor-
respond to the same physical node in the same commu-
nity can contribute positively to the quality function even
when r = 0 [9]. Consequently, we compute single-layer
InfoMap separately and reference it by the data point
"s" on the horizontal axis.
In all experiments in this section, we generate a mul-
tilayer partition using our copying process in Section IV
(see Algorithm 1 in Appendix B) and a multilayer net-
work for a fixed planted partition using a slight variant
of the DCSBM benchmark network model in [43] that
13
avoids generating self-edges and multi-edges (see Algo-
rithm 3 in Appendix D). This produces multilayer net-
works that have only intralayer edges and in which the
connectivity patterns in different layers depend on each
other. Following [43], we parametrize the DCSBM bench-
mark in terms of its distribution of expected degrees and
a community-mixing parameter µ ∈ [0, 1] that controls
the strength of the community structure in the sampled
network edges. For µ = 0, all edges lie within commu-
nities; for µ = 1, edges are distributed independently of
the communities, where the probability of observing an
edge between two state nodes in the same layer depends
only on the expected degrees of those two state nodes.
We use a truncated power law for the distribution of ex-
pected degrees.
The DCSBM benchmark imposes community structure
as an expected feature of an ensemble of networks that
it generates. The definition of the mixing parameter
µ of the DCSBM benchmark ensures that the planted
partition remains community-like -- as intra-community
edges are more likely to be observed and inter-community
edges are less likely to be observed than in a network that
is generated from a single-block DCSBM with the same
expected degrees -- for any value of µ < 1. Consequently,
given sufficiently many samples from the same DCSBM
benchmark (i.e., all samples have the same planted par-
tition and same expected degrees), one should be able to
identify the planted community structure for any value
of µ < 1 (where the necessary number of samples goes
to infinity as µ → 1). This feature makes the DCSBM
benchmark an interesting test for the ability of multilayer
community-detection methods to aggregate information
from multiple layers.
We use NMI [98] to compare the performance of dif-
ferent community-detection algorithms. For each of our
partitions, we compute the mean of the NMI between the
partition induced on each layer by the output partition
and that induced by the planted partition. That is,
(cid:104)NMI(cid:105)(S,T ) =
1
l (cid:88)α∈L
NMI(Sα,T α) .
In particular,
The quantity (cid:104)NMI(cid:105) is invariant under permutations
of the meso-set labels within a layer. Consequently,
(cid:104)NMI(cid:105) is well-suited to comparing multilayer community-
detection methods with single-layer community-detection
methods.
it allows us to test whether
multilayer community-detection methods can exploit de-
pendencies between layers of a multilayer network when
p (cid:29) 0 (see Eq. (6)). In Appendix F, we show numerical
experiments in which we compute NMI between multi-
layer partitions. We denote the NMI between two multi-
layer partitions by mNMI.
In all of our numerical experiments in Sections VI A
and B, we sample the benchmark networks in the fol-
lowing way. For each value of p, we generate 10 sample
partitions. For each sample partition and value of µ, we
generate 10 sample multilayer networks. This yields 100
benchmark instantiations for each pair (p, µ). We run
each community-detection algorithm 10 times on each
In Figs. 6 -- 9, we show (cid:104)NMI(cid:105) between
instantiation.
planted and recovered multilayer partitions averaged over
sample partitions, sample networks, and algorithmic runs
for each value of µ. The results for different planted par-
titions are generally similar. The only exception is In-
foMap on temporal networks with p = 0.99 and p = 1,
where we observe large differences between results for
different partitions for certain values of µ.
Figures 6 -- 9 correspond to different choices for the
interlayer-dependency tensor.
In each figure, rows cor-
respond to different choices for the community-detection
algorithm, and columns correspond to different values
of the copying probabilities (with the strength of inter-
layer dependencies increasing from left to right). All
benchmark instantiations that we use for Figs. 6 -- 10
are available at https://dx.doi.org/10.5281/zenodo.
3304059.
A. Multiplex examples
In this section, we consider two stylized examples of
multiplex networks. Multiplex networks arise in a variety
of different applications, including international relations
and trade [112 -- 114], social networks [115], and ecologi-
cal networks [116]. In our multiplex examples, we con-
sider simple dependency structures in which we expect
multilayer community-detection methods to outperform
single-layer methods by exploiting interlayer dependen-
cies.
In Fig. 6, we consider multiplex networks with uni-
form dependencies between community structure in dif-
ferent layers.
In Fig. 7, we consider multiplex net-
works with nonuniform dependencies between commu-
nity structure in different layers.
In both figures, we
parametrize the amount of interlayer dependency in a
network by the probability p (see Eq. (6)) that a state
node copies its community assignment from a neighbor in
the interlayer-dependency network. All multilayer net-
works have n = 1000 physical nodes and l = 15 lay-
ers. Each node is present in every layer, so there are
a total of 15000 state nodes.
In Fig. 2b, we show the
layer-coupled interlayer-dependency matrix that we use
to generate the uniform multiplex networks. For the
nonuniform multiplex networks, we split the layers into
3 groups of 5 layers each. We use uniform dependencies
between layers in the same group, and layers in differ-
ent groups are independent of each other. The resulting
layer-coupled interlayer-dependency matrix is block diag-
onal with diagonal blocks as in Fig. 2b and 0 entries on
the off-diagonal blocks. We use a Dirichlet null distribu-
tion with nset = 10, θ = 1, and q = 1 (see Appendix A)
to specify expected community sizes and the M-DCSBM
benchmark (see Appendix D) with ηk = 2, kmin = 3, and
kmax = 150 to generate intralayer edges. We perform 200
iterations of our update process (see Appendix B).
Comparing our results for GenLouvain (Figs. 6a -- e
14
and 7a -- c) and GenLouvainRand with reiteration and
post-processing (Figs. 6f -- j and 7d -- f), we see that the
choice of optimization heuristic for a given quality func-
tion can significantly affect the quality of resulting iden-
tified partitions. In particular, for GenLouvain, we ob-
serve two distinct regimes and what appears to be a sharp
transition between them [117]. For ω < 1, we obtain a
partition that is of similar quality to what we obtain by
maximizing single-layer modularity; however, for ω > 1,
we obtain a partition with identical induced partitions
for each layer. We call the latter an aggregate partition.
Although the aggregate partition can be more similar to
the planted partition than the single-layer partition when
p is sufficiently large (e.g., this occurs in Figs. 6d and e),
we do not observe an interval of ω values between the
two regimes in which GenLouvain recovers the planted
partition better than both the single-layer and aggregate
partitions. By contrast, GenLouvainRand with reiter-
ation and post-processing identifies partitions that match
the planted partition more closely than either the single-
layer or aggregate partitions in most cases. The excep-
tions are uniform multiplex networks with p = 0.5 (see
Fig. 6f), where it is unable to exploit the weak depen-
dencies to outperform the single-layer case; and p = 1
(Fig. 6j), where the aggregate partition is always best.
Most of the improvement in the results for GenLou-
vainRand over those for GenLouvain comes from re-
iteration. The additional randomization helps smooth
out the transition at ω ≈ 1. Post-processing only yields
a minor improvement in the value of (cid:104)NMI(cid:105). However,
its effect is more pronounced if we compute the mNMI
between multilayer partitions (see Appendix F).
Multilayer InfoMap exhibits some problematic be-
havior in these benchmark experiments. Our results
for multilayer InfoMap are noticeably worse than those
that we obtain with single-layer InfoMap (correspond-
ing to the data point "s" on the horizontal axis) for
networks with comparatively weak community structure
(specifically, µ ≥ 0.4), unless the planted partition for dif-
ferent layers is very similar (i.e., unless p ≥ 0.99). Fur-
thermore, the methods that are based on multilayer mod-
ularity outperform multilayer InfoMap in our experi-
ments for networks with weak community structure and
similar layers (i.e., when both µ and p are close to 1). In-
foMap does not identify meaningful community struc-
ture in networks with µ > 0.7 in any of our numerical ex-
amples, whereas GenLouvain and GenLouvainRand
identify meaningful structure even for networks with very
weak community structure (e.g., with µ = 0.9) for suffi-
ciently large values of ω.
Our results for nonuniform multiplex networks in Fig. 7
are similar to those that we showed for the uniform
multiplex networks in Fig. 6.
In particular, we see
that GenLouvainRand can exploit dependencies be-
tween community structure in different layers (although
the improvement over single-layer modularity is less pro-
nounced than in the uniform examples), whereas Gen-
Louvain cannot. The main difference is that for p = 1,
GenLouvain
15
(a) p = 0.5
(b) p = 0.85
(c) p = 0.95
(d) p = 0.99
(e) p = 1
GenLouvainRand
(f) p = 0.5
(g) p = 0.85
(h) p = 0.95
(i) p = 0.99
(j) p = 1
InfoMap
(k) p = 0.5
(l) p = 0.85
(m) p = 0.95
(n) p = 0.99
(o) p = 1
Figure 6. Multiplex networks with uniform interlayer dependencies. Effect of interlayer coupling strength ω and relaxation
rate r on the ability of different community-detection algorithms to recover planted partitions as a function of the community-
mixing parameter µ in benchmark networks with uniform multiplex dependencies (see Fig. 2b). Each multilayer network has
1000 nodes and 15 layers, and each node is present in all layers. All NMI values are means over 10 runs of the algorithms and
100 instantiations of the benchmarks. (See the introduction of Section VI.) Each curve corresponds to the mean NMI values
that we obtain for a given value of µ, and the shaded area around a curve corresponds to the minimum and maximum NMI
values that we obtain with the 10 sample partitions for a given value of p.
the aggregate partitions from GenLouvain and Gen-
LouvainRand do not recover the planted partition for
sufficiently large values of ω, because the induced parti-
tions are not identical across layers in the planted parti-
tion. In principle, multilayer InfoMap should have an
advantage over methods based on maximizing multilayer
modularity in this case study, as the former's quality
function is designed to detect breaks in community struc-
ture, whereas multilayer modularity forces some persis-
tence of community labels between any pair of layers. In
practice, however, multilayer InfoMap correctly identi-
fies the planted community structure only when it is par-
ticularly strong (as we show in Appendix F, it correctly
identifies the different groups of layers for networks with
µ ≤ 0.3); and it is outperformed by single-layer InfoMap
for networks with µ ≥ 0.4.
We suspect that at least some of the shortcomings of
multilayer InfoMap in these experiments are due to the
use of a Louvain-like optimization heuristic, rather than
from flaws in InfoMap's quality function. As we have
seen from the results with the heuristics GenLouvain
and GenLouvainRand, seemingly minor adjustments
of an optimization heuristic can have large effects on the
quality of the results. We are thus hopeful that similar
adjustments can also improve the results for multilayer
InfoMap.
B. Temporal examples
Temporal networks arise in many different applications
[2, 3], such as the study of brain dynamics [17], financial-
asset correlations [16], and scientific citations [18]. When
representing a temporal network as a multilayer network,
one orders the layers in a causal way, such that structure
in a particular layer depends directly only on structure
in previous layers (and not in future layers). To gen-
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=110−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIi GenLouvain
(a) p = 0.85, pc = 0
(b) p = 0.95, pc = 0
(c) p = 1, pc = 0
GenLouvainRand
(d) p = 0.85, pc = 0
(e) p = 0.95, pc = 0
(f) p = 1, pc = 0
InfoMap
(g) p = 0.85, pc = 0
(h) p = 0.95, pc = 0
(i) p = 1, pc = 0
Figure 7. Multiplex networks with nonuniform interlayer de-
pendencies. Effect of interlayer coupling strength ω and re-
laxation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a multiplex benchmark
with nonuniform interlayer dependencies. Each multilayer
network has 1000 nodes and 15 layers, and each node is
present in all layers. The layer-coupled interlayer-dependency
matrix is a block-diagonal matrix with diagonal blocks of
size 5 × 5. For each diagonal block, we set the value in the
interlayer-dependency matrix to a value p (so each diagonal
block has the same structure as the matrix in Fig. 2b); for
each off-diagonal block, we set the value to pc = 0, thereby
incorporating an abrupt change in community structure. All
NMI values are means over 10 runs of the algorithms and 100
instantiations of the benchmarks. Each curve corresponds to
the mean NMI values that we obtain for a given value of µ,
and the shaded area around a curve corresponds to the min-
imum and maximum NMI values that we obtain with the 10
sample partitions for a given value of p.
erate multilayer networks that have such structure, we
use our sampling process for fully-ordered multilayer net-
works and the interlayer-dependency tensor of Fig. 2a.
We consider two stylized examples of temporal net-
works. In Fig. 8, we show results for temporal networks
that have uniform dependencies between consecutive lay-
ers (and hence tend to evolve gradually). To generate
these networks, we set pβ = p for all layers in Fig. 2a.
In Fig. 9, we show results for temporal networks with
16
change points. To generate these networks, we set pβ = p
for all layers except layers 25, 50, and 75, for which
pβ = pc = 0, resulting in abrupt changes in community
structure. Each multilayer network in these two exam-
ples has n = 150 physical nodes and l = 100 layers. Each
node is present in every layer, so there are a total of
15000 state nodes. We use a Dirichlet null distribution
to specify expected community sizes and set nset = 5,
θ = 1, and q = 1 (see Appendix A). For the tempo-
ral networks with change points in Fig. 9, we choose the
support of the null distribution so that communities after
a change point have new labels. We use the M-DCSBM
benchmark (see Appendix D) with ηk = 2, kmin = 3, and
kmax = 30 to generate intralayer edges.
Both multilayer-modularity-based algorithms
(i.e.,
GenLouvain and GenLouvainRand) can exploit in-
terlayer dependencies for these temporal benchmark net-
works and identify partitions with significantly larger
(cid:104)NMI(cid:105) than those that we obtain with single-layer mod-
ularity (i.e., with ω = 0). Typically, the peak of (cid:104)NMI(cid:105)
seems to occur when 1 (cid:47) ω (cid:47) 4. When p < 1, one ex-
pects (cid:104)NMI(cid:105) to decrease for sufficiently large values of
ω, as increasing ω further favors "persistence" [16] in
the output partition that is not present in the multilayer
planted partition.
For multilayer InfoMap, the results are less promis-
ing. For most parameter choices, the best result for mul-
tilayer InfoMap is at best similar and often worse than
the result for single-layer InfoMap. (The (cid:104)NMI(cid:105) value
for single-layer InfoMap is the data point that we la-
bel with "s" on the horizontal axis.) An exception are
the results for p = 1 and uniform temporal dependencies
(see Fig. 8o). In this example (where induced partitions
are the same across layers), increasing the value of the
relaxation rate r enhances the recovery for all sampled
planted partitions when µ ≤ 0.4 and for a subset of sam-
pled planted partitions when µ = 0.5 and µ = 0.6. The
results for multilayer InfoMap on temporal benchmarks
with uniform interlayer dependencies with p = 0.99 (see
Fig. 8n for µ = 0.3 and µ = 0.4) and p = 1 (see Fig. 8o
for µ = 0.5 and µ = 0.6) are the only instances where we
observe large differences in results for different partitions
that are generated by our model with the same parameter
values.
Comparing results for GenLouvain (see Figs. 8a -- e
and 9a -- c) and GenLouvainRand with reiteration and
post-processing (see Figs. 8f -- j and 9d -- f), we see that
the difference in results between the two optimization
heuristics for multilayer modularity is even more pro-
nounced in these temporal examples than in the mul-
tiplex examples in Section VI A. In particular, GenLou-
vain exhibits an abrupt change in behavior near ω = 1;
this is related to the transition behavior that was de-
scribed in [16]. As we explained in Section VI A (where
we also observed this phenomenon), this transition occurs
for the following reason: for values of ω that are above
a certain threshold [16], only interlayer merges occur in
the first phase of the Louvain heuristic. This abrupt
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=110−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIi10−210000.20.40.60.81ωhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIi GenLouvain
17
(a) p = 0.5
(b) p = 0.85
(c) p = 0.95
(d) p = 0.99
(e) p = 1
GenLouvainRand
(f) p = 0.5
(g) p = 0.85
(h) p = 0.95
(i) p = 0.99
(j) p = 1
InfoMap
(k) p = 0.5
(l) p = 0.85
(m) p = 0.95
(n) p = 0.99
(o) p = 1
Figure 8. Temporal networks with uniform interlayer dependencies. Effect of interlayer coupling strength ω and relaxation
rate r on the ability of different community-detection algorithms to recover planted partitions as a function of the community-
mixing parameter µ in a temporal benchmark with uniform interlayer dependencies (i.e., pβ = p ∈ [0, 1] for all β ∈ {2, . . . , l}
in Fig. 2a). Each multilayer network has 150 nodes and 100 layers, and each node is present in all layers. All NMI values are
means over 10 runs of the algorithms and 100 instantiations of the benchmark. (See the introduction of Section VI.) Each curve
corresponds to the mean NMI values that we obtain for a given value of µ, and the shaded area around a curve corresponds to
the minimum and maximum NMI values that we obtain with the 10 sample partitions for a given value of p.
transition no longer occurs with the additional random-
ization in GenLouvainRand, which exhibits smooth be-
havior as a function of ω. However, unlike in our multi-
plex experiments, we observe benefits from the behavior
of GenLouvain in some situations, especially for net-
works with weak community structure (specifically, for
µ ≥ 0.8). This phenomenon first becomes noticeable for
networks with p = 0.95, and it becomes more pronounced
for progressively larger values of p. Compare Figs. 8c -- e
with Figs. 8h -- j; and compare Figs. 9b and c with Figs. 9e
and f.
Our results for temporal benchmark networks with
change points (see Fig. 9) are mostly similar to those
for temporal benchmark networks with uniform depen-
dencies when considering (cid:104)NMI(cid:105) to compare planted and
recovered partitions. Only the results for p = 1 are
noticeably different.
(Compare Figs. 9c, f, and i with
Figs. 8e, j, and o.) Because of the change points, the
induced planted partitions for each layer are not identi-
cal, even when p = 1. Consequently, unlike in Fig. 8,
the results for p = 1 in Fig. 9 are qualitatively similar
to those for smaller values of p. In principle, one might
expect multilayer InfoMap to have an advantage over
multilayer-modularity-based methods in this experiment,
as the former's quality function is designed to detect
abrupt changes in community structure, whereas max-
imizing multilayer modularity always favors some persis-
tence of community labels from one layer to the next.
However, this theoretical advantage is not borne out in
our experiment. When comparing the planted and recov-
ered partitions using the NMI between multilayer parti-
tions, we see that multilayer InfoMap can correctly re-
cover the change points only for p = 1 and µ = {0, 0.1}
(see Appendix F).
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=1024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIi 18
(a) p = 0.85, a = 0.1 (b) p = 0.85, a = 0.5 (c) p = 0.85, a = 0.9
(d) p = 0.95, a = 0.1 (e) p = 0.95, a = 0.5 (f) p = 0.95, a = 0.9
GenLouvain
(a) p = 0.85, pc = 0
(b) p = 0.95, pc = 0
(c) p = 1, pc = 0
GenLouvainRand
(d) p = 0.85, pc = 0
(e) p = 0.95, pc = 0
(f) p = 1, pc = 0
InfoMap
(g) p = 0.85, pc = 0
(h) p = 0.95, pc = 0
(i) p = 1, pc = 0
Figure 9. Temporal networks with nonuniform interlayer de-
pendencies. Effect of interlayer coupling strength ω and re-
laxation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a temporal benchmark
with nonuniform interlayer dependencies. Each multilayer
network has 150 nodes and 100 layers, and each node is
present in all layers. Every 25th layer (i.e., for layers 25,
50, and 75), we set the value of pβ in Fig. 2a to pc = 0,
thereby introducing an abrupt change in community struc-
ture; and we set all other values of pβ in Fig. 2a to p. All
NMI values are means over 10 runs of the algorithms and 100
instantiations of the benchmark. Each curve corresponds to
the mean NMI values that we obtain for a given value of µ,
and the shaded area around a curve corresponds to the min-
imum and maximum NMI values that we obtain with the 10
sample partitions for a given value of p.
(g) p = 0.99, a = 0.1 (h) p = 0.99, a = 0.5 (i) p = 0.99, a = 0.9
Figure 10. Multi-aspect networks with uniform interlayer de-
pendencies. Effect of the strength of temporal interlayer cou-
pling ωt and multiplex interlayer coupling ωm on the ability of
maximization of multilayer modularity to recover the planted
partition in two-aspect multilayer networks that we generate
using our framework. The first aspect is temporal (and thus
ordered) and the second aspect is multiplex (and thus un-
ordered). The heat maps show the mean ((cid:104)NMI(cid:105), 0
1)
over all layers of the NMI between induced planted parti-
tions and recovered community structure. A black cross indi-
cates the maximum value of (cid:104)NMI(cid:105). The dependence between
structure in different layers increases from top to bottom, and
the importance of the temporal aspect versus the multiplex
aspect increases from left to right. We average our results for
each pair of parameters over 10 runs of GenLouvainRand
with reiteration. The generated networks have n = 100 nodes
and l = 100 layers with LT = LM = 10, so there are a total
of 10000 state nodes. We fix the community-mixing parame-
ter of the M-DCSBM benchmark to be µ = 0.5.
C. Multi-aspect examples
lows:
In this section, we illustrate the ability of our frame-
work to generate multilayer networks with more than
one aspect. A common situation that gives rise to mul-
tilayer networks with two aspects is a multiplex net-
work that changes over time (e.g., citations patterns that
change over time [118]). In our framework, such a situ-
ation corresponds to the interlayer-dependency tensor in
Fig. 3. For the illustrative examples in this section, we
use uniform multiplex and temporal dependencies; and
we parametrize the interlayer dependency tensor as fol-
αt,αm = δ(αt, βt)(1 − δ(αm, βm))(1 − a)p/(lm − 1)
+ δ(αt + 1, βt)δ(αm, βm)ap ,
(cid:101)P βt,βm
where a controls the balance between multiplex and tem-
poral dependencies and p controls the overall strength of
the interlayer dependencies.
To recover communities, we optimize a multi-aspect
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=1024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIi024681000.20.40.60.81ωhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIis00.20.40.60.8100.20.40.60.81rhNMIi0.7300.20.40.60.810246810ωmωt0.7600.20.40.60.810246810ωmωt0.8300.20.40.60.810246810ωmωt0.7800.20.40.60.810246810ωmωt0.7900.20.40.60.810246810ωmωt0.8800.20.40.60.810246810ωmωt0.8700.20.40.60.810246810ωmωt0.8100.20.40.60.810246810ωmωt0.8600.20.40.60.810246810ωmωt generalization,
Q(S) = (cid:88)(i,α)∈VM
(j,β)∈VM
(cid:20)(cid:16)Aj,β
i,α −
+ C β
i,αkβ
kα
2mβ
j,β
α (cid:17)δ(α, β)
α δ(i, j)(cid:21)δ(Si,α, Sj,β) ,
(8)
of the multilayer modularity function that was intro-
duced in [33].
We specify the interlayer-coupling tensor C (a multi-
aspect generalization of the "interslice coupling" of [33]),
such that it reflects the planted interlayer dependencies,
by writing
C βt,βm
αt,αm = δ(αt, βt)(1 − δ(αm, βm))ωm
+ δ(αt + 1, βt)δ(αm, βm)ωt ,
where ωt denotes the coupling parameter for the tempo-
ral dependencies and ωm denotes the coupling parameter
for the multiplex dependencies.
In Fig. 10, we illustrate the effects of the two inter-
layer coupling parameters, ωt and ωm, on the extent to
which multilayer modularity maximization can recover
a planted partition. We use GenLouvainRand with
reiteration but without post-processing to find a local
optimum of Eq. (8). For all of the values of p and a
that we consider in Fig. 10, we see some improvement
in the performance of multilayer modularity maximiza-
tion with nonzero interlayer coupling compared with the
case where both ωt = 0 and ωm = 0. As the latter
case corresponds to independently maximizing modular-
ity on each layer of a network, this demonstrates that
multilayer modularity is able to exploit the interlayer de-
pendencies to better recover the planted partition. As
expected, increasing a (which gives more weight to tem-
poral dependencies) leads to a shift of the region of the
parameter space with good planted-partition recovery to
larger values of ωt and smaller values of ωm. For pro-
gressively larger p, we observe a small overall increase in
the value of (cid:104)NMI(cid:105), but its dependence on ωt and ωm
remains similar.
VII. CONCLUSIONS
We introduced a unifying and flexible framework for
the construction of generative models for mesoscale struc-
tures in multilayer networks. The three most important
features of our framework are the following: (1) it in-
cludes an explicitly parametrizable tensor P that con-
trols interlayer-dependency structure; (2) it can generate
an extremely general, diverse set of multilayer networks
(including, e.g., temporal, multiplex, "multilevel" [119],
and multi-aspect multilayer networks with uniform or
nonuniform dependencies between state nodes); and (3)
it is modular, as the process of generating a partition is
separate from the process of generating edges, enabling
a user to first generate a partition and then use any
19
planted-partition network model. Along with our pa-
per, we provide publicly available code [14] that users
can modify to readily incorporate different types of null
distributions (see Section IV A), interlayer-dependency
structures (see Section IV B), and planted-partition net-
work models (see Section V).
The ability to explicitly specify interlayer-dependency
structure makes it possible for a user to control which
layers depend directly on each other (by deciding which
entries in the interlayer-dependency tensor are nonzero)
and the strengths of such dependencies (by varying the
magnitude of entries in the interlayer-dependency ten-
sor). One can thereby generate multilayer networks with
either a single aspect or multiple aspects (e.g., tempo-
ral and/or multiplex networks) and vary dependencies
between layers from the extreme case in which induced
partitions in a planted multilayer partition are the same
across layers to the opposite extreme, in which induced
partitions in a planted multilayer partition are gener-
ated independently for each layer from a null distribution
for that layer. To the best of our knowledge, this level
of generality is absent from existing generative models
for mesoscale structures in multilayer networks, as those
models tend to only consider networks with a single as-
pect (e.g., temporal or multiplex) or limited interlayer-
dependency structures (e.g., with the same induced par-
titions in a planted multilayer partition across all layers).
We illustrated several examples of generative models
that one can construct from our framework, with a focus
on a few special cases of interest, rather than on trying
to discuss as many situations as possible. We focused
on community structure in our numerical experiments,
because it is a commonly studied mesoscale structure,
but one can also use our framework to construct genera-
tive models of intra-layer mesoscale structures other than
community structure (e.g., core -- periphery structure, bi-
partite structure, and so on) by taking advantage of the
flexibility of our framework's ability to use any single-
layer planted-partition network model. For our single-
aspect examples, we assumed that interlayer dependen-
cies exist either between all contiguous layers (a special
case of temporal networks), between all layers (a special
case of multiplex networks), or between all contiguous
groups of layers (another special case of multiplex net-
works). For both our temporal and multiplex examples,
we considered both uniform and nonuniform interlayer
dependencies. We also combined some of these scenar-
ios in an example with two aspects, a temporal aspect
and a multiplex aspect. However, our framework's flexi-
bility allows us to construct generative models of multi-
layer networks with more realistic features. For example,
for temporal networks, one can introduce dependencies
between a layer and all layers that follow it (such that
Fig. 2a is an upper triangular matrix with nonzero en-
tries above the diagonal) to incorporate memory effects
[120]. One can also consider interlayer dependencies that
are not layer-coupled. For example, dependencies can be
diagonal but nonuniform, or they can be nonuniform and
exist only between sets of related nodes.
In Section VI, we consider the commonly studied case
of a multilayer network with only intralayer edges and
connectivity patterns in the different layers that depend
on each other. We use a slight variant of the DCSBM
benchmark from [43] to generate edges for each layer (see
Algorithm 3). However, other types of multilayer net-
works are also important [4], and one can readily com-
bine our approach for generating multilayer partitions
with different network generative models that capture
various important features. For example, one can use
an SBM to generate interlayer edges (e.g., using our M-
DCSBM framework, which we discuss in Appendix D), or
one can replace the degree-corrected SBM in Section V
with any other planted-partition network model or other
interesting models (e.g., other variants of SBMs [42, 49]
and models for networks whose structure is affected by
space (and perhaps spatially embedded) [12] or arbitrary
latent features [64]).
In all of these examples, depen-
dencies between connectivity patterns in different layers
arise only from a planted multilayer partition. It is also
possible to modify our network generation process (see
Section V) to incorporate additional dependencies be-
tween layers beyond those that are induced by a planted
mesoscale structure, such as by introducing dependencies
between a node's degree in different layers [102, 121] or
burstiness [104] in inter-event-time distributions of edges.
Our work has the potential for many useful and inter-
esting extensions, and we highlight three of these. First,
although we have given some illustrative numerical exam-
ples in Section VI, the area of benchmarking community-
detection methods in multilayer networks is far from fully
developed. Generative models are useful tools for un-
derstanding the behavior of community-detection meth-
ods in detail and thus for suggesting ways of improving
heuristic algorithms without losing scalability. One can
use our framework to construct benchmark models that
provide a test bed for gaining insight into the advan-
tages and shortcomings of community-detection meth-
ods and algorithms (and, more generally, of meso-set-
detection methods and algorithms). Importantly, we ex-
pect these benchmark models to be very informative for
detecting potential artifacts of algorithms that can some-
times be masked in real-world applications. Second, a
well-understood generative model can be a powerful tool
for statistical inference (i.e., inferring the structure of a
multilayer network rather than generating a multilayer
network with a planted structure) [24]. For temporal
networks, closed forms for the joint distribution of meso-
set assignments have been derived for models that are
special cases of our framework [10, 13]. These results
may be useful for statistical inference. Additionally, it
seems likely that it is possible to adapt Bayesian inference
techniques (including Gibbs sampling [122 -- 124] or varia-
tional methods [125]) that have been developed for SBMs
both for inferring a multilayer partition and for inferring
an interlayer-dependency tensor. A key advantage of the
generality of the framework that we have developed in the
20
present paper is that it may be possible to frame many
model-selection questions in terms of posterior estimation
of the interlayer-dependency tensor. Finally, it is impor-
tant to model interdependent data streams and not just
fixed data sets. For example, for any fully-ordered multi-
layer network, our generative model respects the causal-
ity of layers (e.g., temporal causality), and one can thus
update a multilayer network with a new layer without
the need to update any previous layers. It is critical to
develop generative models that are readily adaptable to
such situations, and our work in this paper is a step in
this direction.
Appendix A: Example null distributions
In this appendix, we discuss parameter choices for a
categorical null distribution and give concrete examples
that can be useful for modeling mesoscale structure in
temporal or multiplex networks.
In Section IV A, we described the general form of a
categorical null distribution:
Pα
0 [s] =(cid:40)pα
s ,
0 ,
s ∈ {1, . . . , nset} ,
otherwise ,
(A1)
where pα
is the probability for a state node in layer
s
α to be assigned to a meso-set s in the absence of in-
terlayer dependencies, nset is the number of meso-sets
in the multilayer partition, and (cid:80)nset
s = 1 for each
α ∈ L. The support Gα of the null distribution Pα
0 is
Gα = {s : Pα
0 [s] (cid:54)= 0}. We say that a label s is active in
a layer α if it is in the support of the null distribution
0 (i.e., if Pα
Pα
0 [s] (cid:54)= 0); and we say that a label is inactive
in layer α if it is in the complement of the support of Pα
(i.e., if Pα
s=1 pα
A natural choice is to sample pα from a Dirichlet dis-
tribution, which is the conjugate prior for the categori-
cal distribution [92, 93]. The Dirichlet distribution over
q variables has q parameters θ1, . . . , θq. Its probability
density function is
0 [s] = 0).
0
p(x1, . . . , xq) =
i=1 θi)
i=1 Γ(θi)
Γ ((cid:80)q
(cid:81)q
xθi−1
i
,
q(cid:89)i=1
where xi ∈ (0, 1) and θi > 0 for each i ∈ {1, . . . , q}.
The case in which all θi are equal is called a "symmet-
ric Dirichlet distribution", which we parametrize by the
common value θ (the so-called "concentration parame-
ter") of the parameters and the number q of variables.
The concentration parameter θ determines the types
of discrete probability distributions that one is likely to
obtain from the symmetric Dirichlet distribution. For
θ = 1, the symmetric Dirichlet distribution is the contin-
uous uniform distribution over the space of all discrete
probability distributions with nset states. As θ → ∞,
the Dirichlet distribution becomes increasingly concen-
trated near the discrete uniform distribution, such that
all entries in pα are approximately equal. As θ → 0,
it becomes increasingly concentrated away from the uni-
form distribution, such that pα tends to have 1 (or a few)
large entries and all other entries are close to 0. Conse-
quently, to have very heterogeneous meso-set sizes, one
would choose θ ≈ 1. To have all meso-sets be of simi-
lar sizes, one would choose a large value of θ. To have a
few large meso-sets and many small meso-sets, one would
choose θ to be sufficiently smaller than 1. The value
of nset also affects the amount of meso-set label overlap
across layers in the absence of interlayer dependencies.
For example, if pα is the same for all layers, then larger
values of nset incentivize less label overlap across layers
(because there are more possible labels for each layer)
and smaller values of nset incentivize more label overlap
across layers (because there are fewer possible labels for
each layer).
In some situations -- e.g., when modeling the birth and
death of communities in temporal networks or the ap-
pearance and disappearance of communities in multiplex
networks -- it is desirable to have meso-set labels that
have a nonzero probability in (A1) only in some layers.
For these situations, we suggest sampling the support of
the distributions before sampling the probabilities pα.
Given the supports for each layer, one then samples the
corresponding probabilities from a symmetric Dirichlet
distribution (or any other distribution over categorical
distributions). That is,
21
One can also allow labels to appear and disappear
when examining multiplex partitions. For example, given
a value for nset, one can generate the support for each
layer by allowing each label s ∈ {1, . . . , nset} to be
present with some probability q and absent with com-
plementary probability 1 − q. This yields a sets of active
and inactive meso-set labels for each layer. One can then
sample the nonzero probabilities in pα that correspond
to active labels from a Dirichlet distribution and set pα
s to
0 for each inactive label s. Because multiplex partitions
are unordered, there is no notion of one layer occurring
after another one, so we do not need to ensure that an
inactive label in a given layer is also inactive in "subse-
quent" layers.
Appendix B: Partition sampling process
In this appendix, we provide a detailed description of
the way in which we sample multilayer partitions, in-
cluding a discussion of the convergence properties of our
sampling process. As we mentioned in Section IV B, we
assume that the multilayer partitions are generated by
a copying process on the meso-set assignment of state
nodes. Recall that the conditional probability distribu-
tion for the updated meso-set assignment of a state node,
given the current state of the copying process (Eq. (4) in
Section IV B), is
α = 0 ,
(A2)
pα
Gα ∼ Dir(θ,Gα) ,
pα
G
= {s ∈ {1, . . . , nset} : Pα
α
where G
0 [s] = 0} is the com-
plement of the support, with nset = maxα∈L max(Gα).
A simple example for a birth/death process for meso-
sets is the following. First, fix a number of meso-sets
and a support (i.e., active community labels) for the first
layer. One then sequentially initializes the supports for
the other layers by removing each meso-set in the sup-
port of the previous layer with probability rd ∈ [0, 1] and
adding a number, sampled from a Poisson distribution
with rate rb ∈ [0,∞), of new meso-sets (with new labels
that are not active in any previous layer). In temporal
networks, for example, this ensures that if a meso-set la-
bel is not in the support of a given layer, then the label
is also not in the support of any subsequent layers. For
this process, the expected number (cid:104)Gα(cid:105) of meso-sets in a
layer approaches rb/rd as one iterates through the layers.
Therefore, one should initialize the size of the support
for the first layer close to this value to avoid transients
in the number of meso-sets. For this process, the life-
time of meso-sets follows a geometric distribution. The
nature of the copying process that we use to introduce de-
pendencies between induced partitions in different layers
typically implies that meso-sets that have been removed
from the support of the null distribution do not lose all
of their members instantly, but instead shrink at a speed
that depends on the values of the copying probabilities
in the interlayer-dependency tensor.
P[Sj,β(τ + 1) = sS(τ )]
P j,β
i,α δ(Si,α(τ ), s)
=(cid:88)(i,α)∈VM
pj,β =(cid:88)(i,α)∈VM
+ (1 − pj,β) Pβ
P j,β
i,α ,
0 [Sj,β = s] ,
(B1)
where P0 is a given set of independent layer-specific null
distributions and P is a given interlayer-dependency ten-
sor. Further recall that a multilayer network can have
both ordered aspects and unordered aspects, where we
denote the set of ordered aspects by O and the set of
unordered aspects by U. Our goal is to sample multi-
layer partitions that are consistent with generation by
Eq. (B1) while respecting the update order from its or-
dered aspects (if present). Recall that two layers, α and
β, are order-equivalent (which we denote by α ∼O β) if
αa = βa for all a ∈ O. Based on this equivalence relation,
we obtain an ordered set of equivalence classes L/∼O by
inheriting the order from the ordered aspects (where, for
definiteness, we consider the lexicographic order [95] over
ordered aspects). We want to simultaneously sample the
meso-set assignment of state nodes within each class of
order-equivalent layers and we want to sequentially sam-
ple the meso-set assignment of state nodes in non-order-
equivalent layers, conditional on the fixed assignment of
state nodes in preceding layers (see Algorithm 1).
function SamplePartition(P , P0, O, nupd)
for (i, α) ∈ Vm do
Si,α ∼ Pα
0
end for
for [α] ∈ L/∼O do
(cid:46) Initialize partition tensor
using null distributions
(cid:46) Loop over ordered aspects
in lexicographic order
(cid:46) Sample uniformly from
order-equivalent layers
(cid:46) Loop over nodes in some
for τ ∈ {1, . . .[α] nupd} do
α ∼ U ([α])
for i ∈ V do
Si,α ∼ P[Si,αS]
order
(cid:46) Update (i, α) according to
Eq. (B1)
end for
end for
end for
return S
end function
ALG. 1. Pseudocode for generating multilayer partitions with
interlayer-dependency tensor P , null distributions Pα
0 , and or-
dered aspects O. The parameter nupd is the expected number
of updates for each state node.
Scan order and compatibility
To sample the induced partitions for each class of
order-equivalent layers, we use a technique that resem-
bles Gibbs sampling [81]. As in Gibbs sampling, we de-
fine a Markov chain on the space of multilayer partitions
by updating the meso-set assignment of state nodes by
sampling from Eq. (B1).
In Gibbs sampling, one as-
sumes that the conditional distributions that are used
to define this Markov chain are compatible, in the sense
that there exists a joint distribution that realizes these
conditional distributions [82 -- 85]. However, for a general
interlayer-dependency tensor, there is no guarantee that
the distributions that are defined by Eq. (B1) are com-
patible. Following [80], we use the term "pseudo-Gibbs
sampling" to refer to Gibbs sampling from a set of po-
tentially incompatible conditional distributions. We can
define different Markov chains by changing the way in
which we select which conditional distribution to apply at
each step (i.e., which state node to update at each step).
The order in which one cycles through the conditional
distributions is known as the scan order [87]. We use the
term sampling Markov chain for the Markov chain that is
defined by pseudo-Gibbs sampling with a specified scan
order. Common scan orders are cycling through con-
ditional distributions in a fixed order and sampling the
update distribution uniformly at random from the set
of conditional distributions. In pseudo-Gibbs sampling,
the choice of scan order deserves careful attention.
In
pseudo-Gibbs sampling, different scan orders correspond
to sampling from different distributions [87]. (By con-
trast, in Gibbs sampling, the choice of scan order influ-
ences only the speed of convergence.) In particular, when
cycling through conditional distributions in a fixed order,
22
conditional distributions that are applied later have an
outsized influence on the stationary distributions of the
sampling Markov chain [87]. Sampling the update distri-
bution uniformly at random mitigates this problem, at
the expense of introducing computational overhead.
We can improve on a purely random sampling strategy
by exploiting the structure of the interlayer-dependency
tensor. In particular, note that the conditional distribu-
tions for state nodes in the same layer are independent
(as we assume that the interlayer-dependency tensor has
no intralayer contributions) and hence compatible. When
updating a set of state nodes from a given layer, we can
update them in any order or even update them concur-
rently. To take advantage of this fact, we first sample a
layer uniformly at random from the current class of order-
equivalent layers and then update all state nodes in that
layer. The sequence of layer updates defines a Markov
chain on the space of multilayer partitions, and its con-
vergence properties determine the level of effectiveness of
our sampling algorithm.
Convergence guarantees
We now discuss two key properties -- aperiodicity and
ergodicity -- that determine the convergence behavior of
finite Markov chains.
First, note that sampling layers uniformly at random
guarantees that the Markov chain is aperiodic. To see
this, note that a sufficient condition for aperiodicity is
that, for all states, there is a nonzero probability to tran-
sition from the state to itself. This clearly holds for our
sampling chain, as it is possible for two consecutive up-
dates to update the same layer, and (because the proba-
bility distributions in Eq. (B1) remain unchanged at the
second update) there is a positive probability that the
second update does not change the partition.
Second, note that if pi,α < 1 for all state nodes
(i, α) ∈ VM such that α ∈ [α], where [α] is a class of
order-equivalent layers, then the sampling Markov chain
for that class is ergodic over a subspace of multilayer par-
titions that includes the support G[α] =(cid:81)α∈[α] Gα of the
null distributions. We have already shown that the sam-
pling Markov chain is aperiodic, so all that remains is to
verify that there exists a sample path from any arbitrary
initial partition S 0 to any arbitrary multilayer partition
S g ∈ G[α] in the support of the null distributions. Such
a sample path clearly exists, as
P(cid:2)Si,α(τ + 1) = Sg
i,α(cid:3)
i,α Si,α(τ ) = S0
(1 − pi,α)Pα
0 [Sg
i,α] > 0 .
≥
1
[α]
Therefore, one can achieve the desired transition using
one update for each state node.
Sample paths of a finite, aperiodic Markov chain con-
verge in distribution to a stationary distribution of the
Markov chain. However, different sample paths can con-
verge to different distributions, and the eventual station-
ary distribution can depend both on the initial condition
and on the transient behavior of the sample path. If the
Markov chain is ergodic, the stationary distribution is
unique, so it is independent of the initial condition and
transient behavior.
After sufficiently many updates (i.e., by specifying a
sufficiently large value for the expected number nupd of
updates for each state node in Algorithm 1), we approx-
imately sample the state of the Markov chains for each
class of order-equivalent layers from their stationary dis-
tributions. For fully-ordered multilayer networks (i.e., for
multilayer networks with U = ∅), setting nupd = 1 is suf-
ficient for convergence, because the equivalence classes
include only a single layer in this case and subsequent
updates are thus independent samples from the same dis-
tribution.
distributions when (cid:80)(i,α)∈VM
Sampling initial conditions in an 'unbiased' way can
be important for ensuring that one successfully explores
the space of possible partitions, as the updating process
is not necessarily ergodic over the support of the null
P j,β
i,α = 1 for some state
nodes. For example, if pi,α = 1 for all (i, α) ∈ VM , then
any partition in which all state nodes in each weakly con-
nected component of the interlayer-dependency network
have the same meso-set assignment is an absorbing state
of the Markov chain.
However, provided the updating process has con-
verged, any partition that one generates should still re-
flect the desired dependencies between induced partitions
for different layers. Therefore, to circumvent the problem
of non-unique stationary distributions when the updating
process is not ergodic, we reinitialize the updating pro-
cess by sampling from the null distribution for each par-
tition that we want to generate. Reinitializing the initial
partition from a fixed distribution for each sample always
defines a unique joint distribution, which is a mixture of
all possible stationary distributions when the updating
process is not ergodic. When the updating process is
ergodic, it is equivalent to sample partitions by reinitial-
izing or to sample multiple partitions from a single long
chain. Using a single long chain is usually more efficient
when one needs many samples and it is not problematic
if there is some dependence between samples [126].
In
our case, however, we usually need only a few samples
with the same parameters, and ensuring independence
tends to be more important than ensuring perfect con-
vergence (provided the generated partitions exhibit the
desired interlayer-dependency structure). Using multiple
chains thus has clear advantages even when the updat-
ing process is ergodic, and it is necessary to use multiple
chains when it is not ergodic.
Convergence tests
It is difficult to determine whether a Markov chain has
converged to a stationary distribution, mostly because it
23
(a) p = 0.5
(b) p = 0.85
(c) p = 0.95
(d) p = 0.99
(e) p = 1
Figure 11. Autocorrelation of the sampling Markov chain,
as measured by mNMI. On the horizontal axis, we show the
value of nupd that we use to generate the first partition in the
comparison. The lag is the number of additional updates per
state node that we use to generate the second partition in the
comparison. We average results over a sample of 100 indepen-
dent chains with different initial partitions. The shaded area
indicates one standard deviation above and below the mean.
Note that the large amount of noise for small values of the
lag is not a result of differences between chains in the sample;
instead, it is an inherent feature of all chains in the sample.
is difficult to distinguish the case of a slow-mixing chain
becoming stuck in a particular part of state space from
the case in which the chain has converged to a station-
ary distribution. There has been much work on trying to
define convergence criteria for Markov chains [127], but
none of the available approaches are entirely successful.
In practice, one usually runs a Markov chain (or chains)
for a predetermined number of steps. One checks manu-
ally for a few examples that the resulting chains exhibit
behavior that is consistent with convergence by examin-
ing autocorrelations between samples of the same chain
and cross-correlations between samples of independent
chains with the same initial state. When feasible, one
can also check whether parts of different, independent
lag=1lag=5lag=10lag=50lag=100lag=20005010015020000.050.10.15nupdmNMI05010015020000.10.2nupdmNMI05010015020000.10.20.30.4nupdmNMI05010015020000.20.40.6nupdmNMI05010015020000.51nupdmNMI 24
pendent of nupd (so the curves in Fig. 11 should become
flat). The results in Fig. 11 suggest that the number of
updates that we need for convergence increases moder-
ately with p, where nupd ≈ 50 is sufficient for convergence
even when p = 1.
In Fig. 12, we examine whether the
dependency pattern between induced partitions in differ-
ent layers -- we characterize such a pattern by the NMI
between induced partitions in different layers -- has con-
verged using the mean absolute distance
2
d(S,T ) =
l(l−1) (cid:88)α<β
NMI(Sα,Sβ) − NMI(T α,T β)
(B2)
between the NMI matrices between induced partitions
in a multilayer partition. (We showed examples of such
NMI matrices in Figs. 4 and 5.) As with mNMI, the
mean absolute distance d should become independent of
nupd once the Markov chain has converged to a station-
ary distribution. This yields estimates for the minimum
number of updates for convergence that are consistent
with those from the results in Fig. 11. It is always safe
to use a larger number of updates to sample partitions,
and the results from Figs. 11 and 12 suggest that our
choice of nupd = 200 is conservative.
Figs. 11 and 12 reveal additional important informa-
tion about the behavior of our sampling process. We can
estimate the mixing time of the sampling Markov chain
at stationarity from Fig. 11, because the mixing time
corresponds to the lag that is necessary for the mNMI
to converge to 0. The mixing time at stationarity in-
creases with p and becomes infinite as p → 1, because
the Markov chain converges to an absorbing state. From
Fig. 12, we observe that, for all values of p, the mean
absolute distance d between NMI matrices converges to
a value near 0 and becomes essentially independent of
the lag. This corroborates our assumption that different
partitions that we generate from our sampling process
have similar interlayer-dependency structure.
Appendix C: A generative model for temporal
networks
(a) p = 0.5
(b) p = 0.85
(c) p = 0.95
(d) p = 0.99
(e) p = 1
Figure 12. Convergence of the dependency pattern be-
tween induced partitions in different layers, as measured by
the mean absolute distance d between NMI matrices (see
Eq. (B2)). On the horizontal axis, we show the value of nupd
that we use to generate the first partition in the comparison.
The lag is the number of additional updates per state node
that we use to generate the second partition in the compar-
ison. We average results over a sample of 100 independent
chains with different initial partitions. The shaded area in-
dicates one standard deviation above and below the mean.
chains or different parts of the same chain are consistent
with sampling from the same distribution.
To estimate the number of updates that we need in
the numerical experiments of Section VI A, we examine
the autocorrelation of the Markov chain in two differ-
ent ways.
In Fig. 11, we consider the multilayer NMI
(mNMI) between sampled partitions at different steps of
the Markov chain. We are interested in the behavior of
the mNMI both as a function of nupd (the expected num-
ber of updates per state node that we use to generate the
first partition in the comparison) and the lag (the num-
ber of additional updates per state node that we use to
generate the second partition in the comparison). As the
Markov chain converges to a stationary distribution, the
value of the mNMI for a given lag should become inde-
In this appendix, we examine the generative model for
temporal networks that we considered in Section IV D.
In this model, we assume uniform dependencies be-
tween contiguous layers (i.e., pβ = p ∈ [0, 1] for all
β ∈ {2, . . . , l} in Fig. 2a). For this choice of interlayer-
dependency tensor, our generative model reduces to Al-
gorithm 2.
In this case (and, more generally, for any
fully-ordered multilayer network), convergence is auto-
matic, because we need only one iteration through the
layers.
A first important feature of the generative model in
Algorithm 2 is that it respects the arrow of time. In par-
ticular, meso-set assignments in a given layer depend only
on meso-set assignments in the previous layer (e.g., the
previous temporal snapshot) and on the null distributions
lag=1lag=5lag=10lag=50lag=100lag=200050100150200024·10−3nupdd0501001502000246·10−2nupdd05010015020000.10.2nupdd05010015020000.20.40.6nupdd05010015020000.51nupdd function TemporalPartition(p, P0)
S = 0
for i ∈ V do
Si,1 ∼ P1
0
end for
for α ∈ 2 . . . l do
for i ∈ V do
(cid:46) Initialize partition tensor of
appropriate size
(cid:46) Loop over nodes in some
order
(cid:46) Initialize induced partition
on first layer using null
distribution
(cid:46) Loop over layers in
sequential order
(cid:46) Loop over nodes in some
if rand() < p then (cid:46) With probability p
order
else
Si,α = Si,α−1
Si,α ∼ Pα
0
(cid:46) Copying update (Step C)
(cid:46) With probability 1 − p
(cid:46) Reallocation update (Step
R)
end if
end for
end for
return S
end function
ALG. 2. Pseudocode for generating partitions of a temporal
network with uniform interlayer dependencies p between suc-
cessive layers (i.e., pβ = p for all β ∈ {1, . . . , l} in Fig. 2a).
P0. That is, for all s ∈ {1, . . . , nl} and all α ∈ {2, . . . , l},
we have
P [ Si,α = s Sα−1] = pδ(Si,α−1, s)
+ (1 − p)Pα
0 [Si,α = s] ,
(C1)
where Sα is the n-node partition that is induced on layer
α by the multilayer partition S. The value of p deter-
mines the relative importance of the previous layer and
the null distribution. When p = 0, meso-set assignments
in a given layer depend only on the null distribution of
that layer (i.e., on the second term of the right-hand side
of Eq. (C1)). When p = 1, meso-set assignments in a
given layer are identical to those of the previous layer
(and, by recursion, to meso-set assignments in all previ-
ous layers).
Using Eq. (C1), we see that the marginal probability
that a given state node has a specific meso-set assignment
is
0 [Si,α = s] ,
= P1
P [Si,α = s] = pP [Si,α−1 = s]
+ (1 − p)Pα
0 [Si,1 = s] pα−1
α−1(cid:88)β=2
+ (1 − p)
+ (1 − p)Pα
Pβ
0 [Si,β = s] pα−β
0 [Si,α = s] , α > 1 .
Computing marginal probabilities can be useful for com-
puting expected meso-set sizes for a given choice of null
distributions.
25
We now highlight how the copying update (i.e., Step
C) and the reallocation update (i.e., Step R) in Algo-
rithm 2 govern the evolution of meso-set assignments be-
tween consecutive layers. Steps C and R deal with the
movement of nodes by first removing some nodes ("sub-
traction") and then reallocating them ("addition"). In
Step C, a meso-set assignment s in layer α can lose a
number of nodes that ranges from 0 to all of them. It can
keep all of its nodes in layer α + 1 (i.e., Ssα+1 = Ssα),
lose some of its nodes (i.e., Ssα+1 ⊂ Ssα), or disappear
entirely (i.e., Ssα+1 = ∅ and Ssα (cid:54)= ∅). The null distri-
bution in Step R is responsible for a meso-set assignment
s gaining new nodes (i.e., Ssα+1 (cid:54)⊂ Ssα) or for the ap-
pearance of a new meso-set label (i.e., Ssα+1 (cid:54)= ∅ and
Ssα = ∅). When defining the null distributions P0, it is
necessary to consider the interplay between Step C and
Step R.
To illustrate how the meso-set-assignment copying pro-
cess and the null distribution in Algorithm 2 can interact
with each other, we give the conditional probability that
a label disappears in layer α and the conditional probabil-
ity that a label appears in layer α. For all s ∈ {1, . . . , nl}
and all α ∈ {2, . . . , l}, the conditional probability that a
label disappears in layer α is
P [Ssα = ∅ Sα−1]
= [(1 − p) (1 − Pα
× [p + (1 − p) (1 − Pα
0 [Si,α = s])](cid:12)(cid:12)Ssα−1(cid:12)(cid:12)
0 [Si,α = s])]n−(cid:12)(cid:12)Ssα−1(cid:12)(cid:12) .
cess and simplifies to (1− p)(cid:12)(cid:12)Ssα−1(cid:12)(cid:12) when Pα
The expression (C2) depends only on our copying pro-
0 [Si,α = s] =
0 (i.e., when the probability of being assigned to label s
is 0 using the null distribution of layer α). Furthermore,
for progressively larger Pα
0 [Si,α = s], the probability that
a label disappears is progressively smaller.
For all s ∈ {1, . . . , nl} and all α ∈ {2, . . . , l}, the con-
ditional probability that a label appears in layer α is
(C2)
P(cid:2) Ssα (cid:54)= ∅ Ssα−1 = ∅(cid:3)
= 1 −(cid:20)p + (1 − p)(cid:18) (cid:88)Srα−1∈Sα−1
Pα
0 [Si,α = r](cid:19)(cid:21)n
.
(cid:80)Srα−1∈Sα−1
larger values of (cid:80)Srα−1∈Sα−1
(C3)
The expression (C3) gives the probability that at least
one node in layer α has the label s, given that
no node in layer α − 1 has the label s. When
Pα
0 [Si,α = r] = 0, the probability that a
label appears depends only on our meso-set-assignment
copying process and is given by 1 − pn. Furthermore,
Pα
0 [Si,α = r] reduce the
probability that a label appears in layer α.
All discussions thus far in this appendix hold for any
choice of P0. In the next two paragraphs, we give two ex-
amples to illustrate some features of the categorical null
distribution from Section IV A. In particular, we focus on
the effect of the support of a categorical null distribution
on a sampled multilayer partition. (See Section IV A for
definitions of some of the notation that we use below.)
The support Gα of a categorical null distribution Pα
0 is
Gα = {s : pα
s (cid:54)= 0}, where s ∈ {1, . . . , nset}. An impor-
tant property of the support for Fig. 2a is that overlap
between Gα and Gα+1 (i.e., Gα+1∩Gα (cid:54)= ∅) is a necessary
condition for a meso-set in layer α to gain new members
in layer α + 1.
s = npα
Let cα denote the vector of expected induced meso-set
sizes in layer α (i.e., cα
s ), and suppose that the
probabilities pα are the same in each layer (i.e., pα = p
for all α). The expected number of meso-set labels is
then the same for each layer, and the expected number
of nodes with meso-set label s is also the same in each
layer and is given by cα
s . This choice produces a temporal
network in which nodes change meso-set labels across
layers in a way that preserves both the expected number
of induced meso-sets in a layer and the expected sizes of
induced meso-sets in a layer.
Now suppose that we choose the pα values such that
their supports do not overlap (i.e., Gα ∩ Gβ = ∅ for all
α (cid:54)= β). At each Step C in Algorithm 2, an existing
meso-set label can then only lose members; and with
probability 1 − pn, at least one new label will appear
in each subsequent layer. For this case, one expects pcα
s
members of meso-set s in layer α to remain in meso-set
s in layer α + 1 and (1 − p)cα
s members of meso-set s in
layer α to be assigned to new meso-sets (because labels
are nonoverlapping) in layer α + 1. This choice thus pro-
duces multilayer partitions in which the expected num-
ber of new meso-set labels per layer is nonzero (unless
p = 1) and the expected size of a given induced meso-set
decreases in time.
Appendix D: M-DCSBM sampling procedure
In this appendix, we discuss a generalization to mul-
tilayer networks of the degree-corrected SBM (DCSBM)
from Ref. [43]. In other words, it is a multilayer DCSBM
(M-DCSBM). The parameters of a general, directed M-
DCSBM are a multilayer partition S (which determines
the assignment of state nodes to meso-sets), a block ten-
sor W (which determines the expected number of edges
between meso-sets in different layers), and a set σ of
state-node parameters (which determine the allocation
of edges to state nodes in meso-sets). The probability of
observing an edge (or the expected number of edges, if
we allow multi-edges) from state node (i, α) to state node
(j, β) with meso-set assignments r = Si,α and s = Sj,β
in a M-DCSBM is
P(cid:104)Aj,β
i,α = 1(cid:105) = σβ
i,α W s,β
r,α σj,β
α ,
(D1)
where W s,β
r,α is the expected number of edges from state
nodes in layer α and meso-set Sr to state nodes in layer
β and meso-set Ss, the quantity σβ
i,α is the probability
for an edge starting in meso-set Sr in layer α and ending
in layer β to be attached to state node (i, α) (note that
function DCSBM(S, σ, W )
A = 0
for α ∈ L do
for r ∈ 1 . . .S do
for s ∈ r . . .S do
26
(cid:46) Initialize adjacency tensor
of appropriate size
(cid:46) Loop over layers
(cid:46) Sample number of edges
from a Poisson distribution
if r = s then
m = Poisson(W r,α
r,α /2)
else
m = Poisson(W s,α
r,α )
end if
e = 0
while e < m do
(cid:46) Count sampled edges
i = Sample(Srα, σSrα,α)
(cid:46) Sample node i from induced
community r in layer α
with probability σi,α
j = Sample(Ssα, σSsα,α)
(cid:46) Sample node j from induced
if i (cid:54)= j & Aj,α
community s in layer α
with probability σj,α
i,α = 0 then
(cid:46) Reject self-edges or
multi-edges
j,α = 1
i,α = 1, Ai,α
Aj,α
e = e + 1
end if
end while
end for
end for
end for
return A
end function
ALG. 3. Sampling multilayer networks from a DCSBM with
community assignments S, node parameters σ, and block ten-
sor W .
i,α), and σj,β
the dependence on Sr is implicit in σβ
α is the
probability for an edge starting in layer α and ending in
meso-set Ss in layer β to be attached to state node (j, β).
(Note that the dependence on Ss is implicit in σj,β
α .) For
an undirected M-DCSBM, both the block tensor W and
the state-node parameters σ are symmetric. That is,
W r,α
s,β = W s,β
The above M-DCSBM can generate multilayer net-
works with arbitrary expected layer-specific in-degrees
and out-degrees for each state node.
(The DCSBM
of [43] can generate single-layer networks with arbitrary
expected degrees.) Given a multilayer network with ad-
jacency tensor A, the layer-α-specific in-degree of state
node (j, β) is
r,α and σi,α
β = σβ
i,α.
and the layer-β-specific out-degree of state node (i, α) is
Aj,β
i,α ,
kj,β
α =(cid:88)i∈V
i,α =(cid:88)j∈V
kβ
Aj,β
i,α .
The layer-α-specific in-degree and out-degree of state
node (i, α) are the "intralayer in-degree" and "intralayer
out-degree" of (i, α).
For an undirected multilayer
network, layer-β-specific in-degrees and out-degrees are
equal (i.e., ki,α
i,α). We refer to their common value
as the "layer-β-specific degree" of a state node. For an
ensemble of networks generated from an M-DCSBM, the
associated means are
β = kβ
(cid:104)kj,β
α (cid:105) = σj,β
α
S(cid:88)r=1
W s,β
r,α ,
s = Sj,β
(D2)
and
(cid:104)kβ
i,α(cid:105) = σβ
i,α
S(cid:88)s=1
W s,β
r,α ,
r = Si,α .
(D3)
As we mentioned in Section V, we consider undirected
multilayer networks with only intralayer edges for our
experiments in Section VI. The block tensor W thus does
not have any interlayer contributions (i.e., W s,β
r,α = 0 if
α (cid:54)= β). Furthermore, we can reduce the number of node
parameters that we need to specify the M-DCSBM to a
single parameter σi,α = σα
α for each state node
(i, α).
We
degrees
ei,α = (cid:104)kα
for the state
nodes from a truncated power law [128] with exponent
ηk, minimum cutoff kmin, and maximum cutoff kmax.
We then construct the block tensor W and state-node
parameters σ for the M-DCSBM from the sampled
expected degrees e and the meso-set assignments S. Let
sample
i,α(cid:105), where kα
expected
i,α =(cid:80)j∈VM
intralayer
Aj,α
i,α,
i,α = σi,α
the
κs,α = (cid:88)i∈Ssα
ei,α ,
Ss ∈ S
be the expected degree of meso-set s in layer α; and let
wα =
1
2(cid:88)i∈V
ei,α
be the expected number of edges in layer α. Conse-
quently,
σi,α =
ei,α
κs,α
,
s = Si,α
is the probability for an intralayer edge in layer α to
be attached to state node (i, α), given that the edge is
attached to a state node that is in layer α and is part of
the community Si,α.
The elements
W s,β
r,α = δ(α, β)(cid:18)(1 − µ)δ(r, s)κs,α + µ
κr,ακs,α
2wα (cid:19)
(D4)
of the block tensor give, for r (cid:54)= s, the expected num-
ber of edges between state nodes in meso-set s in layer
β and state nodes in meso-set r in layer α. For s = r
27
and β = α, the block-tensor element W s,β
r,α gives twice
the expected number of edges. One way to think of the
DCSBM benchmark is that we categorize each edge that
we want to sample as an intra-meso-set edge with prob-
ability 1 − µ or as a "random edge" (i.e., an edge that
can be either an intra-meso-set edge or an inter-meso-set
edge) with probability µ. To sample an edge, we sample
two state nodes (which we then join by an edge). We
call these two state nodes the "end points" of the edge.
We sample the two end points of an intra-meso-set edge
with a frequency that is proportional to the expected de-
gree of their associated state nodes, conditional on the
end points being in the same meso-set. By contrast, we
sample the two end points of a random edge with a fre-
quency proportional to the expected degree of their as-
sociated state nodes (without conditioning on meso-set
assignment). We assume that the total number of edges
in layer α is sampled from a Poisson distribution [129]
with mean wα. Although our procedure for sampling
edges describes a potential algorithm for sampling net-
works from the DCSBM benchmark, it usually is more
efficient to sample edges separately for each pair of meso-
sets.
In Algorithm 3, we show pseudocode for the specific
instance of the model in Eq. (D1) that we use to sam-
ple intralayer network edges (independently for each in-
duced partition) for a given multilayer partition in our
numerical experiments of Section VI. The only difference
between Algorithm 3 and the sampling algorithm of [43]
is that we use rejection sampling to avoid creating self-
edges and multi-edges. (In other words, if we sample an
edge that has already been sampled or that is a self-edge,
we do not include it in the multilayer network; instead,
we resample.) Rejection sampling is efficient provided all
blocks of the network are sufficiently sparse, such that
the probability of generating multi-edges is small. For
dense blocks of a network, we instead sample edges from
independent Bernoulli distributions with success proba-
bility Eq. (D1). This algorithm for sampling networks
from a DCSBM is very efficient; it scales linearly with
the number of edges in a network.
In the above discussion, we defined the parameters
in Eq. (D1) to generate intralayer edges in a multilayer
network by specifying layer-specific in-degrees and out-
degrees. That is, the layer-β-specific in-degree and out-
degree of state node (i, α) are 0 if β (cid:54)= α. We can also
extend the model in Eq. (D1) to generate interlayer edges.
For a directed multilayer network with interlayer edges,
we sample expected layer-β-specific in-degrees ei,α
and
β
out-degrees eβ
i,α for each state node (i, α) and layer β
from appropriate distributions. Given expected layer-
specific degrees e and a multilayer partition S, we then
construct the block tensor W and state-node parameters
σ for the M-DCSBM analogously to the special case that
we described above. Let
κs,α
β = (cid:88)i∈Ssα
ei,α
β ,
κβ
s,α = (cid:88)i∈Ssα
eβ
i,α , Ss ∈ S
28
be the expected layer-β-specific in-degree and out-degree
of meso-set s in layer α; and let
wβ
α =(cid:88)i∈V
eβ
i,α =(cid:88)i∈V
ei,β
α
GenLouvain
be the expected number of edges from layer α to layer β.
(Note the consistency constraint on expected in-degrees
and expected out-degrees.) It then follows that
σi,α
β =
ei,α
β
κs,α
β
,
σβ
i,α =
eβ
i,α
κβ
s,α
,
s = Si,α ,
(a) p = 0.85,
(b) p = 0.95
(c) p = 1
GenLouvainRand, no PP
and
W s,β
r,α = (1 − µ)δ(r, s)
κβ
s,α + κs,β
α
2
+ µ
κβ
r,ακs,β
α
wβ
α
.
For directed networks, the expected layer-specific in-
degrees and out-degrees generated by the above model
do not correspond exactly to the values from the input
parameters e, except when µ = 1.
Appendix E: The Louvain algorithm and its variants
In this appendix, we describe the Louvain algorithm
and variants of the Louvain algorithm that we use in the
numerical experiments in Section VI.
The Louvain algorithm [108] for maximizing (single-
layer or multilayer) modularity proceeds in two phases,
which are repeated iteratively. Starting from an initial
partition, one considers the state nodes one by one (in
some order) and places each state node in a set that re-
sults in the largest increase of modularity. (If there is
no move that improves modularity, a state node keeps
the same assignment.) One repeats this first phase of
the algorithm until reaching a local maximum.
In the
second phase of the Louvain algorithm, one obtains a
new modularity matrix by aggregating the sets of state
nodes that one obtains after the convergence of the first
phase. One then applies the algorithm's first phase to
the new modularity matrix and iterates both phases un-
til one converges to a local maximum. The two Louvain-
like algorithms that we use in Section VI differ in how
they select which moves to make. The first is GenLou-
vain, which always chooses the move that maximally in-
creases modularity; the second is GenLouvainRand,
which chooses modularity-increasing moves at random,
such that the probability of a particular move is pro-
portional to the resulting increase in the quality func-
tion. (The latter is a variant of the algorithm "Louvain-
Rand" in [16]; in that algorithm, one chooses modularity-
increasing moves uniformly at random.) We use "reiter-
ation" and "post-processing" to improve the output of
GenLouvainRand. Reiteration entails taking an out-
put partition of GenLouvainRand as an initial parti-
tion and reiterating the algorithm until the output par-
tition no longer changes. Post-processing entails using
(d) p = 0.85
(e) p = 0.95
(f) p = 1
GenLouvainRand
(g) p = 0.85
(h) p = 0.95
(i) p = 1
InfoMap
(j) p = 0.85
(k) p = 0.95
(l) p = 1
Figure 13. Effect of interlayer coupling strength ω and relax-
ation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a uniform multiplex bench-
mark (see Fig. 2b). Each multilayer network has 1000 nodes
and 15 layers, and each node is present in all layers.
the Hungarian algorithm [130] to optimize the "persis-
tence" [16] (i.e., the number of interlayer edges for which
both end points have the same community assignment) of
each output partition without changing the induced par-
titions on each layer. All Louvain variants and related
functionality are available at [107].
Appendix F: Multilayer NMI experiments
In this appendix, we calculate the multilayer NMI
(mNMI) between the multilayer output partition and
the planted multilayer partition for the numerical experi-
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=110−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMI GenLouvain
GenLouvain
29
(a) p = 0.85, pc = 0
(b) p = 0.95, pc = 0
(c) p = 1, pc = 0
(a) p = 0.85
(b) p = 0.95
(c) p = 1
GenLouvainRand, no PP
GenLouvainRand, no PP
(d) p = 0.85, pc = 0
(e) p = 0.95, pc = 0
(f) p = 1, pc = 0
(d) p = 0.85
(e) p = 0.95
(f) p = 1
GenLouvainRand
GenLouvainRand
(g) p = 0.85, pc = 0
(h) p = 0.95, pc = 0
(i) p = 1, pc = 0
(g) p = 0.85
(h) p = 0.95
(i) p = 1
InfoMap
InfoMap
(j) p = 0.85, pc = 0
(k) p = 0.95, pc = 0
(l) p = 1, pc = 0
(j) p = 0.85
(k) p = 0.95
(l) p = 1
Figure 14. Effect of interlayer coupling strength ω and relax-
ation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a multiplex benchmark
with nonuniform interlayer dependencies. Each multilayer
network has 1000 nodes and 15 layers, and each node is
present in all layers. The layer-coupled interlayer-dependency
matrix is a block-diagonal matrix with diagonal blocks of
size 5 × 5. For each diagonal block, we set the value in the
interlayer-dependency matrix to a value p (so each diagonal
block has the same structure as the matrix in Fig. 2b); for
each off-diagonal block, we set the value to pc = 0, thereby
incorporating an abrupt change in community structure. We
define the null distributions in these experiments so that there
is no overlap in community labels between different groups of
layers.
ments in Figs. 6 -- 9. One needs to be cautious when inter-
preting values of mNMI; they depend on how one deter-
mines the correspondences between communities in dif-
ferent layers, and these correspondences can be ambigu-
Figure 15. Effect of interlayer coupling strength ω and relax-
ation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a temporal benchmark
with uniform interlayer dependencies (i.e., pβ = p ∈ [0, 1]
for all β ∈ {2, . . . , l} in Fig. 2a). Each multilayer network
has 150 nodes and 100 layers, and each node is present in all
layers.
ous. For example, in a temporal network, if a community
in one layer splits into multiple communities in the next
layer, there are multiple plausible ways to define labels
across layers (e.g., all new communities get new labels
or the largest new community keeps the original label).
Moreover, different quality functions reward interlayer-
assignment labeling in different ways, and mNMI con-
flates this issues with how well a quality function recov-
ers structure within layers. In particular, a small value of
mNMI can hide an optimal value of (cid:104)NMI(cid:105) (as (cid:104)NMI(cid:105) = 1
is necessary, but not sufficient, to have mNMI = 1).
We include only a subset of the values of p and p
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=110−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMI10−210000.20.40.60.81ωmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMIµ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=1024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMI GenLouvain
(a) p = 0.85, pc = 0
(b) p = 0.95, pc = 0
(c) p = 1, pc = 0
GenLouvainRand, no PP
(d) p = 0.85, pc = 0
(e) p = 0.95, pc = 0
(f) p = 1, pc = 0
GenLouvainRand
(g) p = 0.85, pc = 0
(h) p = 0.95, pc = 0
(i) p = 1, pc = 0
InfoMap
(j) p = 0.85, pc = 0
(k) p = 0.95, pc = 0
(l) p = 1, pc = 0
Figure 16. Effect of interlayer coupling strength ω and relax-
ation rate r on the ability of different community-detection
algorithms to recover planted partitions as a function of the
community-mixing parameter µ in a temporal benchmark
with nonuniform interlayer dependencies. Each multilayer
network has 150 nodes and 100 layers, and each node is
present in all layers. Every 25th layer (i.e., for layers 25, 50,
and 75), we set the value of pβ in Fig. 2a to pc = 0 (thereby
incorporating an abrupt change in community structure); we
set all other values of pβ in Fig. 2a to p. We define the null
distributions in these experiments so that there is no overlap
in community labels between different groups of layers.
that we considered in Figs. 6 and 8, as these suffice for
the purpose of this appendix.
In the numerical exper-
iments of Section VI, we observed that post-processing
does not have a major effect on (cid:104)NMI(cid:105). This suggests
that any misassignment across layers that underempha-
sizes the "persistence" of a multilayer partition (which
post-processing tries to mitigate [16]) does not affect
an algorithm's ability to identify structure within lay-
30
ers. (Recall that post-processing relabels an assignment
without changing the induced partitions.) We include fig-
ures without post-processing in the experiments of this
appendix, as we expect the effect to be more noticeable
when examining NMI values between multilayer parti-
tions (which accounts for assignments both within and
across layers).
In Figs. 13 and 14, we use the same benchmark net-
works as in the multiplex examples in Figs. 6 and 7, re-
spectively. In Figs. 15 and 16, we use the same bench-
mark networks as in the temporal examples of Figs. 8
and 9, respectively. All mNMI values are means over
10 runs of the algorithms and 100 instantiations of the
benchmark. (See the introduction of Section VI for more
details on how we generate these instantiations.) Each
curve in each figure corresponds to the mean mNMI val-
ues that we obtain for a given value of the community-
mixing parameter µ, and the shaded area around a curve
corresponds to the minimum and maximum mNMI val-
ues that we obtain with the 10 sample partitions for a
given value of p or p.
For experiments with nonuniform interlayer dependen-
cies in Figs. 14 and 16, it seems that InfoMap is better
at detecting abrupt differences ("breaks", such as in the
form of change points for temporal examples) in commu-
nity structure than both GenLouvain and GenLou-
vainRand, especially for the multiplex case (where this
manifests for all examined values of p). However, multi-
layer InfoMap correctly identifies the planted commu-
nity structure only when it is particularly strong (specif-
ically, for µ ≤ 0.3). Recall from our experiments in
Figs. 7 and 9 that both GenLouvain and GenLouvain-
Rand are better than InfoMap at detecting an induced
partition within a layer (especially in cases where the
planted community structure is quite weak). The rea-
son that multilayer modularity maximization does not
perfectly recover breaks in community structure of the
planted multilayer partition is because multilayer modu-
larity maximization with uniform ordinal or categorical
coupling (see the introduction of Section VI) incentivizes
"persistence" of community labels between layers (and
nonzero persistence is achievable even when one inde-
pendently samples induced partitions on different lay-
ers) [16]. However, by design, our benchmark networks
have no persistence whenever there is a break in commu-
nity structure because of how we defined the support of
the null distributions. Therefore, we do not expect mul-
tilayer modularity maximization to recover the planted
labels between layers when pc = 0; this manifests as a de
facto maximum in the mNMI between the output parti-
tion and the planted partition.
ACKNOWLEDGEMENTS
MB acknowledges a CASE studentship award from
the EPSRC (BK/10/41). MB was also supported by
The Alan Turing Institute under the EPSRC grant
µ=0µ=0.1µ=0.2µ=0.3µ=0.4µ=0.5µ=0.6µ=0.7µ=0.8µ=0.9µ=1024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMI024681000.20.40.60.81ωmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMIs00.20.40.60.8100.20.40.60.81rmNMI EP/N510129/1.
LGSJ acknowledges a CASE stu-
dentship award from the EPSRC (BK/10/39). MB,
LGSJ, AA, and MAP were supported by FET-Proactive
project PLEXMATH (FP7-ICT-2011-8; grant #317614)
funded by the European Commission; and MB, LGSJ,
and MAP were also supported by the James S. McDon-
nell Foundation (#220020177). MAP was also supported
by the National Science Foundation (grant #1922952)
through the Algorithms for Threat Detection (ATD) pro-
gram. Computational resources that we used for this re-
search were supported in part by the National Science
Foundation under Grant No. CNS-0521433, in part by
Lilly Endowment, Inc. through its support for the In-
diana University Pervasive Technology Institute, and in
part by the Indiana METACyt Initiative. The Indiana
METACyt Initiative at Indiana University Bloomington
was also supported in part by Lilly Endowment, Inc. We
thank Sang Hoon Lee, Peter Mucha (and his research
group), Brooks Paige, and Roxana Pamfil for helpful
comments.
31
[1] M. E. J. Newman, Networks, 2nd ed. (Oxford University
Press, 2018).
[2] P. Holme and J. Saramaki, "Temporal networks," Phys.
Rep. 519, 97 -- 125 (2012).
in temporal multilayer networks, with an application to
correlation networks," Multiscale Model. Simul. 14, 1 --
41 (2016).
[17] R. F. Betzel and D. S. Bassett, "Multi-scale brain net-
[3] P. Holme, "Modern temporal network theory: A collo-
works," Neuroimage 160, 73 -- 83 (2017).
quium," Eur. Phys. J. B 88, 234 (2015).
[4] M. Kivela, A. Arenas, M. Barth´elemy, J. P. Gleeson,
Y. Moreno, and M. A. Porter, "Multilayer networks,"
J. Complex Netw. 2, 203 -- 271 (2014).
[5] S. Boccaletti, G. Bianconi, R. Criado, C. I. Del Genio,
J. G´omez-Gardenes, M. Romance, I. Sendina-Nadal,
Z. Wang, and M. Zanin, "The structure and dynamics
of multilayer networks," Phys. Rep. 544, 1 -- 122 (2014).
[6] A. Aleta and Y. Moreno, "Multilayer networks in a nut-
shell," Annu. Rev. Condens. Matter Phys. 10 (2019).
[7] G. Bianconi, Multilayer networks: structure and func-
tion (Oxford University Press, 2018).
[8] M. A. Porter, "WHAT IS ... A multilayer network," No-
tices Amer. Math. Soc. 65 (2018).
[9] M. De Domenico, A. Lancichinetti, A. Arenas,
and
M. Rosvall, "Identifying modular flows on multilayer
networks reveals highly overlapping organization in in-
terconnected systems," Phys. Rev. X 5, 011027 (2015).
[10] A. Ghasemian, P. Zhang, A. Clauset, C. Moore, and
L. Peel, "Detectability thresholds and optimal algo-
rithms for community structure in dynamic networks,"
Phys. Rev. X 6, 031005 (2016).
[11] L. Peel, D. B. Larremore, and A. Clauset, "The ground
truth about metadata and community detection in net-
works," Science Advances 3, e1602548 (2017).
[12] M. Sarzynska, E. A. Leicht, G. Chowell, and M. A.
Porter, "Null models for community detection in spa-
tially embedded, temporal networks," J. Complex Netw.
4, 363 -- 406 (2015).
[13] A. R. Pamfil, S. D. Howison, R. Lambiotte,
and
M. A. Porter, "Relating modularity maximization and
stochastic block models in multilayer networks," SIAM
Journal on Mathematics of Data Science 1, 667 -- 698
(2019).
[14] L. G. S. Jeub and M. Bazzi, "A generative model for
mesoscale structure in multilayer networks," https://
github.com/MultilayerGM (2016 -- 2019).
[15] M. De Domenico, A. Sol`e-Ribalta, E. Cozzo, M. Kivela,
Y. Moreno, M. A. Porter, S. G`omez, and A. Arenas,
"Mathematical
formulation of multilayer networks,"
Phys. Rev. X 3, 041022 (2013).
[16] M. Bazzi, M. A. Porter, S. Williams, M. McDonald,
D. J. Fenn, and S. D. Howison, "Community detection
[18] D. Hric, K. Kaski, and M. Kivela, "Stochastic block
model reveals maps of citation patterns and their evo-
lution in time," J. Informetr. 12, 757 -- 783 (2018).
[19] L. Cantini, E. Medico, S. Fortunato, and M. Caselle,
"Detection of gene communities in multi-networks re-
veals cancer drivers," Sci. Rep. 5, 17386 (2015).
[20] C. De Bacco, E. A. Power, D. B. Larremore,
and
C. Moore, "Community detection, link prediction, and
layer interdependence in multilayer networks," Phys.
Rev. E 95, 042317 (2017).
[21] N. Stanley, S. Shai, D. Taylor, and P. J. Mucha, "Clus-
tering network layers with the strata multilayer stochas-
tic block model," IEEE Trans. Network Sci. Eng. 3, 95 --
105 (2016).
[22] It is also possible to have ordinal layering in aspects
other than time.
[23] M. A. Porter, J.-P. Onnela, and P. J. Mucha, "Com-
munities in networks," Notices Amer. Math. Soc. 56,
1082 -- 1097, 1164 -- 1166 (2009).
[24] S. Fortunato and D. Hric, "Community detection in net-
works: A user guide," Phys. Rep. 659 (2016).
[25] G. Rossetti and R. Cazabet, "Community discovery in
dynamic networks: A survey," ACM Comput. Surv. 51,
35 (2018).
[26] H. Cherifi, G. Palla, B. K. Szymanski, and X. Lu, "On
community structure in complex networks: Challenges
and opportunities," arXiv:1908.04901 (2019).
[27] P. Csermely, A. London, L.-Y. Wu,
and B. Uzzi,
"Structure and dynamics of core -- periphery networks,"
J. Complex Netw. 1, 93 -- 123 (2013).
[28] P. Rombach, M. A. Porter, J. H. Fowler, and P. J.
Mucha, "Core -- periphery structure in networks (revis-
ited)," SIAM Rev. 59, 619 -- 646 (2017).
[29] R. A. Rossi and N. K. Ahmed, "Role discovery in net-
works," IEEE Trans. Knowl. Data Eng. 27, 1112 -- 1131
(2015).
[30] S. Fortunato, "Community detection in graphs," Phys.
Rep. 486, 75 -- 174 (2010).
[31] M. Vaiana and S. F. Muldoon, "Multilayer brain
, 1 -- 23
https://doi.org/10.1007/
networks," Journal of Nonlinear Science
(2018),
s00332-017-9436-8.
available
at
[32] P.J. Mucha and M. A. Porter, "Communities in multi-
slice voting networks," Chaos 20 (2010).
[33] P. J. Mucha, T. Richardson, K. Macon, M. A. Porter,
and J.-P. Onnela, "Community structure in time-
dependent, multiscale, and multiplex networks," Sci-
ence 328, 876 -- 878 (2010).
[34] E. Omodei and A. Arenas, "Untangling the role of di-
verse social dimensions in the diffusion of microfinance,"
Applied Network Science 95, 14 (2016).
[35] M. T. Schaub, J.-C. Delvenne, M. Rosvall,
and
R. Lambiotte, "The many facets of community detec-
tion in complex networks," Applied Network Science 2,
4 (2017).
[36] L. G. S. Jeub, P. Balachandran, M. A. Porter, P. J.
Mucha,
and M. W. Mahoney, "Think locally, act
locally: The detection of small, medium-sized, and
large communities in large networks," Phys. Rev. E 91,
012821 (2015).
[37] M. E. J. Newman, "Finding community structure in net-
works using the eigenvectors of matrices," Phys. Rev. E
74, 036104 (2006).
[38] G. Petri and P. Expert, "Temporal stability of network
partitions," Phys. Rev. E 90, 022813 (2014).
[39] R. Lambiotte, J.-C. Delvenne, and M. Barahona, "Ran-
dom walks, Markov processes and the multiscale mod-
ular organization of complex networks," IEEE Trans.
Network Sci. Eng. 1, 76 -- 90 (2015), (see also the precur-
sor paper at arXiv:0812.1770, 2008).
[40] J.-C. Delvenne, S. N. Yaliraki, and M. Barahona, "Sta-
bility of graph communities across time scales," Proc.
Nat. Acad. Sci. U.S.A. 107, 12755 -- 12760 (2010).
[41] M. Rosvall and C. T. Bergstrom, "Maps of random
walks on complex networks reveal community struc-
ture," Proc. Nat. Acad. Sci. U.S.A. 105, 1118 -- 1123
(2008).
[42] P. W. Holland, K. B. Laskey,
and S. Leinhardt,
"Stochastic blockmodels: First steps," Soc. Networks
5, 109 -- 137 (1983).
[43] B. Karrer and M. E. J. Newman, "Stochastic blockmod-
els and community structure in networks," Phys. Rev.
E 83, 016107 (2011).
[44] T. P. Peixoto, "Inferring the mesoscale structure of lay-
ered, edge-valued, and time-varying networks," Phys.
Rev. E 92, 042807 (2015).
[45] I. Kloumann, J. Ugander,
and J. Kleinberg, "Block
models and personalized PageRank," Proc. Nat. Acad.
Sci. U.S.A. 114, 33 -- 38 (2016).
[46] L. G. S. Jeub, M. W. Mahoney, P. J. Mucha, and M. A.
Porter, "A local perspective on community structure
in multilayer networks," Network Science 5, 144 -- 163
(2017).
[47] J. D. Wilson, J. Palowitch, S. Bhamidi, and A. B. No-
bel, "Community extraction in multilayer networks with
heterogeneous community structure," J. Mach. Learn.
Res. 18, 1 -- 49 (2017).
[48] U. Brandes, D. Delling, M. Gaertler, R. Goke, M. Hoe-
fer, Z. Nikoloski, and D. Wagner, "On modularity clus-
tering," IEEE Trans. Knowl. Data Eng. 20, 172 -- 188
(2008).
[49] A. Z. Jacobs and A. Clauset, "A unified view of genera-
tive models for networks: Models, methods, opportuni-
ties, and challenges," in NIPS Workshop on Networks:
From Graphs to Rich Data (2014) arXiv:1411.4070.
32
[50] A. Ghasemian, H. Hosseinmardi,
and A. Clauset,
"Evaluating overfit and underfit in models of network
community structure," IEEE Trans. Knowl. Data Eng.
(2019), available at 10.1109/TKDE.2019.2911585.
[51] S. E. Schaeffer, "Graph clustering," Comput. Sci. Rev.
1, 27 -- 64 (2007).
[52] A. Lancichinetti and S. Fortunato, "Benchmark graphs
for testing community detection algorithms on directed
and weighted graphs with overlapping communities,"
Phys. Rev. E 80, 1 (2009).
[53] A. Lancichinetti, S. Fortunato,
and F. Radicchi,
"Benchmark graphs for testing community detection al-
gorithms," Phys. Rev. E 78, 046110 (2008).
[54] C. Granell, R. K. Darst, A. Arenas, S. Fortunato,
and S. G´omez, "A benchmark model to assess commu-
nity structure in evolving networks," Phys. Rev. E 92,
012805 (2015).
[55] A. Condon and R. M. Karp, "Algorithms for graph
partitioning on the planted partition model," Random
Struct. Algor. 18, 116 -- 140 (2000).
[56] M. Girvan and M. E. J. Newman, "Community struc-
ture in social and biological networks," Proc. Nat. Acad.
Sci. U.S.A. 99, 7821 -- 7826 (2002).
[57] T. P. Peixoto, "Nonparametric weighted stochastic
block models," Phys. Rev. E 97, 012306 (2018).
[58] L. G. S. Jeub, O. Sporns, and S. Fortunato, "Multires-
olution consensus clustering in networks," Sci. Rep. 8,
3259 (2018).
[59] S. Paul and Y. Chen, "Null models and modularity
based community detection in multi-layer networks,"
arXiv:1608.00623 (2016).
[60] A. Decelle, F. Krzakala, C. Moore, and L. Zdeborova,
"Inference and phase transitions in the detection of
modules in sparse networks," Phys. Rev. Lett. 107,
065701 (2011).
[61] D. Taylor, R. S. Caceres, and P. J. Mucha, "Super-
resolution community detection for layer-aggregated
multilayer networks," Phys. Rev. X 7, 031056 (2017).
[62] G. K. Orman, V. Labatut, and H. Cherifi, "Towards
realistic artificial benchmark for community detection
algorithms evaluation," Int. J. Web Based Communities
9, 349 -- 370 (2013).
[63] C. Aicher, A. Z. Jacobs, and A. Clauset, "Learning la-
tent block structure in weighted networks," J. Complex
Netw. 3, 221 -- 248 (2014).
[64] M. E. J. Newman and T. P. Peixoto, "Generalized com-
munities in networks," Phys. Rev. Lett. 115, 088701
(2015).
[65] X. Zhang, C. Moore, and M. E. J. Newman, "Random
graph models for dynamic networks," Eur. Phys. J. B
90, 200 (2017).
[66] Y. Hulovatyy and T. Milenkovi´c, "SCOUT: Simultane-
ous time segmentation and community detection in dy-
namic networks," Sci. Rep. 6, 37557 (2016).
[67] M. Q. Pasta and F. Zaidi, "Network generation model
based on evolution dynamics to generate benchmark
graphs," arXiv:1606.01169 (2016).
[68] S. Paul and Y. Chen, "Consistent community detection
in multi-relational data through restricted multi-layer
stochastic blockmodel," Electron. J. Stat. 10, 3807 --
3870 (2016).
[69] T. Vall`es-Catal`a, F. A. Massucci, R. Guimer`a,
and
M. Sales-Pardo, "Multilayer stochastic block models re-
veal multilayer structure of complex networks," Phys.
Rev. X 6, 011036 (2016).
[70] P. Barbillon, S. Donnet, E. Lazega, and A. Bar-Hen,
"Stochastic block models for multiplex networks: an ap-
plication to networks of researchers," J. Roy. Stat. Soc.
A 180, 295 -- 314 (2016).
[71] Y. Huang, A. Panahi, H. Krim, and L. Dai, "Commu-
nity detection and improved detectability in multiplex
networks," arXiv:1909.10477 (2019).
[72] M. De Domenico, A. Sol`e-Ribalta, S. G`omez,
and
A. Arenas, "Navigability of interconnected networks un-
der random failures," Proc. Nat. Acad. Sci. U.S.A. 111,
8351 -- 8356 (2014).
[73] S. G´omez, A. D´ıaz-Guilera, J. G´omez-Gardenes,
C. J. P´erez Vicente, Y. Moreno, and A. Arenas, "Diffu-
sion dynamics on multiplex networks," Phys. Rev. Lett.
110, 028701 (2013).
[74] Flattening (i.e., 'matricizing') [4] a tensor T entails writ-
ing its entries in matrix form. Consequently, starting
from a tensor T with elements T β
α , one obtain a matrix
(cid:101)T with entries (cid:101)T α, β, where there are bijective mappings
between the indices.
[75] T. G. Kolda and B. W. Bader, "Tensor decompositions
and applications," SIAM Rev. 51, 455 -- 500 (2009).
[76] M. S. Bartlett, An Introduction to Stochastic Processes,
3rd ed. (Press Syndicate of the University of Cambridge,
Cambridge, UK, 1978).
[77] S. van Buuren, Flexible Imputation of Missing Data, 2nd
ed. (Chapman & Hall/CRC Interdisciplinary Statistics,
Boca Raton, FL, USA, 2018).
[78] J. E. Besag, "Spatial interaction and the statistical anal-
ysis of lattice systems," J. Roy. Stat. Soc. B 36, 192 -- 236
(1974).
[79] S. E. Fienberg and A. B. Slavkovic, "Preserving the
confidentiality of categorical statistical data bases when
releasing information for association rules," Data Min.
Knowl. Discov. 11, 155 -- 180 (2005).
[80] D. Heckerman, D. M. Chickering, C. Meek, R. Roundth-
waite, and C. Kadie, "Dependency networks for infer-
ence, collaborative filtering, and data visualization," J.
Mach. Learn. Res. 1, 49 -- 75 (2000).
[81] Alan E. Gelfand, "Gibbs sampling," J. Am. Stat. Assoc.
95, 1300 -- 1304 (2000).
[82] J. P. Hobert and G. Casella, "Functional compatibil-
ity, Markov chains, and Gibbs sampling with improper
posteriors," J. Comput. Graph. Stat. 7, 42 -- 60 (1998).
[83] B. C. Arnold and S. J. Press, "Compatible conditional
distributions," J. Am. Stat. Assoc. 84, 152 -- 156 (1989).
[84] H.-Y. Chen, "Compatibility of conditionally specified
models," Stat. Probab. Lett. 80, 670 -- 677 (2010).
[85] S.-H. Chen and E. H. lp, "Behavior of the Gibbs sampler
when conditional distributions are potentially incom-
patible," J. Stat. Comput. Simul. 85, 3266 -- 3275 (2015).
[86] K.-L. Kuo and Y. J. Wang, "A simple algorithm for
checking compatibility among discrete conditional dis-
tributions," Comput. Stat. Data Anal. 55, 2457 -- 2462
(2011).
[87] K.-L. Kuo and Y. J. Wang, "Pseudo-Gibbs sampler
for discrete conditional distributions," Ann. Inst. Stat.
Math. 71, 93 -- 105 (2017).
[88] D. S. Bassett, N. F. Wymbs, M. A. Porter, P. J. Mucha,
J. M. Carlson, and S. T. Grafton, "Dynamic reconfigu-
ration of human brain networks during learning," Proc.
Nat. Acad. Sci. U.S.A. 118, 7641 -- 7646 (2011).
33
[89] J. M. Buld´u and M. A. Porter, "Frequency-based brain
networks: From a multiplex framework to a full multi-
layer description," Network Neuro. 2, 418 -- 441 (2017).
[90] U. Aslak, M. Rosvall, and S. Lehmann, "Constrained
information flows in temporal networks reveal intermit-
tent communities," Phys. Rev. E 97, 062312 (2018).
[91] E. Valdano, C. Poletto,
and V. Colizza, "Infection
propagator approach to compute epidemic thresholds on
temporal networks: impact of immunity and of limited
temporal resolution," Eur. Phys. J. B 88, 1 -- 11 (2015).
[92] H. Raiffa and R. Schlaifer, Applied Statistical Decision
Theory, Wiley Classics Library (Wiley-Interscience,
New York City, NY, USA, 2000).
[93] A. Agarwal and H. Daum´e III, "A geometric view of
conjugate priors," Mach. Learn. 81, 99 -- 113 (2010).
[94] For a fully-ordered multilayer network (i.e., where
U = ∅), condition (5) is equivalent to the existence
of a matrix representation of the flattened interlayer-
dependency tensor that is upper triangular. (See Fig. 2a
for an example.) In the case of a partially-ordered mul-
tilayer network, the same equivalence holds if we "ag-
gregate" [15] the interlayer-dependency tensor over un-
ordered aspects.
[95] One defines lexicographic order for the Cartesian prod-
uct of ordered sets similarly to the way that words are
ordered in a dictionary. For two elements e, f ∈ E =
E1 × . . . × En, we define e ≺ f if e1 < f1 or there exists
i such that ei < fi and ei(cid:48) = fi(cid:48) for all i(cid:48) < i.
[96] The order in which we update state nodes in the same
layer has no effect on the sampling process, because
their conditional distributions are independent. One can
even update them in parallel.
[97] L. Peel and A. Clauset, "Detecting change points in the
large-scale structure of evolving networks," in Proceed-
ings of the 29th Conference on Artificial Intelligence
(AAAI Press, Palo Alto, California, USA, 2015) pp.
2914 -- 2920.
[98] Different choices for normalizing mutual information
yield different versions of NMI [131]. We use joint-
entropy normalization, which entails that
NMI(S,T ) =
I(S,T )
H(S,T )
,
where S and T denote two multilayer partitions, I de-
notes mutual information, and H denotes joint entropy.
This variant of the NMI has the desirable property that
1 − NMI(S,T ) = NVI(S,T ) =
VI(S,T )
H(S,T )
,
where NVI is a normalized version of variation of in-
formation (VI) [132, 133]. Variation of information is a
metric for comparing partitions, and NVI preserves its
metric properties [134].
[99] R. Gallotti and M. Barth´elemy, "Anatomy and effi-
ciency of urban multimodal mobility," Sci. Rep. 4, 6911
(2014).
[100] M. Zanin and F. Lillo, "Modelling the air transport with
complex networks: A short review," Eur. Phys. J. Spec.
Top. 215, 5 -- 21 (2013).
[101] S. Galhotra, A. Mazumdar, S. Pal, and B. Saha, "The
geometric block model," AAAI-18 Thirty-Second AAAI
Conference on Artificial Intelligence (2018).
[102] K.-M. Lee, J. Y. Kim, W.-K. Cho, K.-I. Goh, , and I.-
M. Kim, "Correlated multiplexity and connectivity of
multiplex random networks," New J. Phys. 14, 033027
(2012).
[103] A. R. Pamfil, S. D. Howison, and M. A. Porter, "Edge
correlations in multilayer networks," arXiv:1908.03875
(2019).
[104] R. Lambiotte, L. Tabourier,
and J.-C Delvenne,
"Burstiness and spreading on temporal networks," Eur.
Phys. J. B 86, 320 (2013).
[105] D. J. Daley and D. Vere-Jones, An Introduction to the
Theory of Point Processes, 2nd ed., Vol. 1 (Springer-
Verlag, Heidelberg, Germany, 2003).
[106] A. G. Hawkes, "Spectra of some self-exciting and mu-
tually exciting point processes," Biometrika 58, 83 -- 90
(1971).
[107] L. G. S. Jeub, M. Bazzi, I. S. Jutla, and P. J. Mucha,
"A generalized Louvain method for community detec-
tion implemented in MATLAB," https://github.com/
GenLouvain/GenLouvain (2011 -- 2019), version 2.2.
[108] V. D. Blondel, J.-L. Guillaume, R. Lambiotte,
and
E. Lefebvre, "Fast unfolding of communities in large
networks," J. Stat. Mech: Theory Exp. 2008, P10008
(2008).
[109] B. K. Fosdick, D. B. Larremore, J. Nishimura,
and
J. Ugander, "Configuring random graph models with
fixed degree sequences," SIAM Rev. 60, 315 -- 355 (2018).
[110] We use version 0.18.2 of the InfoMap code with the
"--two-level" option. The code is available at http:
//mapequation.org/code.
[111] M. Rosvall and C. T. Bergstrom, "An information-
theoretic framework for resolving community structure
in complex networks," Proc. Nat. Acad. Sci. U.S.A. 18,
7327 -- 7331 (2007).
[112] S. J. Cranmer, E. J. Menninga, and P. J. Mucha, "Kan-
tian fractionalization predicts the conflict propensity of
the international system," Proc. Nat. Acad. Sci. U.S.A.
112, 11812 -- 11816 (2015).
[113] M. Barigozzi, G. Fagiolo, and D. Garlaschelli, "Multi-
network of international trade: A commodity-specific
analysis," Phys. Rev. E 81, 046104 (2010).
[114] R. Mastrandrea, T. Squartini, G. Fagiolo, and D. Gar-
laschelli, "Reconstructing the world trade multiplex:
The role of intensive and extensive biases," Phys. Rev.
E 90, 062804 (2014).
[115] W. Fan and A. Yeung, "Similarity between community
structures of different online social networks and its im-
pact on underlying community detection," Commun.
Nonlinear Sci. Numer. Simul. 20, 1015 -- 1025 (2015).
[116] S. K´efi, V. Miele, E. A. Wieters, S. A. Navarrete, and
E. L. Berlow, "How structured is the entangled bank?
The surprisingly simple organization of multiplex eco-
logical networks leads to increased persistence and re-
silience," PLoS Biol. 14, 1 -- 21 (2016).
[117] We observe some erratic behavior at the transition as
the interlayer coupling ω approaches 1 from below. For
values of ω near 1 but smaller than 1, the GenLouvain
algorithm has a tendency to place all state nodes in
a single community. This observation is related to the
transition behavior that was described in [16]. For values
of ω above a certain threshold [16] only interlayer merges
occur in the first phase of Louvain.
[118] M. Starnini, A. Baronchelli, and R. Pastor-Satorras,
"Effects of temporal correlations in social multiplex net-
34
works," Sci. Rep. 7 (2017).
[119] A. Lomi, G. Robins, and M. Tranmer, "Introduction to
multilevel social networks," Soc. Networks 44, 266 -- 268
(2016).
[120] M. Rosvall, A. V. Esquivel, A. Lancichinetti, J. D. We-
stand, and R. Lambiotte, "Memory in network flows
and its effects on spreading dynamics and community
detection," Nat. Commun. 5, 4630 (2014).
[121] V. Nicosia and V. Latora, "Measuring and modeling
correlations in multiplex networks," Phys. Rev. E 92,
032805 (2015).
[122] T. A. B. Snijders and K. Nowicki, "Estimation and pre-
diction for stochastic blockmodels for graphs with latent
block structure," J. Classif. 14, 75 -- 100 (1997).
[123] K. Nowicki and T. A. B. Snijders, "Estimation and pre-
diction for stochastic blockstructures," J. Am. Stat. As-
soc. 96, 1077 -- 1087 (2001).
[124] T. P. Peixoto, "Bayesian stochastic blockmodeling,"
arXiv:1705.10225 (2017).
[125] P. Latouche, E. Birmele, and C. Ambroise, "Variational
Bayesian inference and complexity control for stochastic
block models," Stat. Modelling 12, 93 -- 115 (2012).
[126] L. Tierney, "Markov chains for exploring posterior dis-
tributions," Ann. Statist. 22, 1701 -- 1728 (1994).
[127] M. K. Cowles and B. P. Carlin, "Markov chain Monte
Carlo convergence diagnostics: A comparative review,"
J. Am. Stat. Assoc. 91, 883 -- 904 (1996).
[128] A power law with cutoffs xmin and xmax and exponent τ
is a continuous probability distribution with probability
density function
(cid:40)
p(x) =
C x−η , xmin ≤ x ≤ xmax ,
0 ,
otherwise ,
where
η − 1
− x
is the normalization constant.
−(η−1)
min
C =
x
−(η−1)
max
[129] The choice of a Poisson distribution for the number of
edges ensures that this approach for sampling edges is
approximately consistent with sampling edges indepen-
dently from Bernoulli distributions with success proba-
bilities Eq. (D1). The exact distribution for the number
of edges is a Poisson -- binomial distribution, from which
it is difficult to sample; a Poisson -- binomial distribution
is well-approximated by a Poisson distribution if none
of the individual edge probabilities are too large.
[130] H. W. Kuhn, "The Hungarian method for the assign-
ment problem," Naval Res. Logist. Quart. 2, 83 -- 97
(1955).
[131] Y. Y. Yao, "Information-theoretic measures for knowl-
edge discovery and data mining," in Entropy Measures,
Maximum Entropy Principle and Emerging Applica-
tions, Studies in Fuzziness and Soft Computing, Vol.
119, edited by Karmeshu (Springer-Verlag, Heidelberg,
Germany, 2003) pp. 115 -- 136.
[132] M. Meila, "Comparing clusterings by the variation of
information," in Learning Theory and Kernel Machines,
Lecture Notes in Computer Science, Vol. 2777, edited
by B. Scholkopf and M. K. Warmuth (Springer-Verglag,
Heidelberg, Germany, 2003).
[133] M. Meila, "Comparing clusterings -- An information
based distance," J. Multivar. Anal. 98, 873 -- 895 (2007).
[134] M. Li, J. H. Badger, X. Chen, S. Kwong, P. Kearney,
and H. Zhang, "An information-based sequence distance
and its application to whole mitochondrial genome phy-
logeny," Bioinformatics 7, 194 -- 154 (2001).
35
|
1305.6489 | 2 | 1305 | 2013-12-06T09:19:57 | Social Sensor Placement in Large Scale Networks: A Graph Sampling Perspective | [
"cs.SI",
"physics.soc-ph"
] | Sensor placement for the purpose of detecting/tracking news outbreak and preventing rumor spreading is a challenging problem in a large scale online social network (OSN). This problem is a kind of subset selection problem: choosing a small set of items from a large population so to maximize some prespecified set function. However, it is known to be NP-complete. Existing heuristics are very costly especially for modern OSNs which usually contain hundreds of millions of users. This paper aims to design methods to find \emph{good solutions} that can well trade off efficiency and accuracy. We first show that it is possible to obtain a high quality solution with a probabilistic guarantee from a "{\em candidate set}" of the underlying social network. By exploring this candidate set, one can increase the efficiency of placing social sensors. We also present how this candidate set can be obtained using "{\em graph sampling}", which has an advantage over previous methods of not requiring the prior knowledge of the complete network topology. Experiments carried out on two real datasets demonstrate not only the accuracy and efficiency of our approach, but also effectiveness in detecting and predicting news outbreak. | cs.SI | cs | Social Sensor Placement in Large Scale Networks: A Graph Sampling Perspective
Junzhou Zhao1, John C. S. Lui2, Don Towsley3, Xiaohong Guan1 and Pinghui Wang2
1MOE KLINNS Lab, Xi'an Jiaotong University
2Department of Computer Science & Engineering, The Chinese University of Hong Kong
3Computer Science Department, University of Massachusetts at Amherst
{jzzhao,phwang,xhguan}@sei.xjtu.edu.cn, [email protected], [email protected]
3
1
0
2
c
e
D
6
]
I
S
.
s
c
[
2
v
9
8
4
6
.
5
0
3
1
:
v
i
X
r
a
Abstract
Sensor placement for the purpose of detecting/tracking news
outbreak and preventing rumor spreading is a challenging
problem in a large scale online social network (OSN). This
problem is a kind of subset selection problem: choosing a
small set of items from a large population so to maximize
some prespecified set function. However, it is known to be
NP-complete. Existing heuristics are very costly especially
for modern OSNs which usually contain hundreds of mil-
lions of users. This paper aims to design methods to find
good solutions that can well trade off efficiency and accu-
racy. We first show that it is possible to obtain a high quality
solution with a probabilistic guarantee from a "candidate set"
of the underlying social network. By exploring this candidate
set, one can increase the efficiency of placing social sensors.
We also present how this candidate set can be obtained us-
ing "graph sampling", which has an advantage over previous
methods of not requiring the prior knowledge of the com-
plete network topology. Experiments carried out on two real
datasets demonstrate not only the accuracy and efficiency of
our approach, but aslo effectiveness in detecting and predict-
ing news outbreak.
I. Introduction
The rising popularity of online social networks (OSNs) has
made information sharing and discovery much easier than
ever before. Such social networks shift the role of partici-
pants from few content producers with many consumers to
both producers and consumers of content. While this fun-
damental change enables information diversity in the Inter-
net, it introduces the problem of what information sources
to subscribe to or follow due to users' limited attention ca-
pacities. For example, journalists need to discover breaking
news from OSNs in a timely manner, while government of-
ficers may want to prevent damages caused by destructive
riots arose from OSNs (e.g., the England Riots in 2011), or
twitter users may want to track valuable information by fol-
lowing limited twitter accounts. Hence, it is important to
decide which subset of social network accounts to choose as
information sources such that the total information obtained
is maximized.
The task of selecting a small number of accounts (or
nodes) to cover as much valuable information as possible
Copyright c(cid:13) 2021, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
is feasible in modern OSNs because messages are shared
by users, e.g., a tweet related to breaking news can be
retweeted, reposted or shared by the followers or friends in
social networks. Therefore, one only needs to read one par-
ticipant's tweets and obtain the information about the event.
Here, we say that diffusible tweets triggered by some users
and the associated participants form many information cas-
cades, and the social network accounts which we prefer to
select to monitor are called social sensors. Our problem is
how to select a finite number of social sensors which can
discover as many important cascades as possible.
and
optimal
Challenges: Selecting a small set of items from a large
population to maximize some prespecified set function
is a classical combinatorial optimization problem and
its solution has widespread application, e.g, set cover
problem(Khuller, Moss, and Naor 1999; Fujito 2000),
in-
fluence maximization(Kempe, Kleinberg, and Tardos 2003;
sens-
Li, qiao Zhao, and Lui 2012)
ing(Leskovec et al. 2007; Krause 2008).
However, all
these works assume that the complete data (i.e., network
topology) is available in advance. For modern OSNs, the
network topology is usually not available. This is because
many of these OSNs have hundreds of millions of accounts,
and OSN service providers usually limit the request rates.
This makes the task of discovering the entire network
topology very difficult, if not impossible. Secondly, even
if we know the network topology, the underlying subset
selection problem is NP-complete. When the objective
function is submodular, a greedy algorithm (GA) can find
a solution with the lower bound of 1 − 1/e ≈ 63% to
the optimal solution(Nemhauser, Wolsey, and Fisher 1978),
and its execution time is polynomial to the size of network.
However,
the algorithm does not scale to handle large
graphs like modern OSNs since they usually have hundreds
of millions of nodes. The current state-of-the-art approach
is the Accelerated Greedy (AG)(Minoux 1978). However,
it is not guaranteed to be efficient all the time, i.e., in the
worst case,
it is as inefficient as the greedy algorithm.
Thirdly and most importantly, the purpose of placing social
sensors is to capture "future" important events, i.e., sensors
selected based on historical data should have good predic-
tive capability of future information cascades. However,
users in OSNs are highly dynamic, e.g., everyday many
new users join in and many existing users drop out. This
will lead to a poor predictive capability of sensors selected
based solely on old historical data. To improve sensors' pre-
dictive capability, one has to periodically reselect sensors.
Therefore, we need a computationally efficient social sensor
placement algorithm that can accurately capture important
future events.
Proposed Approach: The above challenges inspire us
to develop computationally efficient, cheaper (no need
to have the topology of an OSN beforehand) methods
that provide quality guarantees.
In this paper, we intro-
duce an approach based on graph sampling(LOVASZ 1993;
Ribeiro and Towsley 2010). The basic idea is that by care-
fully choosing a set of candidate sensors, we can select
the final sensors from this candidate set. By sampling, the
search space can be reduced dramatically and efficiency is
increased. In this study, we show that graph sampling can
be used to find solutions with probabilistic quality guaran-
tees.
Results: We conduct experiments on two real datasets
Sina Weibo and Twitter, which are the two most popular
microblogs in the world, and compare our results with ex-
isting state-of-the-art methods. We not only demonstrate
our approach is computationally efficient, but we also show
that random walk based sampling can produce higher quality
candidate sets than vertex sampling. Finally, we apply our
method to Sina Weibo and show the effectiveness of sensors'
detection and prediction capability on discovering events.
This paper is organized as follows. In Section II, we intro-
duce notations and formulate the problem. The basic frame-
work of the proposed method is introduced in Section III
along with performance guarantee analysis. Experiments are
conducted in Section IV. Section V summarizes the related
works, and Section VI concludes.
II. Problem Statement
Let us first introduce some notations we use in this paper,
and then formally present the problem formulation.
Notations and Problem Formulation
Let G(V, E, C) denote the OSN under study, where V is the
set of nodes, E is the set of edges, and C is the set of infor-
mation cascades. An information cascade (or cascade for
short) c ∈ C is represented by a set of participating times
{tcu : u ∈ V, c ∈ C, tcu ≥ 0}, where tcu denotes the time
that node u first participates cascade c. If u never joins c dur-
ing our observation then tcu = ∞. The size of a cascade is
the number of users with a finite participating time, denoted
by size(c), i.e., size(c) = {u : u ∈ V ∧ tcu < ∞}. Also,
let tc denote the time cascade c begins, i.e., tc = minu tcu.
A summary of these notations is shown in Table 1.
The social sensor placement problem is to select a set of
nodes S ⊂ V as social sensors within budget B, where
S ≤ B ≪ V , so as to maximize a reward function
F (S), which is a set function F : 2V → R≥0 and satis-
fies F (∅) = 0. The optimal sensor set OP T satisfies
OP T = arg max
F (S).
S⊂V ∧S≤B
(1)
Table 1: Frequently used notations.
Notation
G(V, E, C)
tcu
tc
N b(u)
B, B ′
K
F (·)
C, Ck
S, Sk
S ∗
α
OP T
δs(S)
ξp(α)
Description
Social network, V , E and C are node/edge/cascade sets.
The time user u participates cascade c.
The start time of cascade c, tc = minu tcu.
Neighboring nodes of node u.
Budget of finding sensors and candidates respectively.
Greedy algorithm stops after K rounds, where K ≤ B.
Reward function we want to optimize.
Set of candidates or candidates at round k.
Set of sensors or sensors obtained after round k.
Top α% of the nodes ordered by reward gain decreasingly.
The optimal solution of problem (1).
Reward gain of node s with respect to S.
Sample size with confidence p and percentile α.
The reward function, in general, is determined by require-
ments of the problem under study(Leskovec et al. 2007). In
this paper, we want to trade off importance (quantified by
size(c)) against timeliness (quantified by tcu − tc), which
results in the following,
F (S) = Xc∈C
size(c)
1 + minu∈S{tcu − tc}
.
(2)
That is, if the sensor set S can cover as many large size
cascades as early as possible, the reward will be high.
(1)
is
problem
Submodularity and Greedy Algorithm
Optimization
NP-
complete(Kempe, Kleinberg, and Tardos 2003). However,
when F is a) nondecreasing, i.e., if S ⊆ T ⊆ V , then
F (S) ≤ F (T ), and b) submodular, i.e., if S ⊆ T ⊆ V ,
then F (S ∪ {s}) − F (S) ≥ F (T ∪ {s}) − F (T ),
∀s ∈ V \T , the greedy algorithm can obtain an approx-
imate solution that is at least 1 − 1/e ≈ 63% of the
optimal(Nemhauser, Wolsey, and Fisher 1978). It is easy to
show that Eq. (2) possesses these two properties.
The greedy algorithm can be stated as follows. It runs for
at most B rounds to obtain a set S of size S ≤ B. In each
round, it finds a node s ∈ V \S that maximizes the reward
gain δs(S) , F (S ∪ {s}) − F (S), then s is added into S in
this round. This process repeats K rounds until S = B or
δs(S) = 0. The computation complexity of this algorithm is
O(K · V ). Note that this greedy algorithm is not scalable
for large scale OSNs since graphs of these OSNs usually
have large number of nodes (or V is very large).
To speed up the greedy algorithm, Leskovec et
al(Leskovec et al. 2007) use a Cost-Effective Lazy Forward
(CELF) approach, also known as Accelerated Greedy (AG)
proposed in (Minoux 1978) to reduce the computation times
of δs(S) in each round by further utilizing the submodular-
ity of F . The basic idea is that the reward gain of a node in
the current round cannot be better than its reward gain in the
previous round, i.e., if k > l, then δs(Sk) ≤ δs(Sl), ∀s ∈
V \Sk, where Sk, 1 ≤ k ≤ K is the selected nodes after the
k-th round by the greedy algorithm. However, it is important
to note that AG/CELF does not guarantee an improvement
on computational efficiency, i.e., in the worst case, it is as
inefficient as the naive greedy algorithm(Minoux 1978).
III. Search Space Reduction
The inefficiency of the above mentioned algorithms is due to
the large search space and the absence of an efficient search
method. For example, to find at most B sensors from the
node set V , there are PB
n=1(cid:0)V
n (cid:1) = O(2V ) possible solu-
tions. Hence, we modify our problem, and consider how to
find some acceptable good solutions at a much lower com-
putational cost. In the following, we first describe the basic
framework of our approach and present the definition of ac-
ceptable good solution of our algorithm. We then formally
show the performance guarantees of the proposed approach,
along with its variants and cost analysis.
The Basic Framework
In order to reduce search space, we consider using a candi-
date set Ck ⊆ V \Sk−1 to represent the search space at round
k. This forms the basic search space reduction framework
as described in Alg. 1 (We sometimes suppress the subscript
k if there is no ambiguity).
Algorithm 1: The basic framework
Input: Nodes V , budget B.
Output: Sensors S.
1 S = ∅;
2 while S < B, do
3
4
5
6 end
Generate candidate set C ⊆ V \S;
Select a node s∗ such that s∗ = arg maxs∈C δs(S);
S = S ∪ {s∗};
The only difference between Alg. 1 and the original
greedy algorithm is the process at line 3.
In the original
algorithm, one selects a node s∗ from V \S that maximizes
the reward gain δs(S). Here, we select s∗ from a candidate
set C ⊆ V \S to maximize δs(S). Intuitively, the accuracy
of this algorithm should be arbitrary close to the greedy al-
gorithm as C → V \S at each round but with a reduction in
computational cost. Furthermore, if the sample size at each
round is the same, then Alg. 1 becomes V /C times faster
than the original greedy algorithm. In later sub-section, we
will present algorithms on how to generate the candidate set
C. Let us first define what we mean by acceptable good so-
lutions.
Acceptable Good Solutions
At round k, suppose we rank the nodes in V \Sk−1 by reward
gain in a decreasing order, and denote the top α% of the
nodes by S∗
α, we
say s∗ is an acceptable good sensor found in round k. It is
obvious that one needs to set α to be small, say less than 1, so
to find acceptable good sensors. All acceptable good sensors
found after K rounds form the acceptable good solution of
Alg. 1.
α, where α ∈ (0, 100]. If s∗ falls into S∗
C
5000
4500
4000
3500
3000
2500
2000
1500
1000
500
0
p=0.90
p=0.95
p=0.99
p=0.90
p=0.95
p=0.99
100
)
%
(
p
a
l
r
e
v
O
80
60
40
20
0
0.2
0.4
0.6
α
0.8
1
2000 4000 6000 8000
C<k
(a) Candidates size C.
(b) Overlap percentile.
Figure 1: (a) Reduction of C as we increase α. (b) For a
small network of V = 10000 and Ck = 1000 nodes, there
is a high degree of overlap.
The candidate set size C will affect the likelihood that at
α. Formally, we can calculate
least one node in C falls into S∗
the probability that at least k nodes in C falls into S∗
α as
P rob{C ∩ S∗
α ≥ k} =
S ∗
α
Xi=k
α
α
i (cid:1)(cid:0)V \S ∗
(cid:0)S ∗
C−i (cid:1)
(cid:0)V
C(cid:1)
,
(3)
which follows a hypergeometric distribution. For k = 1, this
is equivalent to
P rob{C ∩ S∗
α ≥ 1} = 1 − (1 − α%)C.
(4)
In order to achieve a confidence level that P rob{C ∩ S∗
1} ≥ p, we can determine the lower bound on C as
α ≥
C ≥ (cid:24) ln(1 − p)
ln(1 − α%)(cid:25) .
(5)
Fig. 1a shows the relationship between C and α with
three different confidence levels of p, which we set to 0.90,
0.95 and 0.99. One interesting observation is that, when α
varies from 0.1 to 1 (or top 0.1% to 1%), C drops quickly,
which means that the search space size can be dramatically
reduced. For example, if we want to choose a node in top
1% (or α = 1) of the best nodes from the whole population,
then one can be sure that one of the 458 nodes will be a
good sensor with probability greater than 0.99. There is one
technical issue we need to pay attention:
Dealing with small networks. When the network is small,
the advantage of using our framework is not significant.
We can further improve the efficiency of Alg. 1 by exploit-
ing the submodular property of reward functions as used in
(Minoux 1978; Leskovec et al. 2007). One conclusion from
Eq. (3) is that, when V is small, at round k, the overlap
i=1 Ci and current
candidates Ck, i.e., C<k ∩ Ck, is large and increases as
C<k increasing. An example is illustrated in Fig. 1b. In
the example, we have a small network and its size is set to
V = 10000. At each round, 1000 candidates are selected.
We show the least overlap percentile versus selected candi-
date set size C<k with a given confidence p. For exam-
ple, when 1000 candidates have been selected in the previ-
ous rounds, for a newly 1000 selected candidates, there is a
guarantee that at least 8% of them have been selected in pre-
vious rounds with probability greater than 0.99. For those
between previous candidates C<k = Sk−1
overlapped nodes, we can reduce the calculations of updat-
ing reward gain by utilizing the submodular property. This
will reduce the computational complexity of our framework.
To achieve this, we can store the candidate node u and
its reward gain δu in a tuple hu, δu, #ui. Here #u is the
round during which the reward gain of node u is calculated
or updated. For candidates Ck, we only need to calculate the
reward gain of newly selected nodes and we do not need to
update the reward gain of previously selected nodes imme-
diately. We arrange nodes in Ck as a priority queue Q by
descending order of reward gain. Then we access the head
v of Q to check whether #v equals to the current round k.
If yes, then v is added into S; otherwise, update v's reward
gain, and put v into Q again. This way, the times of updat-
ing reward gain of sampled nodes can be reduced and Alg. 1
will be at least as efficient as AG/CELF when dealing with
small networks.
Performance Analysis
We conclude from our previous discussion that we can prob-
abilistically guarantee in selecting an acceptable sensor at
each round from a small candidate set C. Here, we quantify
the quality of the final sensor set obtained after K rounds.
That is, how close the solution obtained by Alg. 1 is to the
optimal solution of problem (1)? The following theorem an-
swers this question.
Theorem 1. Denote the set of sensors obtained by Alg. 1
after round k as Sk = {s∗
k}, 1 ≤ k ≤ K.
Let λk = δs∗
k =
arg maxs∈V \Sk−1 δs(Sk−1). Let λ = min1≤k≤K λk, then
(Sk−1) ∈ (0, 1], where s∗
(Sk−1)/δs∗
1, · · · , s∗
k
k
F (SK) ≥ (1 −
1
eλ )F (OP T ).
(6)
Proof: Please refer to the Appendix.
Remark: Theorem 1 indicates an important property of
Alg. 1, i.e., if at each step the reward gain δs∗ (S) is bounded
by a factor λ, then the final solution is bounded by an-
other factor 1 − 1/eλ. When λ ≈ 1, our final solution is
1 − 1/e ≈ 63% of the OP T . Hence, Alg. 1 guarantees a
good solution when λ is close to one. Note that in general,
one cannot guarantee that λ is close to one all the time. For
example, the reward gain of the second best node is much
less than the best node in current round.
In such a case,
finding the best node is like finding a needle in a haystack,
which illustrates the intrinsic difficulty due to the reward dis-
tribution. Nevertheless, one can still accept the approximate
solution if we believe the second best solution has a reward
which is not significantly different from the best solution.
r = SK
One can also understand the quality of Alg. 1 from an-
other point of view. We define cover ratio as the frac-
tion of nodes in OP T that are in all candidate sets, i.e.,
k=1 Ck ∩ OP T /K. Intuitively, the higher the cover
ratio, the better the quality of the final solution. Assume the
size of candidate set equal in each round, and denote it by
ξp(α) with confidence p and percentile α. The following
theorem presents a lower bound on the expectation of cover
ratio for fixed p and α.
Theorem 2. If we set α = 100K/V at each round, then
E[r] ≥ 1 − 1/ep.
Proof: Please refer to the Appendix.
Remark: Because p can be very close to 1, 1 − 1/ep will
be very close to 1 − 1/e ≈ 0.63. In fact, the cover ratio
can be further improved by increasing sample size at each
round, e.g., we can derive in a similar manner that if we set
α = 50K/V , then the expected cover ratio will be at least
0.82, or 82% of those sensors in OP T .
Generating C via Graph Sampling Methods
In previous discussion, we did not specify precisely how to
generate the candidate set C. In fact, C can be generated by
randomly selecting nodes in V . This does not exploit any
structural properties of OSNs, hence the performance dis-
cussed in previous sub-section can be viewed as the worst
case guarantee. Here, we show how to use graph sampling
to generate C. The main advantage of using graph sam-
pling is that one does not need to know the complete graph
topology prior to executing the sensor placement algorithm.
This is one main advantage of our framework as compare
with the current state-of-the-art approaches(Minoux 1978;
Leskovec et al. 2007). Furthermore, we also want to explore
how different graph sampling methods may affect sensor
quality. Hence, we make two modification to Alg. 1:
• Instead of selecting new candidates in each round, we
construct one candidate set C at the beginning of the al-
gorithm;
• When a node v is sampled, we select a candidate node,
say s, from v or one of v's neighbors (i.e., {v} ∪ N b(v))
which can maximize the reward gain and we put s into C.
The second change is useful in filtering out noisy tweets in
OSNs (e.g., tweets containing unimportant cascades), and
reduces the candidate set size. Nodes with large reward gain
in the neighborhood are more preferred to be in C.
Vertex sampling (VS) and random walk (RW) are two
popular graph sampling methods. We design two variants
of Alg. 1 based on vertex sampling and random walk re-
spectively, they are illustrated in Algs. 2 and 3.
Algorithm 2: Combining with vertex sampling
Input: Network G, sensor budget B, candidate budget B′.
Output: Sensors S.
1 C = ∅;
2 while C < B′, do
3
4
Choose a node v ∈ V \C;
Select a node s∗ such that
s∗ = arg maxs∈N b(v)∪{v} δs(C);
C = C ∪ {s∗};
5
6 end
7 S = Greedy(C, B) ; /* choose B sensors from C
using the greedy algorithm */
Each of these algorithms contains two steps. In the first
step, the candidate set C with budget B′ are constructed. In
the second step, the final sensors S are chosen. In line 3 of
Algorithm 3: Combining with random walk
Input: Network G, sensor budget B, candidate budget B′.
Output: Sensors S.
1 C = ∅;
2 Set u = u0 which is randomly chosen from V ;
3 while C < B′, do
4
5
Choose a node v ∈ N b(u);
Select the node s∗ such that s∗ = arg maxs∈N b(v) δs(C);
C = C ∪ {s∗};
Set u = v;
6
7
8 end
9 S = Greedy(C, B) ; /* choose B sensors from C
using the greedy algorithm */
Alg. 2 and line 4 of Alg. 3, we can use various attribute infor-
mation within an OSN to bias the selection. Such attributes
of a user can be the number of posts/friends/followers and
so on. For example, when using random walk to build the
candidate set (Alg. 3), we can choose a neighboring node
with probability proportion to its degree or activity (#posts),
which will bias a random walk toward high degree or activ-
ity nodes. Intuitively, large degree nodes are more likely to
be information sources or information hubs, and high activ-
ity nodes are more likely to retweet tweets. The comparison
of using different attributes to bias our node selection will
be discussed in Section IV. Although vertex sampling can-
not be biased without knowing attributes of every node in
advance, in order to study functions of different attributes,
we will assume we know this complete information and let
vertex sampling be biased by different attributes.
Sampling Cost Analysis
The computational cost of social sensor selection is defined
to be the number of times that the reward gain of nodes
is calculated or updated. Because of the second change in
Alg. 2 and Alg. 3, there will be additional cost in obtain-
ing samples C. Here, we analyze the cost of sampling using
uniform vertex sampling (UVS) (for Alg. 2) and uniform
random walk (URW) (for Alg. 3).
For UVS in Alg. 2, the average cost to obtain candidate
set C is
CostUVS = E[
B ′
Xi=1
(di + 1)] = B′(1 + E[dUVS]),
where di is the degree of i-th node in C, and E[dUVS] is
the average node degree obtained by UVS. Let M denote
the maximum degree in the network, and θd the fraction of
nodes with degree d. Then E[dUVS] = PM
d=1 dθd , davg.
For power law networks with degree distribution θd =
1
Z d−a, we obtain
E[dUVS] =
1
Z
M
Xd=1
d1−a.
(7)
For URW, the average cost to obtain candidate set C can
Table 2: Dataset summary
Dataset
Nodes
Edges
Cascades
Sina Weibo Twitter
1.7M
22M
7M
0.3M
1.7M
11M
be derived similarly, that is
CostURW = E[
B ′
Xi=1
di] = B′ · E[dURW],
where E[dURW] is the average node degree under URW, i.e.,
. For power law networks, it be-
E[dURW] = PM
comes
d=1 d dθd
davg
E[dURW] =
1
Zdavg
M
Xd=1
d2−a.
(8)
Equations (7) and (8) reveal the difference between UVS
and URW. Both of them are functions of a, the exponent of
the power law degree distribution. The difference is that,
for UVS, E[dUVS] is always finite when a > 2, which means
that the average cost of obtaining a sample equal the average
degree, and it is usually small and bounded by the Dunbar
number. For URW, E[dURW] is finite if a > 3; otherwise the
average cost can become arbitrarily large as M increases.
To limit the cost of creating candidate set, we can use sim-
ple heuristics to avoid searching a local maximum reward
gain node from all the neighbors. For example, we can limit
the searching scope to the top n most active neighbors. In
our experiments, n can be set very small (e.g., 10), so the
cost of obtaining samples will be bounded by O(B′ · n).
IV. Experiments
To evaluate the performance of previous discussed methods,
we present experimental results on two datasets collected
from Sina Weibo and Twitter, respectively. Then, we con-
duct experiments on Sina Weibo to study the detection ca-
pability, which measures the ability to detect cascades, as
well as prediction capability, which measures the ability to
capture future cascades using the senors selected based on
historial data.
Experiment on Sina Weibo and Twitter
Dataset Sina Weibo is one of the most popular microblog-
ging sites in China. Similar to Twitter, users in Weibo are
connected by the following relationships. Tweets can be
retweeted by one's followers and form cascades. We col-
lected a portion of Weibo network using the Breath First
Search (BFS) method along the following relationships. For
a user, his tweets and neighbors are all collected. We extract
URL links contained in tweets, and consider them as the rep-
resentation of cascades. The Twitter dataset contains tweets
and network, which are from (Link 2011) and (Link 2010),
respectively. Similar to Weibo, URL links and hashtags con-
tained in tweets are extracted to form cascades. Table 2 sum-
marizes the statistics of these two networks.
GAξ
0.9(α)
ξ
0.9(0.5α)
95%
90%
3.0
x106
d
r
a
w
e
R
2.5
2.0
1.5
1.0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
p
u
-
d
e
e
p
S
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
AG/CELF
ξ
0.9(α)
ξ
0.9(0.5α)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
α
d
r
a
w
e
R
3.5
3.0
2.5
2.0
1.5
1.0
x105
95%
90%
p
u
-
d
e
e
p
S
90000
80000
70000
60000
50000
40000
30000
20000
10000
0
GA
Uniform
Degree
Activity
AG/CELF
Uniform
Degree
Activity
200 400 600 800 1000 1200 1400
200 400 600 800 1000 1200 1400
C
C
(a) Reward (Weibo)
(b) Speed-up (Weibo)
(a) Reward (Weibo)
(b) Speed-up (Weibo)
α
α
x105
GAξ
0.9(α)
ξ
0.9(0.5α)
95%
90%
1.6
1.4
1.2
1.0
0.8
0.6
d
r
a
w
e
R
0.4
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
12000
10000
p
u
-
d
e
e
p
S
8000
6000
4000
2000
0
AG/CELF
ξ
0.9(α)
ξ
0.9(0.5α)
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
α
10.0
d
r
a
w
e
R
8.0
6.0
4.0
2.0
0.0
x103
95%
90%
p
u
-
d
e
e
p
S
160000
140000
120000
100000
80000
60000
40000
20000
0
GA
Uniform
Degree
Activity
AG/CELF
Uniform
Degree
Activity
200 400 600 800 1000 1200 1400
200 400 600 800 1000 1200 1400
C
C
(c) Reward (Twitter)
(d) Speed-up (Twitter)
(c) Reward (Twitter)
(d) Speed-up (Twitter)
Figure 2: Evaluating the basic framework on real datasets.
Figure 3: Vertex sampling on real datasets (B = 100).
Cost(algorithm)
Evaluating the basic framework We first evaluate the
performance of Alg. 1 without any network information.
We choose B sensors where B ranges from 0.1% to 1%
of the total number of nodes, and compare the quality of
sensors and speed-up of the algorithm with the greedy al-
gorithm and AG/CELF. Speed-up of an algorithm is defined
as Cost(greedy algorithm)
, where Cost(A) represents the cost of
algorithm A, i.e., number of times of calculating or updat-
ing reward gains. The sample size at each round is fixed to
ξ0.9(α) and ξ0.9(0.5α) respectively, where α = 100B/V .
In the reward curves of Fig. 2, two dashed lines repre-
sent 90% and 95% of the total reward by the greedy algo-
rithm respectively. We show the rewards of sensors with
different sizes. One can observe that the accuracy in to-
tal reward of the sampling approach is within 90% of the
greedy algorithm, and that it is more computational efficient
than AG/CELF from the speed curves. When sample size
increases from ξ0.9(α) to ξ0.9(0.5α), the reward increases
to about 95% of that produced by the greedy algorithm, but
with a slight reduction in speedup. Hence, one can adjust
the sample size to trade off between accuracy and efficiency.
Evaluating the vertex sampling framework Next, we
evaluate the benefit of using attribute information within an
OSN, in particular, in reducing the size of the candidate set
and reducing the computational cost of Alg. 2. For ver-
tex sampling, we consider the following variants: (a) uni-
formly selecting a node from the network; (b) select a node
from the network with a probability proportional to its de-
gree; (c) select the node from the network with a probability
proportional to its activity (say # posts). The aim is to se-
lect B = 100 sensors. Each experiment is run 10 times,
and the averaged results are shown in Fig. 3. From the re-
ward curves, we observe that vertex sampling by degree is
the best follows by sampling by activity, and uniform vertex
sampling is the worst. However, from the speed-up curves
we can see that uniform vertex sampling is the most effi-
cient approach. Sampling by degree is the most expensive
method. But the sampling approaches are in general much
more efficient than AG/CELF.
Evaluating the random walk framework We now eval-
uate the performance of the random walk algorithm as de-
scribed in Alg. 3. Note that this algorithm does not require
the knowledge of the network topology in advance. We have
three variants: (a) select a neighboring node uniformly; (b)
select a neighboring node with a probability proportional to
its degree; (c) select a neigbhoring node with a probabil-
ity proportional to its activity. The other settings are simi-
lar to vertex sampling, and the results are shown in Fig. 4.
Again, we can see that degree is the best attribute for ran-
dom walk, it is also the most expensive one. Comparing
random walk with vertex sampling, we observe that random
walk can achieve higher accuracy but at a higher computa-
tional cost, as is shown in Fig. 5. Both of them are more
efficient than AG/CELF. Furthermore, the random walk al-
gorithm does not require a full topology beforehand.
Application Study on Sina Weibo
For a set of selected sensors, we are interested in its detec-
tion and prediction capabilities. For detection, it means how
many or how timely a set of sensors can discover informa-
tion cascades from a given dataset. For prediction, it means
how well a set of sensors selected based on historical dataset
can generalize to a future events, e.g., a set of sensors are se-
lected based on a past week's data, and we want to know
how well they perform on a future week's data. We conduct
experiments on Sina Weibo to study these two capabilities
of sensors selected by different methods.
Detection Capability Analysis We compare the detection
quality of our algorithm with two baselines: a) randomly
x105
3.5
d
r
a
w
e
R
3.0
2.5
2.0
1.5
10.0
x103
95%
90%
d
r
a
w
e
R
8.0
6.0
4.0
2.0
p
u
-
d
e
e
p
S
35000
30000
25000
20000
15000
10000
5000
0
GA
Uniform
Degree
Activity
AG/CELF
Uniform
Degree
Activity
200 400 600 800 1000 1200 1400
200 400 600 800 1000 1200 1400
C
C
(c) Reward (Twitter)
(d) Speed-up (Twitter)
Figure 4: Random walk on real datasets (B = 100).
x104
5
4
3
2
1
p
u
-
d
e
e
p
S
VS-uniform
VS-degree
VS-activity
RW-uniform
RW-degree
RW-activity
0
2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4
x105
Reward
(a) Weibo
p
u
-
d
e
e
p
S
6
5
4
3
2
1
0
x104
VS-uniform
VS-degree
VS-activity
RW-uniform
RW-degree
RW-activity
3
4
8
9
10
x103
5
7
6
Reward
(b) Twitter
Figure 5: Comparing vertex sampling with random walk in
Speedup and Reward.
choose a set of nodes as sensors; b) choosing sensors by
utilizing friendship paradox(Christakis and Fowler 2010).
Friendship paradox randomly selects a neighbor of a ran-
domly sampled node as a sensor, and it is proved to be
able to sample larger degree nodes than random node se-
lection(Feld 1991).
We then apply the method of Alg. 3 in which random walk
is biased by neighbors' activity and totally 50,000 Weibo ac-
counts are collected, which form the candidate set. We use
the posts between Jan 1, 2012 to Sep. 1, 2012 to evaluate the
quality of a node. From these candidates we choose B sen-
sors where B ranges from 2,000 to 10,000 using the reward
function in Eq. (2). We also introduce two other measures
to evaluate sensor quality: (a) the number of cascades sen-
sors can detect; (b) the detecting lead-time, which measures
the time interval that the sensors first detect a cascade in ad-
vance of the peak time of the cascade. It can be considered as
the warning time for outbreaks a set of sensors can provide.
The results are shown in Fig. 6. We observe that sensors ob-
tained by Alg. 3 can discover around two to six times more
cascades and provide earlier warning time (about two days)
than the other two baseline methods.
95%
90%
GA
Uniform
Degree
Activity
20000
15000
10000
p
u
-
d
e
e
p
S
5000
0
AG/CELF
Uniform
Degree
Activity
RW-activity
Friendship paradox
Random
12
x105
s
t
c
e
t
e
d
#
8
4
200 400 600 800 1000 1200 1400
200
400
C
800
1000
600
C
(a) Reward (Weibo)
(b) Speed-up (Weibo)
0
2000
4000
6000
8000 10000
#sensors
(a) #detects
)
r
u
o
h
(
e
m
i
t
-
d
a
e
l
.
g
v
A
RW-activity
Friendship paradox
Random
400
350
300
250
200
150
100
50
0
100 101 102 103 104 105 106
Cascade size
(b) Lead-time
Figure 6: Detection ability comparison.
)
%
(
s
e
d
a
c
s
a
c
f
o
n
o
i
t
c
a
r
F
50
40
30
20
10
0
CELF on week i
CELF on week 0
0
5 10 15 20 25 30 35
Prediction steps (week)
30
25
20
15
10
)
%
(
s
e
d
a
c
s
a
c
f
o
n
o
i
t
c
a
r
F
CELF
RW-uniform
RW-degree
RW-activity
20
40
60
80 100
#sensors
(a) #detects by T0 (B = 100).
(b) One step prediction.
Figure 7: Prediction ability of sensors.
Prediction Capability Analysis We study prediction in
the scheme of choosing sensors based on historical data and
testing them on future data. We want to answer: Are selected
sensors based on historical data still good for capturing fu-
ture cascades? Here we mainly compare the results with the
method presented in (Leskovec et al. 2007), which chooses
sensors using CELF. For our framework, we use Alg. 3 to
select the sensor. The Sina Weibo data collected by BFS
used in previous section will be our ground truth data, and
we split time into granularity of a week.
For the convenience of description, let Ti, i ≥ 0 denote
the sensors selected by the CELF algorithm using the ith
week data. Fig. 7a has two curves: "CELF on week 0" rep-
resents running CELF on week 0 data only, while "CELF
on week i" represent running CELF on the ith week data.
The figure depicts the fraction of cascade detected at differ-
ent weeks. The black curve corresponds to "CELF on week
0" while the red curve corresponds to "CELF on week i".
We can see that the senors in T0 (or "CELF on week 0") is
reducing its predictive capability as time evolves. However,
if we use Ti (or "CELF on week i"), we can have a much
higher predictive capability. This implies that social network
is highly dynamic and one needs to execute the sensor se-
lection algoirthm more often, rather than relying on sensors
selected based on old histrical data.
Next, we study the prediction performance of our method.
We choose 1,000 candidates by random walk biased by
uniform/degree/activity respectively, from which we select
B sensors where B ranges from 20 to 100, and compare
the one-week prediction capability with aforementioned
"CELF" approach. The result is shown in Fig. 7b. One in-
teresting observation is that our approach is at least as good
as the "CELF" approach. Random walk biased by users'
)
%
(
.
v
o
r
p
m
i
s
t
c
e
e
d
#
t
5
4
3
2
1
0
10
)
%
(
.
v
o
r
p
m
i
s
t
c
e
e
d
#
t
5
4
3
2
1
0
2 3 4 5
)
%
(
.
v
o
r
p
m
i
s
t
c
e
e
d
#
t
5
4
3
2
1
0
1
2
3
#active days a week
#cascades a day
15
12
14
11
#cascades a week
13
Figure 8: One step prediction improvements relative to
"greedy" under different regularization rules.
activity is the best, which can improve about 3% of detects
on the testing data. Since our approach is more computa-
tionally efficient than CELF, hence, one can apply it more
often so to have a good predictive capability on information
cascade, e.g., better than results of "CELF on week i".
To understand why our random walk approach has
good predictive capability, one can consider the candidate-
selecting step as a pruning or regularizing process, which
are common techniques used to avoid overfitting of deci-
sion trees(Witten and Frank 2005, Chapter 6). For example,
when using uniform random walk to collecting candidates,
it prefers to select high degree nodes of the network. This
can be considered as a rule to regularize the search scope,
which can avoid choosing unimportant users who join into
large cascades just by chance. To demonstrate this, we use
the following regularization rules to constraint search scopes
and then apply the CELF to see whether there is any im-
provements,
• #cascades in a week. The search scope is limited to users
who have joined at least x cascades in the history data;
• #active days in a week. The search scope is limited to
users who have been active at least x days in the history
data;
• #cascades per day. The search scope is limited to users
who participates at least x cascades per day.
Fig. 8 depicts the results. We observe some improvements
of prediction performance when using different constraints,
and users' activity in history is the best rule, which is consis-
tent with the results in Fig. 7b. This shows that our random
walk algorithm has the intrinsic property to regularize the
searching scope similar to the above regularization rules.
V. Related Work
In this section, we briefly review some related work, which
are categorized into three groups:
Influence maximization: One important topic related to our
work is how to identify k most influential nodes in a social
network. Domingos et al(Domingos and Richardson 2001)
first posed this problem in the area of marketing.
Kempe et al(Kempe, Kleinberg, and Tardos 2003) proved
the problem is NP-complete under both indepen-
that
dent cascade model and threshold model.
They also
provided an approximation solution which has a con-
stant factor bound. However, computing the influence
spread given a seed set by Monte Carlo simulations is
not scalable. Other works(Chen, Wang, and Yang 2009;
Li, qiao Zhao, and Lui 2012) tried to scale up the Monte
Carlo simulation. In our work, we do not assume any cas-
cade or threshold model, but rather, exploit the attributes of
OSNs to determine the B sensor nodes.
Optimal sensing: Optimal sensing problems(Krause 2008;
Leskovec et al. 2007) are closely related to our work, which
aim to find optimal observers or optimal sensor place-
ment strategies to measure temperature, gas concentra-
tion or detect outbreaks in networks(Leskovec et al. 2007;
Christakis and Fowler 2010). Our work can be considered
as an extension of Leskovec's work(Leskovec et al. 2007) in
which the authors proposed the CELF approach to speedup
the greedy algorithm. It is similar to the AG posed by Mi-
noux(Minoux 1978). There are several differences between
their work and ours. First, we study the problem with-
out assuming the knowledge on the complete OSN topol-
ogy. Second, our work aims to speedup the greedy algo-
rithm via graph sampling, and that we can tradeoff a small
loss of accuracy but obtain large improvement in efficiency.
Furthermore, the performance of our algorithm is guaran-
teed with high probability. The last property is important
since AG/CELF can be as inefficiency as the greedy al-
gorithm in the worst case(Minoux 1978). Third, we ex-
ploit the meta information within in social networks.
In
(Leskovec et al. 2007), authors considered the problem as a
discrete optimizing problem without considering contextual
information in OSNs. In our experiments, we find that con-
textual information can be used to reduce sample size and
improve efficiency.
Graph sampling and optimizing. Graph sampling meth-
ods are used to measure properties of nodes and edges
in graphs,
such as degree distribution(LOVASZ 1993;
Ribeiro and Towsley 2010). To the best of our knowledge,
this is the first work which uses graph sampling methods
to maximize a set function. Lim et al(Lim et al. 2011) and
Maiya et al(Maiya and Berger-Wolf 2010) both use differ-
ent sampling methods to estimate the top k largest centrality
nodes in a graph, i.e., degree centrality, betweenness central-
ity, closeness centrality and eigenvector centrality. However
these top k largest centrality nodes are not the top k optimal
sensors which are defined by a set function F (·). Our work
can be considered as a combination of ordinary optimiza-
tion(Ho, Sreenivas, and Vakili 1992) with greedy algorithm.
It uses the property of greedy algorithm that if the reward
gain at each round is bounded by factor λ, then the final so-
lution obtained by the greedy algorithm is bounded by factor
1 − 1/eλ as shown in Theorem 1.
VI. Conclusions
Many OSNs are large in scale and there is an urgent need
on how to select reliable information sources to subscribe so
one can track/detect information cascades. This can be for-
mulated as a sensor placement problem and previous work
used heuristic greedy algorithms. However, it is impractical
to run these greedy algorithms on a OSN composed of mil-
lions of users. Hence we propose sampling approach to find
these good sensors.
We show that one can significantly reduce the complexity
by sampling the huge search space, and still can guarantee
to have good solutions with high probabilistic guarantees.
Hence, the sampling approach is a good method to trade off
efficiency and accuracy. We evaluate various graph sam-
pling approaches, and find that random walk based sampling
methods perform better than vertex sampling based meth-
ods. This indicates that structural information of OSNs is
important and random walks are suitable for obtaining better
samples, while previous discrete optimization approaches
failed to utilize this information. We apply our framework
on Sina Weibo, and the results demonstrate that using our
algorithm, one can effectively detect information cascade.
Since our algorithm is computationally efficient, one can ex-
ecute it more often so to find new sensor nodes to accurately
predict future information cascade.
References
P.,
and
2001. Mining the network value of
[Chen, Wang, and Yang 2009] Chen, W.; Wang, Y.; and
Yang, S. 2009. Efficient influence maximization in social
networks. In KDD'09, 199 -- 208.
[Christakis and Fowler 2010] Christakis, N. A., and Fowler,
J. H. 2010. Social network sensors for early detection of
contagious outbreaks. PLoS ONE 5.
[Domingos and Richardson 2001] Domingos,
Richardson, M.
customers. In KDD'01.
[Feld 1991] Feld, S. L.
1991. Why your friends have
more friends than you do. American Journal of Sociology
96(6):1464 -- 1477.
[Fujito 2000] Fujito, T. 2000. Approximation algorithms for
submodular set cover with applications. IEICE TRANS. INF.
& SYST.
[Ho, Sreenivas, and Vakili 1992] Ho, Y. C.; Sreenivas, R. S.;
and Vakili, P. 1992. Ordinal optimization of deds. Discrete
Event Dynamic Systems: Theory and Applications 2:61 -- 88.
[Kempe, Kleinberg, and Tardos 2003] Kempe, D.; Klein-
berg, J.; and Tardos, E. 2003. Maximizing the spread of
influence through a social network. KDD'03.
[Khuller, Moss, and Naor 1999] Khuller, S.; Moss, A.; and
Naor, J. S. 1999. The budgeted maximum coverage problem.
Information Processing Letters 70:39 -- 45.
[Krause 2008] Krause, A. 2008. Optimizing Sensing Theory
and Applications. Ph.D. Dissertation, School of Computer
Science Carnegie Mellon University.
[Leskovec et al. 2007] Leskovec, J.; Krause, A.; Guestrin,
C.; Faloutsos, C.; VanBriesen, J.; and Glance, N. 2007.
Cost-effective outbreak detection in networks. KDD'07
420 -- 429.
[Li, qiao Zhao, and Lui 2012] Li, Y.; qiao Zhao, B.; and Lui,
J. C. 2012. On modeling product advertisement in large
scale online social networks. IEEE/ACM Transitions on Net-
working.
[Lim et al. 2011] Lim, Y.; Ribeiro, B.; Menasche, D. S.;
Basu, P.; and Towsley, D. 2011. Online estimating the top k
nodes of a network. IEEE NSW.
social
network
or
is
news
What
a
twitter
tweets.
476 million
twitter,
media?
[Link 2010] 2010.
a
http://an.kaist.ac.kr/traces/WWW2010.html.
[Link 2011] 2011.
http://snap.stanford.edu/data/twitter7.html.
[LOVASZ 1993] LOVASZ, L.
1993. Random walks on
graphs: A survey. BOLYAI SOCIETY MATHEMATICAL
STUDIES 1 -- 46.
[Maiya and Berger-Wolf 2010] Maiya, A. S., and Berger-
Wolf, T. Y. 2010. Online sampling of high centrality in-
dividuals in social networks. PAKDD.
[Minoux 1978] Minoux, M. 1978. Accelerated greedy algo-
rithms for maximizing submodular set functions. Optimiza-
tion Techniques.
[Nemhauser, Wolsey, and Fisher 1978] Nemhauser, G. L.;
Wolsey, L. A.; and Fisher, M. L. 1978. An analysis of
approximations for maximizing submodular set functions.
Mathematical Programming 14:265 -- 294.
[Ribeiro and Towsley 2010] Ribeiro, B., and Towsley, D.
2010. Estimating and sampling graphs with multidimen-
sional random walks. IMC'10.
[Witten and Frank 2005] Witten, I. H., and Frank, E. 2005.
Data Mining: Practical Machine Learning Tools and Tech-
niques. Morgan Kaufmann.
Appendix
Proof of Theorem 1
Proof. By utilizing the non-decreasing property of F , we
have
F (OP T ) − F (Sk−1) ≤F (OP T ∪ Sk−1) − F (Sk−1)
=F (OP T \Sk−1 ∪ Sk−1) − F (Sk−1).
Assume OP T \Sk−1 = {z1, · · · , zm}, m ≤ K, and let j =
1, · · · , m, we have
Zj = F (Sk−1 ∪{z1, · · · , zj})−F (Sk−1∪{z1, · · · , zj−1}).
With some algebraic manipulation, we have
F (OP T ) − F (Sk−1) ≤
m
Xj=1
Zj.
Now notice that
Zj ≤F (Sk−1 ∪ {zj}) − F (Sk−1)
=δzj (Sk−1) ≤
1
λ
δs∗
k
(Sk−1)
=
1
λ
(F (Sk) − F (Sk−1)).
Then we get an iterative formula of F (Sk),
F (OP T ) − F (Sk−1) ≤
m
Xj=1
Zj ≤
K
λ
(F (Sk) − F (Sk−1)).
Iteratively, we can derive the following relationship,
F (Sk) ≥ [1 − (1 −
λ
K
)k]F (OP T ).
Finally, let k = K. We conclude that
F (SK) ≥ [1 − (1 −
λ
K
)K ]F (OP T ) ≥ (1 −
1
eλ )F (OP T ).
Proof of Theorem 2
Proof. Let yi denote the number of nodes in OP T that have
been covered after the i-th round and ¯yi be its expectation.
Further, let qi denote the probability that an uncovered node
in OP T will be covered in the i-th round. Then the expec-
tation of yi is
¯yi = ¯yi−1 + qi,
(9)
For α% = K/V and confidence p, we are pretty sure that
one of the K nodes will fall in ξp(α) with probability at least
p. Hence, the probability of sampling an uncovered node of
OP T in round i is
qi ≥ p
K − ¯yi−1
K
.
Substituting the above equation into Eq. (9), we get
¯yi ≥ ¯yi−1 + p
K − ¯yi−1
K
.
Iteratively, it can be written in the following form,
¯yi ≥ (1 −
p
K
)i ¯y0 + K[1 − (1 −
p
K
)i]
Since ¯y0 = 0, we get
¯yi ≥ K[1 − (1 −
p
K
)i].
After the K-th round, we conclude
E[r] =
¯yK
K
≥ 1 − (1 −
p
K
)K ≥ 1 −
1
ep .
|
1605.07915 | 2 | 1605 | 2017-06-12T14:05:19 | Cross-validation estimate of the number of clusters in a network | [
"cs.SI",
"physics.soc-ph"
] | Network science investigates methodologies that summarise relational data to obtain better interpretability. Identifying modular structures is a fundamental task, and assessment of the coarse-grain level is its crucial step. Here, we propose principled, scalable, and widely applicable assessment criteria to determine the number of clusters in modular networks based on the leave-one-out cross-validation estimate of the edge prediction error. | cs.SI | cs |
Cross-validation estimate of the number of clusters in a network
Tatsuro Kawamoto∗1 and Yoshiyuki Kabashima2
1Artificial Intelligence Research Center, National Institute of Advanced
Industrial Science and Technology, 2-3-26 Aomi, Koto-ku, Tokyo, Japan
2Department of Mathematical and Computing Science, Tokyo Institute of Technology,
4259-G5-22, Nagatsuta-cho, Midori-ku, Yokohama, Kanagawa 226-8502, Japan
(Dated: June 15, 2021)
Network science investigates methodologies that summarise relational data to obtain better inter-
pretability. Identifying modular structures is a fundamental task, and assessment of the coarse-grain
level is its crucial step. Here, we propose principled, scalable, and widely applicable assessment
criteria to determine the number of clusters in modular networks based on the leave-one-out cross-
validation estimate of the edge prediction error.
INTRODUCTION
Mathematical tools for graph or network analysis have wide applicability in various disciplines of science. In fact,
many datasets, e.g., biological, information, and social datasets, represent interactions or relations among elements
and have been successfully studied as networks [1, 2] using the approaches of machine learning, computer science, and
statistical physics. In a broad sense, a major goal is to identify macroscopic structures, including temporal structures,
that are hidden in the data. To accomplish this goal, for example, degree sequences, community and core -- periphery
structures, and various centralities have been extensively studied. Here, we focus on identifying modular structures,
namely, graph clustering [3, 4]. Popular modular structures are the community structure (assortative structure) and
disassortative structure [5 -- 8], although any structure that has a macroscopic law of connectivity can be regarded as
a modular structure. The Bayesian approach using the stochastic block model [9], which we will describe later, is a
powerful tool for graph clustering. In general, graph clustering consists of two steps: selecting the number of clusters
and determining the cluster assignment of each vertex. These steps can be performed repeatedly. Some methods
require the number of clusters to be an input, whereas other methods determine it automatically. The former step is
called model selection in statistical frameworks, and this step is our major focus.
The selection of the number of clusters is not an obvious task. For example, as shown in Fig. 1, more complex
structures can be resolved by using a model that allows a larger number of clusters. However, because the whole
purpose of graph clustering is to coarse grain the graph, the partition with excessively high resolution is not desirable.
This issue is also related to the level of resolution that is required in practice. The role of the model selection
frameworks and criteria is to suggest a possible candidate, or candidates, for the numbers of clusters to be analysed.
Which framework and criterion to use for model selection and its assessment is an inevitable problem that all
practitioners face, as multiple candidates have been proposed in the literature. A classical prescription is to optimise
an objective function that defines the strength of a modular structure, such as the modularity [10, 11] and the map
equation [12, 13]; combinatorial optimisation among all of the possible partitions yields both the number of clusters
and the cluster assignments. Although it might seem to be unrelated to statistical frameworks, it should be noted that
modularity maximisation has been shown to be equivalent to likelihood maximisation of a certain type of stochastic
block model [11, 14]. One can also use the spectral method and count the number of eigenvalues outside of the spectral
band [15 -- 17]; the reason for this prescription is the following: while the spectral band stems from the random nature
of the network, the isolated eigenvalues and corresponding eigenvectors possess distinct structural information about
the network. As another agnostic approach, an algorithm [18] that has a theoretical guarantee of the consistency for
the stochastic block model in sparse regime with sufficiently large average degree was recently proposed; a network is
called sparse when its average degree is kept constant as the size of the network grows to infinity, and typically, it is
locally tree-like. Finally, in the Bayesian framework, a commonly used principle is to select a model that maximises
the model's posterior probability [19 -- 23] or the model that have the minimum description length [24 -- 26], which leads
to Bayesian Information Criterion (BIC)-like criteria.
Minimisation of the prediction error is also a well-accepted principle for model selection, and cross-validation
estimates it adequately [27, 28]. This approach has been applied to a number of statistical models, including those
models that have hidden variables [29, 30]. Although the cross-validation model assessments for the stochastic block
model are also considered in [31 -- 33], they are either not of the Bayesian framework or performed by a brute-force
approach. A notable advantage of performing model selection using prediction error is that the assumed model is not
required to be consistent with the generative model of the actual data, whereas the penalty term in the BIC is derived
q = 2
q = 4
q = 6
2
q = 3
q = 5
Lower complexity
Higher complexity
FIG. 1. Partitions of the network of books about US politics. The inferred cluster assignments for various input
numbers of clusters q. The vertices with the same colour belong to the same cluster. The partition with large q has a higher
resolution and is obtained by fitting a model that has a higher complexity.
by explicitly using the assumption of model consistency. In this study, we propose an efficient cross-validation model
assessment in the Bayesian framework using the belief propagation (BP)-based algorithm.
Generally speaking, the prediction error can be used as a measure for quantifying the generalisation ability of the
model for unobserved data. In a data-rich situation, we can split the dataset into two parts, the training and test
sets, where the training set is used to learn the model parameters and the test set is used to measure the model's
prediction accuracy. We conduct this process for models that have different complexities, and we select the model
that has the least prediction error or the model that is the most parsimonious if the least prediction error is shared by
multiple models. In practice, however, the dataset is often insufficient to be used to conduct this process reliably. The
cross-validation estimate is a way to overcome such a data-scarce situation. In K-fold cross-validation, for example,
we randomly split the dataset into K (> 1) subsets, keep one of them as a test set while the remaining subsets are
used as the training sets, and measure the prediction error. Then, we switch the role of the test set to the training
set, pick one of the training sets as a new test set, and measure the prediction error again. We repeat this process
until every subset is used as a test set; then, we repeat the whole process for different random K-splits. The average
over all the cases is the final estimate of the prediction error of a given model's complexity.
For a dataset with N elements, the N -fold cross-validation is called the leave-one-out cross-validation (LOOCV);
hereafter, we focus on this approach. Note that the method of splitting is unique in the LOOCV. In the context of a
network, the dataset to be split is the set of edges and non-edges, and the data-splitting procedure is illustrated in
Fig. 2: for each pair of vertices, we learn the parameters while using the data without the information of edge existence
between the vertex pair; then, we measure whether we can correctly predict the existence or non-existence of an edge.
At first glance, this approach appears to be inefficient and redundant because we perform the parameter learning
N (N − 1)/2 times for similar training sets, which is true when the LOOCV is implemented in a straightforward
manner. Nevertheless, we show that such a redundant process is not necessary and that the prediction error can be
estimated efficiently using a BP-based algorithm. It is possible to extend the LOOCV estimate that we propose to
the K-fold cross-validation. As we mention in the Methods section, however, such an estimate has both conceptual
and computational issues, i.e., the LOOCV is an exceptionally convenient case.
The BP-based algorithm that we use to infer the cluster assignment was introduced by Decelle et al.
[21, 34].
This algorithm is scalable and performs well for sparse networks; this is favourable, because real-world networks are
typically sparse. Indeed, it is known that the BP-algorithm asymptotically achieves the information-theoretic bound
[35, 36] (when the number of clusters is 2) in the sense of accuracy when the network is actually generated from the
assumed model, the stochastic block model. In the original algorithm, the number of clusters q∗ is determined by
the Bethe free energy, which is equivalent to the approximate negative marginalised log-likelihood; among the models
of different (maximum) numbers of clusters, the most parsimonious model with low Bethe free energy is selected.
3
a
b
i
j
dataset
?
j
i
edge
training set
test set
?
j
i
non-edge
training set
test set
FIG. 2. The LOOCV error estimate of the network data. (a) The original dataset is given as the adjacency matrix A
of a network. (b) In each prediction, we hide a single piece of edge or non-edge information from A: this unobserved edge or
unobserved non-edge is the test set, and the adjacency matrix A\(i,j) without the information on that edge or non-edge is the
training set.
Despite its plausibility, however, it is known that this prescription performs poorly in practice. One of the reasons is
that the estimate using the Bethe free energy relies excessively on the assumption that the network is generated from
the stochastic block model, which is almost always not precisely correct. We show that this issue can be substantially
improved by evaluating cross-validation errors instead of the Bethe free energy while keeping the other parts of the
algorithm the same. With this improvement, we can conclude that the BP-based algorithm is not only an excellent
tool in the ideal case but is also useful in practice.
We denote the set of vertices and edges in the networks as V and E, and their cardinalities as N and L, respectively.
Throughout the current study, we consider undirected sparse networks, and we ignore multi-edges and self-loops for
simplicity.
RESULTS
Stochastic block model
Let us first explain how an instance of the stochastic block model is generated. The stochastic block model is
a random graph model that has a planted modular structure. The fraction of vertices that belong to cluster σ is
specified by γσ; accordingly, each vertex i has its planted cluster assignment σi ∈ {1, . . . , q∗}, where q∗ is the number
of clusters. The set of edges E is generated by connecting pairs of vertices independently and randomly according to
their cluster assignments. In other words, vertices i and j with cluster assignments σi = σ and σj = σ′ are connected
with probability ωσσ′ ; the matrix ω that specifies the connection probabilities within and between the clusters is
called the affinity matrix. Note that the affinity matrix ω is of O(N −1) when the network is sparse.
As the output, we obtain the adjacency matrix A. While the generation of network instances is a forward problem,
graph clustering is its inverse problem. In other words, we infer the cluster assignments of vertices σ or their probability
distributions as we learn the parameters (γ, ω), given the adjacency matrix A and the input (maximum) number of
4
clusters q. After we have obtained the results for various values of q, we select q∗. Note here that the input value q
is the maximum number of clusters allowed and that the actual number of clusters that appears for a given q can be
less than q.
Cross-validation errors
We consider four types of cross-validation errors. All of the errors are calculated using the results of inferences based
on the stochastic block model. We denote A\(i,j) as the adjacency matrix of a network in which Aij is unobserved,
i.e., in which it is unknown whether Aij = 0 or Aij = 1. We hereafter generally denote p(XY ) as the probability of
X that is conditioned on Y or the likelihood of Y when X is observed.
The process of edge prediction is two-fold: first, we estimate the cluster assignments of a vertex pair; then, we
predict whether an edge exists. Thus, the posterior predictive distribution p(Aij = 1A\(i,j)) of the model in which
the vertices i and j are connected given dataset A\(i,j), or the marginal likelihood of the learned model for the vertex
pair, is the following:
p(Aij = 1A\(i,j)) = Xσi,σj
p(Aij = 1σi, σj )p(σi, σjA\(i,j))
= hp(Aij = 1σi, σj)iA\(i,j) ,
(1)
where h· · ·iA\(i,j) is the average over p(σi, σjA\(i,j)) and we have omitted some of the conditioned parameters. The
error should be small when the prediction of edge existence for every vertex pair is accurate. In other words, the
probability distribution in which each element has probability of equation (1) is close, in the sense of the Kullback --
Leibler (KL) divergence, to the actual distribution, which is given as the empirical distribution. Therefore, it is
natural to employ, as a measure of the prediction error, the cross-entropy error function [37]
EBayes(q) = −log p(Aij A\(i,j))
= −
1
LXi<j(cid:20)Aij log p(Aij = 1A\(i,j))
+ (1 − Aij) log p(Aij = 0A\(i,j))(cid:21),
(2)
where X ≡Pi<j X(Aij)/L. Note that we have chosen the normalisation in such a way that EBayes is typically O(1)
in sparse networks. We refer to equation (2) as the Bayes prediction error, which corresponds to the LOOCV estimate
of the stochastic complexity [38]. As long as the model that we use is consistent with the data, the posterior predictive
distribution is optimal as an element of the prediction error because the intermediate dependence (σi, σj) is fully
marginalised and gives the smallest error.
Unfortunately, the assumption that the model that we use is consistent with the data is often invalid in practice.
In that case, the Bayes prediction error EBayes may no longer be optimal for prediction. In equation (2), we employed
− log p(Aij A\(i,j)) as the error of a vertex pair. Instead, we can consider the log-likelihood of cluster assignments
− log p(Aij σi, σj ) to be a fundamental element and measure h− log p(Aij σi, σj )iA\(i,j) as the prediction error of a
vertex pair. In other words, the cluster assignments (σi, σj) are drawn from the posterior distribution, and the error
of the vertex pair is measured with respect to those fixed assignments. Then, the corresponding cross-entropy error
function is
EGibbs(q) = h− log p(Aij σi, σj )iA\(i,j) .
(3)
We refer to equation (3) as the Gibbs prediction error. If the probability distribution of the cluster assignments is
highly peaked, then EGibbs will be close to EBayes, and both will be relatively small if those assignments predict the
actual network well. Alternatively, we can measure the maximum a posteriori (MAP) estimate of equation (3). In
other words, instead of taking the average over p(σi, σjA\(i,j)), we select the most likely assignments to measure the
error. We refer to EMAP(q) as the prediction error.
Finally, we define the Gibbs training error, which we refer to as Etraining. This error can be obtained by taking the
average over p(σi, σjA) instead in equation (3), i.e.,
Etraining(q) = h− log p(Aij σi, σj )iA.
(4)
a
political books
b
5
c
d
metadata
FIG. 3. Model assessments and inferred clusters in the network of books about US politics. (a) Bethe free
energy and (b) prediction and training errors as functions of the number of clusters q. The four data in (b) are, from top to
bottom, Gibbs prediction errors EGibbs (green triangles), MAP estimates EMAP of EGibbs (yellow squares), Bayes prediction
errors EBayes (red circles), and Gibbs training errors Etraining (blue diamonds). For each error, the constant term is neglected
and the standard error is shown in shadow. (c) The learned parameters, ω and γ, are visualised from q = 2 to 8. (d) The
cluster assignments that are indicated in the metadata of the dataset; the vertices with the same colour belong to the same
cluster.
Because we make use of the information regarding the edge existence when we take the average, the result is not a
prediction error, but is the goodness of fit of the assumed model to the given data.
At first glance, the cross-validation errors that we presented above might appear computationally infeasible because
we must know p(σi, σj A\(i,j)) with respect to every vertex pair. In the Methods section, however, we show that we
have analytical expressions of the cross-validation errors in terms of the outputs of BP; therefore, the model assessment
for the sparse networks is very efficient.
In the following subsections, we show the performances of these cross-validation errors for various networks relative
to the performance of the Bethe free energy.
Real-world networks
First, and most importantly, we show the performance of the Bethe free energy and cross-validation errors on
real-world networks. Figures 1 and 3 show the results on the network of books on US politics [39] (which we refer
to as political books). This network is a copurchase network whose vertices are books sold on Amazon.com, and each
edge represents a pair of books that was purchased by the same buyer; the metadata of the dataset has the labels
"conservative", "liberal", and "neutral."
Figure 3b shows that the cross-validation estimate of the Gibbs prediction error EGibbs saturates favourably, while
the Bethe free energy (Fig. 3a), Bayes prediction error EBayes, and Gibbs training error Etraining keep decreasing as
we increase the input number of clusters q. Although the MAP estimate of the Gibbs prediction error EMAP often
exhibits similar behaviour to the Gibbs prediction error EGibbs, we observe in the results of other datasets that it
does not appear to be distinctively superior to EGibbs.
karate club
network science
C. elegans
6
FIG. 4. Results of model assessments for various real-world networks. They are plotted in the same manner as
Figs. 3a and 3b.
Compared with the Gibbs prediction error EGibbs, the Bayes prediction error EBayes appears to be more sensitive
to the assumption that a network is generated by the stochastic block model, so that we rarely observe the clear
saturation of the LOOCV error. In a subsection below, we show how overfitting and underfitting occur by deriving
analytical expressions for the differences among the cross-validation errors.
How should we select the number of clusters q∗ from the obtained plots of cross-validation errors? Although this
problem is well defined when we select the best model, i.e., the model that has the least error, it is common to
select the most parsimonious model instead of the best model in practice. Then, how do we determine the "most
parsimonious model" from the results? This problem is not well defined, and there is no principled prescription
obtained by consensus.
However, there is an empirical rule called the "one-standard error" rule [27] that has often been used. Recall that
each cross-validation estimate is given as an average error per edge; thus, we can also measure its standard error. The
"one-standard error" rule suggests selecting the simplest model whose estimate is no more than one standard error
above the estimate of the best model. In the case of the political books network, the best model is the model in which
q = 6. Because the simplest model within the range of one standard error is q = 5, it is our final choice.
The actual partition for each value of q is shown in Figs. 1 and 3c. The cluster assignments indicated in the
metadata is presented in Fig. 3d for reference. In Fig. 3c, the learned values of the parameters ω and γ are visualised:
a higher value element of the affinity matrix ω is indicated in a darker colour, and the size of an element reflects the
fraction of the cluster size γσ. For q = 3, we can identify two communities and a cluster that is connected evenly to
those communities, as presumed in the metadata. For q = 5, in addition, the sets of core vertices in each community
are also detected. Note that recovering the labels in the metadata is not our goal. The metadata are not determined
based on the network structure and are not the ground-truth partition [54].
It should also be noted that we minimise the Bethe free energy in the cluster inference step in every case. When we
select the number of clusters q∗, i.e., in the model selection step, we propose to use the cross-validation errors instead
of the Bethe free energy.
The results of other networks are shown in Fig. 4. They are the friendship network of a karate club [40], coauthorship
network of scientists who work on networks [41] (see [42] for details of the datasets), and the metabolic network of
C. elegans [43]. We refer to those as karate club, network science, and C. elegans, respectively. The results are
qualitatively similar to the political books network. Note that the initial state dependency can be sensitive when the
input number of clusters q is large. Therefore, the results can be unstable in such a region.
From the results in Figs. 3 and 4, one might conclude that the Gibbs prediction error EGibbs is the only useful
criterion in practice. However, for example, when the holdout method is used instead of the LOOCV (see the Methods
section), it can be confirmed that the Bayes prediction error EBayes also behaves reasonably.
a
Political blogs
b
Political blogs
7
FIG. 5. Results of the model assessments for the network of hyperlinks between blogs on US politics. (a)
The results of the standard model and the learned parameters. (b) The results of the degree-corrected model and the learned
parameters.
The error for q = 1
For the karate club network in Fig. 4, we see that the errors keep decreasing except for the Gibbs prediction error
EGibbs. Recall that, for the stochastic block model with large q and low average degree to be detectable, the network is
required to exhibit sufficiently strong modular structure [21]. Thus, although we do not a priori know which prediction
error should be referred to, because the network has average degree smaller than 5, it is hard to believe that the errors
other than EGibbs are appropriate. The Gibbs prediction error EGibbs has the smallest error at q = 3, and the model
in which q = 2 has an error within the range of one standard error of q = 3. However, one might suspect that the
most parsimonious model is the model with q = 1, i.e., the model that we should assume is a uniform random graph,
and there is no statistically significant modular structure. In this case, the connection probability for an arbitrary
vertex pair is determined by a single parameter, i.e., p(Aij σiσj ) = ω, and it is simply ω = 2L/N (N − 1). Moreover,
there is no difference among the errors that we listed above. Hence,
EBayes = EGibbs = EMAP = Etraining ≈ 1 − log ω.
(5)
Here, we used the fact that ω = O(N −1) because we consider sparse networks. In every plot of cross-validation errors,
the constant and O(N −1) terms are neglected. The number of vertices and edges in the karate club network are
N = 34 and L = 78, respectively, i.e., − log ω ≃ 1.97, which is much larger than the errors for q = 2; we thus conclude
that the number of clusters q∗ = 2.
Degree-corrected stochastic block model
It has been noted that the standard stochastic block model that we explained above is often not suitable for
inferring clusters in real-world networks because many of them have a scale-free degree distribution, while the standard
stochastic block model is restricted to having a Poisson degree distribution. This inconsistency affects both the cluster
inference for a given q∗ and the model selection. For example, as shown in Fig. 5a, we found that all of the criteria
largely overfit for the political blogs network when the standard stochastic block model is used. The political blogs
network is a network of hyperlinks between blogs on US politics (we neglect the directions of the edges), which is
expected to have two clusters.
a
b
c
d
e
8
FIG. 6. Performance of the model assessment criteria on various stochastic block models. The results of (a) Bethe
free energy, (b) Bayes prediction error EBayes, (c) Gibbs prediction error EGibbs, (d) MAP estimate of the Gibbs prediction
error EMAP, and (e) Gibbs training error Etraining are shown. The lines in each plot represent the results of various values of ǫ.
The degree-corrected stochastic block model [44, 45] is often used as an alternative. This model has the degree-
correction parameter θi for each vertex in addition to γ and ω, where θ allows a cluster to have a heterogeneous
degree distribution. (See [44] for details of the model.) The cross-validation model assessment can be straightforwardly
extended to the degree-corrected stochastic block model: for the inference of cluster assignments, the BP algorithm
can be found in [46]; for the cross-validation errors, we only need to replace ωσi,σj for the probability p(Aij = 1σi, σj)
with θiωσi,σj θj in each case. As shown in Fig. 5b, the assessment is reasonable when a degree-corrected stochastic
block model is used. Although the error drops for q ≥ 5, it is better to discard this part because the numbers of
iterations until convergence become relatively large there; we regard it as the "wrong solution" that is mentioned in
Ref. [50].
When q = 1, we do not assume the Erdos-R´enyi random graph, but instead assume the random graph that has
the same expected degree sequence as the actual dataset [7], which is the model often used as the null model in
modularity. Thus, the probability that vertices i and j are connected is didj /2L. After some algebra, assuming that
didj/2L ≪ 1, the error for q = 1 is approximately
1 −
1
L
N
Xi=1
di log di + log(2L),
(6)
which is equal to equation (5) when the network is regular.
In the case of the political blogs network, the cross-
validation error for q = 1 is approximately 3.42. (It is 2.42 in the plot because the constant term is neglected.) Thus,
we obtain q∗ = 2.
Synthetic networks and the detectability threshold
It is important to investigate how well the cross-validation errors perform in a critical situation, i.e., the case in
which the network is close to the uniform random graph. To accomplish this goal, we observe the performance on
the stochastic block model, the model that we assume for the inference itself. As explained in the subsection of the
stochastic block model, the closeness to the uniform random graph is specified with the affinity matrix ω. Here, we
consider a simple community structure: we set ωin for the diagonal elements, and the remaining elements are set
to ωout, where ωin ≥ ωout. We parametrise the closeness to the uniform random graph, with ǫ = ωout/ωin; ǫ = 0
represents that the network consists of completely disconnected clusters, and ǫ = 1 represents the uniform random
graph. There is a phase transition point ǫ∗ above which it is statistically impossible to infer a planted structure. This
point is called the detectability threshold. Our interest here is to determine whether the planted number of clusters
q∗ can be correctly identified using the cross-validation errors when ǫ is set to be close to ǫ∗.
Figure 6 shows the Bethe free energy and cross-validation errors for the stochastic block models. In each plot, we
set q∗ = 4, each cluster has 1, 000 vertices, and the average degree c is set to 8. From bottom to top, the values of ǫ
are 0.1, 0.15, 0.2, and 0.25, while the detectability threshold is ǫ∗ ≃ 0.31 [21]. Note that when the value of ǫ is close
to ǫ∗, the inferred cluster assignments are barely correlated with the planted assignments; thus, the result is close to
a random guess anyway.
In all of the plots, the curves of validation become smoother as we increase the value of ǫ, which indicates that
it is difficult to identify the most parsimonious model. It is clear that the Bethe free energy and Bayes prediction
error EBayes perform better than the Gibbs prediction error EGibbs in the present case because the networks we
9
analyse correspond exactly to the model we assume. Figure 6 shows that the Gibbs prediction error EGibbs and its
MAP estimate EMAP underestimate q∗ near the detectability threshold. This finding is consistent with the results
that we obtained for the real-world networks that the Bethe free energy and Bayes prediction error EBayes tend to
overestimate. Indeed, under certain assumptions, we can derive that the Bayes prediction error EBayes identifies the
planted number of clusters all the way down to the detectability threshold, while the Gibbs prediction error EGibbs
strictly underfits near the detectability threshold. (See the Methods section for the derivation.) Therefore, there is a
trade-off between the Bayes prediction error EBayes and the Gibbs prediction error EGibbs; their superiority depends
on the accuracy of the assumption of the stochastic block model and the fuzziness of the network.
Relation among the cross-validation errors
We show how the model assessment criteria that we consider in this study are related. First, we derive the relation
among the errors EBayes, EGibbs, and Etraining. By exploiting Bayes' rule, we have
log p(Aij A\(i,j)) = log p(Aij σi, σj, A\(i,j))
+ log
p(σi, σjA\(i,j))
p(σi, σj A)
.
(7)
Note here that the left-hand side does not depend on σi and σj. If we take the average with respect to p(σi, σjA\(i,j))
on both sides,
log p(AijA\(i,j)) =Dlog p(Aijσi, σj , A\(i,j))E
+ DKL(cid:16)p(σi, σjA\(i,j)) p(σi, σj A)(cid:17) ,
where DKL(pq) is the KL divergence. Taking the sample average of the edges, we obtain
If we take the average over p(σi, σjA) in equation (8) instead,
EBayes = EGibbs − DKL(cid:0)p(σi, σjA\(i,j))p(σi, σjA)(cid:1).
Because the KL divergence is non-negative, we have
EBayes = Etraining + DKL(cid:0)p(σi, σjA)p(σi, σjA\(i,j))(cid:1).
Etraining ≤ EBayes ≤ EGibbs.
(8)
(9)
(10)
(11)
This inequality follows directly from Bayes' rule and applies broadly. Note, however, that the amounts of the errors
do not indicate the relationships among the numbers of clusters selected.
Under a natural assumption, we can also derive the inequality for the number of clusters selected. Let q be the trial
number of clusters. If the cluster assignment distributions for different values of q constitute a hierarchical structure,
i.e., a result with a small q can be regarded as the coarse graining of a result with a larger q, then the information
monotonicity [47, 48] of the KL divergence ensures that
training ≥ q∗
q∗
Bayes ≥ q∗
Gibbs.
(12)
In other words, the Bayes prediction error EBayes can only overfit when the Gibbs prediction error EGibbs estimates q∗
correctly, while the Gibbs prediction error EGibbs can only underfit when the Bayes prediction error EBayes estimates q∗
correctly. In addition, when the model predicts the edge existence relatively accurately, EMAP is biased in such a way
that the error becomes small. Therefore, EMAP tends to be smaller than EGibbs. As observed for real-world networks,
EGibbs typically performs well while EBayes often overfits in practice; equation (9) implies that detailed information
about the difference in the cluster assignment distribution is often not relevant and simply causes overfitting.
In addition to the relationship among the cross-validation errors, it is also fruitful to seek relationships between the
Bethe free energy and the prediction errors. We found that it is possible to express the Bethe free energy in terms of
the prediction errors. (See the Methods section for the derivation.)
DISCUSSION
10
To determine the number of clusters, we both proposed LOOCV estimates of prediction errors using BP and revealed
some of the theoretical relationships among the errors. They are principled and scalable and, as far as we examined,
perform well in practice. Unlike the BIC-like criteria, the prediction errors do not require the model consistency to
be applicable. Moreover, although we only treated the standard and degree-corrected stochastic block models, the
applicability of our LOOCV estimates is not limited to these models. With an appropriate choice of block models,
we expect that the present framework can also be used in the vast majority of real-world networks, such as directed,
weighted, and bipartite networks and networks that have positive and negative edges. This is in contrast to some
criteria that are limited to specific modular structures.
The code for the results of the current study can be found at [49]. (See the description therein for more details of
the algorithm.)
The selection of the number of clusters q∗ can sometimes be subtle. For example, as we briefly mentioned above,
when the inference algorithm depends sensitively on the initial condition, the results of the LOOCV can be unstable; all
of the algorithms share this type of difficulty as long as the non-convex optimisation problem is considered. Sometimes
the LOOCV curve can be bumpy in such a way that it is difficult to determine the most parsimonious model. For
this problem, we note that there is much information other than the prediction errors that help us to determine the
number of clusters q∗, and we should take them into account. For example, we can stop the assessment or discard
the result when (i) the number of iterations until convergence becomes large [50], as we have observed in Fig. 5b, (ii)
the resulting partition does not exhibit a significant behaviour, e.g., no clear pattern in the affinity matrix, or (iii)
the actual number of clusters does not increase as q increases. All of this information is available in the output of our
code.
It is possible that the difficulty of selecting the number of clusters q∗ arises in the model itself. Fitting with the
stochastic block models is flexible; thus, the algorithm can infer not only the assortative and disassortative structures
but also more complex structures. However, the flexibility also indicates that slightly different models might be able
to fit the data as good as the most parsimonious model. As a result, we can obtain a gradually decreasing LOOCV
curve for a broad range of q. Thus, there should be a trade-off between the flexibility of the model and the difficulty
of the model assessment.
Despite the fact that model assessment based on prediction accuracy is a well-accepted principle, to the best of our
knowledge, the method that is applicable to large-scale modular networks in the Bayesian framework is discussed in
this study for the first time. For better accuracy and scalability, we believe that there is more to be done. One might
investigate whether the LOOCV errors perform, in some sense, better than other BIC-like criteria. The relations
and tendencies of the LOOCV errors compared to other model assessment criteria would certainly provide interesting
future work even though it is unlikely that we can generally conclude the superiority of the criteria over the others
[28]. In practice, if resource allocation allows, it is always better to evaluate multiple criteria [51].
METHODS
Inference of the cluster assignments
Let us first explain the inference algorithm of the cluster assignment using the stochastic block model when the num-
ber of clusters q is given. For this part, we follow the algorithm introduced in [21, 34]: the expectation -- maximisation
(EM) algorithm using BP.
An instance of the stochastic block model is generated based on selected parameters, (γ, ω), as explained in the
subsection of the stochastic block model. Because the cluster assignment of each vertex and the edge generation
between each vertex pair are determined independently and randomly in the stochastic block model, its likelihood is
given by
p(A, σγ, ω, q) =
N
Yi=1
γσiYi<j
ωAij
σiσj (cid:0)1 − ωσiσj(cid:1)1−Aij .
(13)
The goal here is to minimise the free energy defined by f = −N −1 logPσ′ p(A, σ′γ, ω), which is equivalent to
maximising the marginal log-likelihood. Using the identity p(A, σγ, ω) = p(σA, γ, ω)p(Aγ, ω), the free energy f
can be expressed as
f = −
1
N
log
p(A, σγ, ω)
p(σA, γ, ω)
11
(14)
for an arbitrary σ. Taking the average over a probability distribution q(σ) on both sides, we obtain the following
variational expression.
f = −
q(σ) log
p(A, σγ, ω)
q(σ)
q(σ)
p(σA, γ, ω)
1
N Xσ
N"−Xσ
1
=
q(σ) log p(A, σγ, ω) +Xσ
− DKL (q(σ)p(σA, γ, ω))#.
q(σ) log q(σ)
(15)
When q(σ) is equal to the posterior distribution p(σA, γ, ω), the KL divergence disappears, and we should obtain the
minimum. We can interpret the first and second terms as corresponding to the internal energy and negative entropy,
respectively, and equation (15) as the thermodynamic relation of the free energy. For the first term, by substituting
equation (13), we have
−
q(σ) log p(A, σγ, ω)
1
N Xσ
N Xσ
1
= −
q(σ)"Xi
log γσi +Xi<j(cid:18)Aij log ωσiσj
+ (1 − Aij ) log(1 − ωσiσj )(cid:19)#
= −
1
N Xi Xσ
qi
σ log γσ −
1
N X(i,j)∈EXσσ′
qij
σσ′ log ωσσ′
+
c
2
+ O(N −1),
(16)
σ = hδσσi iσ and qij
where qi
h· · ·iσ is the average over q(σ). Therefore, the free energy minimisation in the variational expression requires two
operations: the inference of the marginal probabilities of the cluster assignments, qi
σσ′ , and the learning of the
parameters, (γ, ω). In the EM algorithm, they are performed iteratively.
σσ′ = (cid:10)δσ,σiδσ′σj(cid:11)σ are the marginal probabilities, where δxy is the Kronecker delta and
σ and qij
For the E-step of the EM algorithm, we infer the marginal probabilities of the cluster assignments using the current
estimates of the parameters. Whereas the exact marginalisation is computationally expensive in general, BP yields an
accurate estimate quite efficiently when the network is sparse, i.e., the case in which the network is locally tree-like.
Here, we do not go over the complete derivation of the BP equations, which can be found in [21, 52]. We denote ψi
σ
as the BP estimate of qi
σ, which is obtained by
ψi
σ =
1
Z i γσe−hσ Yk∈∂i Xσk
ψk→i
σk ωσkσ! ,
(17)
where Z i is the normalisation factor with respect to the cluster assignment σ, hσ = PN
σk ωσkσ, which is
due to the effect of non-edges (i, k) /∈ E, and ∂i indicates the set of neighbouring vertices of i. In equation (17), we
also have the so-called cavity bias ψi→j
for an edge (i, j) ∈ E; the cavity bias is the marginal probability of vertex i
without the marginalisation from vertex j. Analogously to equation (17), the update equation for the cavity bias is
obtained by
k=1Pσk
ψk
σ
ψi→j
σ =
1
Z i→j γσe−hσ Yk∈∂i\j Xσk
ψk→i
σk ωσkσ! ,
(18)
where ∂i\j indicates the set of neighbouring vertices of i except for j, and Z i→j is the normalisation factor. For the
BP estimate ψij
σσ′ of the two-point marginal qij
σσ′ for (i, j) ∈ E, we have
12
ψij
σσ′ =
1
Z ij ψi→j
σ ωσσ′ ψj→i
σ′
,
(19)
where Z ij is the normalisation factor with respect to the assignments σ and σ′.
With these marginals in hand, in the M-step of the EM algorithm, we update the estimate of the parameters to γ
and ω as
γσ =
1
N
ωσσ′ =
N
ψi
σ,
Xi=1
N 2γσγσ′ X(i,j)∈E
1
ωσσ′ (ψi→j
σ ψj→i
σ′ + ψj→i
Z ij
σ ψi→j
σ′
)
.
(20)
(21)
Here, we have used the fact that the network is undirected. These update rules can be obtained by simply taking the
derivatives with respect to the parameters in equation (16) with the normalisation constraint Pσ γσ = 1.
We recursively compute equation (18) and the parameter learning, equations (20) and (21), until convergence; then,
we obtain the full marginal using equation (17), which yields the estimates of the cluster assignments of the vertices.
Bethe free energy and derivation of the cross-validation errors
In the algorithm above, because we use BP to estimate the marginal probabilities, we are no longer minimising the
free energy. Instead, as an approximation of the free energy, we minimise the Bethe free energy, which is expressed as
fBethe = −
log Z i +
1
N Xi
1
N X(i,j)∈E
log Z ij −
c
2
,
(22)
where Z i and Z ij are the normalisation factors that appeared in equations (17) and (19). As mentioned in the
Introduction section, to determine the number of clusters q∗ from the Bethe free energy fBethe, we select the most
parsimonious model among the models that have low fBethe. This approach corresponds to taking the maximum
likelihood estimation of the parameters.
We now explain the derivations of the cross-validation errors that we used in the Results section. The LOOCV
estimate of the Bayes prediction error EBayes is measured by equation (2), and its element p(Aij = 1A\(i,j)) is
given as equation (1). The first factor in the sum of equation (1) is simply p(Aij = 1σi, σj) = ωσiσj , by the model
definition. An important observation is that because the cavity bias ψi→j
represents the marginal probability of
vertex i without the information from vertex j, this entity is exactly what we need for prediction in the LOOCV,
σj /Z ij, where
i.e., p(σi, σjA\(i,j)) = ψi→j
Z ij is defined in equation (19). Thus, we have p(Aij = 1A\(i,j)) = Z ij. By using the fact that L = O(N ) and
p(Aij = 1A\(i,j)) = O(N −1), EBayes(q) can be written as
. Additionally, recall that p(σi, σj Aij = 1, A\(i,j)) = ψi→j
σi ωσiσj ψj→i
σi ψj→i
σj
σi
EBayes(q) = 1 −
1
L X(i,j)∈E
log Z ij + O(N −1).
(23)
Equation (23) indicates that the prediction with respect to the non-edges contributes only a constant; thus, EBayes(q)
essentially measures whether the existence of the edges is correctly predicted in a sparse network.
The Gibbs prediction error EGibbs in equation (3) can be obtained in the same manner. Using the approximation
that the network is sparse, it can be written in terms of the cavity biases as
EGibbs(q) ≃ 1 −
1
L X(i,j)∈E Xσi,σj
σi ψj→i
ψi→j
σj
log ωσiσj .
(24)
The (MAP) estimate of the Gibbs prediction error EMAP(q) can be obtained by replacing ψi→j
in equation (24).
σi with δσi,argmax{ψi→j
σ
}
13
For the Gibbs training error Etraining in equation (4), we take the average over p(σi, σjA) instead of p(σi, σjA\(i,j)).
Thus, we have
Etraining(q) ≃ 1 −
1
L X(i,j)∈E Xσi,σj
σi ωσiσj ψj→i
ψi→j
σj
Z ij
log ωσiσj .
(25)
This training error can be interpreted as a part of the internal energy because it corresponds to the BP estimate of
the second term in equation (16).
Note that all of the cross-validation errors, equations (23) -- (25), are analytical expressions in terms of the parameters
and cavity biases. Therefore, we can readily measure those errors by simply running the algorithm once. It should
also be noted that using BP for the LOOCV itself is not totally new; this idea has been addressed in a different
context in the literature, e.g., Ref. [53].
Detectability of the Bayes and Gibbs prediction errors
Let us evaluate the values of the Bayes and Gibbs prediction errors, EBayes and EGibbs, for the stochastic block
model with the simple community structure, i.e., ωσσ′ = (ωin − ωout)δσσ′ + ωout. By substituting this affinity matrix
to equations (23) and (24), we obtain
EBayes(q) = 1 − log ωout
−
1
L X(i,j)∈E
EGibbs(q) = 1 − log ωout
log 1 +
ωin − ωout
ωout
ψi→j
σ ψj→i
σ !
q
Xσ=1
−(cid:18)log
ωin
ωout(cid:19) 1
L X(i,j)∈E
ψi→j
σ ψj→i
σ
.
q
Xσ=1
(26)
(27)
As we sweep the parameter ǫ = ωout/ωin → 1 in the stochastic block model that has 2 clusters, q = 2 is selected as
long as the prediction error for q = 2 is less than the error for q = 1. When the number of vertices N is sufficiently
large, ω = ωout + (ωin − ωout)P2
σ in equation (5). Thus,
σ=1 γ2
EBayes(q = 1) = EGibbs(q = 1)
= 1 − log ωout − log 1 +
ωin − ωout
ωout
γ2
σ! ,
(28)
2
Xσ=1
where we neglected the O(N −1) contribution. Let us presume that the cavity bias ψi→j
structure up to the detectability threshold ǫ∗ and ψi→j
of 2 clusters of equal size. We then have
is correlated with the planted
σ = γσ for ǫ ≥ ǫ∗, which is rigorously proven [35, 36] in the case
σ
EBayes(q = 2) = EBayes(q = 1)
(29)
in the undetectable region. In the case of equal size clusters, EBayes(q = 2) is strictly smaller thatn EBayes(q = 1) in
the detectable region because the last term in (26) is minimised when the distribution is uniform. In contrast, when
ψi→j
σ = γσ, the inequality log(1 + xy) > y log(1 + x) (x > 0, 0 < y < 1) ensures that
EGibbs(q = 2) > EGibbs(q = 1).
(30)
Therefore, assuming that the parameters γ and ω are learned accurately, the Bayes prediction error EBayes suggests
q∗ = 2 all the way down to the detectability threshold, while the Gibbs prediction error EGibbs strictly underfits near
the detectability threshold.
The Bethe free energy in terms of the prediction errors
We can also express the Bethe free energy fBethe in terms of the prediction errors. Ignoring the constant terms and
factors, we observe that the Bayes prediction error EBayes, equation (23), is one component of the Bethe free energy,
equation (22). The remaining part is the contribution of −Pi log Z i that arises as follows.
14
Let us consider hiding a vertex and the edges and non-edges that are incident to that vertex, instead of an edge.
We denote {Aij}j∈V as the set of adjacency matrix elements that are incident to vertex i, and A\i as the adjacency
matrix in which the information of {Aij}j∈V is unobserved. For the prediction of an edge's existence, for example,
because the unobserved vertex receives no information about its cluster assignment from its neighbours, we have
p(Aij = 1A\i) = Xσiσj
= Xσiσj
p(Aij = 1σiσj )p(σiσj A\i)
ωσiσj γσi ψj→i
σj
.
(31)
When we consider the prediction of the set of edges {Aij }j∈V , we find that all of the vertex pairs share the same
cluster assignment for vertex i. Therefore, using the approximation that the network is sparse, we have
p(σiA\i)
p({Aij}j∈V A\i) =Xσi
× Yj∈∂i
Xσj
× Yk /∈∂i Xσk
γσi e−hσi Yj∈∂i
≈Xσi
= Z i.
p(Aij = 1σiσj)p(σj A\i)
p(Aik = 0σiσk)p(σkA\i)!
Xσj
ψj→i
σj ωσj σi
(32)
Again, due to the sparsity, the prediction for the non-edges happened to be neglected, and it is essentially the
prediction of the edges. Analogously to equation (2), we can consider a leave-one-vertex-out version of the Bayes
prediction error, which we refer to as Ev
Bayes, as follows.
1
LXij (cid:20)Aij log p(Aij = 1A\i)
+ (1 − Aij ) log p(Aij = 0A\i)(cid:21)
log
Yj∈∂i
p(Aij = 1A\i) Yj /∈∂i
log p({Aij }j∈V A\i)
p(Aij = 0A\i)
Ev
Bayes(q) = −
= −
= −
= −
1
1
LXi
LXi
LXi
1
log Z i.
(33)
Again, the edges and non-edges that are incident to the unobserved vertex i share the same cluster assignment σi at
one end. In this definition, however, the prediction of an edge's existence is made twice for every edge. If the Bayes
prediction error EBayes is subtracted to prevent this over-counting, it is exactly the Bethe free energy fBethe up to the
constant terms and the normalisation factor.
Derivation of inequality (12)
Recall that the Bayes prediction error EBayes, Gibbs prediction error EGibbs, and Gibbs training error Etraining
are related via equations (9) and (10). We select the number of clusters q as the point at which the error function
saturates (i.e., stops decreasing) with increasing q. For a smaller q to be selected by EBayes than by EGibbs, the
15
FIG. 7. Schematic of the shapes of the error functions. As long as the gap between two error functions is nondecreasing,
the error with the smaller value does not saturate earlier than the other.
gap between them, DKL(cid:0)p(σi, σjA\(i,j))p(σi, σj A)(cid:1), must decrease (see Fig. 7). In this subsection, we explain the
information monotonicity of the KL divergence and when it is applicable in the present context.
Let us consider sets of variables X = {X1, . . . , Xm} and x = {x1, . . . , xn} (n > m). We first set a pair of
probability distributions ¯P (X) and ¯Q(X) on X. Then we define a pair of probability distributions P (x) and Q(x)
on x as refinements of ¯P (X) and ¯Q(X), respectively, if there exists a family of sets {xµ}m
µ=1 that is a partition of x,
i.e., xµ ∩ xµ′
= ∅ for µ 6= µ′ and ∪µxµ = x, that satisfies P (xµ) = ¯P (Xµ) and Q(xµ) = ¯Q(Xµ) for any µ. In other
words, X can be regarded as the coarse graining of x. An example is given in Fig. 8. Note, however, that if X is
actually constructed as the coarse graining of x, the above condition trivially holds. In general, a family that satisfies
the above condition might not exist; even if it exists, it might not be unique.
The information monotonicity of the KL divergence states that for P (x) and Q(x), which are the refinements of
¯P (X) and ¯Q(X), respectively, we have
DKL(P Q) ≥ DKL( ¯P ¯Q),
(34)
which is natural, because the difference between the distributions is more visible at a finer resolution. Equation (34)
is deduced by the convexity of the KL divergence. First, we can rewrite the right-hand side of equation (34) in terms
of P and Q as follows:
Thus, if
DKL( ¯P ¯Q) =
P (xµ) log
P (xµ)
Q(xµ)
;
m
Xµ=1
P (x) log
P (x)
Q(x)
≥ P (xµ) log
P (xµ)
Q(xµ)
Xx∈xµ
(35)
(36)
holds for an arbitrary µ, then equation (34) holds. We split xµ into x1 ∈ xµ and xµ\x1, and we denote the
corresponding probabilities as P1 := P (x1), Q1 := Q(x1), P c
1 := Q(xµ\x1). The right-hand
side of equation (36) is then
1 := P (xµ\x1), and Qc
(P1 + P c
1
Q1 + Qc
1 ) log(cid:18) P1 + P c
1(cid:19)
1 ) log(cid:18) P1
P1 + P c
1
= − (P1 + P c
Q1
P1
+
P c
1
P1 + P c
1
Qc
1
P c
1(cid:19)
≤ P1 log
+ P c
1 log
P1
Q1
P c
1
Qc
1
,
(37)
where we used the convexity of the logarithmic function. By repeating the same argument for the second term of
equation (37), we obtain equation (36). Although the KL divergence is our focus, the information monotonicity holds
more generally, e.g., for f -divergence [48].
We now use the information monotonicity to estimate the error functions.
In the present context, the sets of
variables X and x correspond to the cluster assignments of different q's, (σi, σj) with q and (σ′
j) with q′ (q′ > q),
for a vertex pair i and j. Because the labelling of the clusters is common to all of the vertices, we require that the
refinement condition is satisfied with the common family of sets for every vertex pair. Under this condition, the KL
i, σ′
16
FIG. 8. Example of the refinement of the probability distributions. We can regard P (x) and Q(x) as the refinements
of ¯P (X) and ¯Q(X), respectively, with x1 = {x1, x2}, x2 = {x3}, x3 = {x4, x5, x6}, x4 = {x7}, and x5 = {x8, x9} as a
possible correspondence. Note that the correspondence is not unique. If we refer only to Q(x) and ¯Q(X), the assignment of
{x1, x2}, {x3}, and {x8, x9} is exchangeable within X1, X2, and X5. However, ¯P (X5) does not coincide with P ({x1, x2}) or
P (x3); therefore, only X1 and X2 are exchangeable between {x1, x2} and {x3}. The same argument applies to {x4, x5, x6} and
{x8, x9}.
divergence is nondecreasing as a function of q, which indicates that EBayes does not saturate earlier than EGibbs.
Similarly, Etraining does not saturate earlier than EBayes. Although the refinement condition that we required above is
too strict to be satisfied exactly in a numerical calculation, it is what we naturally expect when the algorithm detects
a hierarchical structure or the same structure with excess numbers of clusters. Moreover, the argument above is only
a sufficient condition. Therefore, we naturally expect that EGibbs suggests a smaller number of clusters than EBayes
and Etraining quite commonly in practice. In addition, note that if we use a different criterion for the selection of q,
e.g., the variation of the slope of the error function, the above conclusion can be violated.
Holdout method and K-fold cross-validation
Other than the LOOCV, it is also possible to measure the prediction errors by the holdout method and the K-fold
cross-validation using BP. They can be conducted by randomly selecting the set of edges to be held out (i.e., the
holdout set) and running BP that ignores the cavity biases sent by the held-out edges. In the holdout method, unlike
the K-fold cross-validation, the test set is never used as the training set and vice versa. The results are listed in Fig. 9.
We observe that the overfitting of the Bayes prediction error EBayes appears to be prevented. Although the resulting
behaviours are somewhat reasonable, we have to bear in mind that they have following conceptual and computational
issues.
First, we cannot expect the cross-validation assessment to work at all when the holdout set is too large, or equiv-
alently, when the training set is too small. It is because, when a significant fraction of edges that are connected to
vertices in certain clusters are held out, it is impossible to learn the underlying block structure correctly [31]. Second,
when the holdout set has more than one edge, we actually need to run the whole algorithm twice for each holdout
set; first to compute the posterior distribution given the training set, and then, to evaluate the predictive distribution
for the holdout set based on the obtained posterior distribution. Then, we need to repeat this process as many times
as the number of holdout sets. Again, none of the above processes are required for the LOOCV, because we have
the analytical expression. Thus, the computational cost for the holdout method and the K-fold cross-validation are
orders of magnitude larger than the LOOCV.
Besides the above issues, a more delicate treatment is required in the K-fold cross-validation. When the unobserved
edges are connected, a prediction error is not given as an independent sum of the error per edge, and we need the
calculation as we have done in equation (32). Therefore, the precise measurement is even more time consuming. In
Fig. 9, we imposed the constraint that the edges are not connected in the holdout method. For the K-fold cross-
validation, we simply ignored the fact that the unobserved edges might be connected, although this discrepancy may
not be negligible.
We lose all the advantages that we had in the LOOCV when the holdout method and the K-fold cross-validation
are considered. However, when the computational efficiency can be compromised, the holdout method with small
holdout set can be a promising alternative.
a
political books
holdout method
10-fold cross-validation
17
b
holdout method
10-fold cross-validation
network science
FIG. 9. Cross-validation errors using the holdout method and the 10-fold cross-validation. As in Fig. 3, the four
data in each plot indicate the Gibbs prediction errors EGibbs (green triangles), MAP estimates EMAP of EGibbs (yellow squares),
Bayes prediction errors EBayes (red circles), and Gibbs training errors Etraining (blue diamonds). In the holdout method, we
randomly selected 1% of the edges (5 edges in (a) political books and 10 edges in (b) network science) as the holdout set. The
solid lines represent the average values obtained by repeating the prediction process 10 times and the shadows represent their
standard errors. For the K-fold cross-validation, we set K = 10. Although it is common to take average over different choices
of K partition, we omitted this averaging process.
Justification of the LOOCV estimates
It is a common mistake that one will make use of the information of the test set (the unobserved edge in the
present context) for the training of the model in the process of cross-validation. Precisely speaking, however, the
information of the unobserved edge does contribute to the training of the model in our procedure; thus, the cross-
validation estimate here is not correct in a rigorous sense. Nevertheless, the treatment that we propose is justified
because the contribution of the unobserved edge in the present setting is negligible. First, when we update the
distribution of the cluster assignment ψi→j
, we make use of the information of the unobserved edge when there are
loops in the network. Because we consider the sparse network, which is locally tree-like, the effect of the loops is
typically negligible. Second, the update equations (20) and (21) of the parameters in the maximisation step of the
EM algorithm contain the unobserved edge. However, because we consider the LOOCV in which only a single vertex
pair is unobserved, this discrepancy results in a difference of only O(N −1) to the parameter estimation. Therefore,
our analytical results in terms of BP are sufficiently accurate to serve as error estimates when using the LOOCV.
σ
This work was supported by JSPS KAKENHI No. 26011023 (TK) and No. 25120013 (YK).
ACKNOWLEDGMENTS
[1] Barabsi, A.-L. Network Science (Cambridge University Press, 2016), 1 edn.
[2] Newman, M. E. J. Networks: An Introduction (Oxford university press, 2010).
[3] Fortunato, S. Community detection in graphs. Phys. Rep. 486, 75 -- 174 (2010).
[4] Leger, J.-B., Vacher, C. & Daudin, J.-J. Detection of structurally homogeneous subsets in graphs. Stat. Comput. 24,
675 -- 692 (2014).
18
[5] Girvan, M. & Newman, M. E. J. Community structure in social and biological networks. Proc. Natl. Acad. Sci. U.S.A.
99, 7821 -- 7826 (2002).
[6] Radicchi, F., Castellano, C., Cecconi, F., Loreto, V. & Parisi, D. Defining and identifying communities in networks. Proc.
Natl. Acad. Sci. U.S.A. 101, 2658 -- 63 (2004).
[7] Newman, M. E. J. Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E 74, 036104
(2006).
[8] Lancichinetti, A., Fortunato, S. & Radicchi, F. Benchmark graphs for testing community detection algorithms. Phys. Rev.
E 78, 046110 (2008).
[9] Holland, P. W., Laskey, K. B. & Leinhardt, S. Stochastic blockmodels: First steps. Soc. Networks 5, 109 -- 137 (1983).
[10] Newman, M. E. J. & Girvan, M. Finding and evaluating community structure in networks. Phys. Rev. E 69, 026113
(2004).
[11] Zhang, P. & Moore, C. Scalable detection of statistically significant communities and hierarchies, using message passing
for modularity. Proc. Natl. Acad. Sci. U.S.A. 111, 18144 -- 18149 (2014).
[12] Rosvall, M. & Bergstrom, C. Maps of random walks on complex networks reveal community structure. Proc. Natl. Acad.
Sci. U.S.A. 105, 1118 -- 1123 (2008).
[13] Rosvall, M. & Bergstrom, C. T. Multilevel compression of random walks on networks reveals hierarchical organization in
large integrated systems. PloS One 6, e18209 (2011).
[14] Newman, M. E. J. Spectral methods for community detection and graph partitioning. Phys. Rev. E 88, 042822 (2013).
[15] Shi, J. & Malik, J. Normalized cuts and image segmentation.
IEEE Transactions on Pattern Analysis and Machine
Intelligence 22, 888 -- 905 (2000).
[16] Luxburg, U. A tutorial on spectral clustering. Stat. Comput. 17, 395 -- 416 (2007).
[17] Krzakala, F. et al. Spectral redemption in clustering sparse networks. Proc. Natl. Acad. Sci. U.S.A. 110, 20935 -- 40 (2013).
[18] Abbe, E. & Sandon, C. Recovering communities in the general stochastic block model without knowing the parameters.
In Cortes, C., Lawrence, N. D., Lee, D. D., Sugiyama, M. & Garnett, R. (eds.) Advances in Neural Information Processing
Systems 28, 676 -- 684 (Curran Associates, Inc., 2015).
[19] Nowicki, K. & Snijders, T. A. B. Estimation and prediction for stochastic blockstructures. J. Amer. Statist. Assoc. 96,
1077 -- 1087 (2001).
[20] Daudin, J. J., Picard, F. & Robin, S. A mixture model for random graphs. Stat. Comput. 18, 173 -- 183 (2008).
[21] Decelle, A., Krzakala, F., Moore, C. & Zdeborov´a, L. Asymptotic analysis of the stochastic block model for modular
networks and its algorithmic applications. Phys. Rev. E 84, 066106 (2011).
[22] Hayashi, K., Konishi, T. & Kawamoto, T. A tractable fully bayesian method for the stochastic block model. arXiv preprint
arXiv:1602.02256 (2016).
[23] Newman, M. E. J. & Reinert, G. Estimating the number of communities in a network. Phys. Rev. Lett. 117, 078301
(2016).
[24] Peixoto, T. P. Parsimonious module inference in large networks. Phys. Rev. Lett. 110, 148701 (2013).
[25] Peixoto, T. P. Hierarchical Block Structures and High-Resolution Model Selection in Large Networks. Physical Review X
4, 011047 (2014).
[26] Peixoto, T. P. Model selection and hypothesis testing for large-scale network models with overlapping groups. Phys. Rev.
X 5, 011033 (2015).
[27] Hastie, T. J., Tibshirani, R. J. & Friedman, J. H. The Elements of Statistical Learning : Data Mining, Inference, and
Prediction. Springer series in statistics (Springer, New York, 2009).
[28] Arlot, S. & Celisse, A. A survey of cross-validation procedures for model selection. Statist. Surv. 4, 40 -- 79 (2010).
[29] Celeux, G. & Durand, J.-B. Selecting hidden markov model state number with cross-validated likelihood. Comput. Stat.
23, 541 -- 564 (2008).
[30] Vehtari, A. & Ojanen, J. A survey of bayesian predictive methods for model assessment, selection and comparison. Stat.
Surv. 142 -- 228 (2012).
[31] Airoldi, E. M., Blei, D. M., Fienberg, S. E. & Xing, E. P. Mixed membership stochastic blockmodels.
In Koller, D.,
Schuurmans, D., Bengio, Y. & Bottou, L. (eds.) Advances in Neural Information Processing Systems 21, 33 -- 40 (Curran
Associates, Inc., 2009).
[32] Hoff, P. Modeling homophily and stochastic equivalence in symmetric relational data. In Platt, J. C., Koller, D., Singer,
Y. & Roweis, S. T. (eds.) Advances in Neural Information Processing Systems 20, 657 -- 664 (Curran Associates, Inc., 2008).
[33] Chen, K. & Lei, J. Network cross-validation for determining the number of communities in network data. arXiv preprint
arXiv:1411.1715 (2014).
[34] Decelle, A., Krzakala, F., Moore, C. & Zdeborov´a, L. Inference and Phase Transitions in the Detection of Modules in
Sparse Networks. Phys. Rev. Lett. 107, 065701 (2011).
[35] Mossel, E., Neeman, J. & Sly, A. Reconstruction and estimation in the planted partition model. Probab. Theory Relat.
Fields 1 -- 31 (2014).
[36] Massouli´e, L. Community detection thresholds and the weak ramanujan property. In Proceedings of the 46th Annual ACM
Symposium on Theory of Computing, STOC '14, 694 -- 703 (ACM, New York, NY, USA, 2014).
[37] Bishop, C. M. Pattern Recognition and Machine Learning (Information Science and Statistics) (Springer-Verlag New York,
Inc., Secaucus, NJ, USA, 2006).
[38] Levin, E., Tishby, N. & Solla, S. A. A statistical approach to learning and generalization in layered neural networks. In
Proceedings of the Second Annual Workshop on Computational Learning Theory, COLT '89, 245 -- 260 (1989).
[39] Newman, M. E. J. Modularity and community structure in networks. Proc. Natl. Acad. Sci. U.S.A. 103, 8577 -- 8582 (2006).
19
[40] Zachary, W. W. An information flow model for conflict and fission in small groups. J. Anthropol. Res. 33, 452 -- 473 (1977).
[41] Newman, M. E. J. Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E 74, 036104
(2006).
[42] Newman, M. E. J. http://www-personal.umich.edu/~mejn/netdata/. (2006) (Date of access: 11/05/2015).
[43] Duch, J. & Arenas, A. Community detection in complex networks using extremal optimization. Phys. Rev. E 72, 027104
(2005).
[44] Karrer, B. & Newman, M. E. J. Stochastic blockmodels and community structure in networks. Phys. Rev. E 83, 016107
(2011).
[45] Zhao, Y., Levina, E. & Zhu, J. Consistency of community detection in networks under degree-corrected stochastic block
models. Ann. Stat. 2266 -- 2292 (2012).
[46] Yan, X. et al. Model selection for degree-corrected block models. J. Stat. Mech. Theor. Exp. 2014, P05007 (2014).
[47] Csisz´ar, I. Axiomatic characterizations of information measures. Entropy 10, 261 (2008).
[48] Amari, S.-i. & Cichocki, A. Information geometry of divergence functions. Bulletin of the Polish Academy of Sciences:
Technical Sciences 58, 183 -- 195 (2010).
[49] Kawamoto, T. https://github.com/tatsuro-kawamoto/graphBIX. (2016) (Date of access: 13/09/2016).
[50] Newman, M. E. J. & Clauset, A. Structure and inference in annotated networks. Nat. Commun. 7, 11863 (2016).
[51] Domingos, P. A few useful things to know about machine learning. Commun. ACM 55, 78 -- 87 (2012).
[52] M´ezard, M. & Montanari, A. Information, Physics, and Computation (Oxford University Press, 2009).
[53] Opper, M. & Winther, O. Mean field approach to bayes learning in feed-forward neural networks. Phys. Rev. Lett. 76,
1964 -- 1967 (1996).
[54] Peel, L., Larremore, D. B. & Clauset, A. The ground truth about metadata and community detection in networks.
arXiv:1608.05878 (2016).
|
1609.03756 | 2 | 1609 | 2017-12-21T21:08:14 | Correlations of consumption patterns in social-economic networks | [
"cs.SI",
"cs.CY",
"physics.soc-ph"
] | We analyze a coupled anonymized dataset collecting the mobile phone communication and bank transactions history of a large number of individuals. After mapping the social structure and introducing indicators of socioeconomic status, demographic features, and purchasing habits of individuals we show that typical consumption patterns are strongly correlated with identified socioeconomic classes leading to patterns of stratification in the social structure. In addition we measure correlations between merchant categories and introduce a correlation network, which emerges with a meaningful community structure. We detect multivariate relations between merchant categories and show correlations in purchasing habits of individuals. Our work provides novel and detailed insight into the relations between social and consuming behaviour with potential applications in recommendation system design. | cs.SI | cs |
Correlations of consumption patterns in social-economic
networks
Yannick Leo1, M´arton Karsai1,*, Carlos Sarraute2 and Eric Fleury1
1
2
*
Univ Lyon, ENS de Lyon, Inria, CNRS, UCB Lyon 1, LIP UMR 5668, IXXI, F-69342, Lyon, France
Grandata Labs, Bartolome Cruz 1818 V. Lopez. Buenos Aires, Argentina
Corresponding author: [email protected]
Abstract
We analyze a coupled anonymized dataset collecting the
mobile phone communication and bank transactions his-
tory of a large number of individuals. After mapping
the social structure and introducing indicators of socioe-
conomic status, demographic features, and purchasing
habits of individuals we show that typical consumption
patterns are strongly correlated with identified socioe-
conomic classes leading to patterns of stratification in
the social structure.
In addition we measure correla-
tions between merchant categories and introduce a cor-
relation network, which emerges with a meaningful com-
munity structure. We detect multivariate relations be-
tween merchant categories and show correlations in pur-
chasing habits of individuals. Our work provides novel
and detailed insight into the relations between social and
consuming behaviour with potential applications in rec-
ommendation system design.
1 Introduction
Social Network Analysis
(SNA) provides one
promising direction to explore such problems [7], due
to its enormous benefit from the massive flow of hu-
man behavioural data provided by the digital data
revolution [8]. The advent of this era was propa-
gated by some new data collection techniques, which
allowed the recording of the digital footprints and in-
teraction dynamics of millions of individuals [9, 10].
On the other hand, although social behavioural data
brought us detailed knowledge about the structure
and dynamics of social
it commonly
failed to uncover the relationship between social and
economic positions of individuals. Nevertheless, such
correlations play important roles in determining one's
socioeconomic status (SES) [11], social tie formation
preferences due to status homophily [12, 13], and in
turn potentially stand behind the emergent stratified
structure and segregation on the society level [4, 14].
However until now, the coupled investigation of indi-
vidual social and economic status remained a great
challenge due to lack of appropriate data recording
such details simultaneously.
interactions,
The consumption of goods and services is a cru-
cial element of human welfare. The uneven dis-
tribution of consumption power among individuals
goes hand in hand with the emergence and reserva-
tion of socioeconomic inequalities in general.
Indi-
vidual financial capacities restrict personal consumer
behaviour, arguably correlate with one's purchas-
ing preferences, and play indisputable roles in deter-
mining the socioeconomic position of an ego in the
larger society [1, 2, 3, 4, 5].
Investigation of rela-
tions between these characters carries a great poten-
tial in understanding better rational social-economic
behaviour [6], and project to direct applications in
personal marketing, recommendation, and advertis-
ing.
As individual economic status restricts one's capac-
ity in purchasing goods and services, it induces diver-
gent consumption patterns between people at differ-
ent socioeconomic positions [6, 1, 2]. This is reflected
by sets of commonly purchased products, which are
further associated to one's social status [15]. Con-
sumption behaviour has been addressed from vari-
ous angles considering e.g. environmental effects, so-
cioeconomic position, or social influence coming from
connected peers [1]. However, large data-driven stud-
ies combining information about individual purchas-
ing and interaction patterns in a society large pop-
ulation are still rare, although questions about cor-
relations between consumption and social behaviour
1
are of utmost interest.
In this study we address these crucial problems
via the analysis of a dataset, which simultane-
ously records the mobile-phone communication, bank
transaction history, and purchase sequences of mil-
lions of inhabitants of a single country over several
months. This corpus, one among the firsts at this
scale and details, allows us to infer the socioeconomic
status, consumption habits, and the underlying social
structure of millions of connected individuals. Using
this information our overall goal is to identify people
with certain financial capacities, and to understand
how much money they spend, on what they spend,
and whether they spend like their friends? More pre-
cisely, we formulate our study around two research
questions:
• Can one associate typical consumption patterns
to people and to their peers belonging to the
same or different socioeconomic classes, and if
yes how much such patterns vary between indi-
viduals or different classes?
• Can one draw relations between commonly pur-
chased goods or services in order to understand
better individual consumption behaviour?
After reviewing the related literature in Section 2,
we describe our dataset in Section 3, and introduce
individual socioeconomic indicators to define socioe-
conomic classes in Section 4. In Section 5 we show
how typical consumption patterns vary among classes
and relate them to structural correlations in the social
network. In Section 6 we draw a correlation network
between consumption categories to detect patterns of
commonly purchased goods and services. Finally we
present some concluding remarks and future research
ideas.
2 Related work
Earlier hypothesis on the relation between consump-
tion patterns and socioeconomic inequalities, and
their correlations with demographic features such as
age, gender, or social status were drawn from spe-
cific sociological studies [16] and from cross-national
social surveys [17]. However, recently available large
datasets help us to effectively validate and draw new
hypotheses as population-large individual level obser-
vations and detailed analysis of human behavioural
data became possible. These studies shown that per-
sonal social interactions, social influence [1], or ho-
mophily [22] in terms of age or gender [20] have strong
effects on purchase behaviour, knowledge which led
to the emergent domain of online social market-
ing [21]. Yet it is challenging to measure correla-
tions between individual social status, social network,
and purchase patterns simultaneously. Although so-
cioeconomic parameters can be estimated from com-
munication networks [18] or from external aggregate
data [19] usually they do not come together with indi-
vidual purchase records. In this paper we propose to
explore this question through the analysis of a com-
bined dataset proposing simultaneous observations of
social structure, economic status and purchase habits
of millions of individuals.
3 Data description
In the following we are going to introduce two
datasets extracted from a corpus combining the mo-
bile phone interactions with purchase history of indi-
viduals.
DS1: Ego social-economic data with
purchase distributions
Communication data used in our study records the
temporal sequence of 7,945,240,548 call and SMS in-
teractions of 111,719,360 anonymized mobile phone
users for 21 consecutive months. Each call detailed
record (CDR) contains the time, unique caller and
callee encrypted IDs, the direction (who initiate the
call/SMS), and the duration of the interaction. At
least one participant of each interaction is a client of a
single mobile phone operator, but other mobile phone
users who are not clients of the actual provider also
appear in the dataset with unique IDs. All unique
IDs are anonymized as explained below, thus indi-
vidual identification of any person is impossible from
the data. Using this dataset we constructed a large
social network where nodes are users (whether clients
or not of the actual provider), while links are drawn
between any two users if they interacted (via call or
SMS) at least once during the observation period. We
filtered out call services, companies, and other non-
human actors from the social network by removing
all nodes (and connected links) who appeared with
either in-degree kin = 0 or out-degree kout = 0.
We repeated this procedure recursively until we re-
ceived a network where each user had kin, kout > 0,
i. e. made at least one out-going and received at
least one in-coming communication event during the
nearly two years of observation. After construction
2
and filtering the network remained with 82,453,814
users connected by 1,002,833,289 links, which were
considered to be undirected after this point.
To calculate individual economic estimators we
used a dataset provided by a single bank. This data
records financial details of 6,002,192 people assigned
with unique anonymized identifiers over 8 consecutive
months. The data provides time varying customer
variables as the amount of their debit card purchases,
their monthly loans, and static user attributes such
as their billing postal code (zip code), their age and
their gender.
A subset of IDs of the anonymized bank and mobile
phone customers were matched1. This way of com-
bining the datasets allowed us to simultaneously ob-
serve the social structure and estimate economic sta-
tus (for definition see Section 4) of the connected in-
dividuals. This combined dataset contained 999,456
IDs, which appeared in both corpuses. However,
for the purpose of our study we considered only the
largest connected component of this graph. This way
we operate with a connected social graph of 992,538
people connected by 1,960,242 links, for all of them
with communication events and detailed bank records
available.
To study consumption behaviour we used purchase
sequences recording the time, amount, merchant cat-
egory code of each purchase event of each individual
during the observation period of 8 months. Purchase
events are linked to one of the 281 merchant cate-
gory codes (mcc) indicating the type of the actual
purchase, like fast food restaurants, airlines, gas sta-
tions, etc. Due to the large number of categories in
this case we decided to group mccs by their types into
28 purchase category groups (PCGs) using the cate-
gorization proposed in [23]. After analyzing each pur-
chase groups 11 of them appeared with extremely low
activity representing less than 0.3% (combined) of the
total amount of purchases, thus we decided to remove
them from our analysis and use only the remaining
K17 set of 17 groups (for a complete list see Fig.2a).
Note that the group named Service Providers (k1
with mcc 24) plays a particular role as it corresponds
to cash retrievals and money transfers and it repre-
sents around 70% of the total amount of purchases.
As this group dominates over other ones, and since
we have no further information how the withdrawn
1 The matching, data hashing, and anonymization proce-
dure was carried out without the involvement of the scientific
partner. After this procedure only anonymized hashed IDs
were shared disallowing the direct identification of individuals
in any of the datasets.
cash was spent, we analyze this group k1 separately
from the other K2-17 = K17\{k1} set of groups.
This way we obtained DS1, which collects the social
ties, economic status, and coarse grained purchase
habit informations of ∼ 1 million people connected
together into a large social network.
DS2: Detailed ego purchase distributions
with age and gender
From the same bank transaction trace of 6,002,192
users, we build a second data set DS2. This dataset
collects data about the age and gender of individu-
als together with their purchase sequence recording
the time, amount, and mcc of each debit card pur-
chase of each ego. To obtain a set of active users we
extracted a corpus of 4,784,745 people that were ac-
tive at least two months during the observation pe-
riod. Then for each ego, we assigned a feature set
P V (u) : {ageu, genderu, SEGu, r(ci, u)} where SEG
assigns a socioeconomic group (for definition see Sec-
tion 4) and r(ci, u) is an ego purchase distribution
vector defined as
r(ci, u) =
.
(1)
u(cid:80)
mci
mci
u
ci
This vector assigns the fraction of mci
u money spent
by user u on a merchant category ci during the obser-
vation period. We excluded purchases corresponding
to cash retrievals and money transfers, which would
dominate our measures otherwise. A minor fraction
of purchases are not linked to valid mccs, thus we
excluded them from our calculations.
This way DS2 collects 3,680,652 individuals, with-
out information about their underlying social net-
work, but all assigned with a P V (u) vector describing
their personal demographic and purchasing features
in details.
4 Measures of socioeconomic position
To estimate the personal economic status we used a
simple measure reflecting the consumption power of
each individual. Starting from the raw data of DS2,
which collects the amount and type of debit card pur-
chases, we estimated the economic position of individ-
uals as their average monthly purchase (AMP). More
precisely, in case of an ego u who spent mu(t) amount
in month t we calculated the AMP as
(cid:80)
t∈T mu(t)
Tu
(2)
Pu =
3
where Tu corresponds to the number of active
months of user u (with at least one purchase in each
month). After sorting people by their AMP values
we computed the normalized cumulative distribution
function of Pu as
(cid:80)f
(cid:80)
f(cid:48)=0 Pu(f(cid:48))
u Pu
C(f ) =
(3)
as a function of f fraction of people. This func-
tion (Fig.1a) appears with high variance and sug-
gests large imbalances in terms of the distribution of
economic capacities among individuals in agreement
with earlier social theory [27].
Fig. 1: Social class characteristics (a) Schematic
demonstration of user partitions into 9 socioe-
conomic classes by using the cumulative AMP
function C(f ). Fraction of egos belonging to
a given class (x axis) have the same sum of
u Pu)/n (y axis) for each class. (b)
Number of egos (green) and the average AMP
(cid:104)P(cid:105) (in USD) per individual (yellow) in differ-
ent classes.
AMP ((cid:80)
was the same equal to ((cid:80)
Subsequently we used the C(f ) function to assign
egos into 9 economic classes (also called socioeco-
nomic classes with smaller numbers assigning lower
classes) such that the sum of AMP in each class sj
u Pu)/n (Fig.1). We de-
cided to use 9 distinct classes based on the common
three-stratum model [25], which identifies three main
social classes (lower, middle, and upper), and for each
of them three sub-classes [26]. There are several ad-
vantages of this classification:
(a) it relies merely
on individual economic estimators, Pu, (b) naturally
partition egos into classes with decreasing sizes for
richer groups and (c) increasing (cid:104)P(cid:105) average AMP
values per egos (Fig.1b).
5 Socioeconomic correlations in
purchasing patterns
In order to address our first research question we
were looking for correlations between individuals in
different socioeconomic classes in terms of their con-
sumption behaviour on the level of purchase category
groups. We analyzed the purchasing behaviour of
people in DS1 after categorizing them into socioeco-
nomic classes as explained in Section 4.
First for each class sj we take every user u ∈ sj
and calculate the mk
u total amount of purchases they
spent on a purchase category group k ∈ K17. Then
we measure a fractional distribution of spending for
each PCGs as:
r(k, sj) =
,
(4)
(cid:80)
(cid:80)
mk
u∈sj
u
u∈s mk
u
where s = (cid:83)
j sj assigns the complete set of users.
In Fig.2a each line shows the r(k, sj) distributions
for a PCG as the function of sj social classes, and
lines are sorted (from top to bottom) by the total
amount of money spent on the actual PCG2. Interest-
ingly, people from lower socioeconomic classes spend
more on PCGs associated to essential needs, such as
Retail Stores (St.), Gas Stations, Service Providers
(cash) and Telecom, while in the contrary, other cat-
egories associated to extra needs such as High Risk
Personal Retail (Jewelry, Beauty), Mail Phone Or-
der, Automobiles, Professional Services (Serv.) (ex-
tra health services), Whole Trade (auxiliary goods),
Clothing St., Hotels and Airlines are dominated by
people from higher socioeconomic classes. Also note
that concerning Education most of the money is spent
by the lower middle classes, while Miscellaneous St.
(gift, merchandise, pet St.) and more apparently En-
tertainment are categories where the lowest and high-
est classes are spending the most.
From this first analysis we can already identify
large differences in the spending behaviour of peo-
ple from lower and upper classes. To further in-
vestigate these dissimilarities on the individual level,
we consider the K2-17 category set as defined in sec-
tion 3 (category k1 excluded) and build a spending
vector SV (u) = [SV2(u), ..., SV17(u)] for each ego u.
2 Note that in our social class definition the cumulative AMP
is equal for each group and this way each group represents the
same economic potential as a whole. Values shown in Fig.2a
assign the total purchase of classes. Another strategy would
be to calculate per capita measures, which in turn would be
strongly dominated by values associated to the richest class,
hiding any meaningful information about other classes.
4
0.00.20.40.60.81.0f0.00.20.40.60.81.0CW(f)CP(f)f(a)Class 1Class 4Class 2Class 3Class 5Class 8Class 6Class 7Class 9(a)(b) Fig. 2: Consumption correlations in the socioeconomic network (a) r(k, si) distribution of spending
in a given purchase category group k ∈ K17 by different classes sj. Distributions are normalised
as in Eq.4, i.e. sums up to 1 for each category. (b) Dispersion σSV (sj) for different socioeconomic
classes considering PCGs in K2-17 (dark blue) and the single category k1 (light blue). (c) (resp.
(d)) Heat-map matrix representation of dSV (si, sj) (resp. dk1(si, sj)) distances between the average
spending vectors of pairs of socioeconomic classes considering PCGs in K2-17 (resp. k1). (e) Shannon
entropy measures for different socioeconomic classes considering PCGs in K2-17 (dark pink) and in
k17 (light pink). (f ) (resp. (g)) Heat-map matrix representation of the average LSV (si, sj) (resp.
Lk1(si, sj)) measure between pairs of socioeconomic classes considering PCGs in K2-17 (resp. k1).
k∈K mk
out of his/her mu =(cid:80)
Here each item SVk(u) assigns the fraction of money
u/mu that user u spent on a category k ∈ K2-17
mk
u total amount of pur-
chases. Using these individual spending vectors we
calculate the average spending vector of a given so-
cioeconomic class as SV (sj) = (cid:104)SV (u)(cid:105)u∈sj . We as-
sociate SV (sj) to a representative consumer of class
sj and use this average vector to quantify differences
between distinct socioeconomic classes as follows.
The euclidean metric between average spending
the richest social classes appear with a very different
behaviour. This is due to their relative underspend-
ing in cash, which can be also concluded from Fig.2a
(first row). On the other hand as going towards lower
classes such differences decrease as cash usage starts
to dominate.
To explain better the differences between socioe-
conomic classes in terms of purchasing patterns, we
introduce two additional scalar measures. First, we
introduce the dispersion of individual spending vec-
tors as compared to their class average as
dSV (si, sj) = (cid:107)SV k(si) − SV k(sj)(cid:107)2,
(5)
σSV (sj) = (cid:104)(cid:107)SV k(sj) − SVk(u)(cid:107)2(cid:105)u∈sj ,
(6)
vectors is:
where (cid:107)(cid:126)v(cid:107)2 =(cid:112)(cid:80)
k v2
k assigns the L2 norm of a vec-
tor (cid:126)v. Note that the diagonal elements of dSV (si, si)
are equal to zero by definition. However, in Fig.2c
the off-diagonal green component around the diag-
onal indicates that the average spending behaviour
of a given class is the most similar to neighboring
classes, while dissimilarities increase with the gap be-
tween socioeconomic classes. We repeated the same
measurement separately for the single category of
cash purchases (PCG k1).
In this case euclidean
distance is defined between average scalar measures
as dk1(si, sj) = (cid:107)(cid:104)SV1(cid:105)(si) − (cid:104)SV1(cid:105)(sj)(cid:107)2.
Interest-
indicates that here
ingly, results shown in Fig.2d.
which appears with larger values if people in a given
class allocate their spending very differently. Second,
we also calculate the Shannon entropy of spending
patterns as
SSV (sj) =
−SV k(sj) log(SV k(sj))
(7)
(cid:88)
k∈K2-17
to quantify the variability of the average spending
vector for each class. This measure is minimal if each
ego of a class sj spends exclusively on the same sin-
gle PCG, while it is maximal if they equally spend on
each PCG. As it is shown in Fig.2b (light blue line
5
(a)(b)(d)(c)(e)(g)(f) with square symbols) dispersion decreases rapidly as
going towards higher socioeconomic classes. This as-
signs that richer people tends to be more similar in
terms of their purchase behaviour. On the other
hand, surprisingly,
in Fig.2e (dark pink line with
square symbols) the increasing trend of the corre-
sponding entropy measure suggests that even richer
people behave more similar in terms of spending be-
haviour they used to allocate their purchases in more
PCGs. These trends are consistent even in case of
k1 cash purchase category (see σSV1(sj) function de-
picted with dark blue line in in Fig.2b) or once we in-
clude category k1 into the entropy measure SSV17(sj)
(shown in Fig.2b with light pink line).
To complete our investigation we characterize the
effects of social relationships on the purchase habits
of individuals. We address this problem through an
overall measure quantifying differences between indi-
vidual purchase vectors of connected egos positioned
in the same or different socioeconomic classes. More
precisely, we consider each social tie (u, v) ∈ E con-
necting individuals u ∈ si and v ∈ sj, and for each
purchase category k we calculate the average absolute
difference of their purchase vector items as
dk(si, sj) = (cid:104)SVk(u) − SVk(v)(cid:105)u∈si,v∈sj .
(8)
Following that, as a reference system we generate a
corresponding configuration network by taking ran-
domly selected edge pairs from the underlying social
structure and swap them without allowing multiple
links and self loops.
In order to vanish any resid-
ual correlations we repeated this procedure in 5×E
times. This randomization keeps the degree, indi-
vidual economic estimators Pu, the purchase vector
SV (u), and the assigned class of each people un-
changed, but destroys any structural correlations be-
tween egos in the social network, consequently be-
tween socioeconomic classes as well. After generating
a reference structure we computed an equivalent mea-
rn(si, sj) but now using links (u, v) ∈ Ern of the
sure dk
randomized network. We repeated this procedure 100
rn(cid:105)(si, sj). In or-
times and calculated an average (cid:104)dk
der to quantify the effect of the social network we
simply take the ratio
Lk(si, sj) =
dk(si, sj)
(cid:104)dk
rn(cid:105)(si, sj)
(9)
and calculate its average LSV (si, sj) = (cid:104)Lk(si, sj)(cid:105)k
over each category group k ∈ K2-17 or respectively k1.
This measure shows whether connected people have
more similar purchasing patterns than one would ex-
pect by chance without considering any effect of ho-
mophily, social influence or structural correlations.
Results depicted in Fig.2f and 2g for LSV (si, sj) (and
Lk1(si, sj) respectively) indicates that the purchas-
ing patterns of individuals connected in the original
structure are actually more similar than expected by
chance (diagonal component). On the other hand
people from remote socioeconomic classes appear to
be less similar than one would expect from the uncor-
related case (indicated by the LSV (si, sj) > 1 values
typical for upper classes in Fig.2f). Note that we
found the same correlation trends in cash purchase
patterns as shown in Fig.2g. These observations do
not clearly assign whether homophily [12, 13] or so-
cial influence [1] induce the observed similarities in
purchasing habits but undoubtedly clarifies that so-
cial ties (i.e. the neighbors of an ego) and socioeco-
nomic status play deterministic roles in the emerging
similarities in consumption behaviour.
6 Purchase category correlations
To study consumption patterns of single purchase
categories PCGs provides a too coarse grained level
of description. Hence, to address our second ques-
tion we use DS2 and we downscale from the category
group level to the level of single merchant categories.
We are dealing with 271 categories after excluding
some with less than 100 purchases and the categories
linked to money transfer and cash retrieval (for a
complete list of IDs and name of the purchase cat-
egories considered see Table 1). As in Section 3 we
assign to each ego u a personal vector P V (u) of four
socioeconomic features: the age, the gender, the so-
cial economic group, and the distribution r(ci, u) of
purchases in different merchant categories made by
the central ego. Our aim here is to obtain an overall
picture of the consumption structure at the level of
merchant categories and to understand precisely how
personal and socioeconomic features correlate with
the spending behaviour of individuals and with the
overall consumption structure.
As we noted in section 5, the purchase spending
vector r(ci, u) of an ego quantifies the fraction of
money spent on a category ci. Using the spending
vectors of n number of individuals we define an over-
all correlation measure between categories as
n((cid:80)
((cid:80)
u r(ci, u))((cid:80)
u r(ci, u)r(cj, u))
.
(10)
u r(cj, u))
ρ(ci, cj) =
6
Fig. 3: Merchant category correlation matrix and graph (a) 163×163 matrix heatmap plot corre-
sponding to ρ(ci, cj) correlation values (see Eq. 10) between categories. Colors scale with the loga-
rithm of correlation values. Positive (resp. negative) correlations are assigned by red (resp. blue)
colors. Diagonal components represent communities with frames colored accordingly.(b) Weighted
G>
ρ correlation graph with nodes annotated with MCCs (see Table 1). Colors assign 17 communities
of merchant categories with representative names summarized in the figure legend.
Fig. 4: Socioeconomic parameters of merchant categories (a) Scatter plot of AF S(ci) triplets (for
definition see Eq. 11 and text) for 271 merchant categories summarized in Table 1. Axis assign
average age and SEG of purchase categories, while gender information are assigned by symbols. The
shape of symbols assigns the dominant gender (circle-female, square-male) and their size scales with
average values. (b) Similar scatter plot computed for communities presented in Fig.3b. Labels and
colors are explained in the legend of Fig.3a.
7
52111711525155335942274159435964411140114112451147225651581359477011412141314789530953315732594859935999792279917999939956917399421547844816519253995734573558115812581459685969597059925994721672307298731173927512752375427933794179967997899959675045504650655085511159957538458252005310554193114812732148997372799459457273598349005039501350725198551155325021571252315719595057337993504780118021806280715722507450945621563156995944597751315441594951225137566151395169517251935714762976356555641545154625973554276225599557156115935594156975681593159717296729778417832721072117932804959217929594059768641594673387221596572777427299799873618099799582118220(a)(b)Car sales and maintenanceHardware storesOffice supply storesIT servicesBooks and newspapersState services and educationHome supply storesNewsstand and duty-free shopsAmusement and recreationTravellingTransportation and commutingLeisureJewellery and gift shopsClothing 1Clothing 2Personal servicesHealth and medical services(a)(b)0 0.5 1female male0 0.5 1female male(a)(b) This symmetric formulae quantifies how much peo-
ple spend on a category ci if they spend on an other
cj category or vice versa. Therefore, if ρ(ci, cj) > 1,
the categories ci and cj are positively correlated and
if ρ(ci, cj) < 1, categories are negatively correlated.
Using ρ(ci, cj) we can define a weighted correlation
graph Gρ = (Vρ, Eρ, ρ) between categories ci ∈ Vρ,
where links (ci, cj) ∈ Eρ are weighted by the ρ(ci, cj)
correlation values. The weighted adjacency matrix
of Gρ is shown in Fig.3a as a heat-map matrix with
logarithmically scaling colors. Importantly, this ma-
trix emerges with several block diagonal components
suggesting present communities of strongly correlated
categories in the graph.
To identify categories which were commonly pur-
chased together we consider only links with positive
correlations. Furthermore, to avoid false positive cor-
relations, we consider a 10% error on r that can in-
duce, in the worst case 50% overestimation of the
correlation values. In addition, to consider only rep-
resentative correlations we take into account category
pairs which were commonly purchased by at least
1000 consumers. This way we receive a G>
ρ weighted
sub-graph of Gρ, shown in Fig.3b, with 163 nodes
and 1664 edges with weights ρ(ci, cj) > 1.5.
To identify communities in G>
ρ indicated by the
correlation matrix in Fig.3a we applied a graph parti-
tioning method based on the Louvain algorithm [28].
We obtained 17 communities depicted with differ-
ent colors in Fig.3b and as corresponding colored
frames in Fig.3a.
Interestingly, each of these com-
munities group a homogeneous set of merchant cat-
egories, which could be assigned to similar types of
purchasing activities (see legend of Fig.3b). In addi-
tion, this graph indicates how different communities
are connected together. Some of them, like Trans-
portation, IT or Personal Serv. playing a central
role as connected to many other communities, while
other components like Car sales and maintenance
and Hardware St., or Personal and Health and med-
ical Serv. are more like pairwise connected. Some
groups emerge as standalone communities like Office
Supp. St., while others like Books and newspapers
or Newsstands and duty-free Shops (Sh.) appear as
bridges despite their small sizes.
Note that the main categories corresponding to
everyday necessities related to food (Supermarkets,
Food St.) and telecommunication (Telecommunica-
tion Serv.) do not appear in this graph. Since they
are responsible for the majority of total spending,
they are purchased necessarily by everyone without
obviously enhancing the purchase in other categories,
thus they do not appear with strong correlations.
An average
categories and personal
Finally we turn to study possible correlations
between purchase
fea-
tures.
set AF S(ci) =
{(cid:104)age(ci)(cid:105),(cid:104)gender(ci)(cid:105),(cid:104)SEG(ci}(cid:105)) is assigned to
each of the 271 categories. The average (cid:104)v(ci)(cid:105) of
a feature v ∈ {age, gender, SEG} assigns a weighted
average value computed as:
feature
(cid:104)v(ci)(cid:105) =
u∈{u}i
αi(vu)vu
u∈{u}u
αi(v)
,
(11)
where vu denotes a feature of a user u from the {u}i
set of individuals who spent on category ci. Here
(cid:80)
(cid:80)
(cid:88)
αi(vu) =
(u∈{u}ivu=v)
r(ci, u)
ni(vu)
(12)
corresponds to the average spending on category ci
of the set of users from {u}i sharing the same value
of the feature v. ni(vu) denotes the number of such
users. In other words, e.g. in case of v = age and c742,
(cid:104)age(c742)(cid:105) assigns the average age of people spent
on Veterinary Services (mcc = 742) weighted by the
amount they spent on it. In case of v = gender we
assigned 0 to females and 1 to males, thus the average
gender of a category can take any real value between
[0, 1], indicating more females if (cid:104)gender(ci)(cid:105) ≤ 0.5
or more males otherwise.
We visualize this multi-modal data in Fig.4a as
a scatter plot, where axes scale with average age
and SEG, while the shape and size of symbols corre-
spond to the average gender of each category. To fur-
ther identify correlations we applied k-means cluster-
ing [29] using the AF S(ci) of each category. The ideal
number of clusters was 15 according to several crite-
ria: Davies-Bouldin Criterion, Calinski-Harabasz cri-
terion (variance ratio criterion) and the Gap method
[30]. Colors in Fig.4a assign the identified k-mean
clusters.
The first thing to remark in Fig.4a is that the av-
erage age and SEG assigned to merchant categories
are positively correlated with a Pearson correlation
coefficient 0.42 (p < 0.01).
In other words, elderly
people used to purchase from more expensive cate-
gories, or alternatively, wealthier people tend to be
older, in accordance with our intuition. At the same
time, some signs of gender imbalances can be also
concluded from this plot. Wealthier people appear to
be commonly males rather than females. A Pearson
correlation measure between gender and SEG, which
8
7296: Clothing Rental
7297: Massage Parlors
7298: Health and Beauty Spas
7299: General Serv.
7311: Advertising Serv.
7941: Sports Clubs
7991: Tourist Attractions
7992: Golf Courses
7993: Video Game Supp.
7994: Video Game Arcades
7995: Gambling
7996: Amusement Parks
7997: Country Clubs
7998: Aquariums
7999: Recreation Serv.
8011: Doctors
8021: Dentists, Orthodontists
8031: Osteopaths
5072: Hardware Supp.
5598: Snowmobile Dealers
5192: Newspapers
5193: Nursery & Flowers Supp.
5198: Paints
5199: Nondurable Goods
5200: Home Supply St.
5211: Materials St.
5231: Glass & Paint St.
5251: Hardware St.
5261: Nurseries & Garden St.
5271: Mobile Home Dealers
5300: Wholesale
5309: Duty Free St.
5950: Glassware, Crystal St.
5960: Dir Mark - Insurance
5962: Direct Marketing - Travel
5611: Men Cloth. St.
5621: Wom Cloth. St.
5963: Door-To-Door Sales
5631: Women?s Accessory Sh. 5964: Dir. Mark. Catalog
5641: Children?s Wear St.
5651: Family Cloth. St.
5655: Sports & Riding St.
5661: Shoe St.
5681: Furriers Sh.
5691: Cloth. Stores
5697: Tailors
5698: Wig and Toupee St.
5699: Apparel Accessory Sh.
5712: Furniture
5713: Floor Covering St.
5714: Window Covering St.
5718: Fire Accessories St.
5719: Home Furnishing St.
5722: House St.
5732: Elec. St.
5733: Music Intruments St.
5734: Comp.Soft. St.
5735: Record Stores
5811: Caterers
5812: Restaurants
5813: Drinking Pl.
5331: Variety Stores
5814: Fast Foods
5399: General Merch.
5912: Drug St.
5411: Supermarkets
5921: Alcohol St.
5422: Meat Prov.
5931: Secondhand Stores
5441: Candy St.
5932: Antique Sh.
5451: Dairy Products St.
5933: Pawn Shops
5462: Bakeries
5935: Wrecking Yards
5499: Food St.
5937: Antique Reproductions 7211: Laundries
5511: Cars Sales
5521: Car Repairs Sales
5940: Bicycle Sh.
5531: Auto and Home Supp. St. 5941: Sporting St.
5532: Auto St.
5533: Auto Access.
5541: Gas Stations
5542: Automated Fuel Dispensers 5945: Toy,-Game Sh.
5551: Boat Dealers
742: Veterinary Serv.
763: Agricultural Cooperative 5074: Plumbing, Heating Equip. 5599: Auto Dealers
5085: Industrial Supplies
780: Landscaping Serv.
5094: Precious Objects/Stones
1520: General Contr.
5099: Durable Goods
1711: Heating, Plumbing
5111: Printing, Office Supp.
1731: Electrical Contr.
5122: Drug Proprietaries
1740: Masonry & Stonework
5131: Notions Goods
1750: Carpentry Contr.
5137: Uniforms Clothing
1761: Sheet Metal
5139: Commercial Footwear
1771: Concrete Work Contr.
1799: Special Trade Contr.
5169: Chemicals Products
2741: Publishing and Printing 5172: Petroleum Products
2791: Typesetting Serv.
2842: Specialty Cleaning
4011: Railroads
4111: Ferries
4112: Passenger Railways
4119: Ambulance Serv.
4121: Taxicabs
4131: Bus Lines
4214: Motor Freight Carriers
4215: Courier Serv.
4225: Public Storage
4411: Cruise Lines
4457: Boat Rentals and Leases 5310: Discount Stores
4468: Marinas Serv. and Supp. 5311: Dep. St.
4511: Airlines
4582: Airports, Flying Fields
4722: Travel Agencies
4784: Tolls/Bridge Fees
4789: Transportation Serv.
4812: Phone St.
4814: Telecom.
4816: Comp. Net. Serv.
4821: Telegraph Serv.
4899: Techno St.
4900: Utilities
5013: Motor Vehicle Supp.
5021: Commercial Furniture
5039: Constr. Materials
5044: Photographic Equip.
5045: Computer St.
5046: Commercial Equipment 5561: Motorcycle Sh.
5571: Motorcycle Sh.
5047: Medical Equipment
5051: Metal Service Centers
5592: Motor Homes Dealers
5065: Electrical St.
5946: Camera and Photo St. 7273: Dating/Escort Serv.
5947: Gift Sh.
5948: Luggage & Leather St. 7277: Counseling Services
5949: Fabric St.
5942: Book St.
5943: Stationery St.
5944: Jewelry St.
7276: Tax Preparation Serv.
5965: Dir. Mark. Retail Merchant 7321: Credit Reporting Agencies
5966: Dir Mark - TV
5967: Dir. Mark.
5968: Dir. Mark. Subscription
5969: Dir. Mark. Other
5970: Artist?s Supp.
5971: Art Dealers & Galleries
5972: Stamp and Coin St.
5973: Religious St.
5975: Hearing Aids
5976: Orthopedic Goods
5977: Cosmetic St.
5978: Typewriter St.
5983: Fuel Dealers (Non Auto)
5992: Florists
5993: Cigar St.
5994: Newsstands
5995: Pet Sh.
5996: Swimming Pools Sales
5997: Electric Razor St.
5998: Tent and Awning Sh.
5999: Specialty Retail
6211: Security Brokers
6300: Insurance
7011: Hotels
7012: Timeshares
7032: Sporting Camps
7033: Trailer Parks, Camps
7210: Laundry, Cleaning Serv.
7333: Graphic Design
7338: Quick Copy
7339: Secretarial Support Serv.
7342: Exterminating Services
7349: Cleaning and Maintenance
7361: Employment Agencies
7372: Computer Programming
7375: Information Retrieval Serv. 8041: Chiropractors
7379: Computer Repair
8042: Optometrists
7392: Consulting, Public Relations 8043: Opticians
7393: Detective Agencies
7394: Equipment Rental
7395: Photo Developing
7399: Business Serv.
7512: Car Rental Agencies
7513: Truck/Trailer Rentals
7519: Mobile Home Rentals
7523: Parking Lots, Garages
7531: Auto Body Repair Sh.
7534: Tire Retreading & Repair
7535: Auto Paint Sh.
7538: Auto Service Shops
7542: Car Washes
7549: Towing Serv.
7622: Electronics Repair Sh.
7623: Refrigeration Repair
7629: Small Appliance Repair
7631: Watch/Jewelry Repair
7641: Furniture Repair
7692: Welding Repair
7699: Repair Sh.
7829: Picture/Video Production
7832: Cinema
7841: Video Tape Rental St.
7911: Dance Hall & Studios
7922: Theater Ticket
7929: Bands, Orchestras
7932: Billiard/Pool
7933: Bowling
8049: Chiropodists, Podiatrists
8050: Nursing/Personal Care
8062: Hospitals
8071: Medical Labs
8099: Medical Services
8111: Legal Services, Attorneys
8211: Elem. Schools
8220: Colleges Univ.
8241: Correspondence Schools
8244: Business Schools
8249: Training Schools
8299: Educational Serv.
8351: Child Care Serv.
8398: Donation
8641: Associations
8651: Political Org.
8661: Religious Orga.
8675: Automobile Associations
8699: Membership Org.
8734: Testing Lab.
8911: Architectural Serv.
8931: Accounting Serv.
8999: Professional Serv.
9211: Courts of Law
9222: Government Fees
9223: Bail and Bond Payments
9311: Tax Payments
9399: Government Serv.
9402: Postal Serv.
7216: Dry Cleaners
7217: Upholstery Cleaning
7221: Photographic Studios
7230: Beauty Sh.
7251: Shoe Repair/Hat Cleaning
7261: Funeral Serv.
7278: Buying/Shopping Serv.
Tab. 1: Codes and names of 271 merchant categories used in our study. MCCs were taken from the Merchant
Category Codes and Groups Directory published by American Express [23]. Abbreviations corre-
spond to: Serv. - Services, Contr. - Contractors, Supp. - Supplies, St. - Stores, Equip. - Equipment,
Merch. - Merchandise, Prov. - Provisioners, Pl. - Places, Sh. - Shops, Mark. - Marketing, Univ. -
Universities, Org. - Organizations, Lab. - Laboratories.
appears with a coefficient 0.29 (p < 0.01) confirmed
it. On the other hand, no strong correlation was ob-
served between age and gender from this analysis.
To have an intuitive insight about the distribution
of merchant categories, we take a closer look at spe-
cific category codes (summarized in Table 1). As
seen in Fig.4a elderly people tend to purchase in spe-
cific categories such as Medical Serv., Funeral Serv.,
Religious Organisations, Motorhomes Dealers, Dona-
tion, Legal Serv.. Whereas categories such as Fast
Foods, Video Game Arcades, Cinema, Record St., Ed-
ucational Serv., Uniforms Clothing, Passenger Rail-
ways, Colleges-Universities are associated to younger
individuals on average. At the same time, wealth-
ier people purchase more in categories as Snowmo-
bile Dealers, Secretarial Serv., Swimming Pools Sales,
Car Dealers Sales, while poorer people tend to pur-
chase more in categories related to everyday neces-
sities like Food St., General Merch., Dairy Products
St., Fast Foods and Phone St., or to entertainment as
Billiard or Video Game Arcades. Typical purchase
categories are also strongly correlated with gender as
categories more associated to females are like Beauty
Sh., Cosmetic St., Health and Beauty Spas, Women
Clothing St. and Child Care Serv., while others are
preferred by males like Motor Homes Dealers, Snow-
mobile Dealers, Dating/Escort Serv., Osteopaths, In-
struments St., Electrical St., Alcohol St. and Video
Game Arcades.
Finally we repeated a similar analysis on commu-
nities shown in Fig.3b, but computing the AF S on a
set of categories that belong to the same community.
Results in Fig.4b disclose positive age-SEG correla-
tions as observed in Fig.4a, together with somewhat
9
intuitive distribution of the communities.
ANR projects.
7 Conclusion
References
In this paper we analyzed a multi-modal dataset col-
lecting the mobile phone communication and bank
transactions of a large number of individuals living
in a single country. This corpus allowed for an in-
novative global analysis both in term of social net-
work and its relation to the economical status and
merchant habits of individuals. We introduced sev-
eral measures to estimate the socioeconomic status of
each individual together with their purchasing habits.
Using these information we identified distinct socioe-
conomic classes, which reflected strongly imbalanced
distribution of purchasing power in the population.
After mapping the social network of egos from mo-
bile phone interactions, we showed that typical con-
sumption patterns are strongly correlated with the
socioeconomic classes and the social network behind.
We observed these correlations on the individual and
social class level.
In the second half of our study we detected corre-
lations between merchant categories commonly pur-
chased together and introduced a correlation network
which in turn emerged with communities grouping
homogeneous sets of categories. We further analyzed
some multivariate relations between merchant cate-
gories and average demographic and socioeconomic
features, and found meaningful patterns of correla-
tions giving insights into correlations in purchasing
habits of individuals.
We identified several new directions to explore in
the future. One possible track would be to better
understand the role of the social structure and inter-
personal influence on individual purchasing habits,
while the exploration of correlated patterns between
commonly purchased brands assigns another promis-
ing directions. Beyond our general goal to better un-
derstand the relation between social and consuming
behaviour these results may enhance applications to
better design marketing, advertising, and recommen-
dation strategies, as they assign relations between co-
purchased product categories.
Acknowledgment
We thank M. Fixman for assistance. We acknowl-
edge the support from the SticAmSud UCOOL
project, INRIA, and the SoSweet (ANR-15-CE38-
0011-01) and CODDDE (ANR-13-CORD-0017-01)
[1] A. Deaton, Understanding Consumption. Claren-
don Press (1992).
[2] A. Deaton and J. Muellbauer, Economics and
Consumer Behavior. Cambridge University Press
(1980).
[3] T. Piketti, Capital in the Twenty-First Century.
(Harvard University Press, 2014).
[4] S. Sernau, Social Inequality in a Global Age.
(SAGE Publications, 2013).
[5] C. E. Hurst, Social Inequality. 8th ed. (Pearson
Education, 2015).
[6] J. E. Fisher, Social Class and Consumer Behavior:
the Relevance of Class and Status", in Advances
in Consumer Research Vol. 14, eds. M. Wallen-
dorf and P. Anderson, Provo, UT : Association
for Consumer Research, pp 492–496 (1987) .
[7] S. Wasserman, K. Faust, Social Network Analy-
sis: Methods and Applications. (Cambridge Uni-
versity Press, 1994).
[8] S. Lohr, The age of big data. (New York Times,
2012).
[9] D. Lazer, et. al. Computational Social Science.
Science 323, 721–723 (2009)
[10] A. Abraham, A-E. Hassanien, V. Smasel (eds.),
Computational Social Network Analysis: Trends,
Tools and Research Advances. (Springer-Verlag,
2010).
[11] P. Bourdieu, Distinction: A Social Critique of
the Judgement of Taste. Harvard University Press
(Cambridge MA) (1984).
[12] M. McPherson, L. Smith-Lovin, J. M. Cook,
Birds of a Feather: Homophily in Social Net-
works. Ann. Rev. Sociol. 27 415–444 (2001).
[13] P. F. Lazarsfeld, R. K. Merton, Friendship as a
Social Process: A Substantive and Methodologi-
cal Analysis. In Freedom and Control in Modern
Society. (New York: Van Nostrand, 1954) pp. 18–
66.
10
[14] D. B. Grusky, Theories of Stratification and In-
equality. In The Concise Encyclopedia of Sociol-
ogy. pp. 622-624. (Wiley-Blackwell, 2011).
[28] V. Blondel, et. al., Fast unfolding of communi-
ties in large networks. J. Stat.l Mech: theory and
experiment P10008 (2008).
[29] C. M. Bishop, Neural Networks for Pattern
Recognition. (Oxford University Press, Oxford,
England) (1995).
[30] R. Tibshirani, G. Walther, T. Hastie, Estimating
the number of clusters in a data set via the gap
statistic. J. Roy. Stat. Soc. B 63, 411-423 (2001).
[15] P. West, Conspicuous Compassion: Why Some-
times It Really Is Cruel To Be Kind. Civitas, In-
stitute for the Study of Civil Society (London)
(2004).
[16] T. W. Chang, Social status and cultural con-
sumption Cambridge University Press (2010)
[17] A. Deaton, The analysis of household surveys: a
microeconometric approach to development pol-
icy. World Bank Publications (1997)
[18] Y. Dong, et. al., Inferring user demographics and
social strategies in mobile social networks. Proc.
of the 20th ACM SIGKDD international confer-
ence on Knowledge discovery and data mining,
15–24 (2014)
[19] N. Eagle, M. Macy, R. Claxton, Network di-
versity and economic development. Science 328,
1029–1031 (2010)
[20] L. Kovanen, et. al., Temporal motifs reveal ho-
mophily, gender-specific patterns, and group talk
in call sequences. Proc. Nat. Acad. Sci., 110,
18070–18075 (2013)
[21] R. Felix, P. A. Rauschnabel, C. Hinsch, Elements
of strategic social media marketing: A holistic
framework. J. Business Res. online 1st (2016)
[22] W. Wood, T. Hayes, Social Influence on con-
sumer decisions: Motives, modes, and conse-
quences. J. Consumer Psych. 22, 324–328 (2012).
[23] Merchant Category Codes and Groups Direc-
tory. American Express @ Work Reporting Ref-
erence (http://tinyurl.com/hne9ct5) (2008) (date
of access: 2/3/2016).
[24] P. Martineau, Social classes and spending behav-
ior. Journal of Marketing 121–130 (1958).
[25] D.F. Brown, Social class and Status. In Mey, Ja-
cob Concise Encyclopedia of Pragmatics. Elsevier
p. 953 (2009).
[26] P. Saunders, Social Class and Stratification.
(Routledge, 1990).
[27] V. Pareto, Manual of Political Economy. Reprint
(New English Trans) edition (1971).
11
|
1405.1001 | 3 | 1405 | 2017-12-15T18:40:04 | Density decompositions of networks | [
"cs.SI",
"cs.DS",
"physics.soc-ph"
] | We introduce a new topological descriptor of a network called the density decomposition which is a partition of the nodes of a network into regions of uniform density. The decomposition we define is unique in the sense that a given network has exactly one density decomposition. The number of nodes in each partition defines a density distribution which we find is measurably similar to the degree distribution of given real networks (social, internet, etc.) and measurably dissimilar in synthetic networks (preferential attachment, small world, etc.). | cs.SI | cs | Density decompositions of networks
Glencora Borradaile1, Theresa Migler2, and Gordon Wilfong3
1 School of EECS Oregon State University
2 Cal Poly, San Luis Obispo
3 Bell Labs
Abstract. We introduce a new topological descriptor of a network called
the density decomposition which is a partition of the nodes of a network
into regions of uniform density. The decomposition we define is unique
in the sense that a given network has exactly one density decomposition.
The number of nodes in each partition defines a density distribution
which we find is measurably similar to the degree distribution of given
real networks (social, internet, etc.) and measurably dissimilar in syn-
thetic networks (preferential attachment, small world, etc.).
1
Introduction
A better understanding of the topological properties of real networks can be
advantageous for two major reasons. First, knowing that a network has certain
properties, e.g., bounded degree or planarity, can sometimes allow for the design
of more efficient algorithms for extracting information about the network or
for the design of more efficient distributed protocols to run on the network.
Second, it can lead to methods for synthesizing artificial networks that more
correctly match the properties of real networks thus allowing for more accurate
predictions of future growth of the network and more accurate simulations of
distributed protocols running on such a network.
We show that networks decompose naturally into regions of uniform density,
a density decomposition. The decomposition we define is unique in the sense that
a given network has exactly one density decomposition. The number of nodes in
each region defines a distribution of the nodes according to the density of the
region to which they belong, that is, a density distribution (Section 2). Although
density is closely related to degree, we find that the density distribution of a
particular network is not necessarily similar to the degree distribution of that
network. For example, in many synthetic networks, such as those generated by
popular network models (e.g. preferential attachment and small worlds), the
density distribution is very different from the degree distribution (Section 3.1).
On the other hand, for all of the real networks (social, internet, etc.) in our
data set, the density and degree distributions are measurably similar (Section 3).
Similar conclusions can be drawn using the notion of k-cores [40], but this suffers
from some drawbacks which we discuss in Section 2.3.
7
1
0
2
c
e
D
5
1
]
I
S
.
s
c
[
3
v
1
0
0
1
.
5
0
4
1
:
v
i
X
r
a
2
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
1.1 Related work
We obtain the density decomposition of a given undirected network by first
orienting the edges of this network in an egalitarian4 manner. Then we partition
the nodes based on their indegree and connectivity in this orientation.
Fair orientations have been studied frequently in the past. These orientations
are motivated by many problems. One such motivating problem is the following
telecommunications network problem: Source-sink pairs (si, ti) are linked by a
directed si-to-ti path ci (called a circuit). When an edge of the network fails, all
circuits using that edge fail and must be rerouted. For each failed circuit, the
responsibility for finding an alternate path is assigned to either the source or
sink corresponding to that circuit. To limit the rerouting load of any vertex, it
is desirable to minimize the maximum number of circuits for which any vertex
is responsible. Venkateswaran models this problem with an undirected graph
whose vertices are the sources and sinks and whose edges are the circuits. He
assigns the responsibility of a circuit's potential failure by orienting the edge to
either the source or the sink of this circuit. Minimizing the maximum number of
circuits for which any vertex is responsible can thus be achieved by finding an
orientation that minimizes the maximum indegree of any vertex. Venkateswaran
shows how to find such an orientation [42]. Asahiro, Miyano, Ono, and Zenmyo
consider the edge-weighted version of this problem [3]. They give a combina-
, (2 − )}-approximation algorithm where wmax and wmin are the
torial { wmax
maximum and minimum weights of edges respectively, and is a constant which
depends on the input [3]. Klostermeyer considers the problem of reorienting
edges (rather than whole paths) so as to create graphs with given properties,
such as strongly connected graphs and acyclic graphs [23]. De Fraysseix and de
Mendez show that they can find an indegree assignment of the vertices given
a particular properties [15]. Biedl, Chan, Ganjali, Hajiaghayi, and Wood give
a 13
8 -approximation algorithm for finding an ordering of the vertices such that
for each vertex v, the neighbors of v are as evenly distributed to the right and
left of v as possible [10]. For the purpose of deadlock prevention [45], Wittorff
describes a heuristic for finding an acyclic orientation that minimizes the sum
over all vertices of the function δ(v) choose 2, where δ(v) is the indegree of vertex
v. This objective function is motivated by a problem concerned with resolving
deadlocks in communications networks [46].
wmin
In our work we show that the density decomposition can isolate the densest
subgraph. The densest subgraph problem has been studied a great deal. Goldberg
gives an algorithm to find the densest subgraph in polynomial time using network
flow techniques [18]. There is a 2-approximation for this problem that runs in
linear time [12]. As a consequence of our decomposition, we find a subgraph
that has density no less than the density of the densest subgraph less one. There
are algorithms to find dense subgraphs in the streaming model [4,17]. There are
algorithms that find all densest subgraphs in a graph (there could be many such
subgraphs) [39].
4 An egalitarian orientation is one in which the indegrees of the nodes are as balanced
as possible as allowed by the topology of the network.
Density decompositions of networks
3
We consider many varied real networks in our study of the density decom-
position. We find our results to be consistent across biological, technical, and
social networks.
2 The density decomposition
In order to obtain the density decomposition of a given undirected network we
first orient the edges of this network in an egalitarian manner. Then we partition
the nodes based on their indegree and connectivity in this orientation.
The following procedure, the Path-Reversal algorithm, finds an egalitarian
orientation [11]. A reversible path is a directed path from a node v to a node u
such that the indegree of v, δ(v), is at least greater than the indegree of u plus
one: δ(v) > δ(u) + 1
Arbitrarily orient the edges of the network.
While there is a reversible path
reverse this path.
Since we are only reversing paths between nodes with differences in indegree
of at least 2, this procedure converges; the running time of this algorithm is
quadratic [11]. The orientation resulting from this termination condition sug-
gests a hierarchical decomposition of its nodes: Let k be the maximum indegree
in the orientation. Ring k, denoted Rk, contains all nodes of indegree k and
all nodes that reach nodes of indegree k. By the termination condition of the
above procedure, only nodes of indegree k or k − 1 are in Rk. Iteratively, given
Rk, Rk−1, . . . , and Ri+1, Ri contains all the remaining nodes with indegree i
along with all the remaining nodes that reach nodes with indegree i. Nodes in
Ri must have indegree i or i − 1 by the termination condition of the procedure.
By this definition, an edge between a node in Ri and a node in Rj is directed
from Ri to Rj when i > j and all the isolated nodes are in R0.
to number of nodes (
Density can be defined in two ways: either as the ratio of number of edges
E
V ) or as the ratio of number of edges to total number
2E
of possible edges (
V (V −1) ). In this discussion we use the former definition.
This definition of density is closely related to node degree (the number of edges
adjacent to a given node): the density of a network is equal to half the average
total degree.
We identify a set S of nodes in a graph by merging all the nodes in S into a
single node s and removing any self-loops (corresponding to edges of the graph
both of whose endpoints were in S). Our partition Rk, Rk−1, . . . , R0 induces
regions of uniform density in the following sense:
Density Property For any i = 0, . . . , k, identifying the nodes in ∪j>iRj and
deleting the nodes in ∪j<iRj leaves a network G whose
density is in the range (i−1, i] (for Ri sufficiently large).
4
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
In particular, Rk isolates a densest region in the network. Consider the net-
work Gi formed by identifying the nodes ∪j>iRj and deleting the nodes in
∪j<iRj; this network has one node (resulting from identifying the nodes ∪j>iRj)
of indegree 0 and Ri nodes of indegree i of i − 1, at least one of which must
have indegree i. Therefore, for any i, the density of Gi is at most i and density
at least
Ri(cid:29)i
−−−−→ i − 1.
(Ri − 1)(i − 1) + i
Ri + 1
In Section 2.1, we observe that this relationship between density and this
decomposition is much stronger.
2.1 Density and the Density Decomposition
In this section we discuss the following three properties:
Property D1 The density of a densest subnetwork is at most k. That is, there
is no denser region Rj for j > k.
Property D2 The density decomposition of a network is unique and does not
depend on the starting orientation.
Property D3 Every densest subnetwork contains only nodes of Rk.
These properties allow us to unequivocally describe the density structure of
a network. We summarize the density decomposition by the density distribution:
(R0,R1, . . .Rk−1,Rk), i.e. the number of nodes in each region of uniform
density. We will refer to a node in Ri as having density rank i.
The subnetwork of a network G induced by a subset S of the nodes of G is
defined as the set of nodes S and the subset of edges of G whose endpoints are
both in S; we denote this by G[S]. First we will note that both the densest sub-
network and the subnetwork induced by the nodes of highest rank have density
between k − 1 and k. Recall that k is the maximum indegree of a node in an
egalitarian orientation of G and that Ri is the set of nodes in the ith ring of the
density decomposition. We will refer to Rk as the densest ring.
We use the following two lemmas to prove Property D1.
Lemma 1. The density of the subnetwork induced by the nodes in Rk is in the
range (k − 1, k].
We could prove Lemma 5 directly with a simple counting argument on the in-
degrees of nodes in Rk (see Appendix A) or by using a network flow construction
similar to Goldberg's and the max flow-min cut theorem [18].
Lemma 2. The density of a densest subnetwork is in the range (k − 1, k].
The upper bound given in Lemma 6 may be proven directly by using a
counting argument for the indegrees of vertices in an egalitarian orientation of
the densest subnetwork (see Appendix A) or by using the relationship between
the density of the maximum density subgraph and the psuedoarboricity [24].
Density decompositions of networks
5
This upper bound proves Property D1 of the density decomposition. Prop-
erty D1 has been proven in another context. It follows from a theorem of Frank
and Gy´arf´as [14] that if (cid:96) is the maximum outdegree in an orientation that min-
imizes the maximum outdegree then the density of the network, d, is such that
(cid:100)d(cid:101) ≤ (cid:96).
Corollary 1. The subgraph induced by the nodes of Rk is at least as dense as
the density of the densest subgraph less one.
Note that the partition of the rings does not rely on the initial orientation,
or, more strongly, nodes are uniquely partitioned into rings, giving Property D2.
Theorem 1. The density decomposition is unique.
We can prove this by noting that the maximum indegree of two egalitarian
orientations for a given network is the same [11,3,42]. For a contradiction, we
consider two different egalitarian orientations of the same graph that yield two
distinct density decompositions. We then compare corresponding rings in each
orientation and find that they are in fact the same (see Appendix A).
The following theorem relies on the fact that the density decomposition is
unique and proves Property D3.
Theorem 2. The densest subnetwork of a network G is induced by a subset of
the nodes in the densest ring of G.
We could prove Theorem 6 directly by comparing the density of the subgraph
induced by the vertices in the densest subgraph intersected with the vertices in
Rk and the density of the densest subgraph (see Appendix A). Or we could use
integer parameterized max flow techniques [3,16].
Note that there are indeed cases where the densest subgraph is induced by
a strict subset of nodes in the top ring. For example, consider the graph, G,
consisting of K3 and K4 with a single edge connecting the two cliques. K4 is the
densest subgraph in G, however all of G is contained in the top ring (R2).
See Figure 1 for an example.
Fig. 1: The orientation shown is an egalitarian orientation. In this graph all nodes
are in the top ring. However, only the nodes in K4 are in the densest subgraph.
6
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
2.2 Interpretation of density rank
We can interpret orientations as assigning responsibility: if an edge is oriented
from node a to node b, we can view node b as being responsible for that con-
nection. Indeed several allocation problems are modelled this way [11,2,42,3,19].
Put another way, we can view a node as wishing to shirk as many of its du-
ties (modelled by incident edges) by assigning these duties to its neighbors (by
orienting the linking edge away from itself). Of course, every node wishes to
shirk as many of its duties as possible. However, the topology of the network
may prevent a node from shirking too many of its duties. In fact, the egalitarian
orientation is the assignment in which every node is allowed to simultaneously
shirk as many duties as allowed by the topology of the network. An example is
given in Figure 2; although nodes a and b both have degree 7, in the star network
(left) a can shirk all of its duties, but in the clique network (right) b can only
shirk half of its duties. There is a clear difference between these two cases that
is captured by the density rank of a and b that is not captured by the degree
of a and b. For example, if these were co-authorship networks, the star network
may represent a network in which author a only co-authors papers with authors
who never work with anyone else whereas the clique network shows that author
b co-authors with authors who also collaborate with others. One may surmise
that the work of author b is more reliable or respected than the work of author
a.
Fig. 2: Two egalitarian orientations for networks with 8 nodes. This example
generalizes to any number of nodes (Theorem 7).
Theorem 3. For a clique on n nodes, there is an orientation where each node
has indegree either (cid:98)n/2(cid:99) or (cid:98)n/2(cid:99) − 1.
A proof for Theorem 7 can be given by construction of such an egalitarian ori-
entation (see Appendix A) or by using a non-linear programming approach [36].
ab Density decompositions of networks
7
2.3 Relationship to k-cores
A k-core of a network is the maximal subnetwork whose nodes all have degree
at least k [40]. A k-core is found by repeatedly deleting nodes of degree less than
k while possible. For increasing values of k, the k-cores form a nesting hierarchy
(akin to our density decomposition) of subnetworks H0, H1, . . . , Hp where Hi is
an i-core and p is the smallest integer such that G has an empty (p + 1)-core.
For networks generated by the Gn,p model, most nodes are in the p-core [29,37]
For the preferential attachment model, all nodes except the initial nodes belong
to the c-core, where c is the number of edges connecting to each new node [1].
These observations are similar to those we find for the density distribution
(Section 3) and many of the observations we make regarding the similarity of
the degree and density distributions of real networks also hold for k-core de-
compositions [33]. However, the local definition of cores (depending only on the
degree of a node) provides a much looser connection to density than the density
decomposition, as we make formal in Lemma 7.
The density of the top core may be less then the density of the top ring. Also,
there are graphs for which the densest subgraph is not contained in the top core
(see Appendix B).
Lemma 3. Given a core decomposition H0, H1, . . . , Hk of a network, the sub-
network formed by identifying the nodes in ∪j>iHj and deleting the nodes in
∪j<iHj has density in the range [ i
2 , i) for Hi sufficiently large.
The proof of the above lemma can be found in Appendix B.
3 The similarity of degree and density distributions
The normalized density ρ and degree δ distributions for three networks (AS
2013, PHYS 2005, and DBLP) are given in Figure 3, illustrating the similarity
of the distributions. We quantify the similarity between the density and degree
distributions of these networks using the Bhattacharyya coefficient, β [8]. For
two normalized p and q, the Bhattacharyya coefficient is:
(cid:88)
i
β(p, q) =
√
pi · qi.
β(p, q) ∈ [0, 1] for normalized, positive distributions; β(p, q) = 0 if and only if
p and q are disjoint; β(p, q) = 1 if and only if p = q. We denote the Bhat-
tacharyya coefficient comparing the normalized density ρ and degree δ distribu-
tions, β(ρ, δ) for a network G by βρδ(G). Specifically,
(cid:88)
(cid:112)ρi · δi
βρδ(G) = β(ρ, δ) =
where ρi is the fraction of nodes in the ith ring of the density decomposition
of G and δi is the fraction of nodes of total degree i in G; we take ρi = 0 for
i
8
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Fig. 3: In the AS network nodes represent autonomous systems and two au-
tonomous systems are connected if there is a routing agreement between
them [49] (44,729 nodes and 170,735 edges). In the PHYS network nodes repre-
sent condensed matter physicists and two physicists are connected if they have at
least one co-authored paper [34] (40,421 nodes and 175,692 edges). In the DBLP
network, nodes represent computer scientists and two computer scientists are
connected if they have at least one co-authored paper [48] (317,080 nodes and
1,049,866 edges). The (truncated) normalized density and degree distributions
are displayed. The degree distributions have long diminishing tails. AS 2013 has
67 non-empty rings, but rings 31 through 66 contain less than 1.5% of the nodes;
ring 67 contains 0.75% of the nodes. DBLP has 4 non-empty rings denser than
ring 30 that are disconnected; rings 32, 40, 52 and 58 contain 0.02%, 0.01%,
0.03% and 0.04% of the nodes, respectively.
i > k where k is the maximum ring index. Refer to Figure 4. For all the networks
in our data set, βρδ > 0.78. Note that if we exclude the Gnutella and Amazon
networks, βρδ > 0.9. We point out that the other networks are self-determining
in that each relationship is determined by at least one of the parties involved.
On the other hand, the Gnutella network is highly structured and designed and
the Amazon network is a is a one-mode projection of the buyer-product network
(which is in turn self-determining).
Perhaps this is not surprising, given the close relationship between density
and degree; one may posit that the density distribution ρ simply bins the degree
distribution δ. However, note that a node's degree is its total degree in the
undirected graph, whereas a node's rank is within one of its indegree in an
egalitarian orientation. Since the total indegree to be shared amongst all the
nodes is half the total degree of the network, we might assume that, if the
density distribution is a binning of the degree distribution, the density rank of a
node of degree d would be roughly d/2. That is, we may expect that the density
05101520253000.10.20.30.4 05101520253000.050.10.150.2 05101520253000.10.20.30.4 AS 2013 density distributionAS 2013 degree distributionDBLP density distributionDBLP degree distributionPHYS 2005 density distributionPHYS 2005 degree distribution distribution is halved in range and doubled in magnitude (ρi ≈ 2δ2i). If this is
the case, then
Density decompositions of networks
9
If we additionally assume that our network has a power-law degree distribution
such as δx ∝ 1/x3,
(cid:112)ρiδi ≈(cid:88)
β(ρ, δ) ≈(cid:88)
(cid:115)
(cid:90) ∞
(cid:18)
d
d
(cid:112)
(cid:19)
β(ρ, δ) ≈
1
2
x3
2
2
(2x)3
2δdδ2d.
dx = 0.5
(after normalizing the distributions and using a continuous approximation of β).
Even with these idealized assumptions, this does not come close to explaining
βρδ being in excess of 0.78 for the networks in our data set. Further, for many
synthetic networks βρδ is small, as we discuss in the next section. We note that
this separation between similarities of density and degree distributions for the
empirical networks and synthetic networks can be illustrated with almost any
divergence or similarity measure for a pair of distributions.
3.1 The dissimilarity of degree and density distributions of random
networks
In contrast to the measurably similar degree and density distributions of real
networks, the degree and density distributions are measurably dissimilar for
networks produced by many common random network models; including the
preferential attachment (PA) model of Barabasi and Albert [6] and the small
world (SW) model of Watts and Strogatz [44]. We use βρδ(M ) to denote the
Bhattacharyya coefficient comparing the expected degree and density distribu-
tions of a network generated by a model M .
Preferential attachment networks In the PA model, a small number, n0, of nodes
seed the network and nodes are added iteratively, each attaching to a fixed
number, c, of existing nodes. Consider the orientation where each added edge
is directed toward the newly added node; in the resulting orientation, all but
the n0 seed nodes have indegree c and the maximum indegree is c. At most cn0
path reversals will make this orientation egalitarian, and, since cn0 is typically
very small compared to n (the total number of nodes), most of the nodes will
remain in the densest ring Rc. Therefore PA networks have nearly-trivial density
distributions: ρc ≈ 1. On the other hand the expected fraction of degree c nodes
is 2/(c + 2) [5]. Therefore βρδ(PA) ≈(cid:112)2/(c + 2).
Small-world networks A small-world network is one generated from a d-regular
networkby reconnecting (uniformly at random) at least one endpoint of every
edge with some probability. For probabilities close to 0, a network generated in
this way is close to d-regular; for probabilities close to 1, a network generated this
way approaches one generated by the random-network model (Gn,p) of Erdos
10
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Fig. 4: Similarity (βρδ) of density and degree distributions for 9 diverse networks.
We introduced AS, PHYS, and DBLP in Figure 3. In the EMAIL network nodes
represent Enron email addresses and two addresses are connected if there has
been at least one email exchanged between them [22] (36,692 nodes and 183,831
edges). In the TRUST network nodes represent epinions.com members and two
members are connected if one trusts the other [41] (75,879 nodes and 405,740
edges). In the SDOT network nodes represent slashdot.org members and two
members are connected if they are friends or foes [28] (82,168 nodes and 504,
230 edges). In the WIKI network nodes represent wikipedia.org users and two
users are connected if one has voted for the other to be in an administrative
role [26] (7,115 nodes and 103,689 edges). In the Amazon network nodes repre-
sent products and two products are connected if they are frequently purchased
together [48] (334,863 nodes and 925,872 edges). In the Gnutella network nodes
represent network hosts and two hosts are connected if they share files [38]
(22,687 nodes and 54,704 edges). EMAIL , TRUST, and WIKI are naturally
directed networks. For these networks, we ignore direction and study the under-
lying undirected networks. Notice that both the Amazon and Gnutella networks
are highly structured. It is not surprising that these networks would have a
weaker connection between the density and degree distributions.
and R´enyi [13]. In the first extreme, βρδ(SW ) = 0 (Lemma 4 below) because
all the nodes have the same degree and the same rank. As the reconnection
probability increases, nodes are not very likely to change rank while the degree
distribution spreads slightly. In the second extreme, the highest rank of a node
is (cid:98)c/2(cid:99) + 1 [47] and, using an observation of the expected size of the densest
subnetwork5, with high probability nearly all the nodes have this rank. It follows
5 e-mail exchange between Glencora Borradaile and Abbas Mehrabian
1041050.750.80.850.90.951Network sizeBhattacharyya coefficient ASPHYSDBLPEMAILTRUSTSDOTWIKIAmazonGnutella Density decompositions of networks
11
(cid:115)
cc/2
e−c(c/2)!
,
that
βρδ(Gn,p) ≈
which approaches 0 very quickly as c grows. We verified this experimentally
finding that βρδ(Gn,p) < 0.5 for c ≥ 5.
Lemma 4. For d ≥ 3, βρδ(G) = 0 for any d-regular network G with d ≥ 3.
Proof. We argue that ρd = 0, proving the lemma since δd = 1 for a d-regular
network. For a contradiction, suppose ρd > 0. Then Rd = x for some x > 0,
where Rd is the set of nodes of G in the dth ring of G's density decomposition.
Note that the highest rank node in G has rank at most d, since there are no
nodes with degree > d. Let H be the subnetwork of H containing all the nodes
of Rd and all the edges of G both of whose endpoints are in Rd. H has at least
one node of indegree d and all other nodes have indegree at least d− 1; therefore
H must have at least d + (x − 1)(d − 1) edges. On the other hand, the total
degree of every node in H is at most d, so H has at most dx/2 edges. We must
have d + (x − 1)(d − 1) ≤ dx/2, which is a contradiction for d ≥ 3 and x > 0.
4 Conclusion
We have introduced the density decomposition and summarized the decomposi-
tion with the density distribution. We found that the hierarchy of vertices within
this decomposition are partitioned according to the density of the induced sub-
graphs. We found that the density and degree distributions are remarkably simi-
lar in real graphs and dissimilar in synthetic networks. In the appendix we discuss
using the density distribution to build more realistic random graph models.
References
1. Jose Alvarez-Hamelin, Luca DallAsta, Alain Barrat, and Alessandro Vespignani. k-
core decomposition of Internet graphs: hierarchies, self-similarity and measurement
biases. Networks and Heterogeneous Media, 3(2):371, 2008.
2. Yuichi Asahiro, Jesper Jansson, Eiji Miyano, and Hirotaka Ono. Upper and lower
degree bounded graph orientation with minimum penalty. In Proceedings in Com-
puting: The Australasian Theory Symposium, pages 139–146, 2012.
3. Yuichi Asahiro, Eiji Miyano, Hirotaka Ono, and Kouhei Zenmyo. Graph orien-
tation algorithms to minimize the maximum outdegree. International Journal of
Foundations of Computer Science, 18(2):197–215, 2007.
4. Bahman Bahmani, Ravi Kumar, and Sergei Vassilvitskii. Densest subgraph in
streaming and mapreduce. Proc. VLDB Endow., 5(5):454–465, January 2012.
5. Albert-L´aszl´o Barab´asi. Network Science. Cambridge University Press, Cambridge,
2016.
6. Albert-L´aszl´o Barab´asi and R´eka Albert. Emergence of scaling in random networks.
Science, 286(5439):509–512, 1999.
12
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
7. Vladimir Batagelj and Matjaz Zaversnik. An O(m) algorithm for cores decom-
position of networks. Advances of Data Analysis and Classification, 5:129–145,
2011.
8. Anil Kumar Bhattacharyya. On a measure of divergence between two statistical
populations defined by their probability distributions. Bull. Calcutta Math. Soc.,
35:99–109, 1943.
9. Ginestra Bianconi and Albert-Laszlo Barabsi. Competition and multiscaling in
evolving networks. EPL (Europhysics Letters), 54(4):436, 2001.
10. Therese Biedl, Timothy Chan, Yashar Ganjali, Mohammad Taghi Hajiaghayi, and
David R. Wood. Balanced vertex-orderings of graphs. Discrete Appl. Math.,
148:27–48, April 2005.
11. Glencora Borradaile, Jennifer Iglesias, Theresa Migler, Antonio Ochoa, Gordon
In Journal of Graph
Wilfong, and Lisa Zhang. Egalitarian graph orientations.
Algorithms and Applications, volume 21, pages 687–708, 2017.
12. Moses Charikar. Greedy approximation algorithms for finding dense components
in a graph. In Proceedings of the Third International Workshop on Approxima-
tion Algorithms for Combinatorial Optimization, pages 84–95, London, UK, 2000.
Springer-Verlag.
13. Paul Erdos and Alfr´ed R´enyi. On random graphs, I. Publicationes Mathematicae
(Debrecen), 6:290–297, 1959.
14. Andr´as Frank and A. Gy´arf´as. How to orient the edges of a graph? Colloquia
Mathematica Societatis J´anos Bolyai, 1:353–364, 1976.
15. Hubert de Fraysseix and Patrice Ossona de Mendez. Regular orientations, arboric-
ity, and augmentation. In Proceedings of the DIMACS International Workshop on
Graph Drawing, pages 111–118, London, UK, 1995. Springer-Verlag.
16. Giorgio Gallo, Michael Grigoriadis, and Robert Tarjan. A fast parametric maxi-
mum flow algorithm and applications. SIAM Journal on Computing, 18(1):30–55,
February 1989.
17. David Gibson, Ravi Kumar, and Andrew Tomkins. Discovering large dense sub-
graphs in massive graphs. In Proceedings of the 31st international conference on
very large data bases, pages 721–732. VLDB Endowment, 2005.
18. Andrew Goldberg. Finding a maximum density subgraph. Technical report, Uni-
versity of California at Berkeley, Berkeley, CA, USA, 1984.
19. Nicholas J. A. Harvey, Richard E. Ladner, L´aszl´o Lov´asz, and Tami Tamir. Semi-
matchings for bipartite graphs and load balancing. Journal of Algorithms, 59:53–
78, 2006.
20. Mark Jerrum and Gregory B. Sorkin. The metropolis algorithm for graph bisection.
Discrete Applied Mathematics, 82(13):155 – 175, 1998.
21. Jeong Han Kim and Van H. Vu. Generating random regular graphs. In Proceedings
of the thirty-fifth annual ACM symposium on Theory of computing, STOC '03,
pages 213–222, New York, NY, USA, 2003. ACM.
22. Bryan Klimt and Yiming Yang. Introducing the Enron Corpus. In First Conference
on Email and Anti-Spam, 2004.
23. William F. Klostermeyer. Pushing vertices and orienting edges. Ars Combinatorial,
24.
51:65–75, 1999.
(cid:32)Lukasz Kowalik. Approximation scheme for lowest outdegree orientation and graph
density measures.
In Proceedings of the 17th International Conference on Al-
gorithms and Computation, ISAAC'06, pages 557–566, Berlin, Heidelberg, 2006.
Springer-Verlag.
Density decompositions of networks
13
25. Silvio Lattanzi and D. Sivakumar. Affiliation networks.
In Proceedings of the
Forty-first Annual ACM Symposium on Theory of Computing, STOC '09, pages
427–434, New York, NY, USA, 2009. ACM.
26. Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. Signed networks in social
media.
In Proceedings of the SIGCHI Conference on Human Factors in Com-
puting Systems, CHI '10, pages 1361–1370, New York, NY, USA, 2010. ACM.
http://snap.stanford.edu/data/.
27. Jure Leskovec, Jon Kleinberg, and Christos Faloutsos. Graphs over time: Den-
sification laws, shrinking diameters and possible explanations. In Proceedings of
the eleventh ACM SIGKDD International Conference on Knowledge Discovery in
Data Mining, pages 177–187. ACM Press, 2005.
28. Jure Leskovec, Kevin Lang, Anirban Dasgupta, and Michael Mahoney. Community
structure in large networks: Natural cluster sizes and the absence of large well-
defined clusters. CoRR, abs/0810.1355, 2008. http://snap.stanford.edu/data/.
29. Tomasz (cid:32)Luczak. Size and connectivity of the k-core of a random graph. Discrete
Mathematics, 91(1):61–68, August 1991.
30. David W. Matula and Leland L. Beck. Smallest-last ordering and clustering and
graph coloring algorithms. J. ACM, 30(3):417–427, July 1983.
31. Julian Mcauley and Jure Leskovec. Discovering social circles in ego networks. ACM
Transactions on Knowledge Discovery from Data, 8(1):4:1–4:28, February 2014.
32. Brendan D. McKay and Nicholas C. Wormald. Uniform generation of random
regular graphs of moderate degree. J. Algorithms, 11(1):52–67, February 1990.
33. Theresa Migler. The Density Signature. PhD thesis, Oregon State University, 2014.
34. Mark Newman.
structure
http://www-
algorithm for detecting
community
Fast
in networks.
Phys. Rev. E,
personal.umich.edu/mejn/netdata/.
69:066133,
Jun 2004.
35. Mark Newman. Networks: An Introduction. Oxford University Press, Inc., New
York, NY, USA, 2010.
36. Jean-Claude Picard and Maurice Queyranne. A network flow solution to some
nonlinear 0-1 programming problems, with applications to graph theory. Networks,
12:141–159, 1982.
37. Boris Pittel, Joel Spencer, and Nicholas Wormald. Sudden emergence of a giant
k-core in a random graph. J. Comb. Theory Ser. B, 67(1):111–151, May 1996.
38. Matei Ripeanu, Ian Foster, and Adriana Iamnitchi. Mapping the Gnutella network:
Properties of large-scale peer-to-peer systems and implications for system design.
IEEE Internet Computing Journal, 6:2002, 2002. http://snap.stanford.edu/data/.
39. Barna Saha, Allison Hoch, Samir Khuller, Louiqa Raschid, and Xiao-Ning Zhang.
Dense subgraphs with restrictions and applications to gene annotation graphs. In
Proceedings of the 14th Annual international conference on Research in Computa-
tional Molecular Biology, RECOMB'10, pages 456–472, Berlin, Heidelberg, 2010.
Springer-Verlag.
40. Stephen Seidman. Network structure and minimum degree. Social Networks,
5(3):269–287, 1983.
41. Maryam Tahajod, Azadeh Iranmehr, and Nasim Khozooyi. Trust manage-
ment for semantic web.
In Computer and Electrical Engineering, 2009.
ICCEE '09. Second International Conference, volume 2, pages 3–6, 2009.
http://snap.stanford.edu/data/.
42. Venkat Venkateswaran. Minimizing maximum indegree. Discrete Appl. Math.,
143:374–378, September 2004.
14
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
43. Fabien Viger and Matthieu Latapy. Efficient and simple generation of random
simple connected graphs with prescribed degree sequence.
In The Eleventh In-
ternational Computing and Combinatorics Conference, pages 440–449. Springer,
2005.
44. Duncan Watts and Steven Strogatz. Collective dynamics of 'small-world' networks.
Nature, 393(6684):409–10, 1998.
45. W. Wimmer.
Ein Verfahren
in
mungen
verfahren-zur-verhinderung-von-verklemmungen-in-vermittlernetzen/,
October 1978.
Vermittlernetzen.
zur Verhinderung
von Verklem-
http://www.worldcat.org/title/
46. Vaughan Wittorff. Implementation of constraints to ensure deadlock avoidance in
networks, 2009. US Patent # 7,532,584 B2.
47. Pu Gao Xavier P´erez-Gim´enez and Cristiane Sato. Arboricity and spanning-tree
packing in random graphs with an application to load balancing. In Proceedings of
the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
'14, pages 317–326. SIAM, 2014.
48. Jaewon Yang and Jure Leskovec. Defining and evaluating network communities
based on ground-truth. In Proceedings of the ACM SIGKDD Workshop on Min-
ing Data Semantics, MDS '12, pages 3:1–3:8, New York, NY, USA, 2012. ACM.
http://snap.stanford.edu/data/.
49. Yu Zhang. Internet AS-level Topology Archive. http://irl.cs.ucla.edu/topology/.
Density decompositions of networks
15
A Proofs
In this section we include proofs for theorems in Sections 2.1 and 2.2.
Lemma 5. The density of the subnetwork induced by the nodes in Rk is in the
range (k − 1, k].
Proof. All nodes in Rk have indegree k or k − 1 in G. Since any edge incident to
a node in Rk but not in G[Rk] is directed out of Rk in G, the indegree of every
node in G[Rk] is k or k − 1. Let nk be the number of nodes of indegree k in
G[Rk] and nk−1 be the number of nodes of degree k − 1 in G[Rk−1]. Therefore,
the number of edges in G[Rk] is knk + (k − 1)nk−1 and:
knk + (k − 1)nk−1
≤ k
density(G[Rk]) =
nk + nk−1
Since there is at least one node of indegree k in G[Rk], nk > 0. Therefore:
knk + (k − 1)nk−1
nk + nk−1
=
(k − 1)(nk + nk−1) + nk
nk + nk−1
> k − 1
Lemma 6. The density of a densest subnetwork is in the range (k − 1, k].
Proof. Let H be a densest subnetwork and let each edge of H inherit the orien-
tation of the same edge in an egalitarian orientation of G. Every node of H has
indegree at most k (when restricted to H). Therefore
density(H) ≤ nH k
nH
≤ k
where nH is the number of nodes in H. Furthermore, by Lemma 5, the density
of G[Rk] is greater than k − 1 and so the densest subnetwork must be at least
this dense.
Theorem 4. The density decomposition is unique.
Proof. The maximum indegree of two egalitarian orientations for a given net-
work is the same [11,3,42]. Suppose, for a contradiction, that there are two
egalitarian orientations (red and blue) for G, resulting in density decomposi-
tions R0, R1, . . . Rk and B0, B1, . . . Bk, respectively. Let i be the largest index
such that Ri (cid:54)= Bi.
We compare the orientation of the edges with one endpoint in S = Ri \ Bi
between the two orientations (illustrated in Figure 5). Since the orientations are
egalitarian:
1. All the edges between Bi and S are directed into S in the blue orientation.
2. All the edges between S and {∪k
j=i+1Rj}\ S are directed into S with respect
to both red and blue orientations.
j=0Rj} \ S are directed out of S with respect to
3. All edges between S and {∪i−1
the red orientation.
16
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Fig. 5: Edges incident to S (grey region) in the proof of Theorem 4 in red and
blue orientations. S is the region in gray. Unoriented edges indicate that the
orientation could be in either direction.
Based on these orientations, we have:
Observation 5 The number of edges directed into S in the blue orientation is
at least the number of edges directed into S in the red orientation.
We will show that Ri ⊆ Bi; symmetrically Bi ⊆ Ri, completing the theorem.
With respect to the blue orientation, all nodes in S have indegree strictly less
than i. Further, by Observation 5, the total indegree shared amongst the nodes
in S with respect to the red orientation is at most that of the blue orientation.
Since all nodes in S have indegree i or i − 1 with respect to the red orientation,
and, by Observation 5, the total indegree shared amongst the nodes in S with
respect to the red orientation is at most that of the blue orientation, all nodes
in S have indegree i − 1 with respect to the red orientation.
In order for every node in S to have indegree i − 1 in the red orientation, all
nodes that are directed into S in the blue orientation, must also be directed into
S in the red orientation; in particular this is true about the edges between S and
Ri \ S. Therefore, none of the nodes in S (which have indegree i − 1) reaches a
node of Ri \ S of indegree i with respect to the red orientation. This contradicts
the definition of Ri; therefore S must be empty.
The following theorem relies on the fact that the density decomposition is
unique and proves Property D3.
Theorem 6. The densest subnetwork of a network G is induced by a subset of
the nodes in the densest ring of G.
... Density decompositions of networks
17
Proof. First note that the densest subnetwork is an induced subnetwork, for
otherwise, the subnetwork would be avoiding including edges that would strictly
increase the density. Let S be a set of nodes that induces a densest subnetwork
of G. Consider a density decomposition of G and let k be the maximum rank of
a node in G. Let Sk = S ∩ Rk and let ¯Sk = S \ Sk.
Let A be the set of edges in G[Sk], let C be the set of edges in G[ ¯Sk], and
let B be the edges of G[S] that are neither in G[Sk] or G[ ¯Sk]. We get
B + C ≤ (k − 1) ¯Sk
(1)
because all the edges in B and C have endpoints in ¯Sk and all the nodes in ¯Sk
have indegree at most k − 1 in the egalitarian orientation of G.
density(G[S]) =
A + B + C
Sk + ¯Sk
(2)
density(G[Sk]) =
=
A
Sk
density(G[S])(Sk + ¯Sk) − (B + C)
using Equation (2)
Sk
density(G[S]) ¯Sk − (B + C)
density(G[S]) ¯Sk − (k − 1) ¯Sk
Sk
Sk
(k − 1) ¯Sk − (k − 1) ¯Sk
Sk
by Lemma 6
by Inequality (1)
= density(G[S]) +
≥ density(G[S]) +
> density(G[S]) +
= density(G[S])
Therefore, removing the nodes of G[S] that are not in Rk produces a network of
strictly greater density.
Theorem 7. For a clique on n nodes, there is an orientation where each node
has indegree either (cid:98)n/2(cid:99) or (cid:98)n/2(cid:99) − 1.
Proof. Give the nodes of the clique an ordering, v1, v2, . . . vn. Orient the edges be-
tween v1 and v2, . . . , v(cid:98)n/2(cid:99)+1 toward v1 and edges between v1 and v(cid:98)n/2(cid:99)+2, . . . vn
toward v(cid:98)n/2(cid:99)+2, . . . vn. Clearly v1 has indegree (cid:98)n/2(cid:99). Similarly, for v2: Orient
the edges between v2 and v3, . . . , v(cid:98)n/2(cid:99)+2 toward v2 and edges between v2 and
v(cid:98)n/2(cid:99)+3, . . . vn toward v(cid:98)n/2(cid:99)+3, . . . vn. Clearly v2 has indegree (cid:98)n/2(cid:99). Continue
in this fashion until vn. It is immediate that v1, v2, . . . v(cid:98)n/2(cid:99) have indegree (cid:98)n/2(cid:99).
Now for the remaining nodes: Consider vi, (cid:98)n/2(cid:99) < i ≤ n. vi has n − i incom-
ing edges from nodes vi+1, . . . vn and also i − (cid:98)n/2(cid:99) − 1 incoming edges from
v1, . . . , vi−(cid:98)n/2(cid:99)−1. Therefore vi has indegree (cid:98)n/2(cid:99) − 1. Therefore all nodes in
the clique have indegree (cid:98)n/2(cid:99) or (cid:98)n/2(cid:99) − 1. Clearly such an orientation is egal-
itarian.
18
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
B k-cores
See Figure 7 for an example of a graph in which the densest subgraph is not
contained in the top core. Further, while the core decomposition of a network
can be found in time linear in the number of edges [30,7,12] as opposed to the
quadratic time required for the density decomposition [11], core decompositions
do not lend themselves to a framework for building synthetic networks, since it
is not clear how to generate a p-core at random, whereas density decompositions
do.
(a)
(b)
(c)
Fig. 6: The top ring (Figure 6b) contains vertices I, K, L, M, C, D, E, H, F, J, N
and O but C, D, N, and O are not in the top core (Figure 6c). The density of
the subgraph induced by the top ring is 21/12 = 1.75 while the density of the
subgraph induced by the top core is 13/8 = 1.625.
BKLJDFIECANPOMHBKLJDFIECANPOMHPBKLJDFIECANOMH Density decompositions of networks
19
Fig. 7: Core 3, the top core, has density 1.5. The densest subgraph is the subgraph
induced by all vertices in core 2 and has density 1.56.
Recall that identifying the nodes in ∪j>iRj and deleting the nodes in ∪j<iRj
leaves a network G whose density is in the range (i − 1, i] (for Ri sufficiently
large). We find that the bound on density for the corresponding cores is much
looser.
Lemma 7. Given a core decomposition H0, H1, . . . , Hk of a network, the sub-
network formed by identifying the nodes in ∪j>iHj and deleting the nodes in
∪j<iHj has density in the range [ i
2 , i) for Hi sufficiently large.
Proof. Let n be the number of nodes in the described subnetwork: n = Hi + 1.
Let d be the degree of the node resulting from the identification of ∪j>iHj.
Since every node in Hi has degree at least i in the subnetwork, the density of
the subnetwork is at least
, from which the lower bound of the lemma
follows since d ≥ 0. This lower bound is also tight when Hi induces an i-regular
network.
2 (i·n+d)
n
1
Further, the i-core is witnessed by iteratively deleting nodes of degree at
most i while such nodes exist. The subnetwork will have the greatest density
(the most edges) if each deletion removes a node of degree exactly i. Then the
subnetwork has density at most i·(n−1)
.
n
20
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Fig. 8
C Random networks with given density distributions
Based on the observation that real networks have similar density and degree
decompositions and that many synthetic networks have dissimilar density and
degree decompositions, we develop an abstract model, that, given a particular
density distribution, produces a network having that density distribution (Sec-
tion C). Applied naıvely, given a density distribution of a real network, this model
generates networks with realistic average path lengths (average number of hops
between pairs of nodes) and degree distributions; that is similar to the given
real network (Section C.2). In addition to having short average path lengths,
large-scale, real networks also tend to have high clustering coefficients [35]. The
clustering coefficient of a node v is the ratio of the number of pairs of neighbors
of v that are connected to the number of pairs of neighbors of v; the clustering
coefficient of a network is the average clustering coefficient of its nodes. Our
model, naıvely applied, unfortunately, but not surprisingly, results in networks
with very low clustering coefficients. However, we show that applying the ab-
stract model in a more sophisticated manner, using ideas from the small world
model of Watts and Strogatz [44], results in much higher clustering coefficients
(Section C.3) suggesting that real networks may indeed be hierarchies of small
worlds.
Our hierarchies of small worlds specification is just one way to tune our ab-
stract model; our model is quite flexible allowing for the easy incorporation of
other network generation techniques, which we discuss at the end of this paper.
A key observation that distinguishes our model from other network models is our
qualitatively different treatment of nodes. That is, our model begins by assigning
nodes to levels of the density decomposition. This sets nodes qualitatively apart
from each other; for example, a node assigned to a dense level of the decompo-
510152025303540455000.050.10.150.20.250.30.350.40.450.5np/2Bhattacharyya coeffi(cid:14)cientrandom graphs on 10,000 vertices Density decompositions of networks
21
sition is treated very differently from a node assigned to a sparse level of the
decomposition.
Unlike our model, other network models, such as the small worlds model and
the classic random graph model (Gn,p), treat each node in the same way [44]. In
the preferential attachment model, in which nodes are attached one at a time to
some fixed number of existing nodes [6], one may argue that nodes are treated
differently since they arrive to the network at different times. However, when
each node arrives, it is treated the same way as nodes before it. Similarly more
recent network models, such as the affiliation network [25], community-guided
attachment and forest-fire models [27], nodes are not distinguished from one
another in a fixed way. We posit that in order to generate realistic networks,
in particular networks exhibiting the rich density hierarchy we have observed
in all the networks we have tested, one must assign nodes to classes and treat
those classes differently. There are existing models that treat nodes qualitatively
differently as we do. In the fitness, or Bianconi-Barab´asi model, each node has
an intrinsic value, or "fitness" [9]. As a network produced by this model evolves,
nodes with higher fitness are more likely to gain additional edges. In the planted
partition model the nodes of the graph are divided into (cid:96) groups of size (cid:96)/V ,
two nodes in the same group are connected with probability p and two nodes
from different groups are connected with probability r < p [20]. Thus nodes in
different groups are treated qualitatively differently.
Data sets We note that our conclusions on the similarity of density and degree
distributions of a given network are stronger for self-determining networks or
those networks that represent relationships, each of which is determined by at
least one of the parties in this relationship. Perhaps the clearest example of a
self-determining network is a social network in which nodes represent people and
an edge represents a friendship between two people. On the other hand a network
representing the transformers and power lines that connect them in a power grid
is clearly not self-determining as the transformers themselves do not determine
which other transformers they are connected to, but a power authority does. For
comparison, we include two non-self-determining networks.
Table 1 provides a list of the real networks that we study.
C.1 Models
Given a density distribution ρ, we can generate a network with n nodes having
this density distribution using the following abstract model:
Input: density distribution ρ and target size n
Output: a network G with n nodes and density distribution ρ
1: Initialize G to be a network with empty node set V
2: for i = ρ, . . . , 0 do
3:
4:
5:
Ri ← set of (cid:98)ρin(cid:99) nodes
add Ri to V
for each node v ∈ Ri do
22
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Self-determining networks
Nodes
autonomous systems
computer scientists
email addresses
Name
AS
DBLP
Enron
Epinions epinions.com members
Facebook Facebook user
PHYS
Slashdot slashdot.org members
Wikivote wikipedia.org users
condensed matter physicists 40,421
82,168
7,115
routing agreements
# Nodes Edges
44,729
317,080 at least one co-authored paper
36,692
75,879
4,039
at least one email exchanged
self-indicated trust
Facebook friends
at least one co-authored paper
indication of friend or foe
votes for administrator role
Non-self-determining networks
# Edges Source
170,735
1,049,866
183,831
405,740
88234
175,692
504,230
103,689
[49]
[48]
[22]
[41]
[31]
[34]
[28]
[26]
Nodes
# Nodes Edges
334,863 pairs of frequently co-purchased items 925,872
22,687
Name
Amazon products
Gnutella network hosts
Table 1: Network data sets. For naturally directed networks (Enron, Epinions
and Wikivote), we ignore the directions and study the underlying undirected
network. We likewise ignore edge annotations (e.g. friend or foe in the Slashdot
network). We use three snapshots of the AS network (from 1999, 2005 and 2011)
and three snapshots of the PHYS network (for papers posted to arxiv.org prior
to 1999, 2003 and 2005). Note that the structure of the Gnutella network is given
by external system design specifications.
connections for file sharing
54,705
# Edges Source
[48]
[38]
6:
connect i nodes of V to v
Using this generic model, we propose two specific models, the random density
distribution model (RDD) and the hierarchical small worlds model (HSW), by
specifying how the neighbors are selected in Step 6. First we show that this ab-
stract model does indeed generate a network with the given density distribution:
Lemma 8. The network resulting from the abstract model has density distribu-
tion ρ.
Proof. We argue that the orientation given by, in Step 6, directing the added
edges into v is egalitarian. For a contradiction, suppose there is a reversible path.
There must be an edge on this path from a node x to a node y such that the
in-degree of y is strictly greater than the degree of x. By construction, then,
x was added after y and so an edge between x and y must oriented into x,
contradicting the direction required by the reversible path.
Finally, since the nodes in set Ri have indegree i according to this orientation,
the orientation is a witness to a density decomposition of the given distribution.
Notice that in this construction, nodes in Ri will have indegree i while a net-
work with the same density decomposition may have nodes in Ri with indegree
i−1. We could additionally specify the number of nodes in Ri that have indegree
i and indegree i − 1; this would additionally require ensuring that there is an
Density decompositions of networks
23
egalitarian orientation in which all the nodes destined to have indegree i − 1 in
Ri reach nodes of indegree i in Ri. We believe this is needlessly over-complicated
and, indeed, over-specification that will have little affect on the generation large
realistic networks.
Further notice that this abstract model may generate a network that is not
simple. Without further constraint, in Step 6, v may connect to itself (introduc-
ing a self-loop) or to a node that v is already connected to (introducing parallel
edges). We adopt a simple technique used for generating d-regular networks [32]:
we constrain the choice in Step 6 to nodes of V that are not v itself nor neigh-
bors of v. McKay and Wormald prove this constraint still allows for uniformity
of sampling of d-regular networks when d is sufficiently small (d = O(n1/3)) [21];
likewise, since i is small compared to Ri for large networks, adopting this tech-
nique should not affect our sampling. In our two specific models, described below,
we ensure the final network will be simple using this technique.
C.2 Random density distribution model
For the RDD model, we choose i nodes from V uniformly at random in Step 6.
We use this to model four networks in our data set (AS, DBLP, EMAIL, and
TRUST). For each given network, we generate another random network hav-
ing the given network's number of nodes and density distribution. Remarkably,
although we are only specifying the distribution of the nodes over a density de-
composition, the resulting degree distributions of the RDD networks are very
similar to the original networks they are modeling. We use the Bhattacharyya
coefficient to quantify the similarity between the normalized degree distribution
of an RDD network and the normalized degree distribution of the original net-
work; we denote this by βδδ (to distinguish from our use of the Bhattacharyya
coefficient to compare degree distributions to density distributions. For all four
models, βδδ > 0.93 (Figure 9). Further, the average path lengths of the RDD net-
works are realistic, within 2 of the average path lengths of the original networks
(Figure 10).
However, the clustering coefficients of the RDD networks are unrealistically
low (Figures 9 and 10). Upon further inspection, we find that, for example, the
PHYS networks have many more edges between nodes of a common ring of
its density decomposition than between rings as compared to the corresponding
RDD model. For the RDD model, we can compute the expected fraction of edges
that will have one endpoint in Ri and one endpoint in Rj. Since there are RjRi
j>i Rj))
such edges to choose from (for j > i) and at most Ri( 1
edges between Ri and Rj (for j > i), we would expect this fraction to be:
2 (Ri− 1) +(cid:80)
2 (Ri − 1) +(cid:80)
Rj
1
2 (Ri − 1) +(cid:80)
2 (Ri − 1)
1
1
j>i Rj for j > i and
j>i Rj for i = j
(3)
In Figure 11 we plot the difference between the actual fraction of edges connect-
ing Ri to Rj in the PHYS networks with this expected fraction for all values of
j − i. We see that when j − i = 0, or for edges with both endpoints in the same
24
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
ring, there is a substantially larger number of edges in the original networks
than is being captured by our model. This provides one explanation for the low
clustering coefficients produced by the RDD model.
Fig. 9: Clustering coefficient versus similarity of degree distribution (between
models and original networks, βδδ) for RDD and HSW models. Measurements
for the SW model networks are not shown as βδδ < 0.4 for all networks generated.
Dotted lines represent the clustering coefficients for the original networks.
C.3 Hierarchical small worlds model
We provide a more sophisticated model which addresses the unrealistically low
clustering coefficients of the RDD model by generating a small world (SW)
network among the nodes of each ring of the density decomposition. Recall that
a SW network on n nodes, average degree d and randomization p network is
created as follows: order the nodes cyclically and connect each node to the d
nodes prior to it; with probability p reconnect one endpoint of each edge to
another node chosen uniformly at random. The SW model provides a trade-
off between clustering coefficient and average path length: as p increases, the
clustering coefficient and the average path length decreases [44].
In the hierarchical small worlds (HSW) model, for nodes in Ri, we create a
SW network on Ri nodes and average degree i in the same way, except for how
we reconnect each edge with probability p. For an edge uv where u is a node
0.80.820.840.860.880.90.920.940.960.98100.10.20.30.40.50.60.70.8βδ δclustering coefficient ASDBLPEMAILTRUSTRDD modelHSW model Density decompositions of networks
25
Fig. 10: Clustering coefficient versus average path length for RDD, HSW, SW
and DS models. Colors indicate the network being modelled. Squares denote the
data for the original networks.
within d nodes prior to v in the cyclic order, we select a node x uniformly at
random from ∪j>iRj and replace uv with xv. For the densest ring, we select a
node uniformly at random from the densest ring.
This process is exactly equivalent to the following: order Ri cyclically; for
each v ∈ Ri, with probability p, connect each of the i nodes before v in this
order to v; if c ≤ i neighbors for v are selected in this way, select i − c nodes
uniformly at random from ∪j>iRj (or Ri if this is the densest ring) and connect
these to v. Clearly, this is a specification of neighbor selection for Step 6 of the
abstract model.
For the AS, DBLP, EMAIL, and TRUST networks in our data set, we gener-
ate a random network according to the HSW model that is of the same size and
density distribution of the original network. We do so for p = 0.1, 0.2, . . . , 0.9. As
with the SW model, the HSW model provides a similar trade-off between clus-
tering coefficient and average path length (Figure 10), although the relationship
is less strong. In addition, we observe a similar trade-off between p and degree
distribution: as p increases, the degree distribution approaches that of the orig-
inal network (Figure 9). This is in sharp contrast to the SW model which have
degree distributions far from the original (normal vs. close to power law).
246810121400.10.20.30.40.50.60.70.8average path lengthclustering coefficient ASDBLPEMAILTRUSTRDD modelHSW modelSW modelDS model 26
Glencora Borradaile, Theresa Migler, and Gordon Wilfong
Fig. 11: Range of difference between the actual fraction of edges connecting Ri
to Rj and expected fraction (Equation (3)) over all j ≥ i as a function of j−i for
three PHYS networks. Error bars show max and min values of these differences
and dots indicate the average.
C.4 Comparing to the degree sequence model
We also compare our models (RDD and HSW) to a degree sequence (DS) model.
For a given degree distribution or sequence (assignment of degree to each node),
a DS model will generate a graph, randomly, having that degree sequence. We
use the model of Viger and Latapy which generates a connected, simple graph
by iteratively selecting neighbors for nodes (from highest remaining degree to be
satisfied to lowest) and randomly shuffling to prevent the process from getting
stuck (if no new neighbor exists that has not yet fulfilled its prescribed de-
gree) [43]. As with RDD and HSW we generate a network using this DS model
corresponding to the degree sequence of the AS, EMAIL, and TRUST networks.
The clustering coefficients of the resulting networks are much lower than in the
real networks (Figure 10); in the case of the AS network, this mismatch is less
extreme, most likely because this network has an extremely long tail with a
node with degree 4,171; many nodes would connect to these high degree nodes,
providing an opportunity for clustering. The average path lengths are close to
the original networks. Notably, the density distributions of the networks gener-
0246810121416−0.2−0.100.10.20.30.40.50.60.70.8difference in ring numberbias from expected fraction of edges 200320051999 Density decompositions of networks
27
ated by the DS model are very similar to their degree distribution, all having
βρδ > 0.9.
These observations for the DS model add evidence to our proposal that in or-
der to generate realistic networks, one must distinguish between types of nodes;
doing so results in networks that resemble real networks. However, we must note
that the DS model suffers from two drawbacks. First, the algorithms for gener-
ating such networks are much less efficient than our models (RDD and HSW,
which run in linear time); in order to guarantee simplicity and connectivity, the
reshuffling required incurs a large computational overhead, particularly when the
degree sequence includes very high degree nodes (such as in the AS network).
Second, the amount of information required to specify network generation via
the DS model is much greater than our abstract model. In the former, the de-
gree of every node must be specified, or at least the number of nodes having
each degree. For example, the SLASH network has 457 unique degrees (and a
maximum degree of 2553) while only having 61 non-empty rings in the density
decomposition.
|
1603.04466 | 1 | 1603 | 2016-03-14T20:48:43 | Sequential Voting Promotes Collective Discovery in Social Recommendation Systems | [
"cs.SI",
"cs.AI",
"cs.HC",
"cs.IR",
"physics.soc-ph"
] | One goal of online social recommendation systems is to harness the wisdom of crowds in order to identify high quality content. Yet the sequential voting mechanisms that are commonly used by these systems are at odds with existing theoretical and empirical literature on optimal aggregation. This literature suggests that sequential voting will promote herding---the tendency for individuals to copy the decisions of others around them---and hence lead to suboptimal content recommendation. Is there a problem with our practice, or a problem with our theory? Previous attempts at answering this question have been limited by a lack of objective measurements of content quality. Quality is typically defined endogenously as the popularity of content in absence of social influence. The flaw of this metric is its presupposition that the preferences of the crowd are aligned with underlying quality. Domains in which content quality can be defined exogenously and measured objectively are thus needed in order to better assess the design choices of social recommendation systems. In this work, we look to the domain of education, where content quality can be measured via how well students are able to learn from the material presented to them. Through a behavioral experiment involving a simulated massive open online course (MOOC) run on Amazon Mechanical Turk, we show that sequential voting systems can surface better content than systems that elicit independent votes. | cs.SI | cs | Sequential Voting Promotes Collective Discovery
in Social Recommendation Systems
L. Elisa Celis†, Peter M. Krafft‡, Nathan Kobe†
† ´Ecole Polytechnique F´ed´ereal de Lausanne, Switzerland
‡Massachusetts Institute of Technology, USA
[email protected], [email protected], [email protected]
6
1
0
2
r
a
M
4
1
]
I
S
.
s
c
[
1
v
6
6
4
4
0
.
3
0
6
1
:
v
i
X
r
a
Abstract
One goal of online social recommendation systems is to har-
ness the wisdom of crowds in order to identify high qual-
ity content. Yet the sequential voting mechanisms that are
commonly used by these systems are at odds with existing
theoretical and empirical literature on optimal aggregation.
This literature suggests that sequential voting will promote
herding-the tendency for individuals to copy the decisions
of others around them-and hence lead to suboptimal con-
tent recommendation. Is there a problem with our practice,
or a problem with our theory? Previous attempts at answer-
ing this question have been limited by a lack of objective
measurements of content quality. Quality is typically defined
endogenously as the popularity of content in absence of so-
cial influence. The flaw of this metric is its presupposition
that the preferences of the crowd are aligned with underlying
quality. Domains in which content quality can be defined ex-
ogenously and measured objectively are thus needed in order
to better assess the design choices of social recommendation
systems. In this work, we look to the domain of education,
where content quality can be measured via how well students
are able to learn from the material presented to them. Through
a behavioral experiment involving a simulated massive open
online course (MOOC) run on Amazon Mechanical Turk, we
show that sequential voting systems can surface better con-
tent than systems that elicit independent votes.
Introduction
Many social media sites, news aggregators, Q&A websites,
massive open online courses (MOOCs), and other online
platforms such as Amazon, Yelp and YouTube rely on so-
cial recommendation in order to aggregate the opinions of
their many users. Content is then sorted or ranked using this
information. Recommendations are solicited by, e.g., allow-
ing users to like, rate, or, when presented with items in a
ranked order, upvote content, and aggregate opinions are of-
ten displayed publicly. Such sequential voting systems, in
which individual or aggregated previous votes are publicly
displayed and continuously updated, are now widespread on
the internet.
Yet literature on collective intelligence and the wisdom
of crowds suggests that these sequential voting systems
Copyright © 2016, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
may be suboptimal at surfacing high quality content. Se-
quential voting produces an avenue through which previ-
ous decisions can influence future decisions. Social influ-
ence through this avenue has been shown to exist across
a diversity of online systems, including social recommen-
dation systems (Salganik, Dodds, and Watts 2006; Much-
nik, Aral, and Taylor 2013; van de Rijt et al. 2014). How-
ever, the classic Condorcet jury theorem and related results
suggest that independent votes will yield better aggregation
than correlated ones (Ladha 1995). Further theoretical work
on "herding" and "information cascades" also suggests that
sequential votes can lead to group convergence on subopti-
mal outcomes (Bikhchandani, Hirshleifer, and Welch 1992;
Chamley 2004). Empirical work on aggregation for point es-
timation tasks has also lent support to these theoretical re-
sults (Lorenz et al. 2011). In summary, online ratings can ul-
timately be very different when votes are sequential instead
of independent, and the prevailing hypothesis is that these
differences may be detrimental in the sequential setting.
However, recent work has begun to question the universal-
ity of this prediction. Researchers have investigated whether
social influence actually leads to miscalibrated rankings
(Stoddard 2015) as well as whether social influence can
sometimes help improve rankings (Lerman and Hogg 2014;
Abeliuk et al. 2015). The results thus far support the idea
that sequential votes may not do much harm to collective
outcomes in social recommendation systems and leave open
the possibility that mechanisms that allow social influence
may in fact accelerate the discovery of good content and al-
low for effective rankings.
In order to assess the outcomes of sequential voting sys-
tems, this line of work makes one simple but potentially
problematic assumption. They assume that content quality is
equivalent to popularity as measured by independent private
votes. However, there is no reason a priori to privilege pop-
ularity in the independent context over the sequential one,
and doing so precludes the possibility that social interaction
might ultimately surface qualitatively better content. Do-
mains in which content quality can be defined exogenously
and measured objectively are thus needed in order to better
assess the design choices of social recommendation systems.
Our Contribution
To address this problem, we select a domain in which we
have an objective measure of underlying quality: the edu-
cation domain. Here, quality is measured by how well stu-
dents are able to learn from the material presented to them.
We measure learning in two ways: by testing students on
material related to the content they consume, and by ask-
ing students to self-report their skill level before and after
consuming content.1 This gave us two exogenous metrics of
quality, which allow us to objectively compare different con-
tent ranking methods.
We then designed an experiment to test the effect of
sequential voting on the quality of surfaced content. We
compare against independent voting, in which participants
cannot view previous votes. We further benchmark against
content curated by experts and randomly selected content.
Based on the convergence of the web industry on the use
of sequential voting systems, we posit that sequential voting
can actually lead to better content ranking than independent
voting.
We find that self-reported learning of content consumers
was significantly higher when the content was chosen by
sequential votes compared to all other conditions, includ-
ing the expert condition. Consistent with this trend, average
test scores are also always higher in the sequential condition,
though not statistically so. Hence, sequential voting can al-
low for the discovery of high-quality content without expert
intervention, suggesting a beneficial effect of social influ-
ence.
Approach
Our goal is to determine if sequential voting can select
better answers in a student forum than other types of cu-
ration methods, including independent votes and expert-
grading. The type of student forum we consider is common
in MOOCs and other online Q&A sites. It begins with a
single question (often posted by a user) followed by a se-
quence of answers posted by peers. These answers are then
voted on by other students. Rather than using an existing live
MOOC, we take an approach that has been used recently in
the education literature to better isolate aspects of learning,
e.g., (Coetzee et al. 2015) and use a simulated MOOC using
workers from Amazon Mechanical Turk (AMT) as students.
In this manner we simulated a MOOC forum in which con-
tent was generated, ranked and consumed by AMT workers
as participants.
Our goal was to quantitatively measure each participant's
skill after being exposed to filtered peer-generated content.
This was measured by 1) an exam on the material and 2)
by two survey questions answered at the end of the entire
course that asked participants to self-report their knowledge
before and after course completion. Each course progressed
through three phases:
• Content generation – which simulates students answering
questions in a forum,
1Self-reported learning can be a proxy for motivation (Schunk
2003), and is correlated with actual learning (Anaya 1999).
• Vote collection – which simulates the discovery of con-
tent from the generation phase via either sequential or in-
dependent upvotes, and
• Testing – which allows us to measure students' knowl-
edge after consuming content surfaced in the voting
phase.
See the Experiment section for more details.
Importantly, as is the case in many social recommenda-
tion systems, having a content generation phase allowed us
to present material generated by peer participants. We were
then able to study the effect that sequential vs independent
voting in the collection phase had on the testing phase. Par-
ticipants were allowed to engage in at most one of the three
phases in each course in order to isolate the learning that
occurs within each phase.
We hypothesized that students shown content discovered
by sequential votes would outperform those shown con-
tent discovered by independent votes due to an underlying
herding effect that allows for greater efficiency in identify-
ing good explanations.2 In addition, we compared against
benchmarks consisting of a random condition, in which ran-
dom answers were shown, and an expert condition, in which
the authors graded each explanation and explanations with
top grades were shown.
Experiment
We ran two simulated MOOC courses; one on introductory
programming, and another on introductory art styles. Each
course had five lessons: the Computer Science course taught
Assignment, If-Else Statements, For Loops, Functions and
Recursion, while the Art course taught the defining fea-
tures of Photorealism, Realism, Expressionism, Impression-
ism, Abstract, Cubism, Color-Field, Hard-Edge, Nanga and
Mughal. The authors are experts in computer science, and
one of the authors is knowledgeable in art. The lessons and
questions for the Computer Science course were developed
jointly, and the lessons and questions for the Art course were
developed by the knowledgeable author. In order to validate
the authors' decisions, two professors in Computer Science
and one professor in Art, each with more than 15 years ex-
perience teaching introductory courses in their respective ar-
eas, were consulted, as described in the External Validators
section.
Each of the two course began with a content phase in
which 10 participants answered a set of questions for each
lesson in the course and gave explanations for their answers.
This phase simulates populating the content of a forum. This
phase was followed by a voting phase that had 4 conditions
with 25 participants in each condition. In the voting phase,
participants voted on which explanations they thought were
good. One condition in this phase had independent votes and
three conditions had votes that were public to others present
in that condition. Finally, there was a testing phase that had 6
2Formal models for describing and predicting herd behavior
have emerged in many fields (see (Raafat, Chater, and Frith 2009)
for a summary); the exact mechanism through which herding or
social influence occurs is immaterial to our study, we simply allow
for it (or not) by showing (or hiding) votes.
Figure 1: The experimental design for both courses followed
the depiction above. Each panel (representing a phase)
was run as a separate batch of experiments. Within each
batch, participants were randomly assigned across condi-
tions within the phase. Participants were only allowed to par-
ticipate in one phase (and hence one condition) per course.
Arrows indicate dependencies between conditions across
phases. Blue conditions did not require participants.
conditions with 100 participants in each condition, with one
condition for each of the voting phase conditions in addition
to an expert graded condition and a random condition. Each
phase is described in more detail below; see Figure 1 for a
flowchart depicting the different phases and conditions.
Participants
Our experiment participants were recruited on Amazon Me-
chanical Turk (AMT), an online crowdsourcing platform
in which requesters can post discrete, typically short, tasks
with replicates called HITs. If an AMT worker finds and ac-
cepts a HIT, the worker gets paid upon successful comple-
tion. In our setting, each phase of each lesson corresponded
to a task, and each task had many HITs falling into each of
our different experimental conditions. We conducted our ex-
periments over a two-week period in May of 2015. All HITs
within a single phase of each lesson were posted simultane-
ously and completed in under 12 hours.
Workers within a course participated in at most one phase,
but were allowed to participate in both courses if they so de-
sired. We placed weak selection criteria on the AMT work-
ers, allowing workers with a minimum of 50 HITs com-
pleted and at least 90% of completed HITs approved (i.e.,
marked as correct by the job requester). The workers were
payed $3.5, $3.5 and $2 per HIT for the content phase, vot-
ing phase and testing phase respectively, leading to an effec-
tive hourly wage of $6.49 as measured by AMT.3 In total,
3We expect the hourly wage to actually be higher due to known
factors such as skimming, in which workers accept and work on
multiple HITs at once (Gong 2012).
Figure 2: Depicts a sample lesson from the Art course. Each
lesson contained a brief description of two styles of art along
with a sample painting corresponding to each, as depicted.
The sample lesson for the Computer Science course con-
tained a brief description of the topic along with a sample
piece of pseudocode and its correct output.
1211 unique workers participated in the experiment.
Experiment Phases
Content Phase During the content phase, participants
were presented with brief 3-5 sentence lessons (as illustrated
in Figure 2). In each lesson, the participants were prompted
to read the lesson, answer a question on the material, and
provide an explanation as to why they chose their answers.
In the Art course participants were shown a painting and
asked to identify the art style of the painting given two
choices (as illustrated in Figure 3). In the Computer Science
course, the participants were asked to give the output of a
short piece of pseudocode (as illustrated in Figure 4).
Voting Phase
In the voting phase, participants were pre-
sented with the same brief lessons as in the content phase.
The first part of the question was identical to the content
phase, but instead of being asked to provide an explanation,
participants were asked to upvote the explanations generated
in the content phase that they thought were good (as illus-
trated in Figure 5). The format was modeled after common
Q&A platforms and MOOC forums.
This phase had two types of conditions, one in which the
numbers of upvotes from previous participants were made
public and the other in which these numbers were kept hid-
den. Public upvotes allow for social influence as participants
can incorporate the information into their decisions to up-
vote. This type of condition forms our sequential condition.
We compared against a condition where the number of up-
votes were private (same as Figure 5, except the green boxes
were not displayed), which we call the independent condi-
tion.
Testing Phase During the testing phase, the goal was to
measure a participant's ability to learn solely from explana-
tions generated in the content phase and chosen by the voting
phase. No lesson was presented and vote information was
Figure 4: Depicts a sample question from the content phase
of the Computer Science course (for the lesson on For
Loops). The participant is shown a short piece of pseu-
docode, asked to type the output of the code, and asked to
give an explanation for their answer.
sisted of a short piece of pseudocode and asked for its out-
put, again without any explanation. As all questions had an
objective right answer, each question was worth 0 or 1 point,
depending on whether the answer was incorrect or correct.
Self-Reported Learning
Participants in all conditions were presented with the same
optional survey at the end of the course, which included a
self-assessment of learning. Participants reported their skill
levels on the topic (Computer Science or Art) before and af-
ter the course on a scale from 0 (none) to 4 (expert). We mea-
sure self-reported learning by taking the difference between
the before and after self-reported skill levels. While individ-
ual estimates of learning may be inflated overall, differences
across conditions must be real since participants were ran-
domly assigned to conditions.
External Validators
We further had external validators in art and computer sci-
ence grade the student explanations in order to indepen-
dently measure the quality of the content curated by the dif-
ferent conditions. This additional step was conducted due
to the unexpected finding (discussed below) that the expert-
curated condition did not outperform the other conditions,
and was sometimes significantly worse. The external valida-
tion confirmed that the expert condition had selected good
content as evaluated by external graders and lends weight to
the finding. Moreover, the topics and question format were
confirmed by the external validators to be similar to what is
used in their courses.
The external validators were professors who had no previ-
ous knowledge of the experiment and no prior collaboration
with the authors. Each has more than 15 years experience
teaching introductory courses in their field, and either has
developed or is in the process of developing a MOOC on
such material. The external validators were asked to grade
Figure 3: Depicts a sample question from the content phase
of the Art course (for the lesson on Photorealism and Real-
ism). The participant was shown a painting, asked whether
it represented a painting of type A or B, and asked to give an
explanation for their answer.
not shown. Instead, the participants were presented with the
questions asked in content and voting phases, along with two
explanations from the content phase (as illustrated in Figure
6). The two explanations the participants were shown varied
across conditions.
The participants were randomly assigned to one of four
different types of conditions: sequential, independent, ran-
dom and expert. The sequential conditions displayed the
top two most upvoted explanations from the sequential con-
ditions in the voting phase. We had three parallel repli-
cates of the sequential condition. The independent condi-
tion displayed the top two most upvoted explanations from
the independent condition of the voting phase. The random
condition displayed two explanations selected uniformly at
random from all the explanations generated in the content
phase. In the expert condition, the authors independently
manually graded the content. The authors' grades were av-
eraged into a final grade for each explanation, and the two
explanations with the best average grades were presented.
Ties were broken at random so that exactly two explanations
were presented in every condition.
The tests given to participants consisted of four multiple-
choice questions per lesson in the Art course similar to the
one depicted in Figure 3 but without the prompt for an expla-
nation, and one numerical question per lesson in the Com-
puter Science course. The computer science question con-
Figure 6: Depicts a sample testing phase: The student ex-
planations, which were generated in the content phase and
selected in the voting phase, are displayed in lieu of a les-
son.
man 2009; Szabo and Huberman 2010; Cheng et al. 2014),
which also strives to quantify the effect of social influence
on popularity (Salganik, Dodds, and Watts 2006; Krumme
et al. 2012; Stoddard 2015). Other work has examined the
impact of social influence on various online user behav-
iors more generally (Salganik and Watts 2008; Chen 2008;
Muchnik, Aral, and Taylor 2013; van de Rijt et al. 2014;
Wang, Wang, and Wang 2014). Our work contributes to
these areas by identifying the effect that social influence has
on the discovery of high quality content in a domain where
quality can be objectively defined.
There are many mechanisms used for social recommen-
dation systems. Variations, both on the allowable user in-
put and on the form of aggregation, exist. These variations
have been studied widely in theory and practice, and an
overview is outside the scope of this work (see, e.g., (Aska-
lidis and Stoddard 2013) and references cited within). We
select a specific form of of user input (upvotes on ranked
content) and the simplest and most common form of aggre-
gation (sorting by number of upvotes) since these choices
are prevalent in the education context. Most popular MOOC
forums, including Coursera, MIT-X, Harvard-X, and Stan-
ford Online, as well as Q&A websites such as Yahoo! An-
swers, Stack Overflow, Baidu Knows and Quora use varia-
tions of this type of sequential voting mechanism.
Analysis
We now present an overview of the participant responses fol-
lowed by the results of our experiments using both quality
Figure 5: Depicts a sample voting phase in the social condi-
tion. The independent condition does not display the green
boxes containing the previous upvotes. Only a portion of the
display is shown here. 10 explanations were presented at a
given time, sorted according to the number of upvotes re-
ceived thus far in the sequential condition and sorted ran-
domly in the independent condition.
on the "usefulness of the explanation to other students who
will have to answer similar questions", and assigned a grade
between 0-10 to each explanation generated in the content
phase (100 explanations in Art, and 50 explanations in CS).
Related Work
There are a number of avenues of research in computa-
tional social science related to our study. Our work is re-
lated to the large literature on identifying content quality,
which primarily strives to develop automated techniques for
quality prediction in online settings. One way to identify
the quality of content is based on the contributor's repu-
tation and past work (Chen et al. 2006; Dom and Paran-
jpe 2008). Specific content features can also be used, such
as the inclusion of references to external resources, the
length, the utility or the verifiability (Adamic et al. 2008;
Kelly et al. 2007; Kim and Oh 2009). Another growing
related area involves popularity prediction (Hogg and Ler-
Figure 7: Explanation grades given by external validators
normalized per-lesson to a 0-1 scale. Error bars are standard
errors of the mean. The plot shows that from the perspec-
tive of the external validators, explanations selected by the
sequential and independent conditions are indistinguishable,
while the explanations selected in the expert condition tend
to be better and the random condition tends to be worse; de-
spite this, we later see that the sequential condition tends
to outperform the rest with respect to self-reported learning
and demonstrated skill.
metrics, self-reported learning and test scores.4
Descriptive Statistics
The conditions for the testing phase were the sequential con-
ditions (3 replicates), the independent condition (1 repli-
cate), the random condition (1 replicate) and the expert-
curated condition (1 replicate). We used three runs of the
sequential condition because prior work suggests that so-
cial information may lead to different outcomes across repli-
cates (Salganik, Dodds, and Watts 2006). These different
outcomes occur because of herding amplifying the first few
votes in each replicate.
100 HITs for each testing condition were posted on AMT;
i.e., 600 HITs for the Art course and 600 HITs for the Com-
puter Science course. The dataset was filtered so that only
the results from workers who had completed all five lessons
were retained. In some cases we had more participants than
HITs, presumably from workers who clicked on our ad but
did not accept the HIT. The number of completed lessons per
condition ranged from 97 to 106 with an average of 100.92.
Explanation Quality
Our external validators independently graded all student ex-
planations on a 0-10 scale. The combined score is an expla-
nation's raw grade. If we think of the external validators as
instructors in the course, this would be the grade they assign
to an explanation posted on the course forum.
The overall average explanation grade in the Art course
was 6.13, and the average best explanation grade (across
lessons) in this course was 8.45. The average explanation
Figure 8: Self-reported skill levels before and after task com-
pletion, versus test scores. The test scores are more corre-
lated with self-reported skill after a lesson, suggesting that
self-reported learning can be thought of as a proxy for ac-
tual learning. Error bars are standard errors of the means.
Regression lines are fitted to the raw data.
grade in the Computer Science course was approximately
2.36, and the average best explanation grade in the Com-
puter Science course was 5.7.
We use these grades to evaluate the quality of the con-
tent selected by the different conditions. We focus on the
average grades of the two explanations selected in each con-
dition of the voting phase. We normalize the grades so that,
for each lesson, they lie on a 0-1 scale by dividing by the
maximum grade in a given lesson.5 We then compared the
average grade of the top explanations selected by each con-
dition (see Figure 7).
In both the CS and the Art courses the explanation grades
in the expert condition are significantly higher than the ex-
planation grades in the sequential condition (two-sided t-
test with p = 0.009976 for Computer Science and p =
0.0001374 for Art). We include this result to illustrate the
differences in explanation quality as judged by an exter-
nal validator. In particular, this shows that there is a range
4The code and data used for our analysis are available online:
https://github.com/pkrafft/Sequential-Voting-Promotes-Collective-
Discovery-in-Social-Recommendation-Systems.
5This is necessary as some lessons are naturally harder than
others. Without normalization the data across courses would not be
comparable.
0.000.250.500.751.00ArtCSCourseAverage Explanation GradesExperiment ConditionExpertSequentialIndependentRandomlllll0.000.250.500.751.0001234Self−Reported Level BeforeTest Scorelllll0.000.250.500.751.0001234Self−Reported Level AfterTest Score Figure 9: Comparison of self-reported learning scores. Error
bars are standard errors of the mean. The sequential condi-
tion produces a significant increase in self-reported learning
compared to the other conditions.
Figure 10: Comparison of test scores. Error bars are stan-
dard errors of the mean. The sequential condition remains
on average the best in terms of performance in both courses,
though not always significantly.
in explanation quality (especially in the Computer Science
course). This result also validates the authors' grades as used
in the expert-curated condition. Though not statistically sig-
nificantly, independent votes achieve identify higher graded
explanations than sequential votes in both courses, as would
be suggested by the literature on social influence. However,
our desired metric was the effect on the student, i.e., the test
scores and self-reported learning of a content consumer, not
the grades given by experts.
Learning
In the testing phase we took explanations selected in the
voting phase and assessed how well participants learned
when consuming that material as instruction. We measure
the change in self-reported skill level from before and after
the lessons. In addition, we measure the participants' knowl-
edge with a test on the subject material after the lesson was
completed.
The self-reported skill level before and after the course
was submitted on a 0-4 scale with 0 being "none" and 4
being "expert". The test scores are normalized to be from
0 to 1. The self-reported skill level after task completion
has a much higher correlation with test scores than the
self-reported skill level from before the task (0.31186 and
0.11505 respectively; see Figure 8). Hence, self-reported
skill after the course is a better proxy for actual skill level
than self-reported skill level before the course.
We find that self-reported learning in the sequential condi-
tion is significantly higher than in the independent condition
(two-sided t-test p < 0.0005312), in the random condition
(two-sided t-test p = 0.002109), and in the expert condition
(two-sided t-test p = 0.04359). These differences remain
significant controlling for the course type and the experi-
ment instance using mixed effects models (with fixed effects
for the course type and random effects for the particular ex-
periment instances). Breaking the analysis up by course, the
trends are the same in all comparisons (see Figure 9), how-
ever the comparison to the expert condition is not significant
in the Computer Science course, and the comparison to the
random condition is not significant in the Art course.
Consistent with these trends, average test scores were
higher in the sequential condition than in the independent,
random, and expert conditions, though these comparisons
were not statistically significant (see Figure 10). The sequen-
tial condition is significantly better than the expert condition
(two-sided t-test p = 0.02682) and random (two-sided t-test
p < 0.0001) in the Art course, though not the Computer Sci-
ence course. Those single comparisons are significant. Us-
ing the mixed effects model to look at both courses jointly,
the comparison to the expert condition, but not to random,
remains significant. The effect is weaker in one context (CS)
than the other (Art) because the sequential and independent
conditions were more similar in the CS test phase. Unlike
in the Art course, the CS voting phase resulted in similar
top explanations across sequential and independent condi-
tions, likely because the content-generation phase resulted
in fewer good explanations.
Herding
We hypothesized that the differences we observed in learn-
ing across conditions was due to a beneficial herding effect
which allows for greater efficiency in identifying good ex-
planations. To test this hypothesis, we examined the skew-
ness of the distribution of the ultimate number of votes each
explanation got per question. Following previous work (Gini
1912; Salganik, Dodds, and Watts 2006), we first take the
number of votes each explanation got and divide those num-
bers by the total number of votes per question. These nor-
malized popularity values are the "market shares" of popu-
larity of each explanation. We then use the average differ-
ence in market share between items as a measure of the
skewness of the distribution, or the "inequality" in popu-
larity. Higher inequality values in the sequential condition
compared to the independent condition indicates that herd-
ing may be occurring.
We find in the Art course that average inequality across
questions is significantly higher in the sequential condition
than the independent condition (two-sided t-test p < 1e-5),
suggesting a herding effect may be at play. We also find a
trend in the same direction in the Computer Science course,
0.000.250.500.751.001.25ArtCSCourseSelf−Reported LearningConditionExpertSequentialIndependentRandom0.000.250.500.75ArtCSCourseAverage Test ScoresConditionExpertSequentialIndependentRandom opposed to a true pretest-posttest) was used so that partic-
ipants would have the same context with which to address
their skill. Having a (quasi)-pretest-posttest design was fea-
sible for self-reported learning but not for actual test scores,
since taking a test requires more time than completing a sur-
vey. This is one of the main limitations of our work, and fur-
ther investigation is required in order to determine whether
the results on test scores are indeed significant.
In our statistical tests, we used a p-value cutoff of 0.05 for
declaring significance. If we instead use a Bonferroni cor-
rection, then the p-value cutoff would become 0.001. This
correction accounts for 50 tests, which is larger than the 30
tests in our final script for analyzing the data and larger than
the 17 that we report. Using this substantially more conser-
vative p-value, we still find that the sequential condition has
significantly higher self-reported learning than the indepen-
dent condition, but notably, the comparisons to the expert
condition are no longer significant.
Domain Differences. As is readily apparent in Figures 9
and 10, both the test scores and self-reported learning are
lower in the Computer Science course than in the Art course.
The disparity could be accounted for by the fact that the ex-
planations in the Computer Science course were much worse
than those in the Art course according to the grades assigned
by the external validators (see the section on Explanation
Quality); in other words, the content consumers had overall
lower-quality explanations to help them learn, which could
explain the lower learning scores overall.
Inadequacy of Experts.
In some cases, our evidence sug-
gests that the sequential condition is better than the expert
condition; in fact, in some cases even the random condition
was on average (although not significantly) better than the
expert condition. A priori, this result is very surprising, and
initially we were concerned that one reason for it could be
that the expert grading was performed poorly. The external
validators were brought on at this point to grade the con-
tent in order to independently validate the experts' decisions.
The fact that the external validation agreed with the author's
decisions gives weight to the conclusion that sequential vot-
ing can in fact outperform experts; hence we do not believe
this is a limitation of our work, but rather a point for further
reflection.
One possible explanation for why expert-selected content
could be worse is the mechanism used for curation: The ex-
perts did not vote on which pairs of explanations would be
best. They graded each explanation individually, and the ex-
planations with the top two grades were selected for use in
the expert condition. It is possible that sequential voting al-
lows for better joint selection of explanations compared to a
traditional grading method.
Herding. We hypothesized that the differences in learning
between the sequential and independent conditions was due
to a beneficial herding effect that led to better ranked con-
tent. Higher inequality values, measured with the Gini index
Figure 11: Average inequality (Gini index) over the mar-
ket share of upvotes. Error bars are standard errors of the
mean. Higher inequality values in the sequential condition
compared to the independent condition suggest that herding
may be occurring.
but the difference is not statistically significant (see Figure
11).
Discussion
Main Results. The results using our objective measure-
ments of quality paint a different picture than the grades
of the external validators. Participants' self-reported assess-
ments suggest that the sequential condition curates better
content than all other conditions. This finding is notable
since the sequential condition outperforms the independent
condition (which had equivalent grades from the external
validators), and since the sequential condition is not out-
performed by the expert condition (which had significantly
higher grades from the external validators). The differences
in self-reported learning across conditions cannot be due to
participants having a tendency to tell us what we want to
hear (i.e., report learning when none occurred) because of
the randomization of participants to different conditions.
Self-reported learning is an important metric since per-
ceived progress is strongly correlated to motivation (Schunk
2003). Moreover, self-reported learning can often be used
as a valid proxy for more direct measures of learn-
ing (Anaya 1999). Therefore, observing an improvement in
self-reported learning alone is enough reason to adopt the
sequential voting procedure.
Limitations. Participants' actual test scores follow the
trend observed in the self-reported assessments, though not
always with statistical significance. While we did not expect
to be under-powered, we believe that, due to the differences
in experimental design between the two measurements, we
were underpowered on test scores and not on self-reported
learning. Test scores used a standard between-subjects
design while self-reported learning used a quasi-pretest-
posttest between-subjects design (where the "pretest" and
"posttest" were both end-survey questions). The latter de-
sign should yield lower variance in the dependent variable,
and thus will have higher power. A quasi-pretest-posttest (as
0.00.10.20.30.4ArtCSCourseAverage Inequality (Gini 1912), in the sequential condition compared to inde-
pendent conditions indicate that herding may be occurring.
However our findings were not conclusive in the CS course.
The participants in the CS course only generated a few good
explanations, while the participants in the Art course gener-
ated many. This factor likely led to high "inequality" in the
distribution of votes for CS even in the independent condi-
tion, thereby diminishing the effect of herding.
Conclusion & Future Work
Despite many models of collective intelligence clearly pre-
dicting deficits in aggregate outcomes due to social influ-
ence, our work suggests that sequential voting may indeed
improve ranking and collective discovery in online social
recommendation systems. However, the mechanism through
which sequential voting has this positive effect remains un-
clear. While there is some evidence that herding may be
playing a role, for example, by emphasizing content that
might have been overlooked in the independent voting set-
ting, another possibility is that people are responding to oth-
ers with anticorrelated votes in an intelligent manner; be-
havior that has recently been observed empirically (Sipos,
Ghosh, and Joachims 2014). Our understanding of collec-
tive intelligence, even in this simple interaction context, is
clearly in a nascent stage; better models and further analy-
ses are required.
Future work that uses the education domain to achieve
objective measurements of underlying quality should also
address some of the limitations of our present work. For
example, we see teasing out the differences between self-
reported learning and test scores as an important topic for fu-
ture work. More comprehensive tests, delays between learn-
ing and testing, a true pre-test / post-test design, and larger
sample sizes of students could be employed in order to iden-
tify whether our null results on test scores are actually a
false negative. Interesting high-level questions in this vein
include: are participants implicitly distinguishing between
what they were taught as opposed to what they learned in
the self-assessment? How do test scores and self-reported
learning correlate, if at all, with reported confidence? What
type of model (if any) of social influence is warranted by the
results we observed, and can we increase the effects to pro-
duce better outcomes? And what other forms of information
can we make public in a way that improves learning?
Our work also leaves open the mechanism through which
information assimilation benefits learning. Understanding
this mechanism would allow us to predict how student
knowledge varies as a function of the type and amount of
information revealed, and could lead to designing systems
which maximize beneficial information, e.g., ranking con-
tent (as in (Lerman and Hogg 2014)) using learning as a
quality metric.
The difference we observed between the expert and se-
quential conditions also warrants further investigation. This
finding, if general, suggests that sequential voting might lead
not just to the crowd being able to discover better content
than it could through standard wisdom-of-crowds aggrega-
tion, but also better content than selected via expert grading.
Identifying the conditions under which crowds can outper-
form experts is an important open area of research in collec-
tive intelligence, and our work suggests that even a simple
interaction mechanism may at times achieve this goal.
Acknowledgments
We would like to thank the anonymous reviewers for their
insightful comments which have aided in the exposition of
this work. We would also like to thank our three external val-
idators: Prof. Nicolas Bock (University of Lausanne), Prof.
Jean-C´edric Chappelier (EPFL) and Prof. Djamila Sam-
Haroud (EPFL), and the anonymous Amazon Mechanical
Turk workers who participated in our study. The authors ob-
tained approval for the experiment from the ´Ecole Polytech-
nique F´ed´ereal de Lausanne (EPFL) Human Research Ethics
Committee (HREC) in Switzerland. This material is based
upon work supported by the National Science Foundation
Graduate Research Fellowship under Grant No. 1122374.
Any opinion, findings, and conclusions or recommendations
expressed in this material are those of the authors(s) and
do not necessarily reflect the views of the National Science
Foundation
References
[Abeliuk et al. 2015] Abeliuk, A.; Berbeglia, G.; Cebrian,
M.; and Van Hentenryck, P. 2015. The benefits of social
influence in optimized cultural markets. PLoS ONE 10(4).
[Adamic et al. 2008] Adamic, L. A.; Zhang, J.; Bakshy, E.;
and Ackerman, M. S. 2008. Knowledge sharing and Yahoo
Answers: Everyone knows something. In Proceedings of the
International Conference on the World Wide Web (WWW),
1013–1022.
[Anaya 1999] Anaya, G. 1999. College impact on student
learning: Comparing the use of self-reported gains, stan-
dardized test scores, and college grades. Research in Higher
Education 40(5):499–526.
[Askalidis and Stoddard 2013] Askalidis, G., and Stoddard,
G. 2013. A theoretical analysis of crowdsourced content cu-
ration. In Proceedings of the 3rd Workshop on Social Com-
puting and User Generated Content.
[Bikhchandani, Hirshleifer, and Welch 1992] Bikhchandani,
S.; Hirshleifer, D.; and Welch, I. 1992. A theory of fads,
fashion, custom, and cultural change as informational
cascades. Journal of Political Economy 992–1026.
[Chamley 2004] Chamley, C. 2004. Rational Herds: Eco-
nomic Models of Social Learning. Cambridge University
Press.
[Chen et al. 2006] Chen, W.; Zeng, Q.; Wenyin, L.; and Hao,
T. 2006. A user reputation model for a user-interactive ques-
tion answering system. In Proceedings of the Second Inter-
national Conference on Semantics, Knowledge, and Grid.
[Chen 2008] Chen, Y.-F. 2008. Herd behavior in purchasing
books online. Computers in Human Behavior 24(5):1977–
1992.
[Cheng et al. 2014] Cheng, J.; Adamic, L.; Dow, P. A.;
Kleinberg, J. M.; and Leskovec, J. 2014. Can cascades be
self-fulfilling prophecies in an artificial cultural market. So-
cial Psychology Quarterly 71(4):338–355.
[Salganik, Dodds, and Watts 2006] Salganik, M. J.; Dodds,
P. S.; and Watts, D. J. 2006. Experimental study of in-
equality and unpredictability in an artificial cultural market.
Science 311(5762):854–856.
[Schunk 2003] Schunk, D. H. 2003. Self-efficacy for reading
and writing: Influence of modeling, goal setting, and self-
evaluation. Reading & Writing Quarterly 19(2):159–172.
[Sipos, Ghosh, and Joachims 2014] Sipos, R.; Ghosh, A.;
and Joachims, T. 2014. Was this review helpful to you?:
it depends! context and voting patterns in online content. In
Proceedings of the 23rd International Conference on World
Wide Web (WWW).
[Stoddard 2015] Stoddard, G. 2015. Popularity dynamics
and intrinsic quality in Reddit and Hacker News. In AAAI
Conference on Web and Social Media (ICWSM).
[Szabo and Huberman 2010] Szabo, G., and Huberman,
B. A. 2010. Predicting the popularity of online content.
Communications of the ACM 53:184–192.
[van de Rijt et al. 2014] van de Rijt, A.; Kang, S. M.;
Restivo, M.; and Patil, A.
2014. Field experiments of
success-breeds-success dynamics. Proceedings of the Na-
tional Academy of Sciences 111(19):6934–6939.
[Wang, Wang, and Wang 2014] Wang, T.; Wang, D.; and
Wang, F. 2014. Quantifying herding effects in crowd wis-
dom. In Proceedings of the ACM International Conference
on Knowledge Discovery and Data Mining (KDD).
Turk:
Market
and
1995.
1912.
2012.
Skimming
failure in
cherry-picking.
predicted? In Proceedings of the 23rd International Confer-
ence on the World Wide Web (WWW), 925–936. ACM.
[Coetzee et al. 2015] Coetzee, D.; Lim, S.; Fox, A.; Hart-
mann, B.; and Hearst, M. A. 2015. Structuring interactions
for large-scale synchronous peer learning. In Proceedings
of the 18th ACM Conference on Computer-Supported Co-
operative Work and Social Computing (CSCW), 1139–1152.
ACM.
[Dom and Paranjpe 2008] Dom, B., and Paranjpe, D. 2008.
A Bayesian technique for estimating the credibility of ques-
tion answerers. In Proceedings of the Society for Industrial
and Applied Mathematics.
Variabilit`a e mutabilit`a.
[Gini 1912] Gini, C.
Reprinted in Memorie di metodologica statistica (Ed. Pizetti
E, Salvemini, T). Rome: Libreria Eredi Virgilio Veschi 1.
[Gong 2012] Gong, A.
Mechanical
http://compsocsci.blogspot.ch/2012/03/market-failure-
in-mechanical-turk.html.
[Hogg and Lerman 2009] Hogg, T., and Lerman, K. 2009.
In Pro-
Stochastic models of user-contributory web sites.
ceedings of the Third AAAI Conference on Weblogs and So-
cial Media (ICWSM).
[Kelly et al. 2007] Kelly, D.; Wacholder, N.; Rittman, R.;
Sun, Y.; Kantor, P.; Small, S.; and Strzalkowski, T. 2007.
Using interview data to identify evaluation criteria for in-
teractive, analytical question-answering systems. Journal of
the American Society for Information Science and Technol-
ogy 58(8):80–88.
[Kim and Oh 2009] Kim, S., and Oh, S. 2009. User's rele-
vance criteria for evaluating answers in a social Q&A site.
Journal of the American Society for Information Science and
Technology 60:716–727.
[Krumme et al. 2012] Krumme, C.; Cebrian, M.; Pickard,
G.; and Penetland, S. 2012. Quantifying social influence
in an online cultural market. PLoS ONE 7(5).
[Ladha 1995] Ladha, K. K.
Information pooling
through majority-rule voting: Condorcet's jury theorem with
correlated votes. Journal of Economic Behavior & Organi-
zation 26(3):353–372.
[Lerman and Hogg 2014] Lerman, K., and Hogg, T. 2014.
Leveraging position bias to improve peer recommendation.
PLoS One 9(6).
[Lorenz et al. 2011] Lorenz, J.; Rauhut, H.; Schweitzer, F.;
and Helbing, D. 2011. How social influence can undermine
the wisdom of crowd effect. Proceedings of the National
Academy of Sciences 108(22):9020–9025.
[Muchnik, Aral, and Taylor 2013] Muchnik, L.; Aral, S.; and
Taylor, S. J. 2013. Social influence bias: A randomized
experiment. Science 341(6146):647–651.
[Raafat, Chater, and Frith 2009] Raafat, R. M.; Chater, N.;
and Frith, C. 2009. Herding in humans. Trends in Cog-
nitive Sciences 13(10):420–428.
[Salganik and Watts 2008] Salganik, M. J., and Watts, D. J.
2008. Leading the herd astray: An experimental study of
|
1911.09534 | 1 | 1911 | 2019-11-21T15:20:09 | Active Re-identification Attacks on Periodically Released Dynamic Social Graphs | [
"cs.SI"
] | Active re-identification attacks pose a serious threat to privacy-preserving social graph publication. Active attackers create fake accounts to build structural patterns in social graphs which can be used to re-identify legitimate users on published anonymised graphs, even without additional background knowledge. So far, this type of attacks has only been studied in the scenario where the inherently dynamic social graph is published once. In this paper, we present the first active re-identification attack in the more realistic scenario where a dynamic social graph is periodically published. The new attack leverages tempo-structural patterns for strengthening the adversary. Through a comprehensive set of experiments on real-life and synthetic dynamic social graphs, we show that our new attack substantially outperforms the most effective static active attack in the literature by increasing the success probability of re-identification by more than two times and efficiency by almost 10 times. Moreover, unlike the static attack, our new attack is able to remain at the same level of effectiveness and efficiency as the publication process advances. We conduct a study on the factors that may thwart our new attack, which can help design graph anonymising methods with a better balance between privacy and utility. | cs.SI | cs |
Active Re-identification Attacks on Periodically
Released Dynamic Social Graphs
Xihui Chen1, Ema Kepuska2, Sjouke Mauw1,2 and Yunior Ram´ırez-Cruz1
1SnT, 2CSC, University of Luxembourg
6, av. de la Fonte, L-4364 Esch-sur-Alzette, Luxembourg
{xihui.chen, sjouke.mauw, yunior.ramirez}@uni.lu, [email protected]
November 22, 2019
Abstract
Active re-identification attacks pose a serious threat to privacy-preserving social
graph publication. Active attackers create fake accounts to build structural pat-
terns in social graphs which can be used to re-identify legitimate users on published
anonymised graphs, even without additional background knowledge. So far, this type
of attacks has only been studied in the scenario where the inherently dynamic social
graph is published once. In this paper, we present the first active re-identification
attack in the more realistic scenario where a dynamic social graph is periodically
published. The new attack leverages tempo-structural patterns for strengthening the
adversary. Through a comprehensive set of experiments on real-life and synthetic
dynamic social graphs, we show that our new attack substantially outperforms the
most effective static active attack in the literature by increasing the success proba-
bility of re-identification by more than two times and efficiency by almost 10 times.
Moreover, unlike the static attack, our new attack is able to remain at the same level
of effectiveness and efficiency as the publication process advances. We conduct a
study on the factors that may thwart our new attack, which can help design graph
anonymising methods with a better balance between privacy and utility.
Keywords: privacy-preserving social graph publication, re-identification attack, active
adversary, dynamic social networks
1
Introduction
Social graphs have proven to be a valuable data source for conducting sociological studies,
market analyses, and other forms of complex data analysis. This creates a strong incentive
for the establishment of a mutually beneficial relation between analysts and data owners.
For analysts, it is of paramount importance to have access to abundant, reliable social graph
1
data in order to conduct their studies. For data owners, making these data available to third
parties opens a number of additional business opportunities, as well as opportunities for
improving their social perception by contributing to the advancement of research. However,
releasing social network data raises serious privacy concerns, due to the sensitive nature of
much of the information implicitly or explicitly contained in social graphs. Consequently,
the data needs to be properly sanitised before publication.
It has been shown that some forms of sanitisation, e.g. removing users' identities and
personally identifying information from the released data, a process known as pseudonymi-
sation, are insufficient for protecting sensitive information. This is due to the fact that
a majority of users can still be unambiguously re-identified in the pseudonymised graph
by means of simple structural patterns [14, 22, 2, 23]. The re-identification subsequently
facilitates inferring relations between users, group affiliations, etc. A method allowing a
malicious agent, or adversary, to re-identify (a subset of) the users in a sanitised social
graph is called a re-identification attack.
A large number of anonymisation methods have been proposed for publishing sanitised
social graphs that effectively resist re-identification attacks. The largest family of graph
anonymisation methods (e.g. [14, 16, 3, 5, 33, 17, 30, 29, 4, 38, 39, 34, 18, 19]) follows a
common strategy of editing the vertex and/or edge set of the pseudonymised graph in order
to satisfy some formal privacy properties. These privacy properties rely on an adversary
model, which encodes a number of assumptions about the adversary capabilities. In the
context of social graph publication, there are two classes of adversaries. On the one hand,
passive adversaries depend on publicly available information, in case that it can be obtained
from online resources, public records, etc., without interacting with the social network
before publication. On the other hand, active adversaries interact with the network before
the sanitised dataset is released, in order to force the existence of structural patterns.
Then, when the sanistised graph is published, they leverage these patterns for conducting
the re-identification. Active adversaries have been shown to be a serious threat to social
graph publication [2, 20], as they remain plausible even if no public background knowledge
is available. Active attackers have the capability of inserting fake accounts in the social
network, commonly called sybil nodes, and creating connection patterns between these fake
accounts and a set of legitimate users, the victims. After the publication of the sanitised
graph, the attacker uses these unique patterns for re-identifying the victims.
Social networks are inherently dynamic. Moreover, analysts require datasets containing
dynamic social graphs in order to conduct numerous tasks such as community evolution
analysis [6], link prediction [15] and link persistence analysis [25], among others. Despite
the need for properly anonymised dynamic social graphs, the overwhelming majority of
studies on graph anonymisation have focused on the scenario of a social graph being re-
leased only once. The rather small number of studies on dynamic social graph publication
have provided only a partial understanding of the field, as they have exclusively focused on
passive adversary models. Consequently, the manners in which active adversaries can profit
from a dynamic graph publication scenario remain unknown. In this paper, we remedy this
situation by formulating active re-identification attacks in the scenario of dynamic social
graphs. We consider a scenario where the underlying dynamic graph is periodically sampled
2
for snapshots, and sanitised versions of these snapshots are published. We model an active
adversary whose knowledge consists in tempo-structural patterns, instead of exclusively
structural patters as those used by the original (static) active adversary. Moreover, in our
model the adversary knowledge is incremental, as it grows every time a new snapshot is
released, and the adversary has the opportunity to adapt along the publication process.
Under the new model, we devise for the first time a dynamic active re-identification attack
on periodically released dynamic social graphs. The new attack is more effective than
the alternative of executing independent static attacks on different snapshots. Further-
more, it is also considerably more efficient than the previous attacks, because it profits
from temporal patterns to accelerate the search procedures in the basis of several of its
components.
Our contributions. The main contributions of this paper are listed in what follows:
• We formulate, for the first time, active re-identification attacks in the scenario of
periodically released dynamic social graphs.
• Based on the new formulation, we present, to the best of our knowledge, the first dy-
namic active re-identification attack on periodically released dynamic social graphs,
which constructs and leverages tempo-structural patterns for re-identification.
• We conduct a comprehensive set of experiments on real-life and synthetic dynamic
social graphs, which demonstrate that the dynamic active attack is more than two
times more effective than the alternative of repeatedly executing the strongest active
attack reported in the literature for the static scenario [20].
• Our experiments also show that, as the number of published snapshots grows, the dy-
namic active attack runs almost 10 times faster than the static active attack from [20].
• We analyse the factors that affect the effectiveness of our new attack. The conclusions
of this study serve as a starting point for the development of anonymisation methods
for the new scenario.
Structure of the paper. We discuss the related work in Sect. 2, focusing in similarities
and differences with respect to our new proposals. Then, we describe the periodical graph
publication scenario, accounting for active adversaries, in Sect. 3; and we introduce the
new dynamic active attack in Sect. 4. Finally, our experimental evaluation is presented in
Sect. 5 and we give our conclusions in Sect. 6.
2 Related Work
Re-identification attacks are a relevant threat for privacy-preserving social graph publica-
tion methods that preserve a mapping between the real users and a set of pseudonymised
nodes in the sanitised release [14, 16, 3, 5, 33, 17, 30, 29, 4, 38, 39, 34, 18, 19].
3
Depending on the manner in which the attacker obtains the knowledge used for re-
identification, these attacks can be divided into two classes: passive and active attacks.
Passive adversaries collect publicly available knowledge, such as public profiles in other
social networks, and searches the sanitised graph for vertices with an exact or similar
profile. For example, Narayanan and Shmatikov [22] used information from Flickr to re-
identify users in a pseudonymised subgraph of Twitter. A considerable number of passive
attacks have been proposed, e.g. [22, 21, 35, 26, 24, 9, 10, 8]. On the other hand, active
adversaries interact with the real network before publication, and force the existence of
the structural patterns that allow re-identification after release. The earliest examples of
active attacks are the walk-based attack and the cut-based attack, introduced by Backstrom
et al. in [2]. Both attacks insert sybil nodes in the network, and create connection patterns
between the sybil nodes that allow their efficient retrieval in the pseudonymised graph. In
both attacks, the connection patterns between sybil nodes and victims are used as unique
fingerprints allowing re-identification once the sybil subgraph is retrieved. Due to the low
resilience of the walk-based and cut-based attacks, a robust active attack was introduced
by Mauw et al. in [20]. The robust active attack introduces noise tolerant sybil subgraph
retrieval and fingerprint mapping, at the cost of larger computational complexity. The
attack proposed in this paper preserves the noise resiliency of the robust active attack,
but puts a larger emphasis on temporal consistency constraints for reducing the search
space. As a result, for every run of the re-identification, our attack is comparable to the
original walk-based attack in terms of efficiency and to the robust active attack in terms
of resilience against modifications in the graph.
Notice that, by itself, the use of connection fingerprints as adversary knowledge does
not make an attack active. The key feature of an active attack is the fact that the adversary
interacts with the network to force the existence of the fingerprints. For example, Zou et
al. [39] describe an attack that uses as fingerprints the distances of the victims to a set of
hubs. This is a passive attack, since hubs exist in the network without intervention of the
attacker.
The attacks discussed so far assume a single release scenario. A smaller number of
works have discussed re-identification in a dynamic scenario. Some works assume an ad-
versary who can exploit the availability of multiple snapshots, although they only give a
coarse overview of the increased adversary capabilities, without giving details on attack
strategies. Examples of these works are [31], which models a passive adversary that knows
the evolution of the degrees of all vertices; and [39], which models another passive adver-
sary that knows the evolution of a a subgraph around the victims. An example of a full
dynamic de-anonymization method is given in [7]. Although they do not model an active
adversary, the fact that the method relies on the existence of a seed graph makes it po-
tentially extensible with an active first stage for seed re-identification, as done for example
in [27, 28]. Our attack differs from the methods above in the fact that it uses an evolving
set of sybil nodes that dynamically interact with the network and adapt to its evolution.
4
3 Periodical Graph Publication in the Presence of Ac-
tive Adversaries
In this section we describe the scenario where the owner of a social network periodically
publishes sanitised snapshots of the underlying dynamic social graph, accounting for the
presence of active adversaries. We describe this scenario in the form of an attacker-defender
game between the data owner and the active adversary. We first introduce the basic
notation and terminology, and then give an overview of the entire process and a detailed
description of its components.
3.1 Notation and Terminology
We represent a dynamic social graph as a sequence G = (G1, G2, . . . , Gi, . . .), where each
Gi is a static graph called the i-th snapshot of G. Each snapshot of G has the form
Gi = (Vi, Ei), where Vi is the set of vertices (also called nodes indistinctly throughout the
paper) and Ei ⊆ Vi × Vi is the set of edges.
We will use the notations VG and EG for the vertex and edge sets of a graph G. In this
paper, we assume that graphs are simple and undirected. That is, G contains no edges
of the form (v, v) and, for every pair v, w ∈ VG, (v, w) ∈ EG ⇐⇒ (w, v) ∈ EG. The
neighbourhood of a vertex v in a graph G is the set NG(v) = {w ∈ V (v, w) ∈ E}, and
its degree is δG(v) = NG(v). For the sake of simplicity, in the previous notations we drop
the subscript when it is clear from the context and simply write N (v), δ(v), etc.
For a subset of nodes S ⊆ VG, we use (cid:104)S(cid:105)G to represent the subgraph of G induced by
S, i.e. (cid:104)S(cid:105)G = (S, EG ∩ (S × S)). Similarly, the subgraph of G weakly induced by S is
defined as (cid:104)S(cid:105)w
G = (S ∪ NG(S), EG ∩ (S × (S ∪ NG(S)))). For every graph G and every
S ⊆ VG, (cid:104)S(cid:105)G is a subset of (cid:104)S(cid:105)w
G additionally contains the neighbourhood of S
and every edge between elements of S and their neighbours. Also notice that (cid:104)S(cid:105)w
G does
not contain the edges linking pairs of elements of NG(S).
An isomorphism between two graphs G = (V, E) and G(cid:48) = (V (cid:48), E(cid:48)) is a bijective
function ϕ : V → V (cid:48) such that ∀v,w∈V (v, w) ∈ E ⇐⇒ (ϕ(v), ϕ(w)) ∈ E(cid:48). Additionally, we
denote by ϕ(S) the restriction of ϕ to a vertex subset S ⊆ V , that is ϕ(S) = {ϕ(v) v ∈ S}.
G, as (cid:104)S(cid:105)w
3.2 Overview
Fig. 1 depicts the process of periodical graph publication in the presence of an active
adversary. We model this process as a game between two players, the data owner and the
adversary. From a practical perspective, it is implausible for the data owner to release a
snapshot of the dynamic graph every time a small amount of changes occurs, hence the
periodical nature of the publication process. The data owner selects a set of time-stamps
T = {t1, t2, . . . , ti, . . .}, t1 < t2 < . . . < ti < . . ., and incrementally publishes the sequence
G(cid:63) = (G(cid:63)
t1, G(cid:63)
t2, . . . , G(cid:63)
ti
, . . .)
5
Figure 1: Overview of the process of periodical graph publication in the presence of active
adversaries.
of sanitised snapshots of the underlying dynamic social graph. The adversary's goal is
to re-identify, in a subset T (cid:48)
⊆ T of the releases, a (possibly evolving) set of legitimate
users referred to as the victims. To achieve this goal, the active adversary injects an (also
evolving) set of fake accounts, commonly called sybils, in the graph. The sybil accounts
create connections among themselves, and with the victims. The connection patterns
between each victim and some of the sybil nodes is used as a unique fingerprint for that
victim. The likely unique patterns built by the adversary with the aid of the sybil nodes will
enable her to effectively and efficiently re-identify the victims in the sanitised snapshots.
At every re-identification attempt, the adversary first re-identifies the set of sybil nodes,
and then uses the fingerprints to re-identify the victims.
The data owner and the adversary have different partial views of the dynamic social
graph. On the one hand, the data owner knows the entire set of users, both legitimate
users and sybil accounts, but she cannot distinguish them. The data owner also knows
all relations. On the other hand, the adversary knows the identity of her victims and the
structure of the subgraph weakly induced by the set of sybil nodes, but she does not know
the structure of the rest of the network. In this paper we conduct the analysis from the
perspective of an external observer who can view all of the information. For the sake of
simplicity in our analysis, we will differentiate the sequence
G+ = (G+
t1, G+
t2, . . . , G+
ti
, . . .),
which represents the view of the network according to the data owner, which is the real
network, i.e.
the one containing the nodes representing all users, both legitimate and
6
t1t2...ti...Gt1Gt2...Gti...connectupdateconnection...updateconnectionG+t1G+t2G+ti...anonymiseanonymiseanonymise...G?t1G?t2G?ti...re-identifyre-identifyre-identify...sybilsubgraphsybilsubgraphsybilsubgraphcreateupdateupdateφt1:VG?t1→Yt1φt2:VG?t2→Yt2φti:VG?ti→Yti.....................evolveevolveevolveLegit.useracti-vitySybilinjectionRe-identificationPublicationAdversaryDataOwnerAdversary malicious, from the sequence
G = (Gt1, Gt2, . . . , Gti, . . .),
ti = (Vti ∪ Sti, Eti ∪ E+
which represents the view of the unattacked network, that is the view of the dynamic
subgraph induced in G+ by the nodes representing legitimate users.
In the original formulation of active attacks, a single snapshot of the graph is released,
so all actions executed by the sybil nodes are assumed to occur before the publication. This
is not the case in the scenario of a periodically released dynamic social graph. Here, the
adversary has the opportunity to schedule actions in such a way that the subgraph induced
by the sybil nodes evolves, as well as the set of fingerprints. In turn, that allows her to
use temporal patterns in addition to structural patterns for re-identification. Additionally,
the adversary can target different sets of victims along the publication process and adapt
the induced tempo-structural patters to the evolution of the graph and the additional
knowledge acquired in each re-identification attempt.
In the new scenario, the actions
performed by the adversary and the data owner alternate as follows before, during and
after each time-stamp ti ∈ T .
Before ti: The adversary may remain inactive, or she can modify the set of sybil nodes,
as well the set of sybil-to-sybil and sybil-to-victim edges. The result of these actions is the
graph G+
ti ), where Vti is the current set of legitimate users, Sti is the
current set of sybil nodes, Yti ⊆ Vti is the current set of victims, Eti = EGti ⊆ Vti × Vti is
the set of connections between legitimate users, and E+
ti ⊆ (Sti × Sti)∪ (Sti × Yti) is the set
, weakly induced in G+
of connections created by the sybil accounts. The subgraph (cid:104)Sti(cid:105)w
ti
by the set of sybil nodes, is the sybil subgraph. We refer to the set of modifications of the
sybil subgraph executed before the adversary has conducted any re-identification attempt
as sybil subgraph creation. If the adversary has conducted a re-identification attempt on
earlier snapshots, we refer to the modifications of the sybil subgraph as sybil subgraph
update.
During ti: The data owner applies an anonymisation method to G+
ti to obtain the sanitised
version G(cid:63)
, which is then released. The anonymisation must preserve the consistency of the
ti
pseudonyms. That is, every user must be labelled with the same pseudonym throughout the
sequence of snapshots where it appears. Consistent annotation is of paramount importance
for a number of analysis tasks such as community evolution analysis [6], link prediction [15],
link persistence analysis [25], among others, that require to track users along the sequence
of releases. The data owner anonymises every snapshot exactly once.
After ti: The adversary adds G(cid:63)
to her knowledge. At this point, she can remain inactive,
ti
or she can execute a re-identification attempt on G(cid:63)
. The result of a re-identification
ti
attempt is a mapping φti : VG(cid:63)
ti → Yti determining the pseudonyms assigned to the victims
by the anonymisation method. Here, the adversary can additionally modify the results of
a previous re-identification attempt conducted on some of the preceding releases.
G+
ti
7
3.3 Components of the Process
To discuss in detail the different actions that the data owner or the adversary execute, we
follow the categorisation given in above in terms of the time where each action may occur
for every time-stamp ti. We first discuss sybil subgraph creation and update, which occur
before ti, then graph publication, which occurs at ti, and finally re-identification, which
occurs after ti.
3.3.1 Sybil subgraph creation and update
As we mentioned above, sybil subgraph creation is executed before the adversary has
attempted re-identification for the first time; whereas sybil subgraph update is executed
in the remaining time-steps.
y
G+
ti
Sybil subgraph creation: In the dynamic scenario, the adversary can build the initial
sybil subgraph along several releases. This allows the creation of tempo-structural pat-
terns, incorporating information about the first snapshot where each sybil node appears,
to facilitate the sybil subgraph retrieval stage during re-identification. As in all active
attacks, the patterns created must ensure that, with high probability, (cid:104)Sti(cid:105)w
is unique.
We denote by Fti(y) the fingerprint of a victim y ∈ Yti in terms of Sti. Throughout this
paper we consider that Fti(y) is uniquely determined by the neighbourhood of y in Sti,
. We denote by Fti the set of fingerprints of all victims in G+
ti .
that is Fti(y) = Sti ∩ NG+
Sybil subgraph update: In this step, the adversary can modify the set of sybil nodes, by
adding new sybil nodes or replacing existing ones. The latter action helps to make every
individual sybil node's behaviour appear more normal, which in turn makes it more likely
to succeed in establishing the necessary relations and less likely to be detected by sybil
defences. The adversary can also modify the inter-sybil connections and the fingerprints.
Since sybil subgraph update is executed after at least one re-identification attempt has
been conducted, the adversary can use information from this attempt, such as the level of
uncertainty in the re-identification, to decide the changes to introduce in the sybil subgraph.
Finally, if the number of fingerprints that can be constructed using the new set of sybil
nodes is larger than the previous number of targeted victims, that is 2Sti
− 1 > Yti−1, the
adversary can additionally target new victims, either new users that joined the network in
the last inter-release interval, or previously enrolled users that had not been targeted so
far. In the latter case, even if these victims had not been targeted before, the consistency
of the labelling in the sequence of sanitised snapshots entails that a re-identification in the
ti-th snapshot can be traced back to the previous ones.
3.3.2 Graph publication
As we discussed above, at time step ti, the data owner anonymises G+
ti and publishes
the sanitised version G(cid:63)
. From a general point of view, we treat the anonymisation as
ti
a two-step process. The first step is pseudonymisation, which consists in building the
8
ti → V (cid:48)(cid:63)
ti
, with V (cid:48)(cid:63)
isomorphism ϕti : VG+
for a pseudonym. The pseudonymised graph is denoted as ϕtiG+
are freshly generated.
vertices are kept, and fresh pseudonyms are assigned to new vertices.
= ∅, that replaces every real identity in G+
ti
ti . If i = 1, all pseudonyms
In the remaining cases, the pseudonyms for previously existing
ti ∩ VG+
ti
Releasing the pseudonymised graph has been proven to be insufficient for preventing re-
identification [22, 14, 32, 2, 20]. Thus, the second step of the anonymisation process consists
in applying a perturbation method Φti : ϕtiG+
) to the pseudonymised
graph. Perturbation consists in editing the vertex and/or edge sets of the pseudonymised
graph in such a way that the resulting graph satisfies some privacy guarantee against re-
identification. For the case of active adversaries, the relevant perturbation methods are
the ones based on random vertex/edge flipping and those based on the notions of (k, Γ(cid:96))-
(adjacency) anonymity [32, 18, 19, 20]. Finally, the data owner releases the graph G(cid:63)
ti
obtained as the result of applying pseudonymisation on G+
ti and perturbation on ϕtiG+
ti ,
that is G(cid:63)
ti
ti → (V (cid:63)
= Φti(ϕtiG+
ti × V (cid:63)
, V (cid:63)
ti ).
ti
ti
3.3.3 Re-identification
In the new scenario, re-identification is dynamic, as it occurs along several time-steps,
leveraging the increase of the adversary knowledge after the release of every new snapshot.
Considering the applicable techniques, we differentiate the first re-identification attempt,
which can be executed immediately after the publication of G(cid:63)
, from the remaining at-
ti
tempts, which we refer to as refinements, and can be executed after the release of every
other G(cid:63)
tj
, j > i.
First re-identification attempt: The first re-identification attempt is composed of two
steps: sybil subgraph retrieval and fingerprint matching. From a general point of view, the
procedure consists in the following steps:
1. Sybil subgraph retrieval:
(a) Find in G(cid:63)
ti
a set Xti = {X1, X2, . . . , Xp}, Xj ⊆ VG(cid:63)
, of candidate sybil sets. For
is a candidate sybil subgraph. Every specific
every X ∈ Xti, the graph (cid:104)X(cid:105)w
attack defines the conditions under which a candidate is added to Xti.
G(cid:63)
ti
ti
(b) Filter out elements of Xti that fail to satisfy temporal consistency constraints
. The specific constraints to
with respect to (cid:104)St1(cid:105)w
enforce depend on the instantiation of the attack strategy.
, . . . , (cid:104)Sti−1(cid:105)w
, (cid:104)St2(cid:105)w
G+
t2
G+
t1
ti−1
G+
(c) If Xti = ∅, the attack fails. Otherwise, proceed to fingerprint matching (step 2).
2. Fingerprint matching:
(a) Select one element X ∈ Xti. As in the previous steps, every specific attack
defines how the selection is made.
9
(b) Using X and Fti, find a set of candidate mappings YX = {φ1, φ2, . . . , φq}, where
ti \ Sti → Yti. Every element of YXti
every φj (1 ≤ j ≤ q) has the form φj : VG(cid:63)
represents a possible re-identification of the victims in G(cid:63)
ti
.
(c) Filter out elements of YXti
tency constraints with respect to Ft1,Ft2, . . . ,Fti−1.
that fail to satisfy (attack-specific) temporal consis-
(d) If YX (cid:48)
= ∅, the attack fails. Otherwise, select one element of YX (cid:48)
and give it as
the result of the re-identification. As in the previous steps, every specific attack
defines how the selection is made.
ti
ti
In an actual instantiation of the attack, steps 1.a-c, as well as steps 2.a-d, are not
necessarily executed in that order, nor independently. As we will show in Sects. 4 and 5,
combining temporal consistency constraints with structural similarity allows for higher
effectiveness and considerable speed-ups in several steps.
Re-identification refinement: As we discussed above, the first re-identification attempt
on G(cid:63)
can be executed immediately after the snapshot is published. Then, after the publi-
ti
cation of G(cid:63)
, j > i, the re-identification refinement step allows to improve the adversary's
tj
certainty on the previous re-identification, by executing the following actions:
1. Filter out elements of Xti that fail to satisfy additional temporal consistency con-
straints with respect to (cid:104)Sti(cid:105)w
G+
ti
and (cid:104)Stj(cid:105)w
G+
tj
.
2. Repeat the fingerprint matching step.
only once, waiting for the release of several G(cid:63)
tj1
Note that, in a specific attack, the adversary may choose to run the re-identification
on G(cid:63)
, and combining all
ti
temporal consistency checks of the first attempt and the refinements in a single execution
of step 1.b described above. However, we keep the main re-identification attempt separable
from the refinements considering that in a real-life attack the actual time elapsed between
G(cid:63)
ti
can be considerably large, e.g. several months.
, . . . , G(cid:63)
tjr
and G(cid:63)
tj
, G(cid:63)
tj2
4 A Novel Dynamic Active Attack
In this section we present what is, to the best of our knowledge, the first active attack on
periodically released dynamic social graphs. The novelty of our attack lies in its ability
to exploit the dynamic nature of the social graph being periodically published, and the
fact that the publication process occurs incrementally. Our attack benefits from temporal
information in two fundamental ways. Firstly, we define a number of temporal consistency
constraints, and use them in all stages of the re-identification process. In sybil subgraph
retrieval, consistency constraints allow us to obtain considerably small sets of plausible can-
didates, which increases the likelihood of the attacker selecting the correct one. A similar
situation occurs in fingerprint matching. Moreover, the incremental publication process al-
lows the adversary to refine previous re-identifications by applying new consistency checks
10
based on later releases. In all cases, temporal consistency constraints additionally make
the re-identification significantly fast, especially when compared with comparably noise-
resilient methods reported in the literature for the static publication scenario. The second
manner in which our attack benefits from the dynamicity of the publication process is
by adapting the set of fingerprints in the interval between consecutive re-identification
attempts in such a way that the level of uncertainty in the previous re-identification is
reduced.
In the remainder of this section, we will describe out new attack in detail. We will first
introduce the notions of temporal consistency. Then, we will describe the manner in which
temporal consistency is exploited for dynamic re-identification. Finally, we will describe
how tempo-structural patterns are created and maintained.
4.1 Temporal consistency constraints
As we discussed in Sect. 3.2, the data owner must assign the same pseudonym to each user
throughout the subsequence of snapshots where it appears, to allow for analysis tasks such
as community evolution analysis [6], link prediction [15], link persistence analysis [25], etc.
Since the data owner cannot distinguish between legitimate users (including victims) and
sybil accounts, she will assign time-persistent pseudonyms to all of them. Additionally,
since the adversary receives all sanitised snapshots, she can determine when a pseudonym
was used for the first time, whether it is still in use, and in case it is not, when it was used
for the last time.
t6
t5. After the release of G(cid:63)
t2. Additionally, she inserted s3 before the publication of G(cid:63)
In our attack, the adversary exploits this information in all stages of the re-identification
process. For example, consider the following situation. The set of sybil nodes at time-step
t6 is St6 = {s1, s2, s3, s4}. The adversary inserted s1 and s2 in the interval preceding
the publication of G(cid:63)
t3 and s4
before the publication of G(cid:63)
t6, during the sybil subgraph retrieval
phase of the first re-identification attempt, the adversary needs to determine whether a
set X ⊆ VG(cid:63)
, say X = {v1, v2, v3, v4}, is a valid candidate. Looking at the first snapshot
where each of these pseudonyms was used, the adversary observes that v1 and v3 were first
used in G(cid:63)
t2, so they are feasible matches for s1 and s2, in some order. Likewise, v2 was
first used in G(cid:63)
t5, so it is a feasible match for s4. However, she observes that v4 was first
used in G(cid:63)
t4, unlike any element of St6. From this observation, the adversary knows that
X is not a valid candidate, regardless of how structurally similar (cid:104)X(cid:105)w
are.
The previous example illustrates how temporal consistency constraints are used for
discarding candidate sybil sets. We now formalise the different types of constraints used in
our attack. To that end, we introduce some new notation. The function α+ : ∪ti∈T VG+
ti → T
yields, for every vertex v ∈ ∪ti∈T VG+
, the order of the first snapshot where v exists, that is
ti
G(cid:63)
t6
and (cid:104)St6(cid:105)w
G+
t6
α+(v) = min{{ti ∈ T v ∈ VG+
ti}}.
Analogously, the function α(cid:63) : ∪ti∈T VG(cid:63)
ti → T yields the order of the first snapshot where
11
each pseudonym is used, that is
α(cid:63)(x) = ti ⇐⇒ ∃v∈V
G+
ti
α+(v) = ti ∧ ϕti(v) = x.
Clearly, the adversary knows the values of the function α(cid:63) for all pseudonyms used by
the data owner. Additionally, she knows the values of α+ for all of her sybil nodes. Thus,
the previous functions allow us to define the notion of first-use-as-sybil consistency, which
is used by the sybil subgraph retrieval method.
Definition 1. Let X ⊆ VG(cid:63)
be a set of pseudonyms such that X = Sti and let φ : Sti →
be a mapping from the set of real sybil nodes to the elements of X. We say that X
VG(cid:63)
ti
and Sti satisfy first-use-as-sybil consistency according to φ, denoted as X (cid:39)φ Sti, if and
only if ∀s∈Sti
α+(s) = α(cid:63)(φ(s)).
Note that first-use-as-sybil consistency depends on the order in which the elements of
the candidate set are mapped to the real sybil nodes, which is a requirement of the sybil
subgraph retrieval method.
ti
We define an analogous notion of first use consistency for victims. In this case, the
adversary may or may not know the value of α+.
In our attack, we assume that she
does not, and introduce an additional function to represent the temporal information the
adversary must necessarily have about victims. The function β+ : ∪ti∈T Yti → T yields,
for every v ∈ ∪ti∈T Yti, the order of the snapshot where v was targeted for the first time,
that is
The new function allows us to define the notion of first-time-targetted consistency, which
β+(v) = min{{ti ∈ T v ∈ Yti}}.
is used in the fingerprint matching method.
Definition 2. Let v ∈ VG(cid:63)
be a victim candidate and let y ∈ Yti be a real victim. We
say that v and y satisfy first-time-targetted consistency, denoted as v (cid:39) y, if and only if
α(cid:63)(v) ≤ β+(y).
This temporal consistency notion encodes the rationale that the adversary can ignore
during fingerprint matching those pseudonyms that the data owner used for the first time
after the corresponding victim had been targeted.
ti
Finally, we define the notion of sybil-removal-count consistency, which is used by the
re-identification refinement method to encode the rationale that a sybil set candidate X,
for which no temporal inconsistencies were found during the ti-th snapshot, can be removed
from Xti when the ti+1-th snapshot is released, if the number of sybil nodes removed by the
adversary in the interval between these snapshots does not match the number of elements
of X that cease to exist in G(cid:63)
Definition 3. We say that a set of pseudonyms X ⊆ VG(cid:63)
satisfies sybil-removal-count
consistency with respect to the pair (Sti, Sti+1), which we denote as X (cid:39) (Sti, Sti+1), if and
only if X \ VG(cid:63)
ti+1 = Sti \ Sti+1.
ti+1
.
ti
Defs. 1 to 3 play an important role in the new dynamic re-identification methods in-
troduced as part of the new attack, as we will discuss in what follows.
12
4.2 Dynamic re-identification
In what follows we describe the methods for sybil subgraph retrieval and fingerprint
matching, which are conducted during the first re-identification attempt, as well as the
re-identification refinement method.
In all cases, we pose the emphasis on the manner
in which they use the notions of temporal consistency for maximising effectiveness and
efficiency.
4.2.1 Sybil subgraph retrieval
The sybil subgraph retrieval method is a breath-first search procedure, which shares the
philosophy of analogous methods devised for active attacks on static graphs [2, 20], but
differs from them in the use of temporal consistency constraints for pruning the search
space. To establish the order in which the search space is traversed, our method relies on
the existence of an arbitrary (but fixed) total order ≺ among the set of sybil nodes, which
is enforced by the sybil subgraph creation method and maintained by the sybil subgraph
update method.
Let s1 ≺ s2 ≺ . . . ≺ sSti be the order established on the elements of Sti. The search
procedure first builds a set of cardinality-1 partial candidates
Xti,1 = {{vj1} vj1 ∈ VG(cid:63)
ti}.
(cid:12)(cid:12)(cid:12)δG(cid:63)
(cid:48)
ti,1 by removing from Xti,1 all elements
(cid:12)(cid:12)(cid:12) > θ. The first condition verifies
Then, it obtains the pruned set of candidates X
{vj1} such that α(cid:63)(vj1) (cid:54)= α+(s1), or
that the first-use-as-sybil consistency property {vj1} (cid:39)φ {s1} holds, with φ = {(s1, vj1)}.
The second condition is analogous to the one applied in the noise-tolerant sybil subgraph
retrieval method introduced in [20] as part of the so-called robust active attack. It aims to
exclude from the search tree all candidates X such that ∆((cid:104)X(cid:105)w
) > θ, where ∆
G(cid:63)
ti
is a structural dissimilarity function and θ is a tolerance threshold.
(vj1) − δG+
,(cid:104)Sti(cid:105)w
(s1)
G+
ti
ti
ti
After pruning Xti,1, the method builds the set of cardinality-2 partial candidates
Xti,2 = {{vj1, vj2} {vj1} ∈ Xti,(cid:96), vj2 ∈ VG(cid:63)
ti \ {vj1}}.
Similarly, Xti,2 is pruned by removing all elements {vj1, vj2} such that {vj1, vj2} (cid:54)(cid:39)φ {s1, s2},
with φ = {(s1, vj1), (s2, vj2)}, and ∆((cid:104){vj1, vj2}(cid:105)w
,(cid:104){s1, s2}(cid:105)w
) > θ.
G+
ti
G(cid:63)
ti
In general, for (cid:96) ≤ Sti, the method builds the set of partial candidates
Xti,(cid:96) = {{vj1, . . . , vj(cid:96)} {vj1 . . . , vj(cid:96)−1} ∈ Xti,(cid:96)−1,
vj(cid:96) ∈ VG(cid:63)
ti \ {vj1, . . . , vj(cid:96)−1}}
and obtains the pruned candidate set X
such that
(cid:48)
ti,(cid:96) by removing from Xti,(cid:96) all elements {vj1, . . . , vj(cid:96)}
{vj1, . . . , vj(cid:96)} (cid:54)(cid:39)φ {s1, . . . , s(cid:96)},
13
with φ = {(s1, vj1), . . . , (s(cid:96), vj(cid:96))}, and
∆((cid:104){vj1, . . . , vj(cid:96)}(cid:105)w
G(cid:63)
ti
) > θ.
G+
ti
,(cid:104){s1, . . . , s(cid:96)}(cid:105)w
(cid:96)(cid:88)
(cid:12)(cid:12)(cid:12)δ(cid:48)
(cid:12)(cid:12)(cid:12)
(sk) − δ(cid:48)
G(cid:63)
ti
G+
ti
(vjk)
In our attack, we use the structural dissimilarity measure defined in [20], which makes
∆((cid:104){vj1, . . . , vj(cid:96)}(cid:105)w
k=1
(sk) is is the number of neighbours of sk in G+
,(cid:104){s1, . . . , s(cid:96)}(cid:105)w
) = D +
G+
ti
G(cid:63)
ti
(vjk) is is the number of neighbours of vjk in G(cid:63)
ti
where δ(cid:48)
G+
ti
δ(cid:48)
G(cid:63)
ti
ti that are not in {s1, . . . , s(cid:96)},
that are not in {vj1, . . . , vj(cid:96)}, and
D = {(jk, jk(cid:48)) EG(cid:63)
ti ∩ {(vjk, vjk(cid:48) )} + EG+
ti ∩ {(sk, sk(cid:48))} = 1}
is an efficiently computable estimation of the edge-edit distance between (cid:104){vj1, . . . , vj(cid:96)}(cid:105)G(cid:63)
and (cid:104){s1, . . . , s(cid:96)}(cid:105)G+
Sti candidates, that is
Finally, the sybil subgraph retrieval method gives as output the pruned set of cardinality-
ti
ti
.
Xti = X
(cid:48)
ti,Sti.
In other words, our method gives as output the set of temporally consistent subsets whose
weakly induced subgraphs in G(cid:63)
are structurally similar, within a tolerance threshold θ,
ti
to that of the original set of sybil nodes in G+
ti .
4.2.2 Fingerprint matching
After Xti is obtained, a candidate X = {vj1, vj2, . . . vjSti
} is randomly selected from Xti,
with probability 1Xti, for conducting the fingerprint matching step. Let vj1 ≺ vj2 ≺ . . . ≺
be the order established on the elements of X by the sybil subgraph retrieval method.
vjSti
Our fingerprint matching method is a depth-first search procedure, which gives as
output a set YX = {φ1, φ2, . . . , φq}, where every φ ∈ YX has the form φ : Yti → NG(cid:63)
(X).
Every element of YX maximises the pairwise similarities between the original fingerprints
of the victims and the fingerprints, with respect to X, of the corresponding pseudonymised
vertices.
ti
u = NG(cid:63)
ti
The method first finds all equally best matches between the (real) fingerprint Fj of a
(X) with respect to X,
victim yj ∈ Yti and that of a temporally consistent vertex u ∈ NG(cid:63)
that is F (cid:63)
(u) ∩ X. Then, for every such match, it recursively applies the search
procedure to match the remaining real victims to other temporally consistent candidate vic-
tims. For every victim yj and every candidate match u, the similarity function sim(F (cid:63)
u , Fj)
integrates the verification of the temporal consistency and the structural fingerprint, and
is computed as
ti
sim(F (cid:63)
u , Fj) =
u , Fj)
if u (cid:39) yj
and simc(F (cid:63)
otherwise.
u , Fj) ≥ η
simc(F (cid:63)
0
14
where η is a tolerance threshold allowing to ignore insufficiently similar matches and the
function simc(F (cid:63)
u , Fj) is defined as
Sti(cid:88)
with
µk(F (cid:63)
u , Fj) =
simc(F (cid:63)
u , Fj) =
µk(F (cid:63)
u , Fj)
(cid:26) 1
0
k=1
if vjk ∈ F (cid:63)
otherwise.
u and sk ∈ Fj
Our method is similar to the one introduced in [20] in the fact that it discards match-
ings whose structural similarity is insufficiently high. Moreover, temporal consistency con-
straints allow our method to considerably reduce the number of final candidate mappings.
4.2.3 Re-identification refinement
ti+1
ti \ VG(cid:63)
After the ti+1-th snapshot is released, the adversary obtains additional information that
can improve the re-identification at the ti-th snapshot. In specific, the adversary learns
the set VG(cid:63)
of pseudonyms corresponding to users that ceased to be members
of the social network in the interval between the ti-th and the ti+1-th snapshots. If the
adversary removed some sybil nodes sj1, sj2, . . . , sjr in this interval, then she knows
that {ϕti(sj1), ϕti(sj2), . . . , ϕti(sjr)} ⊆ VG(cid:63)
. This information allows the adversary
to refine the set Xti obtained in the first re-identification attempt on G(cid:63)
. Certainly, a
ti+1 (cid:54)= Sti \ Sti+1 is not a valid match for Sti.
candidate X such that X \ VG(cid:63)
that violate the sybil-removal-count consistency notion, that is
, the adversary refines Xti by removing the candidates
Thus, after the publication of G(cid:63)
ti \ VG(cid:63)
ti+1
ti+1
ti
(cid:8)X X (cid:54)(cid:39) (Sti, Sti+1)(cid:9) ,
(cid:48)
ti
X
= Xti \
and re-runs the fingerprint matching step with X
(cid:48)
ti
.
4.3 Sybil Subgraph Creation and Update
Here we describe how the necessary tempo-structural patterns for dynamic re-identification
are created and maintained. Both the sybil subgraph creation and the sybil subgraph
update stages contribute to this task. Additionally, the sybil subgraph update addresses
other aspects of the dynamic behaviour of the new attack, e.g. targeting new victims.
4.3.1 Sybil subgraph creation
In the dynamic attack, the initial sybil graph is not necessarily created before the first
snapshot is released. Let G(cid:63)
be the first snapshot where the adversary conducts a re-
ti
identification attempt. Then, the sybil sugbraph creation is executed during the entire time
window preceding ti. The adversary initially inserts a small number of sybil nodes, no more
15
(cid:106)
(cid:16)
(cid:17)(cid:107)
log2
VG+
ti
. This makes the sybil subgraph very unlikely to be detected by sybil
than
defences [37, 36, 2, 19, 20], while allowing to create unique fingerprints for a reasonably
large number of potential initial victims. Spreading sybil injection over several snapshots
helps create temporal patterns that reduce the search space during sybil subgraph retrieval.
As sybil nodes are inserted, they are connected to other sybil nodes and to some of the
victims. Inter-sybil edges are created in a manner that has been shown in [2] to make the
sybil subgraph unique with high probability, which helps in accelerating the breath-first
search procedure in the basis of sybil subgraph retrieval. First, an arbitrary (but fixed)
order is established among the sybil nodes. In our case, we simply take the order in which
the sybils are created. Let s1 ≺ s2 ≺ . . . ≺ sSti represent the order established among
the sybils. Then, the edges (s1, s2), (s2, s3), . . . , (sSti−1, sSti) are added to force the
existence of the path s1s2 . . . sSti. Additionally, every other edge (sj, sk), j − k ≥ 2, is
added with probability 0.5. The initial fingerprints of the elements of Yti are randomly
generated by connecting each victim to each sybil node with probability 0.5, checking that
all fingerprints are unique.
4.3.2 Sybil subgraph update
ti−1
and G(cid:63)
ti
Let G(cid:63)
be two consecutive releases occurring after the first snapshot where the
adversary conducted a re-identification attempt (G(cid:63)
itself may have been this snapshot).
In the interval between G(cid:63)
, the adversary updates the sybil subgraph by adding
and/or removing sybil nodes and inter-sybil edges, updating the fingerprints of (a subset of)
the victims, and possibly targeting new victims. The changes made in the sybil subgraph
aim to improve the sybil subgraph retrieval and fingerprint matching steps in future re-
identification attempts. We describe each of these modifications in detail in what follows.
and G(cid:63)
ti
ti−1
ti−1
Adding and replacing sybil nodes
log2
(cid:106)
(cid:16)
(cid:17)(cid:107)
In our attack, the adversary is conservative regarding the number of sybil nodes, balancing
the capacity to target more victims with the need to keep the likelihood of being detected by
sybil defences sufficiently low. Thus, the number of sybil nodes is increased as the number
. If the graph growth
of nodes in the graph grows, but keeping Sti ≤
rate between releases is small, this strategy translates into not increasing the number of
sybils during many consecutive releases. However, this does not mean that new sybil nodes
are not created, since the attack additionally selects a (small) random number of existing
sybil nodes and replaces them for fresh sybil nodes. The purpose of these replacements is
twofold. First, they allow to keep the activity level of each individual sybil node sufficiently
low, and thus make it less distinguishable from legitimate nodes. Secondly, the frequent
modification of the set of sybils helps reduce the search space for sybil subgraph retrieval, by
increasing the number of potential candidates violating the first-use-as-sybil consistency
constraint, and enables the re-identification refinement process to detect more sybil set
candidates violating the sybil-removal-count consistency constraint.
VG(cid:63)
ti−1
16
1, s(cid:48)
ti , and let s(cid:48)
node is replaced. Let S(cid:48) = {s(cid:48)
We now discuss the changes that the adversary does in the set of inter-sybil edges to
handle sybil node addition and replacement. Let Sti−1 = {s1, s2, . . . , sSti−1} be the set
of sybil nodes present in G+
ti−1, and let s1 ≺ s2 ≺ . . . ≺ sSti−1 be the order established
among them. We first consider the case of sybil node addition, where no existing sybil
2, . . . , sq} be the set of new sybil nodes that will be
2 ≺ . . . ≺ sq be the order established on them. A structure
ti the edges
q−1, s(cid:48)
q), which results in extending the path s1s2 . . . sSti−1 into
ti every node (x, y), x ∈ S(cid:48),
1 ≺ s(cid:48)
added to G+
analogous to that of the previous sybil subgraph is enforced by adding to G+
(sSti−1, s(cid:48)
1), (s(cid:48)
1, s(cid:48)
2), . . . , (s(cid:48)
1s(cid:48)
s1s2 . . . sSti−1s(cid:48)
2 . . . sq. Additionally, the adversary adds to G+
y ∈ (Sti−1 ∪ S(cid:48)) \ NG+
Now, we describe the modifications made by the adversary for replacing a sybil node
sj ∈ Sti−1 for a new sybil node s (s /∈ S(cid:48)).
In this case, the adversary adds to G+
ti
the edges (sj−1, s) and (s, sj+1), where sj−1 and sj+1 are the sybil nodes immediately
preceding and succeeding sj according to ≺. The order ≺ is updated accordingly to make
s1 ≺ s2 ≺ . . . ≺ sj−1 ≺ s ≺ sj+1 ≺ . . . ≺ sSi−1. These modifications ensure that the path
s1s2 . . . sSti−1 guaranteed to exist in G+
ti for s1s2 . . . sj−1ssj+1 . . . sSi−1.
Additionally the new sybil node s is connected to every other sybil node with probability
0.5. In our attack, every sybil node removal is part of a replacement, so the number of
sybil nodes never decreases.
ti−1 is replaced in G+
(x), with probability 0.5.
ti
Updating fingerprints of existing victims
After replacing a sybil node s ∈ Sti−1 for a new sybil node s(cid:48)
to G+
s for s(cid:48) does not render any pair of fingerprints identical in G+
ti .
ti the edge (s(cid:48), y) for every y ∈ Yti−1 ∩ NG+
ti−1
∈ Sti\Sti−1, the adversary adds
(s), to guarantee that the replacement of
Finally, if the adversary has conducted a re-identification attempt on G(cid:63)
Additionally, if new sybil nodes were added, the fingerprints of all previously targeted
victims in Yti−1 are modified by creating edges linking them to a subset of the new sybil
nodes. For each new sybil node s ∈ Sti \ Sti−1 and every victim y ∈ Yti−1, the edge (s, y) is
added with probability 0.5.
, she makes
additional changes in the set Fti of fingerprints in G+
ti based on the outcomes of the re-
identification, specifically on the level of uncertainty suffered because of Fti−1. To that
end, she selects a subset Y (cid:48)
of victims whose fingerprints were the least useful during the
re-identification attempt, in the sense that they were the most likely to lead to a larger
number of equally likely options after fingerprint mapping. The adversary modifies the
by randomly flipping one edge of the form (y, s), s ∈ Sti−1,
checking that the new fingerprint does not coincide with a previously existing fingerprint.
is obtained as follows. For every victim yj ∈ Yti−1, let pj : Rj → [0, 1] be
fingerprint of every y ∈ Y (cid:48)
The set Y (cid:48)
ti−1
ti−1
ti−1
ti−1
a probability distribution where
(cid:110)
(cid:111)
Rj =
v ∈ VG(cid:63)
ti−1 ∃X∈Xti−1∃φ∈YX φ(v) = yj
17
is the set of vertices mapped to yj according to some X ∈ Xti−1 and the corresponding YX.
That is, for every element v ∈ Rj, pj(v) represents the probability that v has been mapped
to yj in the previous re-identification attempt according to some sybil subgraph candidate
and some of the resulting fingerprint matchings, and is defined as
where
τj(φ, v) =
pj(v) =
(cid:26) 1
0
(cid:80)
(cid:80)
X∈Xti−1
τj (φ,v)
φ∈YX
YX
Xti−1
if v ∈ dom(φ) and φ(v) = yj
otherwise.
Finally, the set Y (cid:48)
ti−1
is obtained by making
where H(pj) is the entropy of the distribution pj, that is
Y (cid:48)
ti−1
= arg max
yj∈Yti−1 {H(pj)},
(cid:88)
pj(v) log pj(v).
H(pj) = −
v∈Rj
If the maximum entropy value is reached for more than one victim, all of their fingerprints
are modified. We chose to use entropy for obtaining Y (cid:48)
because it is a well established
quantifier of uncertainty.
ti−1
Targeting new victims
(cid:110)
After the fingerprints of existing victims have been updated, the final step of sybil subgraph
update consists in targeting new victims. To that end, the adversary chooses a set of
− Yti−1 − 1
vertices y1, y2, . . . , yr ∈ Vti\Yti−1, with 0 ≤ r ≤ min
and adds them to Yti. The new victims can either be fresh vertices, that is vertices that
first appear in G+
ti , or previously untargeted ones. The fingerprints of the new victims are
created in the same manner as those of the initial victims. That is, for every yj, 1 ≤ j ≤ r,
the vertex (yj, s), s ∈ Sti, is added to G+
ti with probability 0.5, checking that the newly
generated fingerprint is different from those of all previously targeted victims.
ti−1 − Yti−1, 2Sti−1
VG(cid:63)
(cid:111)
5 Experimental Evaluation
In this section, we experimentally evaluate our new dynamic active attack. Our evaluation
has three goals. First, we show that our attack outperforms Mauw et al.'s static robust ac-
tive attack [20] in terms of both effectiveness and efficiency. For simplicity, throughout this
section we will use the acronym D-AA for our attack and S-RAA for the static robust active
attack. Secondly, we determine the factors that affect the performance of our new attack,
18
and evaluate their impact. From this analysis, we derive a number of recommendations
allowing data owners to balance privacy preservation and utility in random perturbation
methods for periodical social graph publication. Due to the scarcity of real-life temporally
labelled social graphs, we run the aforementioned experiments on synthetic dynamic social
graphs. We developed a synthesiser that can be flexibly configured to generate synthetic
dynamic social graphs with specific properties, e.g. the initial number of nodes and the
growth rate. Using this synthesiser, we run each instance of our experiments in a collec-
tion of 100 synthetic datasets, which allows to mitigate the impact of random components
of our attack and the synthesiser itself. To conclude, we replicate some of the previous
experiments on a real-life dataset, to show that some of the findings obtained on synthetic
data remain valid in practical scenarios.
5.1 Experimental Setting
We implemented an evaluation tool based on the model described in Sect. 3. A dynamic
social graph simulator loads a real-life dataset, or uses the synthesiser, to generate the se-
quence G = (Gt1, Gt2, . . . , Gti, . . .) containing only legitimate users. Each snapshot is then
processed by a second module that simulates sybil subgraph creation or update. The out-
put, which is the data owner's view of the social graph, is processed by graph perturbation
module. In this module, we implement a simple perturbation method consisting in the
addition of cumulative noise. Finally, a fourth module simulates the re-identification on
the perturbed graph and computes the success probability of the attack. Sybil subgraph
creation and update, as well as re-identification, are discussed in detail in Sect. 4. We
describe in what follows the implementation of the remaining modules.
5.1.1 Dynamic social graph simulator
Our simulator allows us to conduct experiments on temporally annotated real-life datasets,
as well as synthetic datasets.
In the first case, the simulator extracts from the dataset
the graph snapshots by using a specific handler. The simulator is parameterised with
a sequence of the time-stamps indicating when each snapshot should be taken. Every
snapshot is built by taking all vertices and edges created at a moment earlier or identical
to the corresponding time-stamp and still not eliminated.
As we mentioned, we include in our simulator a synthesiser for generating periodically
released dynamic social graphs, which is based on the Barab´asi-Albert (BA) generative
graph model [1]. We use BA because it preserves the properties of real social graphs, namely
power-law degree distribution [1], shrinking diameter [13], and preferential attachment.
The BA model generates scale-free networks by iteratively adding vertices and creating
connections for the newly added vertices using a preferential attachment scheme. This
means that the newly added vertices are more likely to be connected to previously existing
nodes with larger degrees. The BA model has two parameters: the number of nodes n0 of
a (small) seed graph, and the initial degree Me (Me ≤ n0) of every newly added node. The
initial seed graph can be any graph. In our case we use a complete graph Kn0. Every time a
19
new node v is added to the current version G of the BA graph, Me edges are added between
v and randomly selected vertices in VG. The probability of selecting a vertex w ∈ VG for
creating the new edge (v, w) is
δG(x), as prescribed by preferential attachment.
To generate the graph sequence, the synthesiser takes four parameters as input:
δG(w)
x∈VG
(cid:80)
• The parameter n0 of the BA model.
• The parameter Me of the BA model.
• The number of vertices of the first snapshot nv.
• The growth rate r∆, defined as the proportion of new edges compared to the previous
number.
The parameters nv and r∆ determine when snapshots are taken. The first snapshot is
taken when the number of vertices of the graph generated by the BA model reaches nv,
and every other snapshot is taken when the ratio between the number of new edges and
that of the previous snapshot reaches r∆.
5.1.2 Graph perturbation via cumulative noise addition
To the best of our knowledge, all existing anonymisation methods against active attacks
based on formal privacy properties [18, 19] assume a single release scenario, and are thus
insufficient for handling multiple releases. Proposing formal privacy properties that take
into account the specificities of the multiple release scenario is part of the future work.
In our experiments, we adapted the other known family of perturbation methods, random
noise addition, to the multiple release scenario.
, the noise incrementally added on G(cid:63)
is re-applied on the pseudonymised graph ϕtiG+
, and then fresh noise is added on G(cid:63)
ti
To account for the incrementality of the publication process, the noise is added in a
t1, G(cid:63)
cumulative manner. That is, when releasing G(cid:63)
t2,
ti
. . . , G(cid:63)
ti to obtain an intermediate noisy
ti−1
graph G(cid:63)
that is released.
ti
,
In re-applying the old noise, all noisy edges incident in a vertex v ∈ VG(cid:63)
ti−1 \ Vϕti G+
ti
removed after the release of G(cid:63)
, are forgotten. The fresh noise addition consists in
randomly flipping a number of edges of G(cid:63)
is
ti
, the edge is removed, otherwise it is added. The
uniformly selected and, if (v, w) ∈ E G(cid:63)
cumulative noise addition method has one parameter: the amount of fresh noise to add in
each snapshot, called noise ratio and denoted Ωnoise. It is computed with respect to E G(cid:63)
ti,
the number of edges of the pseudonymised graph after restoring the accumulated noise.
. For every flip, a pair (v, w) ∈ V G(cid:63)
to obtain the graph G(cid:63)
ti
ti × V G(cid:63)
ti
ti−1
ti
5.1.3 Success probability
As in previous works on active attacks for the single release scenario [18, 19, 20], we evaluate
the adversary's success in terms of the probability that she correctly re-identifies all victims,
20
P r(ti)
succ =
X∈Xti
Xti
p
(ti)
X
if Xti (cid:54)= ∅
otherwise
(cid:80)
0
(cid:26) 1YX
0
Figure 2: Effectiveness comparison between S-RAA and D-AA.
which in our scenario is computed by the following formula for the ti-th snapshot:
where
p(ti)
X =
if ∃φ∈YX φ−1 = ϕtiYti
otherwise
and, as discussed in Sect. 3.3, ϕti is the isomorphism applied on G+
nymised graph ϕtiG+
re-identification refinement is executed.
ti . For every snapshot G(cid:63)
ti
ti to obtain the pseudo-
, we compute success probability after the
5.2 Results and Discussion
We begin our discussion with the comparison of D-AA and S-RAA. Then, we proceed to
study the factors that affect the effectiveness of our attack, and characterise their influence.
Finally, we use the real-life dataset Petster [12] to illustrate the effectiveness of our attack
in practice. For the first two sets of results, we use synthetic dynamic graphs generated by
our synthesiser. Table 1 summarises the different configurations used for the generation.
Furthermore, before each release, we select at least 1 and maximum 5 random legitimate
vertices as new victims. For each parameter combination, we generated 100 synthetic
dynamic graphs, and the results shown are the averages over each subcollection.
5.2.1 Comparing D-AA and S-RAA
The goal of this comparison is to show that our dynamic active attack outperforms the
original attack in both effectiveness and efficiency. We use six settings for the dynamic
21
24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=200D-RAAS-RAA24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=400D-RAAS-RAA24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=800D-RAAS-RAA24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=200D-RAAS-RAA24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=400D-RAAS-RAA24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=800D-RAAS-RAA Sect. n0 Me
5.2.1 30
5.2.2 30
5, 10
5, 10
nv
200, 400,800
2000, 4000, 8000
r∆ Ωnoise(%)
5% 0.5
5% 0.5, 1.0, 1.5, 2.0
Table 1: Combinations of parameters for the simulator.
graph synthesiser. For each value of Me (5 and 10), we set the initial number of vertices
nv at 200, 400 and 800. In all our experiments, sybil subgraph creation spans the first
and second snapshots, and the re-identification is executed for the first time on the second
snapshot.
Effectiveness comparison. In Fig. 2 we show the success probabilities of the two attacks
on graphs with different initial sizes and amounts of changes between consecutive releases
(determined by Me with r∆ fixed as 5%).
We have three major observations from the results. First, we can see that D-AA
significantly outperforms S-RAA in terms of success probability. The improvement becomes
larger when smaller numbers of changes occur between releases. D-AA outperforms S-RAA
by at least twice, even up to three times for the first few snapshots. When graphs grow
slowly (Me = 5), our attack always displays an average success probability larger than 0.5
which S-RAA never reaches. Second, the success probability of the original S-RAA has a
general trend to drop along with time, while D-AA displays a large increase from the first
to the second snapshot, and then remains stable or degrades very slowly. This is the result
of the reduced uncertainty enabled by the temporal consistency constraints. The frequent
modification of the set of sybil nodes allows our attack to offset the noises accumulated
over time and maintain an acceptable success probability even at the later releases. When
Me is set to 10, the success probability remains above 0.5 until the sixth snapshot, even
though noise grows faster in this case. Third, S-RAA is more likely to be influenced by
the randomness of graph structures and noise, as shown by the large fluctuations of the
success probability, while our D-AA displays smaller variance and smoother curves.
Efficiency comparison. Fig. 3 shows the average amount of time consumed by S-RAA
and D-AA in different scenarios. We can see that D-AA takes an almost constant amount
of time at all snapshots, whereas the time consumption of S-RAA grows considerably
along time. This clearly shows that the use of temporal information in dynamic social
graphs helps D-AA to effectively avoid the computation overhead. We highlight the fact
that D-AA runs at least 10 times faster in almost all cases, especially in late snapshots.
An interesting observation is that the running time of S-RAA decreases when Me = 10
and the size of the initial graph is 800. Rather than an improvement, this is in fact the
consequence of the repeated failure of the sybil subgraph retrieval algorithm to find any
candidates. This problem is in turn caused by the small tolerance threshold required by
S-RAA to complete runs in reasonable time.
22
Figure 3: Efficiency comparison between S-RAA and D-AA.
5.2.2 Factors influencing our attack
We intend this analysis to serve as a guide for customising the settings of privacy-preserving
publication methods for dynamic social graphs, in particular for determining the amount
of perturbation needed to balance the privacy requirements and the utility of published
graphs. Since our attack is considerably efficient, we will use for this evaluation dynamic
graphs featuring 2000, 4000 and 8000 initial vertices. In addition, we set large tolerance
thresholds for structural dissimilarity in the sybil subgraph retrieval method, which in-
creases the probability that it finds ϕti(Sti) as a candidate. In our experiment, we set the
threshold θti used at the i-th snapshot to θti = min(1500, 16 + 250× (i− 2)2). Three factors
may possibly impact the effectiveness of re-identification attacks on dynamic graphs: the
amount of noise, the size of graphs and the speed of growth between two releases. As a
result, we analyse three parameters which determine these three factors in our simulator:
Ωnoise, Me and nv. The number of vertices in the initial snapshots nv determines the scale
of the released graphs, while the parameter Me of the BA model controls the number of
new nodes and edges added before the next release. The noise ratio Ωnoise determines the
amount of noise.
Effectiveness. Fig. 4 shows the success probability of our attack when different noise
ratios are applied on dynamic graphs with different initial sizes and growth speeds. First,
we can see that the success probability decreases when more noise is applied. This is
natural as more perturbation makes it more difficult to find the correct sybil sugraph,
either because the sybil graph has been too perturbed to be found as a candidate or because
edge perturbation generates more subgraphs similar to the original sybil subgraph. When
Me = 5 and the noise ratio is set to 0.5%, success probability always remains above 0.5. For
this value of Me, even with Ωnoise at 2.0%, the attack still displays success probability above
0.5 in the first three snapshots. Second, the rate at which success probability decreases
slows down as we increase the value of the noise ratio. The largest drop occurs when
23
24681012141618snapshot01020304050time (s)Me=5, nv=200D-RAAS-RAA24681012141618snapshot01020304050time (s)Me=5, nv=400D-RAAS-RAA24681012141618snapshot01020304050time (s)Me=5, nv=800D-RAAS-RAA24681012141618snapshot01020304050time (s)Me=10, nv=200D-RAAS-RAA24681012141618snapshot01020304050time (s)Me=10, nv=400D-RAAS-RAA24681012141618snapshot01020304050time (s)Me=10, nv=800D-RAAS-RAA Figure 4: Factors influencing the success probability.
Figure 5: Factors influencing the utility of released graphs.
we increase Ωnoise from 0.5 to 1.0. This suggests that keeping increasing the level of
perturbation may not necessarily guarantee a better privacy protection, but just damage
the utility of the released graphs. Third, the sequences of success probability values show a
very small dependence on the initial size of the graphs, with other parameters fixed. Last,
the success probability decreases when dynamic graphs grow faster. From the figure, we
can see that the probability reduces by about 10% when we increase the value of Me from
5 to 10.
Summing up, we observe that the risk of re-identification decreases when more pertur-
bation is applied and when the graphs grow faster, whereas the initial size of the graphs
has a relative small impact on this risk.
Utility. We evaluate the utility of released graphs in terms of three measures: the per-
centage of edge editions, the variation of the average local clustering coefficient, and the
KL-divergence of degree distributions. The first measure quantifies the percentage of edge
flips with respect to the total number of edges. In fact, it quantifies the amount of noise
24
24681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=200024681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=400024681012141618snapshot0.00.20.40.60.81.0success probabilityMe=5, nv=800024681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=200024681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=400024681012141618snapshot0.00.20.40.60.81.0success probabilityMe=10, nv=8000noise=0.5%noise=1.0%noise=1.5%noise=2.0%24681012141618snapshot0510152025edge difference (%)24681012141618snapshot510152025avg lcc difference (%)24681012141618snapshot1.001.011.021.031.041.051.061.07KL-divergencenoise=0.5%,Me=5noise=1.0%,Me=5noise=1.5%,Me=5noise=2.0%,Me=5noise=0.5%,Me=10noise=1.0%,Me=10noise=1.5%,Me=10noise=2.0%,Me=10 accumulated so far. For the ti-snapshot, the percentage of edge editions is computed as
1
EG+
ti {(v, v(cid:48)) ∈ EG+
ti (ϕ(v), ϕ(v(cid:48))) (cid:54)∈ G(cid:63)
ti} ∪ {(v, v(cid:48)) ∈ EG(cid:63)
ti (ϕ−1(v), ϕ−1(v(cid:48))) (cid:54)∈ G+
ti} .
The local clustering coefficient (LCC) of a vertex measures the proportion of pairs of
mutual neighbours of the vertex that are connected by an edge. We calculate the average
to the LCCs of all vertices and take the proportion between the differences of the value of
the original graph and that of the anonymised graph as, that is
avgLcc(G+
ti ) − avgLcc(G(cid:63)
avgLcc(G+
ti )
ti
)
.
where avgLcc(G) is the average local clustering coefficient of graph G. Our last measure
uses the KL-divergence [11] as an indicator of the difference between the degree distribution
of the original graph and that of the perturbed graph.
As all the three measures present almost identical patterns for different values of nv, we
only show the results for nv = 8000. We have two major observations. First, as expected,
the values of all three measures increase as the noise accumulates over time, indicating
that the utility of released graphs deteriorates. Even with Ωnoise set to just 1.0%, at the
tenth snapshot we can have up to 10% of edges flipped and changes in edge density around
15%. At this point, we can say that the utility of released graphs has already been greatly
damaged. Second, when dynamic graphs grow faster, the impact of noise becomes smaller,
as more legitimate edges offset the impact of noisy edges.
Together with the finding that larger growth speed results in smaller success probability,
we can conclude that the social networks that grow fast among releases display a better
balance between re-identification risk and the utility of the released graphs.
5.2.3 Results on a real-life dynamic social graph
We make use of a publicly available graph collected from Petster, a website for pet owners to
communicate [12], to validate to what extent the results reported in the previous subsection
remain valid in a more realistic domain. The Petster dataset is an undirected graph whose
vertices represent the pet owners. The vertices are labelled by their joining date, which
span from January 2004 to December 2012. The graph is incremental, which means no
vertices are removed. It contains 1898 vertices and 16, 750 edges. We take a snapshot every
six months.
We present in Fig. 6(a) the success probability of our D-AA attack on the Petster
dataset when the noise ratio is set to 0.5%, 1.0% and 1.5%, respectively. Compared to the
success probabilities discussed above on simulated graphs, the curves have different shapes
and more fluctuations. This is because, instead of a fixed growth speed (determined by r∆
and Me in our simulator), the real-life graph grows at different speeds in different periods,
as shown in Fig. 6(b). We can see that the number of new vertices varies before each
release. After the first few years of steady growth, Petster gradually lost its popularity,
25
(a) Re-identification risk on Petster.
(b) Numbers of new vertices added be-
fore release.
Figure 6: Evaluation on Petster.
especially with few new vertices added in the last three years. By cross-checking the two
figures, we can see that the success probability changes with the amount of growth before
the corresponding release. It first increases steadily due to the steady growth of the graph
until the fifth snapshot, which suddenly has the largest number of new vertices. Then when
the growth slows down, the success probability also recovers; and when the growth stops
(e.g., from the 12 snapshot), it starts increasing again, even though the noise continues
to accumulate. These observations validate our findings on synthetic graphs, that is, the
speed of growth is the dominating factor that affects the re-identification risk.
6 Conclusions
In this paper, we have presented the first dynamic active re-identification attack on pe-
riodically released social graphs. Unlike preceding attacks, the new attack exploits the
inherent dynamic nature of social graphs by leveraging tempo-structural patterns for re-
identification. Compared to existing (static) active attacks, our new dynamic attack sig-
nificantly improves success probability, by more than two times, and efficiency, by almost
10 times. Through comprehensive experimental evaluation on synthetic data, we analysed
the factors influencing the success probability of our attack, namely the growth rate of
the graph and the amount of noise injected. These findings can subsequently be used
to develop graph anonymisation methods that better balance privacy protection and the
utility of the released graphs. For instance, for a given noise level, the decision to publish
a new snapshot should be determined by taking into account the number of changes that
have occurred from the last release. Similarly, if the time for the next release is given,
the amount of noise should be customised according to the number of changes that have
occurred. Additionally, we evaluated our attack on Petster, a real-life dataset. This evalu-
ation showed that some of the findings obtained on synthetic data remain valid in practical
scenarios.
Acknowledgements: The work reported in this paper received funding from Luxem-
bourg's Fonds National de la Recherche (FNR), via grant C17/IS/11685812 (PrivDA).
26
2468101214snapshot0.00.20.40.60.81.0success probabilitynoise=0.5%noise=1.0%noise=1.5%01234567891011121314151617snapshot050100150200250300350400#new users References
[1] R´eka Albert and Albert-L´aszl´o Barab´asi. Statistical mechanics of complex networks.
Review of Modern Physics, 74:47 -- 97, 2002.
[2] Lars Backstrom, Cynthia Dwork, and Jon M. Kleinberg. Wherefore art thou r3579x?:
anonymized social networks, hidden patterns, and structural steganography. Commu-
nications of the ACM, 54(12):133 -- 141, 2011.
[3] Jordi Casas-Roma, Jordi Herrera-Joancomart´ı, and Vicen¸c Torra. An algorithm for
k-degree anonymity on large networks. In Procs. of the 2013 IEEE/ACM Int'l Conf.
on Advances in Social Networks Analysis and Mining, pages 671 -- 675, 2013.
[4] Jordi Casas-Roma, Jordi Herrera-Joancomart´ı, and Vicen¸c Torra. k-degree anonymity
improving data utility in large networks. Knowledge and Infor-
and edge selection:
mation Systems, 50(2):447 -- 474, 2017.
[5] Sean Chester, Bruce M Kapron, Ganesh Ramesh, Gautam Srivastava, Alex Thomo,
and S Venkatesh. Why waldo befriended the dummy? k-anonymization of social
networks with pseudo-nodes. Social Network Analysis and Mining, 3(3):381 -- 399, 2013.
[6] Narimene Dakiche, Fatima Benbouzid-Si Tayeb, Yahya Slimani, and Karima Be-
natchba. Tracking community evolution in social networks: A survey. Information
Processing & Management, 56(3):1084 -- 1102, 2019.
[7] Xuan Ding, Lan Zhang, Zhiguo Wan, and Ming Gu. De-anonymizing dynamic social
networks. In 2011 IEEE Global Telecommunications Conference-GLOBECOM 2011,
pages 1 -- 6. IEEE, 2011.
[8] Shouling Ji, Weiqing Li, Mudhakar Srivatsa, and Raheem Beyah. Structural data
de-anonymization: Quantification, practice, and implications. In Proceedings of the
2014 ACM SIGSAC Conference on Computer and Communications Security, pages
1040 -- 1053. ACM, 2014.
[9] Shouling Ji, Weiqing Li, Mudhakar Srivatsa, Jing Selena He, and Raheem Beyah.
In
Structure based data de-anonymization of social networks and mobility traces.
International Conference on Information Security, pages 237 -- 254. Springer, 2014.
[10] Nitish Korula and Silvio Lattanzi. An efficient reconciliation algorithm for social
networks. Proceedings of the VLDB Endowment, 7(5):377 -- 388, 2014.
[11] S. Kullback and R.A. Leibler. On information and sufficiency. The Annals of Mathe-
matical Statistics, 22(1):79 -- 86, 1951.
[12] J´erome Kunegis. KONECT: the koblenz network collection. In Proc. 22nd Interna-
tional World Wide Web Conference (WWW), pages 1343 -- 1350. ACM Press, 2013.
27
[13] Jure Leskovec, Jon M. Kleinberg, and Christos Faloutsos. Graph evolution: Densi-
fication and shrinking diameters. ACM Transactions on Knowledge Discovery from
Data (TKDD), 1(1):2, 2007.
[14] Kun Liu and Evimaria Terzi. Towards identity anonymization on graphs. In Proc.
2008 ACM SIGMOD International Conference on Management of Data (SIGMOD),
pages 93 -- 106. ACM Press, 2008.
[15] Linyuan Lu and Tao Zhou. Link prediction in complex networks: A survey. Physica
A: statistical mechanics and its applications, 390(6):1150 -- 1170, 2011.
[16] Xuesong Lu, Yi Song, and St´ephane Bressan. Fast identity anonymization on graphs.
In Procs. of the Int'l Conf. on Database and Expert Systems Applications, pages 281 --
295, 2012.
[17] Tinghuai Ma, Yuliang Zhang, Jie Cao, Jian Shen, Meili Tang, Yuan Tian, Abdullah
Al-Dhelaan, and Mznah Al-Rodhaan. Kdvem: a k-degree anonymity with vertex and
edge modification algorithm. Computing, 97(12):1165 -- 1184, 2015.
[18] Sjouke Mauw, Yunior Ram´ırez-Cruz, and Rolando Trujillo-Rasua. Anonymising social
graphs in the presence of active attackers. Transactions on Data Privacy, 11(2):169 --
198, 2018.
[19] Sjouke Mauw, Yunior Ram´ırez-Cruz, and Rolando Trujillo-Rasua. Conditional adja-
cency anonymity in social graphs under active attacks. Knowledge and Information
Systems, 61(1):485 -- 511, 2018.
[20] Sjouke Mauw, Yunior Ram´ırez-Cruz, and Rolando Trujillo-Rasua. Robust active at-
tacks on social graphs. Data Mining and Knowledge Discovery, 33(5):1357 -- 1392, 2019.
[21] Arvind Narayanan, Elaine Shi, and Benjamin IP Rubinstein. Link prediction by
In The 2011
de-anonymization: How we won the kaggle social network challenge.
International Joint Conference on Neural Networks, pages 1825 -- 1834. IEEE, 2011.
[22] Arvind Narayanan and Vitaly Shmatikov. De-anonymizing social networks. In Proc.
30th IEEE Symposium on Security and Privacy (S&P), pages 173 -- 187. IEEE Com-
puter Society, 2009.
[23] Shirin Nilizadeh, Apu Kapadia, and Yong-Yeol Ahn. Community-enhanced de-
anonymization of online social networks. In Proc. 2014 ACM SIGSAC Conference on
Computer and Communications Security (CCS), pages 537 -- 548. ACM Press, 2014.
[24] Shirin Nilizadeh, Apu Kapadia, and Yong-Yeol Ahn. Community-enhanced de-
anonymization of online social networks. In Proceedings of the 2014 acm sigsac con-
ference on computer and communications security, pages 537 -- 548. ACM, 2014.
28
[25] Fragkiskos Papadopoulos and Kaj-Kolja Kleineberg. Link persistence and conditional
distances in multiplex networks. Physical Review E, 99(1):012322, 2019.
[26] Pedram Pedarsani, Daniel R Figueiredo, and Matthias Grossglauser. A bayesian
method for matching two similar graphs without seeds. In 2013 51st Annual Allerton
Conference on Communication, Control, and Computing (Allerton), pages 1598 -- 1607.
IEEE, 2013.
[27] Wei Peng, Feng Li, Xukai Zou, and Jie Wu. Seed and grow: An attack against
anonymized social networks.
In Procs. of the 9th Annual IEEE Communications
Society Conf. on Sensor, Mesh and Ad Hoc Communications and Networks, pages
587 -- 595, 2012.
[28] Wei Peng, Feng Li, Xukai Zou, and Jie Wu. A two-stage deanonymization attack
against anonymized social networks. IEEE Transactions on Computers, 63(2):290 --
303, 2014.
[29] Fran¸cois Rousseau, Jordi Casas-Roma, and Michalis Vazirgiannis. Community-
preserving anonymization of graphs. Knowledge and Information Systems, 54(2):315 --
343, 2017.
[30] Juli´an Salas and Vicen¸c Torra. Graphic sequences, distances and k-degree anonymity.
Discrete Applied Mathematics, 188:25 -- 31, 2015.
[31] Chih-Hua Tai, Peng-Jui Tseng, S Yu Philip, and Ming-Syan Chen.
Identities
anonymization in dynamic social networks. In 2011 IEEE 11th International Con-
ference on Data Mining, pages 1224 -- 1229. IEEE, 2011.
[32] Rolando Trujillo-Rasua and Ismael Gonz´alez Yero. k-metric antidimension: A privacy
measure for social graphs. Information Sciences, 328:403 -- 417, 2016.
[33] Yazhe Wang, Long Xie, Baihua Zheng, and Ken CK Lee. High utility k-anonymization
for social network publishing. Knowledge and Information Systems, 41(3):697 -- 725,
2014.
[34] Wentao Wu, Yanghua Xiao, Wei Wang, Zhenying He, and Zhihui Wang. K-symmetry
model for identity anonymization in social networks. In Procs. of the 13th Int'l Conf.
on Extending Database Technology, pages 111 -- 122, 2010.
[35] Lyudmila Yartseva and Matthias Grossglauser. On the performance of percolation
graph matching. In Proceedings of the first ACM conference on Online social networks,
pages 119 -- 130. ACM, 2013.
[36] Haifeng Yu, Phillip B Gibbons, Michael Kaminsky, and Feng Xiao. Sybillimit: A
near-optimal social network defense against sybil attacks. In Procs. of the 2008 IEEE
Symposium on Security and Privacy, pages 3 -- 17, Oakland, CA, USA, 2008.
29
[37] Haifeng Yu, Michael Kaminsky, Phillip B Gibbons, and Abraham Flaxman. Sybil-
guard: defending against sybil attacks via social networks. In Procs. of the 2006 Conf.
on Applications, Technologies, Architectures, and Protocols for Computer Communi-
cations, pages 267 -- 278, Pisa, Italy, 2006.
[38] Bin Zhou and Jian Pei. Preserving privacy in social networks against neighborhood
In Procs. of the 2008 IEEE 24th Int'l Conf. on Data Engineering, pages
attacks.
506 -- 515, Washington, DC, USA, 2008.
[39] Lei Zou, Lei Chen, and M. Tamer Ozsu. K-automorphism: A general framework for
privacy preserving network publication. PVLDB, 2(1):946 -- 957, 2009.
30
|
1204.3860 | 1 | 1204 | 2012-04-17T17:57:58 | Macroscopes: models for collective decision making | [
"cs.SI",
"cs.CC"
] | We introduce a new model of collective decision making, when a global decision needs to be made but the parties only possess partial information, and are unwilling (or unable) to first create a globalcomposite of their local views. Our macroscope model captures two key features of many real-world problems: allotment structure (how access to local information is apportioned between parties, including overlaps between the parties) and the possible presence of meta-information (what each party knows about the allotment structure of the overall problem). Using the framework of communication complexity, we formalize the efficient solution of a macroscope. We present general results about the macroscope model, and also results that abstract the essential computational operations underpinning practical applications, including in financial markets and decentralized sensor networks. We illustrate the computational problem inherent in real-world collective decision making processes using results for specific functions, involving detecting a change in state (constant and step functions), and computing statistical properties (the mean). | cs.SI | cs | PROCEEDINGS, CI 2012
2
1
0
2
r
p
A
7
1
]
I
S
.
s
c
[
1
v
0
6
8
3
.
4
0
2
1
:
v
i
X
r
a
MACROSCOPES: MODELS FOR COLLECTIVE DECISION MAKING
Subramanian Ramamoorthy
School of Informatics
University of Edinburgh
10 Crichton Street, Edinburgh
EH8 9AB, United Kingdom
[email protected]
Andr´as Z. Salamon
School of Informatics
University of Edinburgh
10 Crichton Street, Edinburgh
EH8 9AB, United Kingdom
[email protected]
Rahul Santhanam
School of Informatics
University of Edinburgh
10 Crichton Street, Edinburgh
EH8 9AB, United Kingdom
[email protected]
ABSTRACT
We introduce a new model of collective decision mak-
ing, when a global decision needs to be made but the
parties only possess partial information, and are unwill-
ing (or unable) to first create a global composite of their
local views. Our macroscope model captures two key fea-
tures of many real-world problems: allotment structure
(how access to local information is apportioned between
parties, including overlaps between the parties) and the
possible presence of meta-information (what each party
knows about the allotment structure of the overall prob-
lem). Using the framework of communication complex-
ity, we formalize the efficient solution of a macroscope.
We present general results about the macroscope model,
and also results that abstract the essential computa-
tional operations underpinning practical applications,
including in financial markets and decentralized sensor
networks. We illustrate the computational problem in-
herent in real-world collective decision making processes
using results for specific functions, involving detecting a
change in state (constant and step functions), and com-
puting statistical properties (the mean).
INTRODUCTION
We consider collective decision making processes such as
a market that acts as a central mechanism for coordinat-
ing the actions of autonomous participants. We address
the questions: how does one measure the quality of the
collective decision making process, and how weak can
the central market mechanism be? In many applications,
there is significant interest in decentralizing computation
while still being able to arrive at results that cannot be
computed entirely locally. We use a simple model to cap-
ture the informational complexity of computing global
functions by aggregating results from participants1 who
are endowed with arbitrary allotments of local infor-
mation. This allows us to draw conclusions about the
requirements on allotments and protocols, for efficient
collective information processing. A key aspect of our
model is the specification of meta-information based on
distinguishing perfect information, single-blind arrange-
ments, and double-blind arrangements. Our technical
framework is built on the notions of communication com-
plexity. We assume that participants possess informa-
tion which is not available to other participants; we call
this the private information.
1We use the common English words agent, party, participant,
and player synonymously, ignoring more specific usage.
This work is motivated by several applications.
A rather timely application is found in the domain of
participants in electronic markets. Often, such as in fi-
nancial markets, participating agents would benefit from
an understanding of the global system dynamics (Darley
& Outkin 2007).2 For instance, agents might like to have
signals that indicate the presence of herding, bubbles and
other aggregate phenomena. Typically, the local view of
a single agent does not provide sufficient information to
reliably detect this. Moreover, in such a domain, one is
tightly constrained by what information can be revealed,
incentives to reveal this information, and other aspects
related to privacy in computation.
If we seek efficient
decentralized information processing mechanisms under
these constraints, then we would like to be able to deter-
mine what is or is not possible, employing only a coarse
characterization of resources and endowment of informa-
tion. Recent studies such as in anonymized financial chat
rooms (Lu & Mizrach 2011) provide interesting insights
into the behaviour of such collectives, such as the charac-
terization of equilibria in which a subset of traders profit
from the information of others. This is but one example
of a larger body of economic literature related to phe-
nomena in networked markets (Hurwicz & Reiter 2006).
However, in that literature, it is not typical to investigate
our question of how the allotment structure and com-
munication protocols relate to the efficiency with which
specific types of computation are achieved. For instance,
change detection (Basseville & Nikiforov 1993) is of fun-
damental importance in financial markets -- how weak a
protocol is sufficient to decide a change has occurred?
Recent work on the topic of complexity of financial com-
putations by Arora, Barak, Brunnermeier & Ge (2011)
indicates that this is a fertile direction to pursue.
Similar issues arise in many other application domains,
such as mobile sensor networks and distributed robotics.
Leonard, Paley, Lekien, Sepulchre, Fratantoni & Davis
(2007) describe a mobile sensor network for optimal data
gathering, using a combination of underwater and sur-
face level sensing robots to optimally gather informa-
tion such as chemical and thermal dynamics in a large
2See
http://www.bankofengland.co.uk/publications/
speeches/2009/speech386.pdf for a discussion of
this
issue from the perspective of financial regulation, and
http://www.bis.gov.uk/foresight/our-work/projects/
current-projects/computer-trading for information on
a major study by the UK government, under the Foresight
Project.
PROCEEDINGS, CI 2012
volume of water (typically measured in square miles).
Similar systems have been utilized for tracking oil spills
and algal blooms. A key computational method utilized
by such distributed robotic networks involves distributed
optimization (Bullo, Cort´es & Mart´ınez 2009). The de-
ployment of modules in such a network needs to satisfy
a spatial coverage requirement, described as a cost func-
tion, so that each module plans trajectories to optimize
this criterion. The sensor fusion problem, to determine
a combined estimate of an uncertain parameter, may
also be posed an as optimization problem in the sense
of maximizing information gain. Despite this rigorous
approach, relatively little is known about how to com-
pare different formulations of these optimization prob-
lems -- given that we are interested in a certain type
of global function (say, number of peaks in a chemical
concentration profile or some distributional aspect of the
overall field) using weak local sensing and the ability to
move sensing nodes, how does one compare or otherwise
characterize protocols and other aspects of the problem
formulation?
A line of work that begins to touch upon some of these
questions is that of Ghrist and collaborators. Barysh-
nikov & Ghrist (2009) use tools from algebraic topology
to solve the problem of counting targets using very weak
local information (such as unlabelled counts in a local
neighbourhood). de Silva & Ghrist (2007) present an
approach to detection of holes in coverage through decen-
tralized computation of homology. Here again, the focus
being on aspects of the specific function being computed,
the authors do not address the relationship between the
protocols and problem formulation, and the efficiency of
computation.
Extending the idea of decentralized computation in so-
cial systems, consider the problem faced by a program
committee, such as one that might review this paper. We
seek a decentralized computation of a ranking problem.
Similar ranking problems also occur in executive decision
making such as the hiring decision in academic depart-
ments. The key issue here is that of parsimonious infor-
mation sharing, coupled implicitly or explicitly with the
meta-information problem. These challenges arise due
to limitations on the capacities of the decision makers to
exchange information with each other.
The common theme underlying all of these applications
is the computation of a function based on allotment of
portions of the information to parties who have reason-
able amounts of computational resource but would like
to keep exchange of information limited. We wish to
understand how weak the corresponding protocols can
be, for various types of functions. Major categories of
functions of interest include change detection and rank-
ing. We model change detection by an abstract version
of the key underlying problem, of determining whether
the data forms a constant or a step function. The main
statistical property we consider here is the computation
of the mean.
We are interested in understanding just how much com-
munication must occur to answer various questions of
interest. Therefore,
instead of working with detailed
models for the questions about market behaviour or sen-
sor networks discussed previously, we have deliberately
kept the models we study as simple as possible. This
makes our lower bound results stronger. Determining an
answer to any more realistic question will require even
more information to be exchanged than in these simple
models, as long as the more realistic model includes the
simpler problem at its core. It therefore makes sense in
our setting to study the simplest possible embodiment
of each of the core problems. For the upper bounds, our
results are a first step and will need to be extended to
more realistic models.
MODEL
Our model is based on the notion of communication
complexity (Kushilevitz & Nisan 1997), which has been
highly influential in computer science. A Boolean func-
tion models yes/no decisions, by requiring that the func-
tion take either the value 0 or the value 1. A quantity
with value either 0 or 1 is known as a bit, and quantities
that are drawn from a larger range of values can be ex-
pressed by using multiple bits; a function that is defined
over a domain containing 2n different values is said to
have an n-bit input. Say two players Alice and Bob wish
to compute a Boolean function f on a 2n-bit input, but
Alice only has access to the first n bits and Bob to the
other n bits. Alice and Bob are not computationally
constrained, but they are informationally constrained.
The question now is: how many bits of information do
Alice and Bob need to exchange to compute f on a given
2n-bit input? A protocol for this problem specifies, given
the inputs to the players and the communication so far,
which player is the next to send information, as well
as what information is actually sent. There is a trivial
protocol where Alice merely sends her part of the input
to Bob. Bob now has all the information he needs to
compute f , and he sends back the 1-bit answer to Alice.
The cost of a protocol is the total number of bits that
are exchanged; this simple protocol has a cost of n + 1
for any function. The field of two-party communication
complexity studies, for various functions f of interest,
whether more efficient protocols exist. As an example,
for the Equality function which tests whether Alice and
Bob's inputs are exactly the same, it is known that the
communication upper bound of n + 1 is tight for de-
terministic protocols, but there is an improved protocol
with cost O(log(n)) when the players' messages are al-
lowed to be randomized and it is sufficient for the final
answer to be correct with high probability.
The notion of communication complexity can be general-
ized from the two-party setting to the multi-party setting
(Chandra, Furst & Lipton 1983). Here the number of
players is not limited to two, each player has some infor-
mation about the global input, and they wish to compute
some Boolean function of the global input. There are two
PROCEEDINGS, CI 2012
standard models for how the input is distributed among
the players: the number-in-hand model (NIH) and the
number-on-forehead (NOF) model. Suppose there are k
players and the global input is N bits long. In the NIH
model, there is some fixed partition of the global input
into k parts, and each player gets one of these parts. In
the NOF model, again there is a fixed partition into k
parts, but the i-th player gets all the parts except the
i-th part.
The main motivation for our model is that in many sit-
uations, such as financial markets or sensor networks,
information is distributed among players in a more com-
plex fashion than in the NOH or NIF models. Moreover,
the players might not have control over which pieces of
information they have access to -- the allotment of inputs
to players might be arbitrary, perhaps even adversarial.
As an example, creators of financial instruments may
decide which assets to bundle into pools that are then
offered for sale. Purchasers of such instruments might
wish to check fairness of allocation, without revealing
to each other their precise holdings; or regulators might
wish to check that sellers behaved impartially but with-
out relying on full disclosure.
Yet, the players might still wish to compute some func-
tion of their global input in this less structured setting.
Now different kinds of question arise than in the standard
communication complexity setting. For a given function,
which kinds of allotment structures allow for protocols?
Does the meta-knowledge of what the allotment struc-
ture actually is make a difference to whether there is an
efficient protocol or not? These questions are interesting
even for simple functions which have been thoroughly
investigated in the standard setting.
To be more formal, let f be a function which k play-
ers wish to compute on a global input x of size N bits.
An (N, k) allotment structure is a sequence of k subsets
S1, S2 . . . Sk of [N ] = {1, 2, . . . , N}. An allotment struc-
ture corresponds to an allotment of input bits among
players in the following way: Player i receives all bits xj
for j ∈ Si. Note that unlike in the NOH and NIF mod-
els, this allotment of input bits is completely general --
it might be the case that two players receive the same
set of bits, for example. This is the main novelty of our
approach. Our intention here is to model two kinds of
situations. In the first, the players have little control over
which pieces of information they can access -- they have
to do the best they can, with the available information.
In the second, the allotment is made by a centralized
authority, and it is of interest to study which allotment
would most facilitate the computation in question.
A k-player macroscope on N bits is simply a function f
on N input bits together with an (N, k) allotment struc-
ture. We will abuse notation and sometimes use a macro-
scope to refer to a sequence of functions fN , N = 1 . . .∞,
where each function fN depends on N bits. This will en-
able us to pose and study the question of the asymptotic
efficiency of protocols for macroscopes.
x
x1
x2
x3
x4
...
xN
S1 = {1, 2, N}
S2 = {1, 3, 4, N}
...
Sk = {2, 3, 4}
Figure 1. Example (N, k) allotment structure. We wish to
compute f (x) with a k-player protocol, in two situations.
(Single-blind) Player i knows {S1, . . . , SN} and xj for j ∈ Si.
(Double-blind) Player i only knows xj for j ∈ Si.
The generalized modelling of the allotment of inputs
raises the issue of meta-information -- how much do play-
ers know about the allotment of inputs, and how can
they take advantage of this? In the case of the NIH and
NOF models, the allotment is implicitly known to all
players because it is fixed in advance. However, in our
setting, there are two different kinds of situations -- the
single-blind situation and the double-blind situation. In
a single-blind macroscope, all players know the allotment
structure, however Player i does not know the values of
any input bits apart from the ones whose indices are in
Si. In a double-blind macroscope, the players are more
hampered in that they do not even know the allotment
structure, however they do know the indices of the bits
they receive.
It remains to formally define what a protocol is in our
model. To keep things simple, we focus on simultaneous-
message protocols, where each player broadcasts a se-
quence of bits to all players; this is often presented fig-
uratively as each player writing their bit string on a
universally viewable blackboard. A protocol solves a
macroscope if each player can determine the value of
the function on the global input simply by looking at its
own input bits as well as the information written on the
blackboard. The cost of a protocol for a macroscope is
PROCEEDINGS, CI 2012
then the total length of strings written by the players.
In protocols for single-blind macroscopes, the message
of Player i is a function of the values of bits whose in-
dices are in Si as well as of the allotment structure. For
double-blind macroscopes, the message of Player i is a
function only of the values of bits whose indices are in
Si.
We make a deliberate choice in our modelling to be
highly general in terms of the allotment structure, and to
be specific in terms of the structure of the actual commu-
nication. This is because our main aim is to understand
the impact of the allotment structure on efficiency of
communication. Our model can be extended to allow
more degrees of freedom with regard to the communica-
tion structure. One way in which this can be done is to
allow multiple-round protocols, where players communi-
cate in turns, with the protocol specifying whose turn it
is to communicate. Another is by allowing randomness
-- here each player is assumed to have access to a private
source of randomness, on which its message can depend.
A third way is to restrict communication to take place
between specified pairs of players, i.e., there is an implicit
topology of communication. This third approach is taken
in the field of distributed algorithms (Lynch 1996), where
however input allotment is not modelled in a flexible way.
We are interested in protocols which have communica-
tion as low as possible. This is desirable not just in terms
of efficiency, such as meeting bandwidth constraints, but
also in terms of privacy. In applications such as financial
markets, the players would like to obtain some global
knowledge without revealing their own inputs. Thus,
Player i has more than one reason for not following the
trivial protocol of publishing the values of all bits in Si.
The lower the communication, the less the information
revealed about the values of bits held by individual play-
ers; we will rely on this link between parsimony of com-
munication and the weakness of the coordination mecha-
nism. Privacy requirements are modelled more explicitly
in sub-areas of cryptography such as secure multi-party
computation (Yao 1982, Goldreich 2004). We prefer not
to model these requirements explicitly so as not to com-
plicate our model too much.
We make no assumption about the relationship between
the number of players and the number of bits in the
global input. In an application such as sensor networks,
there might be few players (sensors), each having a large
amount of information, whereas in the financial markets
application, there are typically many players each having
few pieces of information. Our model deals equally well
with both extremes.
A first observation is that to compute a non-trivial func-
tion over the global input, i.e. a function that depends
on all the input bits, the allotment structure must sat-
isfy the covering property -- each index j ∈ [N ] lies in at
least one set Si of the allotment structure. If the cov-
ering property did not hold, consider an index j which
does not belong to the allotment, and an input X such
j
j
), where X f lip
that f is sensitive to X at index j, meaning that f (X) is
different from F (X f lip
is X with the value
of the jth bit flipped. By the non-triviality of f , such
an input X must exist. Clearly any protocol outputs the
same answer for X as for X f lip
since j does not belong to
the allotment, and hence the protocol cannot be correct.
Henceforth, we automatically assume that a macroscope
has the covering property.
j
There are no general necessary conditions on the allot-
ment structure beyond the covering property for compu-
tation of non-trivial functions. But intuitively, the more
"even" the allotment is, in the sense of each bit being
allotted to the same number of players, the easier it is
to compute a symmetric function of the inputs. We de-
fine an even (N, k) allotment structure as an allotment
structure for which there is a number C such that each
index i ∈ [N ] belongs to exactly C distinct subsets Si,
and each subset Si is of the same size. Clearly, for such
an allotment structure, each set Si is of size N C/k.
RESULTS
Our first results address the question of what we can say
in general about the cost of single-blind and double-blind
macroscopes.
Theorem 1. Every single-blind macroscope on N bits
has a protocol with cost N . Moreover, this bound is op-
timal.
Note that the upper bound does not depend on the num-
ber of players. The proof of the upper bound in Theo-
rem 1 takes advantage of the global knowledge the play-
ers have about the allotment structure.
Proof. Consider a protocol in which each input bit
Xi has a player "responsible" for it -- Player j is re-
sponsible for input bit Xi iff i ∈ Sj and i (cid:54)∈ Sk for
k < j. It follows from the covering property that each
input bit has a player responsible for it. It should also
be clear that at most one player is responsible for any
given input bit. The protocol consists of players sending
the values of all the bits they are responsible for. Since
the macroscope is single-blind, each player knows who is
responsible for which input bits. Hence each player can
reconstruct the input from the information sent in the
protocol, and therefore also compute the function on the
input. The cost of the protocol is N , since each bit has
exactly one player responsible for it.
To see that this bound is optimal, consider the macro-
scope consisting of the Parity function3 together with
an (N, N ) allotment structure which allots each input
bit to a distinct player. Suppose there is a protocol for
this macroscope where one of the players does not send
a message. Assume, without loss of generality, that this
player holds the i-th bit. Then the Parity function can-
not depend on the i-th bit, which is a contradiction.
3The Parity function returns 1 if an odd number of the input
bits are 1, and 0 otherwise.
PROCEEDINGS, CI 2012
Theorem 2. Every k-player double-blind macroscope
on N bits has a protocol with cost at most 2N k.
Proof. The protocol giving the upper bound of The-
orem 2 is a simple one. The j-th player sends an N -bit
string specifying its allotment Sj, as well as at most N
bits which specify the values of the bits whose indices are
in the allotment Sj. Once this information is made pub-
lic by each player, the players can each reconstruct the
input and hence compute the function on that input.
We do not know whether the bound in Theorem 2 is
optimal in general, but we believe this to be the case for
protocols that use only one round of communication.
Our primary focus is on studying the complexity of
macroscopes for problems which arise in contexts such as
electronic markets and distributed sensor networks. For
each of these problems, we are interested in issues such as
the optimal cost of solving a macroscope, the differential
cost of meta-information (the reduction in cost when us-
ing a single-blind macroscope rather than a double-blind
macroscope), and for single-blind macroscopes, the de-
pendency of the cost on the allotment structure.
A fundamental problem in the context of electronic mar-
kets is the Constancy problem of detecting whether a
given function is constant or not.4 We model this prob-
lem as a Boolean function on [D]N , which is 1 if all the
inputs are equal, and 0 otherwise. This requires a slight
adaptation of our model to inputs which are D-ary rather
than binary, but this adaptation can be done in a natural
way. We are able to characterize the cost of protocols for
single-blind Constancy macroscopes optimally in terms
of the allotment structure.
Given an (N, k) allotment structure, we define the inter-
section graph of the structure as follows. The graph has
N vertices, and there is an edge between vertex i and
vertex j for i, j ∈ [N ] iff Si ∩ Sj (cid:54)= ∅.
Theorem 3. Every k-player single-blind Constancy
macroscope on N D-ary inputs can be solved with cost
r((cid:100)log(D)(cid:101)) + k, where r is the number of connected
components of the intersection graph of the allotment
structure associated with the macroscope. Moreover, this
bound is optimal to a constant factor.
Note that the bound does not actually depend on N , the
number of inputs!
Proof. Since the macroscope is single-blind, each
player knows the identities of all the other players in the
same connected component in the intersection graph of
the allotment structure. This is because these identities
depend only on the allotment structure and not on the
input.
4Although the actual values may be constantly changing, we
use the abstract Boolean version of this problem to repre-
sent a more significant shift from some statistical notion of
the normal behaviour. For instance, we might ask if the ob-
served deviations from some base value are significant at a
5% confidence level.
The protocol is as follows: For each connected compo-
nent, it is only the player with the smallest index in that
connected component who sends a "long" message of
(cid:100)log(D)(cid:101) bits long, specifying a value in [D] that occurs
in its portion of the input. In addition, each player sends
a 1-bit message saying whether its portion of the input is
constant or not. The players know that Constancy holds
if each 1-bit message encodes "yes", and in addition, the
values in [D] sent in the long messages are all the same.
Indeed, if the function is constant, it is clear that all the
1-bit messages are "yes", and that the values in the long
message are the same. To argue the converse, just notice
that if a fixed player in a connected component of the
intersection graph sees a constant value l ∈ [D], and if
all other players in the connected component see a con-
stant value, it follows that all players in the connected
component see the same constant value l.
We argue that this bound is optimal up to a factor of
2. We will give separate arguments that k bits of com-
munication are required and that r(cid:100)log(D)(cid:101) bits of com-
munication are required. From these separate bounds, it
follows that max(k, r(cid:100)log(D)(cid:101)) ≥ (k + r(cid:100)log(D)(cid:101))/2 bits
of communication are required.
To see that k bits of communication are necessary, con-
sider any allotment structure such that Si \ (∪j(cid:54)=iSj) is
non-empty for each player i. Define a function f : [k] →
[N ] such that f (i) ∈ Si \ ∪j(cid:54)=iSj for each player i. Sup-
pose there is a player i who does not send a message.
Then the protocol gives the same answer on both the all
1's input and the input that is all 1 except at f (i), since
the communication pattern is the same for both of these
inputs. However the Constancy function differs on these
inputs.
Next we show the r(cid:100)log(D)(cid:101) lower bound. Consider
any allotment structure whose intersection graph con-
tains r connected components. Suppose that fewer than
r(cid:100)log(D)(cid:101) bits of communication are sufficient for solv-
ing the macroscope on this allotment structure. Then
there is some connected component C of the intersection
graph such that the corresponding players send less than
log(D) bits of communication in all. This implies that
there are two values v1, v2 ∈ [D] such that the communi-
cation pattern of players in C is exactly the same when
the players in C all receive the input v1 as when they all
receive the input v2. Now consider two inputs -- input x
in which all co-ordinates are the constant v1 and input
y in which all co-ordinates outside C have value v1 and
co-ordinates in C have value v2. The communication
pattern of the protocol is the same for x and y, however
the Constancy function is true for x and false for y. This
is a contradiction.
Thus, for single-blind Constancy macroscopes, the crit-
ical property of the allotment structure is the number
of connected components of the intersection graph. The
fewer the number of connected components, the more
efficiently the macroscope can be solved. We next study
the situation for double-blind macroscopes.
PROCEEDINGS, CI 2012
Theorem 4. Every k-player double-blind Constancy
macroscope on N D-ary inputs can be solved with cost
k(cid:100)log(D + 1)(cid:101). Moreover, there are k-player double-blind
Constancy macroscopes which require cost k(cid:100)log(D)(cid:101).
Proof. The protocol giving the upper bound is sim-
ple. Each player sends a message encoding one of D + 1
possibilities: either the players' portion of the input is
non-constant, or if it is constant, which of the D possible
values it is. The protocol accepts if each message encodes
the same value v ∈ [D].
For the lower bound, since each player is unaware of
players' allotments other than its own, the lower bound
of r(cid:100)log(D)(cid:101) in the proof of Theorem 3 holds with the
maximum possible value of r, namely r = k.
Thus, in the case of the Constancy function, the differ-
ential cost of meta-information can be quite significant,
especially when the intersection graph of the allotment
structure is connected.
Next we consider a formalization of the change detection
problem. The Boolean Step Function (BSF) problem is
defined as follows: a string x ∈ {0, 1}N evaluates to 1 if
there is an index i such that xj = 0 for j ≤ i and xj = 1
for j > i, or to 0 otherwise.
. . .
x1
x2
x3
x4
x5
. . .
xN−1 xN
Figure 2. Boolean step function (BSF).
Possibilities for different structures of allotment lead to a
twofold challenge in a BSF macroscope. First, no party
may see the step, so parties need to share some informa-
tion about the values they see. Second, if several parties
detect a step, they then need to determine whether they
are observing the same step.
. . .
x1
x2
x3
x4
x5
. . .
xN−2 xN−1 xN
Figure 3. A Boolean function that is not a step function.
scope on N bits can solved with cost 2k(cid:100)log(N )(cid:101).
Theorem 5. Every k-player double-blind BSF macro-
Proof. The protocol is as follows: for each i ∈ [k],
Player i sends two indices l(i) and m(i), where l is the
largest index in Si for which xl = 0 and m is the smallest
index in Si for which xm = 1. Given all these messages,
each player can calculate the value of the smallest index
m for which xm = 1 simply by taking the minimum of
m(i) over all players i, as well as the largest index l for
which xl = 0, simply by taking the maximum of l(i) over
all players i. Note that BSF (x) = 0 iff l = m − 1.
We conjecture that the bound of Theorem 5 is tight
for double-blind BSF macroscopes for protocols that
use only one round of communication. For single-blind
macroscopes, however, we can do better.
Theorem 6. Every k-player single-blind BSF macro-
scope on N bits can be solved with cost k(cid:100)log(N )(cid:101) + 2k.
Proof. The protocol witnessing the upper bound is
as follows. Each player sends a message consisting of two
parts. The first part is 2 bits long, and specifies which of
the following is the case: (1) the player's portion of the
input is constant, (2) there is a single transition from 0 to
1 in the player's input, (3) neither (1) nor (2) holds. The
second part is (cid:100)log(N )(cid:101) bits long. The interpretation of
the second part of the message is as follows: if case (1)
holds for the first part, then the second part encodes
which constant (either 0 or 1) the player is given.
If
case (2) holds, then the second part encodes the index
at which a transition occurs, i.e., a number j ∈ [N ] such
that j ∈ Si (assuming that the player in question is
Player i) and such that for all l ∈ Si, l ≤ j =⇒ xj = 0
and l > j =⇒ xj = 1. If case (3) holds, the contents of
the second part of the message are irrelevant.
From the messages, the players can either reconstruct
the input x, if case (1) or (2) holds for each player, or
conclude directly that BSF (x) = 0, if case (3) holds for
any player. From the input x, each player can compute
BSF (x) on its own.
In this case, too, the advantage of using single-blind
macroscopes can be seen, though a matching lower
bound in Theorem 5 is needed to prove this.
Next, we attempt to model the averaging function. Dis-
tributional statistics of different kinds need to be com-
puted in a decentralized way in various contexts such as
sensor networks. To model this, we again depart from
the framework of bit strings as inputs. The input is
now a sequence of real numbers {xi}, xi ∈ [0, 1] for each
1 ≤ i ≤ N . Given a parameter > 0, we study the cost
of protocols for -Averaging macroscopes, where each
player needs to arrive at an -additive approximation
to the average of the numbers xi by using the protocol
to communicate.
Theorem 7. Let > 0 be fixed. Every k-player
single-blind -Averaging macroscope on N inputs can be
solved with cost k(cid:100)log(k/)(cid:101).
Notice again that there is no dependence of the cost on
N , merely on the number of players and the approxima-
tion error.
Proof. We again use critically the meta-information
of players about the allotment structure. Each player
i knows, for each index j ∈ Si, the number Nj of dis-
tinct players receiving input xj. The message sent by
Player i is an /k-additive approximation to the quan-
tity Σj∈Sixj/Nj. Since the quantity is between 0 and
PROCEEDINGS, CI 2012
1, the approximation can be specified using (cid:100)log(k/)(cid:101)
bits. Given the messages of all players, each player can
compute an -approximation to the average simply by
summing all the individual approximation. Since the
individual approximations are /k-additive approxima-
tions, the sum will be an -approximation to the aver-
age.
In the case of Averaging macroscopes, we can show that
the double-blind restriction is a significant one, in that
it leads to a dependence of the cost on the number of
players. The proof uses a dimensionality argument.
Theorem 8. Let > 0 be a parameter. There are
2-player double-blind -Averaging macroscopes on N in-
puts which require cost N log(1/(N )).
Proof. Consider a 2-player double-blind macroscope
for -Averaging, with Player 1 receiving an allotment
S1 ⊆ [N ] and Player 2 receiving an allotment S2 ⊆ [N ].
Our assumption that the allotment structure is covering
implies S1∪S2 = [N ], but neither player knows anything
about the other player's allotment beyond this fact.
We show that a player's message must essentially reveal
its input. Consider Player 1, and let S1 = l. Now
for each i ∈ S1, consider S2,i = [N ] \ {i}.
If Player
1 has allotment S1 and Player 2 has allotment S2,i,
then in a correct protocol, Player 2 eventually knows
an -approximation to Σj∈[N ]xj/N . Since it also knows
all values except xi, this implies that it knows an N -
approximation to xi. Thus an N -approximation to xi
should be extractable from Player 1's message. This
should hold for every i, which means Player 1 must send
at least l log(1/(N )) bits. By a symmetric argument,
Player 2 must send at least (N − l) log(1/(N )) bits,
which means that at least N log(1/(N )) bits are com-
municated in all.
Theorem 8 has the disadvantage that it does not say
much about , which may be large in comparison to
N . However, we believe that a refinement of the proof
which argues about information revealed about subsets
of inputs rather than individual inputs can be used to
establish an improved lower bound.
DISCUSSION
The previous two sections have focused on the details of
a model for collective information processing and prop-
erties of this model. We now take a step back to discuss
why these results are of relevance to the motivating ex-
amples identified in the Introduction.
Consider the problem of agents in financial markets. It
is increasingly the case that, with the emergence of di-
verse communication methods and agents deciding at
time scales ranging from microseconds to days, com-
mon knowledge isn't so commonly available in prac-
tice. This has significant implications for dynamics
(Rubinstein 1989) and much of the modern discus-
sion regarding markets is related to such issues (Arora
et al. 2011). In this setting, there is a need for diagnostic
tools that could provide useful signals -- by computing
global properties, i.e. functions, based on local informa-
tion that can be used subject to limitations on protocols.
For instance, has there been a change from a 'constant'
level in a global sense? Or, is there a significant differ-
ence -- in the form of a step change -- between segments
of a networked market? We illustrate the use of con-
cepts from complexity theory to address abstract ver-
sions of such questions. Indeed, our techniques could be
used to answer further such questions -- about statistical
distributions, ranking queries, etc. A key novelty, even
in comparison to the state of the art in communication
complexity, is that we consider an arbitrary endowment
of inputs and meta-information such as single and double
blind protocols.
An important general direction for future work in this
area would be to extend our analysis to more directly
address the subtleties of the above mentioned dynam-
ics as they occur in applications of interest. Also, we
would like to better understand the relationships be-
tween our model of decentralized computation under in-
formational constraints and previous established models,
such as (Hurwicz & Reiter 2006), which employ different
methods and focus more on temporally extended sequen-
tial protocols such as auctions.
In terms of more specific questions, our model could
be extended in various ways. We have considered the
case of a simple structure of communication, with si-
multaneous messages sent in one round of communica-
tion, and a possibly complex allotment structure of the
inputs. We could allow more sophisticated communi-
cation structure. For example, double-blind protocols
with two rounds of communication can emulate single-
round single-blind protocols with some loss in efficiency,
simply by using the first round to share publicly infor-
mation about the input allotment, and then running the
single-blind protocol in the second round. More gen-
erally, communication might be restricted to occur only
between specific pairs of players. In the context of sensor
networks, for instance, it is natural to model both the
location of information and the structure of communica-
tion as governed by the topology of the ambient space.
Even studying very simple functions such as Constancy
in such general models appears to be interesting.
We emphasize that we are interested in understand-
ing the communication requirements of even very sim-
ple functions in modelling frameworks that render them
non-trivial. This distinguishes our work from the exist-
ing research on communication complexity, where func-
tions such as Constancy and BSF are trivial because
the model of communication is so simple. We are espe-
cially interested in modelling some aspects of collective
information processing, such as information overlap and
meta-information, which have been neglected so far and
which we believe can have a significant impact on the
efficiency of communication. The way we capture these
notions is quite flexible and can be used both to model
PROCEEDINGS, CI 2012
computation of continuous quantities such as Average
and computation of discrete quantities such as Boolean
functions, as we have illustrated with our results.
Bullo, F., Cort´es, J. & Mart´ınez, S. (2009), Distributed
Control of Robotic Networks, Applied Mathematics
Series, Princeton University Press.
Chandra, A. K., Furst, M. L. & Lipton, R. J. (1983),
Multi-party protocols, in 'STOC 1983: Proceedings of
the fifteenth annual ACM symposium on Theory of
computing', ACM, pp. 94 -- 99.
doi:10.1145/800061.808737
Darley, V. & Outkin, A. V. (2007), A NASDAQ market
simulation: insights on a major market from the
science of complex adaptive systems, Vol. 1 of Complex
systems and interdisciplinary science, World Scientific.
de Silva, V. & Ghrist, R. (2007), 'Homological sensor
networks', Notices Amer. Math. Soc. 54(1), 10 -- 17.
Available from: https:
//www1.ams.org/notices/200701/fea-ghrist.pdf
Goldreich, O. (2004), Foundations of Cryptography -
Volume 2, Cambridge University Press.
Hurwicz, L. & Reiter, S. (2006), Designing Economic
Mechanisms, Cambridge University Press.
doi:10.2277/0521836417
Kushilevitz, E. & Nisan, N. (1997), Communication
Complexity, Cambridge University Press.
Leonard, N., Paley, D., Lekien, F., Sepulchre, R.,
Fratantoni, D. & Davis, R. (2007), 'Collective motion,
sensor networks, and ocean sampling', Proceedings of
the IEEE 95(1), 48 -- 74.
doi:10.1109/JPROC.2006.887295
Lu, J. & Mizrach, B. (2011), 'Strategic Interaction in a
Stock Trading Chat Room', SSRN eLibrary . Available
from: http://ssrn.com/paper=1746464
Lynch, N. (1996), Distributed Algorithms, Morgan
Kaufmann.
Rubinstein, A. (1989), 'The electronic mail game:
Strategic behavior under "Almost Common
Knowledge"', American Economic Review
79(3), 385 -- 391. Available from:
http://www.jstor.org/stable/1806851
Yao, A. (1982), Protocols for secure computations, in
'FOCS 1982: Proceedings of 23rd Annual Symposium
on Foundations of Computer Science', IEEE Computer
Society, pp. 160 -- 164. doi:10.1109/SFCS.1982.38
Another direction that we find compelling is modelling
meta-information in a more sophisticated way. As of
now, we have the single-blind model and the double-
blind model. But there are various intermediate notions
that are reasonable to study. For example, each player
might know the number of players and the number of in-
puts but nothing about which inputs are given to which
other players. Or a player might know which other play-
ers also receive the inputs it receives, but nothing about
inputs it does not receive. Or some global property of
the allotment structure, such as that the allotment struc-
ture is even, might be known. Notice that in the case
of Averaging macroscopes, there is an efficient proto-
col whose cost doesn't depend on the number of players
if the allotment structure is even and the size of each
allotment is known. The protocol simply involves the
players summing all their inputs and dividing by a uni-
versal constant. In general, one can ask: assuming that
there are efficient single-blind protocols known, what is
the minimal information about the allotment structure
required to give efficient protocols?
Ranking is an important subject that we have not ad-
dressed in this paper. Some interesting problems can
be captured as ranking macroscopes, and we leave their
study for further work.
We have seen that in some cases more information hin-
ders making a global decision, rather than helping. More
generally, why is allotment structure important? We are
striving to fully understand how the allotment structure
of information affects our ability to efficiently answer
questions that require global information.
ACKNOWLEDGEMENTS
S.R. would like to acknowledge the support of the
UK Engineering and Physical Sciences Research Coun-
cil (grant number EP/H012338/1) and the European
Commission (TOMSY Grant Agreement 270436, un-
der FP7-ICT-2009.2.1 Call 6). A.S. acknowledges the
support of EPSRC via platform grant EP/F028288/1.
R.S. acknowledges the support of EPSRC via grant
EP/H05068X/1.
REFERENCES
Arora, S., Barak, B., Brunnermeier, M. & Ge, R.
(2011), 'Computational complexity and information
asymmetry in financial products', Communications of
the ACM 54(5), 101 -- 107. doi:10.1145/1941487.1941511
Baryshnikov, Y. & Ghrist, R. (2009), 'Target
enumeration via Euler characteristic integrals', SIAM
J. Appl. Math. 70(3), 825 -- 844. doi:10.1137/070687293
Basseville, M. & Nikiforov, I. V. (1993), Detection of
Abrupt Changes - Theory and Application,
Prentice-Hall, Inc.
|
1109.2215 | 2 | 1109 | 2011-11-20T19:21:15 | Finding missing edges and communities in incomplete networks | [
"cs.SI",
"physics.data-an",
"physics.soc-ph"
] | Many algorithms have been proposed for predicting missing edges in networks, but they do not usually take account of which edges are missing. We focus on networks which have missing edges of the form that is likely to occur in real networks, and compare algorithms that find these missing edges. We also investigate the effect of this kind of missing data on community detection algorithms. | cs.SI | cs | Finding missing edges and communities in incomplete
networks
Bowen Yan and Steve Gregory
Department of Computer Science, University of Bristol, Bristol BS8 1UB, England
E-mail: [email protected] +44 117 954 5142
Abstract. Many algorithms have been proposed for predicting missing edges in networks, but
they do not usually take account of which edges are missing. We focus on networks which have
missing edges of the form that is likely to occur in real networks, and compare algorithms that
find these missing edges. We also investigate the effect of this kind of missing data on
community detection algorithms.
1. Introduction
Many complex systems can be described as networks, in which vertices represent individuals, and
edges denote relations between pairs of vertices. Researchers analyse these networks in order to
understand the features and properties of them, but often neglect to consider the correctness and
completeness of the dataset itself. In most cases, they assume that the data are complete and accurate.
Most real-world networks represent the results of investigations and experiments, which are often
incomplete and inaccurate [1]. For example, in a 1996 survey of sexual behaviour, only 59% of
individuals responded to the interview [2].
Missing edges might affect the properties of networks very much. For example, the experimental
results of protein-protein interactions have found different properties by using different methods,
because 80% of the interactions of proteins are unknown [3]. Also, community detection algorithms
might place some vertices in incorrect communities if some edges are missing.
It is not easy to find missing edges in networks because we do not know where they exist. The
properties of networks will be affected in different ways by different types of missing edges. Many
edge prediction methods have been proposed, but they all consider different issues, and are normally
tested only on networks in which edges are missing at random. We believe that it would be beneficial
to know how or why edges are missing.
There are several common reasons for an edge to be missing; for example:
1. Edges may be missing at random [4].
2. Edges may be missing because of a limit on the number of neighbours of a vertex: the
problem of right censoring by vertex degree in social networks [5].
3. Many real-world networks are too large to analyse, so often a sample is extracted from them;
for example, by snowball sampling, where a crawling algorithm performs (e.g.) a breadth-first
search from an initial vertex to collect edges. Since the sample is a subgraph of the whole
network, many edges are missing at the periphery.
4. Edges may be missing when a vertex does not have many neighbours: the cold ends problem
in biological and information networks [6].
In this paper, we define three types of incomplete network based on the first three of the above
forms of missing data, and use them for comparing algorithms that find missing edges. We also
investigate the effect of the same types of missing data on community detection algorithms.
2. Related work
Researchers’ attention has increasingly focused on the evolution of networks. Networks are dynamic
objects which grow and change, so many new vertices and edges appear in an original network over
time. Although the study of missing edges usually focuses on static networks, both problems
concentrate on vertex similarity and the structural properties of networks [7].
Many methods have already been used to predict edges in numerous fields; for example, proximity
measures that are based on network topological features [8]; supervised learning methods [9]; and
relational learning methods [10,11], which consider relational attributes of elements in a relational
dataset. These and others are reviewed in the excellent survey by Lü and Zhou [12]. In our paper, we
restrict our attention to proximity measures in undirected, unweighted, and unipartite networks.
One of the most popular concepts of proximity between two vertices, structural similarity, is based
on the structure information of a network. This concerns the common features that pairs of vertices
have: for example, the more common neighbours two vertices have, the greater the chance that they
include Common neighbours and Jaccard coefficient
know each other. These methods
[8,13,14,15,16,17,18]; or the path information that two vertices share [19,20,21,22,23]. If two vertices
do not have any common neighbours, some extended algorithms can be used to calculate their
similarity [7,24], taking account of the similarity between the neighbours of two vertices. In Ref. [25],
the transferring similarity, which contains all high-order correlations between vertices, has been
proposed.
Also, Kossinets [5] has analysed the effect of missing data on social networks via some statistical
methods. These network-level statistical properties include mean vertex degree, clustering coefficient,
assortativity and average path length. In Ref. [5], Kossinets assumed that the experimental networks
were complete first, and removed some vertices and edges randomly for statistical analysis.
The point that network data are not reliable has also been addressed in [26], who emphasized that it
is easy to mislead the results of the experiments with incorrect data. Costenbader and Valente [27]
have analysed the stability of centrality measures on sampling networks with missing and spurious
data. Borgatti et al. [28] have explored the robustness of centrality measures with missing data. Other
work includes research on sampling networks with missing data [29].
Understanding the reasons for missing data is a significant goal of edge prediction, because the
structure and features of networks depends on the type of missing data. For example, in scale-free
networks, the mean degree is unaffected even if there is a large amount of random missing data, but is
severely affected by the removal of vertices with high degree [30].
3. Methodology
3.1. Types of incomplete network
We simulate three types of missing edges in both artificial networks and real-world networks, each
with a different cause.
With our first type, edges are missing at the boundary. Generally, this problem is caused by
crawling to collect a sample of a network, e.g., by a breadth-first search, so we call it a crawled
network. With the second type, edges are missing at random, e.g., because of the non-response
problem [3,4,5,27], so we name it a random-deletion network. With the third type, edges are missing
by reason of the fixed-choice problem (right-censoring by vertex degree [5,31]). Since this problem
corresponds to limiting the vertex degree, we call it a limited-degree network.
3.2. Testing edge prediction methods
For our edge prediction experiments, we construct incomplete copies of an initial network as follows:
1. Crawled network: We first choose a starting vertex – one of the vertices with the smallest
maximum distance from all other vertices – and then use a breadth-first search from the start
vertex to collect a sampled network containing the required number of edges. Figure 1(a)
shows the famous karate club network [32] while figure 1(b) shows its crawled network with
50% of the original network’s edges. This network includes all edges within distance 2 of the
starting vertex (shown in red) as well as some of the edges within distance 3, so that the total
number of edges is as required.
2. Random-deletion network: An edge is chosen randomly and deleted, repeatedly until the
required number of edges remain.
3. Limited-degree network: At each step, one of the vertices with the maximum degree, v, is
chosen randomly, and one of its edges {u,v} is chosen randomly for deletion, irrespective of
the degree of u.
(a)
(b)
(c)
(d)
(e)
(f)
Figure 1. (a) Karate network. (b) Crawled karate network, showing missing vertices (those present in
the original network but not the crawled network). (c) Induced subnetwork of karate network, showing
missing vertices (those present in the original network but not the subnetwork). (d) Crawled karate
network, showing missing edges (those present in the induced subnetwork but not the crawled
network). (e) Random-deletion karate network, showing missing edges. (f) Limited-degree karate
network, showing missing edges.
3.3. Testing the effect of missing data on community detection algorithms
The above methods of constructing incomplete networks may cause the network to become
disconnected. This is not a problem for our evaluation of edge prediction methods but it would have a
severe impact on the comparison of community detection algorithms, because these generally work
only on connected networks. Therefore, for our community detection experiments in Section 4.2, we
use somewhat different strategies to generate incomplete networks, which ensure that they are all
connected and all contain the same set of vertices as well as the same number of edges.
1. Crawled network: This is constructed in the same way as above (shown in figure 1(b)). To
construct our other incomplete networks, we first construct the subnetwork of the original
network induced by the vertices in our crawled network. Figure 1(c) shows the induced
subnetwork of the karate network, while figure 1(d) shows the difference between this and the
crawled network.
2. Random-deletion network: We start with the induced subnetwork (to ensure comparability
with the crawled network) and randomly delete edges repeatedly until the required number of
edges remain. When choosing an edge to delete, we never choose an edge whose removal
would make the network disconnected. Figure 1(e) shows a random-deletion network derived
from our induced subnetwork.
3. Limited-degree network: Again, we start with the induced subnetwork, and then randomly
choose edges {u,v} such that v has the maximum degree in this network and the deletion of
{u,v} would not make the network disconnected. If we cannot find an appropriate u, then we
choose a v with less than the maximum degree. Figure 1(f) shows a limited-degree network
and the induced subnetwork.
3.4. Network datasets used
We have experimented with three types of network. First, we use artificial random (Erdős-Rényi)
networks [33], because we expect edge prediction methods to be unsuccessful on these.
Second, we use the benchmark networks of Lancichinetti et al. [34]; these are artificial networks
that are claimed to reflect the important aspects of real-world networks. The networks have several
parameters:
1. n is the number of vertices.
2. k is the average degree. kmax is the maximum degree.
3. τ1 is the exponent of the power-law distribution of vertex degrees.
4. τ2 is the exponent of the power-law distribution of community sizes.
5. μ is the mixing parameter: each vertex shares a fraction μ of its edges with vertices in other
communities.
6. cmin is the minimum community size. cmax is the maximum community size.
Finally, some real-world networks have been tested, listed in Table 1.
Name
scientometrics
c. elegans
email
karate
terrorists
grassland species
Table 1. Real-world networks.
Vertices
Type
Ref.
2678
citation
[35]
453
metabolic
[36]
[37]
social
1133
34
social
[32]
62
social
[38]
[39]
food web
75
Edges
10368
2025
5451
78
152
113
4. Experiments
4.1. Edge prediction methods on incomplete networks
In this section we apply a few edge prediction methods to find the “missing” edges in an incomplete
network. We define score(u,v) to be the value of relationship between u and v. The higher the score,
the more likely they are to be neighbours.
v
( ) .
Common neighbours (CN). The number of common neighbours that two vertices have is a basic
idea that suggests a mutual relationship between them. For example, it may be more likely that two
people know each other if they have one or more acquaintances in common in a social network [40].
The function is defined as [8]:
score u v
u
( , )
( )
( )u
( )v
represent the set of neighbours of vertex u and v, respectively.
where
and
Jaccard, Meet/Min, and Geometric. These three coefficients have a similar definition, related to the
probability of triangles in all connected edges of any two vertices [14,16,17,18]. They are defined as
[14]:
u
score u v
u
v
( )
( )
( , )
( ) /
score u v
u
v
u
( ) ,
( ) / min(
( )
( , )
2
score u v
u
v
u
( , )
( )
( )
/ (
( )
v
( ) .
v
( ) ).
v
( ) ).
(2)
(3)
(4)
(1)
Adamic and Adar (AA). Adamic and Adar [13] proposed a similarity measure to define the
similarity between two vertices in terms of the neighbours of the common neighbours of the two
vertices. It is translated into [8]:
1
s
log ( )
score u v
( , )
s
u
v
( )
( )
Resource allocation index (RA). This is a variant of the method of Adamic and Adar, which
assumes that the common neighbours could transmit resources from one vertex to the other one, and is
defined as [15]:
(5)
.
1
s
( )
.
score u v
( , )
s
v
u
( )
( )
Preferential attachment (PA). This shows that the probability of a connection between two arbitrary
vertices is related to the number of neighbours of each vertex. The assumption of this method is that
vertices prefer to connect with other vertices with a large number of neighbours. It is described as
[40,41]:
score u v
( , )
v
( ) .
(6)
(7)
u
( )
Hierarchical structure method (HRG) [42]. This method combines a maximum-likelihood method
with a Markov chain Monte Carlo method to sample the hierarchical structure models with probability
proportional to their likelihood from the given network. This model is a binary tree with n leaves (the
vertices from the given network) and n-1 internal nodes, and a probability pr is associated with each
internal node r. The probability, pr, of the deepest common ancestor of two vertices represents the
probability of a connection between them.
Stochastic block model method (BM) [3]. This method uses the stochastic block model, in which
vertices are placed into communities and a matrix Q includes the probabilities of connections between
communities. Qαβ is the probability of a connection between two vertices that are in community α and
community β. Then, the probability of an edge can be calculated as explained in Ref. [3].
Because the HRG and BM methods have a higher computational complexity than other methods,
we can only evaluate these on some small networks in our experiments reported below.
To determine the accuracy of edge prediction methods, a common measure is the AUC: the area
under the ROC (receiver-operating characteristic) curve [42,43,44]. The interpretation of AUC is the
probability that the score of a randomly-chosen missing edge is higher than that of a randomly-chosen
pair of unconnected vertices.
Figure 2 shows a comparison of edge prediction methods on three incomplete versions of ER
networks. Since ER networks are random networks, which do not have any vertex-to-vertex similarity,
it is impossible to find missing edges by using edge prediction methods based on common neighbours.
In figure 2, these methods have no effect with any type of missing edges. In contrast, the PA method
relies on the number of neighbours of each vertex to find the relation between them. For example, PA
performs worse in crawled networks, because the vertices at the boundary of crawled networks always
have a low degree; in limited-degree networks, we limit vertices which have a high degree, so PA
performs better than other methods on limited-degree networks.
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(b)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(c)
(d)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(e)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(f)
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0.2
0.4
0.6
Fraction of edges observed
0.8
1
0
0
0.2
0.4
0.6
Fraction of edges observed
0.8
1
Figure 2. Performance of edge prediction methods on incomplete ER networks, with average degree
10. (a) Crawled networks, n=100. (b) Random-deletion networks, n=100. (c) Limited-degree
networks, n=100. (d) Crawled networks, n=1000. (e) Random-deletion networks, n=1000. (f)
Limited-degree networks, n=1000.
Figures 3-5 show comparisons of edge prediction methods on three incomplete versions of LFR
networks (named LFR1, LFR2, and LFR3, respectively), using these parameters:
1. n = 100, k = 5, kmax = 15, τ1 = 2, τ2 = 1, µ = 0.3, cmin = 10, cmax = 20.
2. n = 1000, k = 10, kmax = 25, τ1 = 2, τ2 = 1, µ = 0.3, cmin = 10, cmax = 20.
3. n = 1000, k = 10, kmax = 25, τ1 = 2, τ2 = 1, µ = 0.3, cmin = 20, cmax = 40.
The results of CN, AA, and RA are similar, because they are all based on common neighbours.
They perform best on LFR2 networks because these have a higher clustering coefficient than LFR1
and LFR3 (see Table 2). Jaccard, Meet/Min, and Geometric also consider common neighbours, but are
affected by other conditions. For example, Jaccard performs best in limited-degree networks because
the degrees of many vertices are similar, especially when there are more missing edges. Jaccard is not
appropriate when there is a missing edge between a high-degree vertex and a low-degree one.
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(b)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(c)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
Figure 3. Performance of edge prediction methods on incomplete LFR networks: n=100, k=5,
kmax=15, τ1=2, τ2=1, µ=0.3, cmin=10, cmax=20. (a) Crawled networks. (b) Random-deletion networks.
(c) Limited-degree networks.
Table 2. Clustering coefficient of LFR networks and real-world networks.
Clustering coefficient
Network
LFR1
0.1460
0.2915
LFR2
0.1394
LFR3
0.2557
karate
terrorists
0.3609
0.1741
grassland species
0.1663
email
0.0993
scientometrics
c. elegans
0.1244
PA performs well in limited-degree networks, and is not affected by the community size and the
clustering coefficient, because PA is appropriate to the mechanism of limited-degree networks. On the
contrary, vertices at the periphery usually have a low degree in crawled networks, and it is impossible
to know whether the endvertices of missing edges in the crawled and random-deletion networks have
high degree.
The HRG and BM methods are better than other edge prediction methods in random-deletion
networks, but they do not perform well in crawled networks. In limited-degree networks, these two
methods perform similarly to other edge prediction methods.
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(b)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(c)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(b)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(c)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
Figure 5. Performance of edge prediction
methods on
incomplete LFR networks:
n=1000, k=10, kmax=25, τ1=2, τ2=1, µ=0.3,
cmin=20, cmax=40. (a) Crawled networks. (b)
Random-deletion networks. (c) Limited-degree
networks.
Figure 4. Performance of edge prediction
methods on incomplete LFR networks: n=1000,
k=10, kmax=25, τ1=2, τ2=1, µ=0.3, cmin=10,
cmax=20.
(a)
Crawled
networks.
(b)
Random-deletion networks. (c) Limited-degree
networks.
Figure 6 shows comparisons of edge prediction methods on three incomplete versions of small
real-world networks. Basically, BM and HRG work better than other methods in random-deletion and
limited-degree networks. HRG performs well in all types of incomplete networks if the original
network is a hierarchical structure (see figures 6(g)-6(i)). PA performs well in limited-degree networks,
and sometimes it is even close to or better than BM. CN, RA and AA perform better than BM and
HRG in crawled networks if the original network has a high clustering coefficient (see “terrorists” in
table 2).
Figure 7 shows comparisons of edge prediction methods on three incomplete versions of larger
real-world networks; these are too large for BM and HRG to process in a reasonable time. The results
of RA, AA and CN are similar, and they are better than other prediction methods in crawled and
random-deletion networks. PA performs well on limited-degree networks.
In general, BM and HRG perform better than other methods in random-deletion and limited-degree
networks. However, they are too slow to apply to large networks. In addition to these two methods,
CN, RA and AA perform well, while PA usually works well in limited-degree networks. All methods
do better on limited-degree networks than on crawled and random-deletion networks.
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(f)
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
0.2
(g)
0.4
0.6
Fraction of edges observed
0.8
1
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(h)
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(i)
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
CN
AA
RA
Jaccard
MeetMin
Geom etric
PA
HRG
BM
1
0.8
0.6
0.4
0.2
C
U
A
0
0
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
HRG
BM
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(b)
CN
AA
RA
Jaccard
MeetMin
Geom etric
PA
HRG
BM
0.8
1
0.2
0.6
0.4
Fraction of edges observed
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(c)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(d)
(e)
CN
AA
RA
Jaccard
MeetMin
Geom etric
PA
HRG
BM
0.8
1
0.2
0.6
0.4
Fraction of edges observed
Figure 6. Performance of edge prediction methods on incomplete small real networks. (a) Crawled
karate network. (b) Random-deletion karate network. (c) Limited-degree karate network. (d) Crawled
terrorists network. (e) Random-deletion terrorists network. (f) Limited-degree terrorists network. (g)
Crawled grassland species network. (h) Random-deletion grassland species network.
(i)
Limited-degree grassland species network.
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
(a)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(b)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(c)
CN
AA
RA
Jaccard
MeetMin
Geom etric
PA
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(d)
1
0.8
0.6
0.4
0.2
C
U
A
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
0
0
1
C
U
A
0.8
0.6
0.4
0.2
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
CN
AA
RA
Jaccard
MeetMin
Geometric
PA
0.2
0.6
0.4
Fraction of edges observed
0.8
1
0
0
(e)
(f)
CN
AA
RA
Jaccard
MeetMin
Geom etric
PA
0.2
0.6
0.4
Fraction of edges observed
0.8
1
(g)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(h)
0.2
0.4
0.6
Fraction of edges observed
0.8
1
(i)
0.2
0.6
0.4
Fraction of edges observed
0.8
1
Figure 7. Performance of edge prediction methods on incomplete large real networks. (a) Crawled
email network. (b) Random-deletion email network. (c) Limited-degree email network. (d) Crawled
scientometrics network.
(e) Random-deletion scientometrics network.
(f) Limited-degree
scientometrics network. (g) Crawled c. elegans network. (h) Random-deletion c. elegans network. (i)
Limited-degree c. elegans network.
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
(a)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(d)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
0.6
0.7
Fraction of edges observed
0.8
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
(b)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(e)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
0.6
0.7
Fraction of edges observed
0.8
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
(c)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(f)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
0.6
0.7
Fraction of edges observed
0.8
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
Figure 8. Performance of community detection algorithms on various types of incomplete LFR
networks. (a) COPRA. (b) CliqueMod. (c) CNM. (d) Walktrap. (e) Wakita and Tsurumi. (f) Louvain
method.
4.2. The effect of missing data on community detection algorithms
We now investigate the effect of missing data on community detection algorithms: COPRA [45],
CliqueMod [46], CNM [47], Wakita and Tsurumi [48], Walktrap [49], and the Louvain method [50]. In
this part, we construct the incomplete networks on the induced subnetworks to keep them connected,
as explained in Section 3.3, and we compare all of them, including the induced subnetworks, with the
original networks. For artificial networks, we use the normalized mutual information (NMI) measure
[51] to compare the known partition with the partition found by each algorithm. For real-world
networks, since we do not know the real community structure, we use two community detection
algorithms (COPRA and the Louvain method), which do not require to be told the number of
communities. We run these two algorithms on the original network first, to obtain a partition, and
choose the one with the maximum modularity as the “real” partition. Then, we compare the partition
obtained by each algorithm for all incomplete networks with that “real” partition.
Figure 8 shows the results of different community detection algorithms on the induced subnetworks
and on three incomplete versions of LFR networks (parameters n=1000, k=20, kmax=50, τ1=2, τ2=1,
=0.1, cmin=50, cmax=100; the results are averaged over 100 random networks with the same
parameters). All of the results of the induced subnetworks are better than those of the incomplete
networks. We can also see that the results for crawled networks are usually worse than for other
incomplete networks. To understand the reason for this, we calculated the number of removed
intercommunity edges, intracommunity edges, and the remaining intercommunity edges in these
incomplete networks (see table 3). This shows that random-deletion and limited-degree networks
usually lose more intercommunity edges and fewer intracommunity edges than crawled networks. This
explains why results are worst for crawled networks. For the other incomplete networks, communities
can be found well even when there are many missing edges.
Finally, we examine three real networks: email, scientometrics and c. elegans. Figure 9 shows the
performance of COPRA and the Louvain method on induced networks and the three incomplete
networks, compared with the original networks. The results are broadly similar to those of artificial
networks: the results of random-deleteion and limited-degree networks are better than those of crawled
networks.
Table 3. Number of intracommunity and intercommunity edges removed in incomplete networks.
# of other
%
Type of
# of
# of
inter
removed
removed
observed
missing edge
edges
intra
inter
edges
edges
edges
85
866
99
852
868
83
158
1767
199
1726
199
1726
2730
298
319
2709
312
2716
504
3400
487
3417
3415
489
744
4111
788
4067
4067
788
Crawled
Random-deletion
Limited-degree
Crawled
Random-deletion
Limited-degree
Crawled
Random-deletion
Limited-degree
Crawled
Random-deletion
Limited-degree
Crawled
Random-deletion
Limited-degree
90%
90%
90%
80%
80%
80%
70%
70%
70%
60%
60%
60%
50%
50%
50%
915
901
917
842
801
801
702
681
688
496
513
511
256
212
212
5. Conclusions
Most edge prediction methods have not been evaluated on different types of networks and different
forms of missing data. We have found that the performance of these methods is strongly affected by
both factors. In particular, we have found that one algorithm, PA, performs well on limited-degree
incomplete networks, whereas it has obtained poor results on random-deletion networks in previous
research. BM and HRG perform better than other methods on random-deletion and limited-degree
networks, even though they are not suitable for large networks. CN, RA, and AA work well in crawled
and random-deletion networks. Therefore, to deal with the limited-degree problem in reality, we could
choose the PA method, or BM and HRG for small networks. CN, RA, and AA are still good methods
for finding other types of missing edges.
We have also investigated the effect of missing data on community detection algorithms. The
results show that community detection algorithms perform surprisingly well in the presence of missing
edges. The performance depends on the type of missing edges and is worst for crawled networks,
probably because of the number of intercommunity edges missing in this type of incomplete network.
Acknowledgements
We are grateful to Roger Guimerà for sending us the code for the BM method, and we would like to
thank the anonymous referees for their useful comments and valuable suggestions.
References
Sprinzak E, Sattath S and Margalit H 2003 How reliable are experimental protein-protein
[1]
interaction data? J. Mol. Biol 327 919
[2] Liljeros F, Edling C R, Amaral L A N, Stanley H E and Aaberg Y 2001 The web of human
sexual contacts Nature 411 907
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0
0.5
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
0.6
0.7
Fraction of edges observed
0.8
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
0.6
0.7
Fraction of edges observed
0.8
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(a)
(b)
(c)
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(d)
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(e)
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
Induced subnetwork
Craw led network
Random -deletion network
Lim ited-degree network
(f)
1
0.8
0.6
0.4
0.2
n
o
i
t
a
m
r
o
f
n
i
l
a
u
t
u
m
d
e
z
i
l
a
m
r
o
N
0.6
0.7
Fraction of edges observed
0.8
0.9
0
0.5
0.6
0.7
Fraction of edges observed
0.8
0.9
Figure 9. Performance of COPRA and the Louvain method on various types of incomplete real
networks. (a) Louvain method on incomplete email network. (b) Louvain method on incomplete
scientometrics network. (c) Louvain method on incomplete c. elegans network. (d) COPRA on
incomplete email network. (e) COPRA on incomplete scientometrics network. (f) COPRA on
incomplete c. elegans network.
[3] Guimerà R and Sales-Pardo M 2009 Missing and spurious interactions and the reconstruction of
complex networks Proc. Natl. Acad. Sci. USA 106 22073
[4] Allison P D 2001 Missing Data (Thousand Oaks, CA: Sage Publications)
[5] Kossinets G 2006 Effect of missing data in social networks Soc. Netw. 28 247
[6] Zhu Y-X, Lü L, Zhang Q-M and Zhou T 2011 Uncovering missing links with cold ends arXiv:
1104.0395
[7] Leicht E A, Holme P and Newman M E J 2006 Vertex similarity in networks Phys. Rev. E 73
026120
[8] Liben-Nowell D and Kleinberg J 2007 The link-prediction problem for social networks J. Am.
Soc. Inform. Sci. Technol. 58 1019
[9] Hasan M A, Chaoji V, Salem S and Zaki M 2006 Link prediction using supervised learning
Proc. Workshop on Link Analysis (Counterterrorism and Security with SIAM Data Mining Conf.,
Bethesda, MD)
[10] Popescul A and Ungar L H 2003 Statistical relational learning for link prediction Proc. IJCAI03
Workshop on Learning Statistical Models from Relational Data p 81
[11] Taskar B, Wong M-F, Abbeel P and Koller D 2003 Link prediction in relational data Proc.
Neural Information Processing Systems Conf. p 659
[12] Lü L and Zhou T 2011 Link prediction in complex networks: a survey Physica A 390 1150
[13] Adamic L A and Adar E 2003 Friends and neighbours on the web Soc. Netw. 25 211
[14] Goldberg D S and Roth F P 2003 Assessing experimentally derived interactions in a small world
Proc. Natl. Acad. Sci. USA 100 4372
[15] Zhou T, Lü L and Zhang Y 2009 Predicting missing links via local information Eur. Phys. J. B
71 623
[16] Ravasz E, Somera A L, Mongru D A, Oltvai Z N and Barabási A L 2002 Hierarchical
organization of modularity in metabolic networks Science 297 1551
[17] Jaccard P 1912 The distribution of flora in the alpine zone New Phytologist 11 37
[18] Ihmels J, Friedlander G, Bergmann S, Sarig O, Ziv Y and Barkai N 2002 Revealing modular
organization in the yeast transcriptional network Nature Genet. 31 370
[19] Katz L 1953 A new status index derived from sociometric analysis Psychometrika 18 39
[20] Gobel F and Jagers A A 1974 Stoch. Proc. Appl. 2 311
[21] Jeh G and Widom J 2002 SimRank: a measure of structural-context similarity Proc. ACM
SIGKDD Int. Conf. on Knowledge Discovery and Data Mining p 538
[22] Lü L, Jin C-H and Zhou T 2009 Similarity index based on local paths for link prediction of
complex networks Phys. Rev. E 80 046122
[23] Fouss F, Pirotte A, Renders J-M and Saerens M 2007 Random-walk computation of similarities
between nodes of a graph with application to collaborative recommendation IEEE Trans. Knowl.
Data Eng. 19 355
[24] Borgatti S P and Everett M G 1993 Two algorithms for computing regular equivalence Soc.
Netw. 15 361
[25] Sun D, Zhou T, Liu Z-G, Liu R-R, Jia C-X and Wang B-H 2009 Information filtering based on
transferring similarity Phys. Rev. E 80 017101
[26] Marsden P V 1990 Network data and measurement Annu. Rev. Sociol. 16 435
[27] Costenbader E and Valente T W 2003 The stability of centrality measures when networks are
sampled Soc. Netw. 25 283
[28] Borgatti S P, Carley K M and Krackhardt D 2006 On the robustness centrality measure under
conditions of imperfect data Soc. Netw. 28 124
[29] Gile K and Handcock M S 2006 Model-based assessment of the impact of missing data on
inference for networks Working Paper No 66 (Center for Statistics and the Social Sciences,
University of Washington)
[30] Albert R, Jeong H and Barabási A L 2000 Error and attack tolerance of complex networks
Nature 406 378
[31] Holland P W and Leinhard S 1973 Structural implications of measurement error in sociometry J.
Math. Sociol. 3 85
[32] Zachary W W 1977 An information flow model for conflict and fission in small groups J.
Anthropol. Res. 33 452
[33] Erdös P and Rényi A 1959 On random graphs I. Publ. Math. 6 290
[34] Lancichinetti A, Fortunato S and Radicchi F 2008 Benchmark graphs for testing community
detection algorithms Phys. Rev. E 78 046110
[35] Pajek, Network from Pajek datasets http://vlado.fmf.uni-lj.si/pub/networks/data/
[36] Jeong H, Tombor B, Albert R, Oltvai Z and Barabási A L 2000 The large-scale organization of
metabolic networks Nature 407 651
[37] Guimera R, Danon L, Diaz-Guilera A, Giralt F and Arenas A 2003 Self-similar community
structure in a network of human interactions Phys. Rev. E 68 065103
[38] Krebs V 2002 Mapping networks of terrorist cells Connections 24 43
[39] Dawah H A, Hawkins B A and Claridge M F 1995 Structure of the parasitoid communities of
grass-feeding chalcid wasps J. Anim. Ecol. 64 708
[40] Newman M E J 2001 Clustering and preferential attachment in growing networks Phys. Rev. E
64 025102
[41] Barabasi A L, Jeong H, Neda Z, Ravase E, Schubert A and Vicsek T 2002 Evolution of the
social network of scientific collaboration Physica A 311 590
[42] Clauset A, Moore C and Newman M E J 2008 Hierarchical structure and the prediction of
missing links in networks Nature 453 98
[43] Fawcett T 2006 An introducation to ROC analysis Pattern Recognit. Lett. 27 861
[44] Hanley J A and McNeil B J 1982 The meaning and use of the area under a receiver operating
characteristic (ROC) curve Radiology 143 29
[45] Gregory S 2010 Finding overlapping communities in networks by label propagation New J.
Phys. 12 103018
[46] Yan B and Gregory S 2009 Detecting communities in networks by merging cliques, Proc. IEEE
Int. Conf. on Intelligent Computing and Intelligent Systems p 832
[47] Clauset A, Newman M E J and Moore C 2004 Finding community structure in very large
networks Phys. Rev. E 70 066111
[48] Wakita K and Tsurumi T 2007 Finding community structure in a megascale social networking
service, Proc. IADIS Intl Conf. on WWW/Internet 2007 p 153
[49] Pons P and Latapy M 2006 Computing communities in large networks using random walks J.
Graph Algorithms Appl. 10 191
[50] Blondel V D, Guillaume J L, Lambiotte R and Lefebvre E 2008 Fast unfolding of community
hierarchies in large networks J. Stat. Mech. P10008
[51] Lancichinetti A, Fortunato S and Kertész J 2009 Detecting the overlapping and hierarchical
community structure of complex networks New J. Phys. 11 033015
|
1408.1542 | 4 | 1408 | 2015-05-25T02:28:39 | Measuring and Optimizing Cultural Markets | [
"cs.SI",
"physics.soc-ph"
] | Social influence has been shown to create significant unpredictability in cultural markets, providing one potential explanation why experts routinely fail at predicting commercial success of cultural products. To counteract the difficulty of making accurate predictions, "measure and react" strategies have been advocated but finding a concrete strategy that scales for very large markets has remained elusive so far. Here we propose a "measure and optimize" strategy based on an optimization policy that uses product quality, appeal, and social influence to maximize expected profits in the market at each decision point. Our computational experiments show that our policy leverages social influence to produce significant performance benefits for the market, while our theoretical analysis proves that our policy outperforms in expectation any policy not displaying social information. Our results contrast with earlier work which focused on showing the unpredictability and inequalities created by social influence. Not only do we show for the first time that dynamically showing consumers positive social information under our policy increases the expected performance of the seller in cultural markets. We also show that, in reasonable settings, our policy does not introduce significant unpredictability and identifies "blockbusters". Overall, these results shed new light on the nature of social influence and how it can be leveraged for the benefits of the market. | cs.SI | cs | Measuring and Optimizing Cultural Markets
A. Abeliuk1
G. Berbeglia2
M. Cebrian1
P. Van Hentenryck3
June 14, 2021
1 The University of Melbourne and National ICT Australia.
2 Melbourne Business School (The University of Melbourne) and National ICT Australia.
3 The Australian National University and National ICT Australia.
Corresponding author: Pascal Van Hentenryck ([email protected])
Abstract
Social influence has been shown to create significant unpredictability in cultural
markets, providing one potential explanation why experts routinely fail at predicting
commercial success of cultural products. To counteract the difficulty of making ac-
curate predictions, "measure and react" strategies have been advocated but finding a
concrete strategy that scales for very large markets has remained elusive so far.
Here we propose a "measure and optimize" strategy based on an optimization policy
that uses product quality, appeal, and social influence to maximize expected profits in
the market at each decision point. Our computational experiments show that our policy
leverages social influence to produce significant performance benefits for the market,
while our theoretical analysis proves that our policy outperforms in expectation any
policy not displaying social information.
Our results contrast with earlier work which focused on showing the unpredictabil-
ity and inequalities created by social influence. Not only do we show for the first time
that dynamically showing consumers positive social information under our policy in-
creases the expected performance of the seller in cultural markets. We also show that,
in reasonable settings, our policy does not introduce significant unpredictability and
identifies "blockbusters".
Overall, these results shed new light on the nature of social influence and how it
can be leveraged for the benefits of the market.
5
1
0
2
y
a
M
5
2
]
I
S
.
s
c
[
4
v
2
4
5
1
.
8
0
4
1
:
v
i
X
r
a
1
1
Introduction
Prediction in cultural markets is a wicked problem: On the one hand, experts routinely fail
at predicting commercial success for items such as movies, books, and songs [1, 2, 3, 7, 14]
while, on the other hand, hit products are generally considered qualitatively superior to
the less successful ones [24]. If those hit products are of superior quality, why can't experts
detect them a priori? As Duncan Watts puts it, in cultural markets "everything is obvious
... once you know the answer," as experts find it extremely hard to benefit from their
specific knowledge [4].
In their seminal paper [17], Salganik, Dodds, and Watts offer a satisfactory explanation
for this paradox: The impact of social influence on the behavior of individuals (word-of-
mouth), a process a priori invisible to experts, may distort the quality perceived by the
customers, making quality and popularity out of sync and expert predictions less reliable.
To investigate this hypothesis experimentally, they created an artificial music market called
the MusicLab. Participants in the MusicLab were presented a list of unknown songs from
unknown bands, each song being described by its name and band. The participants were
divided into two groups exposed to two different experimental conditions: The independent
condition and the social influence condition. In the first group (independent condition),
participants were provided with no additional information about the songs. Each partic-
ipant would decide which song to listen to from a random list. After listening to a song,
the participant had the opportunity to download it. In the second group (social influence
condition), each participant was provided with an additional information: The number of
times the song was downloaded by earlier participants. Moreover, these participants were
presented with a list ordered by the number of downloads. Additionally, to investigate
the impact of social influence, participants in the second group were distributed in eight
"worlds" evolving completely independently. In particular, participants in one world had
no visibility about the downloads and the rankings in the other worlds.
The MusicLab provides an experimental testbed of measuring the unpredictability
of cultural markets. By observing the evolution of different words given the same initial
conditions, the MusicLab provides unique insights on the impact of social influence and
the resulting unpredictability. In particular, Salganik et al suggested that social influence
contributes to both unpredictability and inequality of success, with follow-up experiments
confirming these initial findings [19, 18, 13, 23] (see also [6] for a perspective on the Mu-
sicLab paper). These experimental results received further theoretical support by the
development of a generative model that reproduces the experimental data [10]. The Musi-
cLab model, which is presented below, teased out the role of the quality and appeal of the
songs and position bias in the unpredictability and inequality of the experimental market.
Together these results painted a bleak scenario for cultural markets. From the stand-
point of the market, e.g., producers and consumers of cultural products, this is an inefficient
situation [21, 22, 8]: Fewer cultural products are consumed overall and some high-quality
products may go unnoticed, while lower-quality ones are oversold. It is thus an important
2
open question to determine whether this situation can be remedied and how. In [24], Watts
advocates the use of "measure and react" strategies to address the difficulty in making cor-
rect predictions: "Rather than predicting how people will behave and attempting to design
ways to make customers behave in a particular way [...] we can instead measure directly how
they respond to a whole range of possibilities and react accordingly" [24]. The MusicLab
provides a unique setting to validate the benefits of a "measure and react" strategy.
This paper presents a "measure and optimize" (M&O) algorithm, that is a "measure
and react" algorithm which maximizes expected downloads at each decision point. In the
MusicLab, the only degree of freedom is the playlist presented to an incoming participant.
Each ranking policy for the playlist produces a "measure and react" algorithm. The original
MusicLab experiments presented one such policy: To rank the songs by number of down-
loads. Our "measure and optimize" algorithm in contrast uses a performance ranking that
maximizes the expected downloads globally for the incoming participant, exploiting the
quality and appeal of each song, the position visibilities in the playlist, and social influence.
The performance ranking is based on two insights.
1. It uses the generative MusicLab model that describes the probability of sampling a
song given its quality, appeal, position in the playlist, and the downloads of all songs
at some point in time.
2. It solves the resulting global optimization problem in strongly polynomial time by a
reduction to the Linear Fractional Assignment Problem, meaning that the algorithm
scales to large instances and avoid exploring the n! possible rankings.
In addition, since the quality of the songs is not known a priori, our M&O algorithm learns
them over time, using an often overlooked insight by Salganik et al.
[17]: The fact that
the popularity of a song in the independent condition is a natural measure of its quality,
capturing both its intrinsic "value" and the preferences of the participants.
The M&O algorithm was evaluated on a simulated version of the MusicLab based on
the model proposed in [10]; It was compared to the download ranking policy used in the
original MusicLab experiments [17] and a random ranking policy. These ranking policies
were evaluated both under the social influence and independent conditions.
Regarding efficiency, the computational results contain two important highlights:
1. The M&O algorithm produces a significant increase in expected downloads compared
to other tested policies.
2. The M&O algorithm leverages both position bias and social influence to maximize
downloads, providing substantial gains compared to ranking policies under the inde-
pendent condition.
These performance results are robust, not an artifact of distributions and parameters used
in experimental setting. Indeed, our theoretical results show that a performance-ranking
3
policy always benefits from social influence and outperforms all policies not using social
influence in expectation. The performance results also shed light on recent results that
demonstrated the benefits of position bias in the absence of social influence [11]. Our
computational results show that a performance-ranking policy can leverage both social
influence and position visibility and that the improvements in expected downloads are
cumulative.
Regarding unpredictability, our results are also illuminating. They highlight that the
performance-ranking policy often identifies "blockbusters", which are the primary sources
of the increased market efficiency. This provides a sharp contrast to the download rank-
ing, which cannot identify these blockbusters reliably. Moreover, even in that worst-case
scenario, our computational results show that the performance ranking brings significant
benefits over the download ranking.
Finally, it is important to emphasize that the performance-ranking policy is general-
purpose and can be applied to a variety of cultural markets, including books and movies,
where product appeal, position visibility, and social influence can be optimized jointly.
2 Results
We first describe the MusicLab model proposed in [10], the experimental setting, and a
presentation of the various ranking policies. We then show how to estimate the qualities of
the songs. We then report the core computational results about the performance and
unpredictability of the various ranking policies. We finally prove that social influence
always helps the performance ranking in maximizing the expected downloads. As a result,
the performance ranking under social influence outperforms all ranking policies under the
independent condition.
2.1 The MusicLab Model
The descriptive model for the MusicLab [10] is based on the data collected during the
actual experiments and is accurate enough to reproduce the conclusions in [17] through
simulation. The model is defined in terms of a market composed of n songs. Each song
i ∈ {1, . . . , n} is characterized by two values:
1. Its appeal Ai which represents the inherent preference of listening to song i based only
on its name and its band;
2. Its quality qi which represents the conditional probability of downloading song i given
that it was sampled.
The MusicLab experiments present each participant with a playlist π, i.e., a permutation
of {1, . . . , n}. Each position p in the playlist is characterized by its visibility vp which is
the inherent probability of sampling a song in position p. Since the playlist π is a bijection
4
from the positions to the songs, its inverse is well-defined and is called a ranking. Rankings
are denoted by the letter σ in the following, πi denotes the song in position i of the playlist
π, and σi denotes the position of song i in the ranking σ. Hence vσi denotes the visibility
of the position of song i. The model specifies the probability of listening to song i at time
k given a playlist σ as
(cid:80)n
pi,k(σ) =
vσi(αAi + Di,k)
j=1 vσj (αAj + Dj,k)
,
where Di,k is the number of downloads of song i at time k and α > 0 is a scaling factor
which is the same for all songs. Observe that the probability of sampling a song depends
on its position in the playlist, its appeal, and the number of downloads at time k. As an
experiment proceeds, the number of downloads dominates the appeal of the song at a rate
dictated by α.
When simulating the independent condition, the download counts Di,k in the probability
pi,k(σ) becomes zero and hence the probability of listening to song i only depends on the
appeals and visibilities of the songs. We obtain a probability
(cid:80)n
vσiAi
j=1 vσj Aj
pi(σ) =
2.2 The Experimental Setting
The experimental setting in this paper uses an agent-based simulation to emulate the Mu-
sicLab. Each simulation consists of N iterations and, at each iteration k,
1. the simulator randomly selects a song i according to the probability distribution
specified by the pj.k's.
2. the simulator randomly determines, with probability qi, whether selected song i is
downloaded; If the song is downloaded, then the simulator increases the download
counts of song i, i.e., Di,k+1 = Di,k + 1.
The "measure and react" algorithms studied in this paper monitors the above process (e.g.,
to estimate song quality). Every r iterations, the algorithms compute a new playlist σ using
one of the ranking policies described below. For instance, in the social influence condition
of the original MusicLab experiments, the policy ranks the songs in decreasing order of
download counts. The parameter r ≥ 1 is called the refresh rate.
The experimental setting tries to be close to the MusicLab experiments and considers
50 songs and simulations with 20,000 steps. The songs are displayed in a single column.
The analysis in [10] indicated that participants are more likely to sample songs higher in
the playlist. More precisely, the visibility decreases with the playlist position, except for
a slight increase at the bottom positions. Figure 2.1 depicts the visibility profile based on
these guidelines, which is used in all computational experiments in this paper.
5
Figure 2.1: The visibility vp (y-axis) of position p in the playlist (x-axis) where p = 1 is the
top position and p = 50 is the bottom position of the playlist which is displayed in a single
column.
Figure 2.2: The Quality qi (blue) and Appeal Ai (red) of Song i in the two settings. In the
first setting, the quality and the appeal for the songs were chosen independently according
to a Gaussian distribution normalized to fit between 0 and 1. The second setting explores
an extreme case where the quality is negatively correlated with the appeal.
The paper also uses two settings for the quality and appeal of each song, which are
depicted in Figure 2.2:
1. In the first setting, the quality and the appeal for the songs were chosen independently
according to a Gaussian distribution normalized to fit between 0 and 1.
2. The second setting explores an extreme case where the quality is negatively correlated
with the appeal.
The experimental results were obtained by running W = 400 simulations.
2.3 The Ranking Policies
This section reviews a number of ranking policies used to influence the participants. Each
policy updates the order of the songs in the playlist and thus defines a different "measure
and react" algorithm that uses the policy to present a novel playlist after each refresh.
6
The first policy is called the download ranking [17]: At iteration k, it simply orders
the songs by the number of downloads Di,k and assigns the positions 1..n in that order to
obtain the σd
k. Note that downloads do not necessarily reflect the inherent quality of the
songs, since they depend on how many times the songs were listened to.
The second policy is the performance ranking which maximizes the expected number
of downloads at each iteration, exploiting all the available information globally, i.e., the
appeal, the visibility, the downloads, and the quality of the songs.
Definition 2.1. The performance ranking σ∗
k at iteration k is defined as
n(cid:88)
i=1
σ∗
k = arg max
σ∈Sn
pi,k(σ) · qi.
The performance ranking uses the probability pi,k(σ) of sampling song i at iteration k given
ranking σ, as well as the quality qi of song i. Obviously, in practical situations, the song
qualities are unknown. However, as mentioned in the introduction, Salganik et al.
[17]
argue that the popularity of a song (at least in the independent condition) can be used as
a measure of its quality. Hence, the quality qi in the above expression can be replaced by
its approximation qi,k at iteration k as discussed later in the paper.
It is not obvious a priori how to compute the performance ranking: In the worst case,
there are n! rankings to explore. Fortunately, the performance ranking can be computed
efficiently (in strongly polynomial time) through a reduction to the Linear Fractional As-
signment Problem.
Theorem 2.2. The performance ranking can be computed in strongly polynomial time.
Proof. We show that the performance ranking problem can be polynomially reduced to the
Linear Fractional Assignment Problem (LFAP). The LFAP, which we state below, is known
to be solvable in polynomial time in the strong sense [20, 9].
A LFAP input is a bipartite graph G = (V1, V2, E) where V1 = V2 = n, and a cost
cij and a weight dij > 0 for each edge (i, j) ∈ E. The LFAP amounts to finding a perfect
matching M minimizing
(cid:80)
(cid:80)
(i,j)∈M cij
(i,j)∈M dij
.
−(cid:80)
(cid:80)
To find the performance ranking, it suffices to solve the LFAP problem defined over a
bipartite graph where V1 represents the songs, V2 represents the positions, ci,j = −vj(αAi +
Di)qi and dij = vj(αAi + Di). This LFAP instance determines a perfect matching M
such that
is minimized. The result follows since this is a polynomial
reduction.
In the following, we use D-rank and P-rank to denote the "measure and react" algo-
rithms using the download and performance rankings respectively. We also annotate the
(i,j)∈M vj (αAi+Di)qi
(i,j)∈M vj (αAi+Di)
7
algorithms with either SI or IN to denote whether they are used under the social influ-
ence or the independent condition. For instance, P-rank(SI) denotes the algorithm with
the performance ranking under the social influence condition, while P-rank(IN) denotes
the algorithm with the performance ranking under the independent condition. We also
use rand-rank to denote the algorithm that simply presents a random playlist at each
iteration. Our proposed M&O algorithm is P-rank(SI).
Under the independent condition, the optimization problem is the same at each iter-
ation as the probability pi(σ) does not change over time. Since the performance ranking
maximizes the expected downloads at each iteration, it dominates all other "measure and
react" algorithms under the independent condition.
Proposition 2.3 (Optimality of Performance Ranking). Given specified songs appeal and
quality, P-rank(IN) is the optimal "measure and react" algorithm under the independent
condition.
2.4 Recovering the Songs Quality
This section describes how to recover songs quality in the MusicLab. It shows that "mea-
sure and react" algorithms quickly recover the real quality of the songs.
As mentioned several times already, Salganik et al.
[17] stated that the popularity of
a song in the independent condition is a natural measure of its quality and captures both
its intrinsic "value" and the preferences of the participants. To approximate the quality of
a song, it suffices to sample the participants. This can be simulated by using a Bernoulli
sampling based on real quality of the songs. The predicted quality qi of song i is obtained
by running m independent Bernoulli trials with probability qi of success, i.e.,
qi =
k
m
,
where k is the number of successes over the m trials. The law of large numbers states that
the prediction accuracy increases with the sampling size. For a large enough sampling size,
qi has a mean of qi and a variance of qi(1−qi). This variance has the desirable property that
the quality of a song with a more 'extreme' quality (i.e., a good or a bad song) is recovered
faster than those with average quality. In addition, "measure and react" algorithms merge
additional information about downloads into the prediction as the experiment proceeds: At
step k, the approximate quality of song i is given by
qi,k =
qi,0 · m + Di,k
m + Si,k
,
where m is the initial sample size and Si,k the number of samplings of song i up to step k.
Figure 2.3 presents experimental results about the accuracy of the quality approximation
for different rankings, assuming an initial sampling set of size 10. More precisely, the figure
8
for the top 10 quality songs. The figure reports the average squared difference(cid:80)n
Figure 2.3: Average Squared Difference of Inferred Quality over Time for Different Rankings
(qi,k−qi)2
between the song quality and their predictions for the various ranking policies under the
social influence and the independent conditions.
i=1
n
reports the average squared difference
n(cid:88)
(qi,k − qi)2
i=1
n
between the song quality and their predictions for the various ranking policies under the
social influence and the independent conditions. In all cases, the results indicate that the
songs quality is recovered quickly and accurately. Observe that the performance ranking
recovers the song quality faster than the other rankings, both in the social influence and in
the independent conditions.
2.5 Expected Performance
This section studies the expected performance of the various rankings under the social
influence and the independent conditions. The performance ranking uses the quality ap-
proximation presented in the last section (instead of the "true" quality). Figures 2.4 and
2.5 depict the average number of downloads for the various rankings given the quality and
appeal data depicted in Figure 2.2.
The experimental results highlight three significant findings:
1. The performance ranking provides substantial gains in expected downloads compared
to the download ranking and the random ranking policies, both in the social influence
and the independent conditions.
9
Figure 2.4: The Number of Downloads over Time for the Various Rankings (Independent
Gaussian Setting). The x-axis represents the number of song samplings and the y-axis
represents the average number of downloads over all experiments. The left figure depicts the
results for the social influence condition and the right figure for the independent condition.
2. The benefits of the performance ranking are particularly striking under the social
influence condition. Social influence creates a Matthew effect [16] for the number of
expected downloads.
3. The gain of the performance ranking is also more significant on the worst-case sce-
nario, where the song appeal is negatively correlated with the song quality. The
expected number of downloads is lower in this worst-case scenario, but the shapes
of the curves are in fact remarkably similar in the social influence and independent
conditions.
In summary, the performance ranking produces significant improvements in expected down-
loads over the download ranking, which also dominate the random ranking. The benefits
come both from position bias and social influence which provide cumulative improvements.
2.6 Market Unpredictability
Figure 2.6 depicts the unpredictability results using the measure proposed in [17]. The
unpredictability ui of song i is defined as the average difference in market share for that
song between all pairs of realizations, i.e.,
ui =
mi,w − mi,w(cid:48)/
(cid:18)W
(cid:19)
,
2
W(cid:88)
W(cid:88)
w=1
w(cid:48)=w+1
10
Figure 2.5: The Number of Downloads over Time for the Various Rankings (Negative
Correlation Setting). The x-axis represents the number of song samplings and the y-axis
represents the average number of downloads over all experiments. The left figure depicts the
results for the social influence condition and the right figure for the independent condition.
where mi,w is the market share of song i in world w, i.e.,
n(cid:88)
k=1
mi,j = Dj
i,N /
Dj
i,N
where Dw
dictability measure is the average of this measure over all n songs, i.e.,
i,t is the number of downloads of song i at step t in world w. The overall unpre-
n(cid:88)
U =
ui/n.
The figure highlights two interesting results.
j=1
1. Exploiting social influence does not necessarily create an unpredictable market. In
the first setting, the performance ranking introduces only negligible unpredictability
in the market. This obviously contrasts with the download ranking, which introduces
significant unpredictability.
2. An unpredictable market is not necessarily an inefficient market. In the second setting,
the performance ranking exploits social influence to provide significant improvements
in downloads although, as a side-effect, it creates more unpredictability. Subsequent
results will explain the source of this unpredictability.
11
(cid:80)W
Figure 2.6: The Unpredictability of the Rankings for the Unpredictability Measure of [17].
The measure of unpredictability ui for song i is defined as the average difference in mar-
k=j+1 mi,j −
ket share for that song between all pairs of realizations, i.e., ui = (cid:80)W
mi,k/(cid:0)W
measure is the average of this measure over all n songs, i.e., U = (cid:80)n
(cid:1), where mi,j is the market share of song i in world j. The overall unpredictability
j=1
2
j=1 ui/n.. The left
figure depicts the results for the first setting, while the right figure depicts the results for
the second setting.
2.7 Downloads Versus Quality
Figures 2.7 and 2.8 report the correlation between number of downloads of a song (y-axis)
and its true quality (x-axis) for each of the rankings. More precisely, the figures show the
distributions of the downloads of every song over the simulations. The songs are ranked
in increasing quality on the x-axis and there are 400 dots for each song, representing the
number of downloads in each of the experiments. As a result, the figures give a clear visual
representation of what happens in the MusicLab for the various rankings, including how
the downloads are distributed and the sources of unpredictability.
The results are particularly interesting. In the first setting depicted in Figure 2.7, P-
rank(SI) clearly isolates the best song, which has an order of magnitude more downloads
than other songs. When comparing with P-rank(SI) and P-rank(IN), we observe that the
additional market efficiency produced by P-rank(SI) primarily comes from the downloads
of this best song. In contrast, the download ranking has a lot more unpredictability and
has significant download counts for many other songs, even for some with considerably less
quality. The variance of downloads for the best song is substantial, indicating that there
are simulations in which this best song has very few downloads. This never happens with
the performance ranking.
The results in Figure 2.8 highlight the source of unpredictability in the performance
ranking. These results are for the second setting where the appeal is negatively correlated
with the quality. The results indicate that the performance ranking is not capable of isolat-
ing the best song in this setting. There are significant downloads of the next high-quality
songs however. This also contrasts with the download ranking where lower-quality songs
12
Figure 2.7: The Distribution of Downloads Versus the True Song Qualities (First Setting).
The songs on the x-axis are ranked by increasing quality from left to right. Each dot is the
number of downloads of a song in an experiment.
can have substantial downloads. It is interesting to point out that the unpredictability of
the performance ranking in this setting comes from songs which are not easy to distinguish:
These songs have very similar values for quality and appeal.
These results shed interesting light on social influence. First, the download ranking
clearly shows that social influence can transform average songs into "hits". However, this
behavior is entirely eliminated by the performance ranking in the first setting: The per-
formance ranking clearly leverages social influence to isolate the "blockbuster". The second
setting highlights a case where even the performance ranking cannot recover the best song.
This setting is of course a worst-case scenario: A terrific product with a poor appeal. How-
ever, the performance ranking still leverages social influence to promote high-quality songs:
Social influence under the performance ranking is much less likely to turn an average song
into a "hit" compared to the download ranking.
13
Figure 2.8: The Distribution of Downloads Versus the True Song Qualities (Second Setting).
The songs on the x-axis are ranked by increasing quality from left to right. Each dot is the
number of downloads of a song in an experiment.
2.8 The Benefits of Social Influence
We now prove that social influence always increases the number of expected downloads of
the performance ranking. We show that this is true regardless of the number of songs, their
appeal and quality, and the visibility values. The proof assumes that the quality of the
songs has been recovered, i.e., we use qi instead of qi,k. Also, for simplifying the notations,
we use ai = αAi + Di,t and assume without loss of generality that v1 ≥ ··· ≥ vn > 0.
We are interested in showing that the expected number of downloads increases over time
for the performance ranking under the social influence condition. In state t, the probability
that song i is downloaded for ranking σ is
vσiaiqi(cid:80) vσiai
.
14
Hence, the expected number of downloads at time t for ranking σ is defined as
(cid:80)
(cid:80)
E[Dσ
t ] =
i vσiaiqi
i vσiai
.
Under the social influence condition, the number of expected downloads over time can be
considered as a Markov chain where state t + 1 only depends on state t. Therefore the
expected number of downloads at time t + 1 conditional to time t if ranking σ and σ(cid:48) are
used at time t and t + 1 respectively is
(cid:32)
(cid:88)
j
vσj ajqj(cid:80) vσiai
·
E[Dσ,σ(cid:48)
t+1 ] =
(cid:80)
(cid:80)
(cid:80)
(cid:80)
i(cid:54)=j vσ(cid:48)
i
aiqi + vσ(cid:48)
ai + vσ(cid:48)
j
j
(aj + 1)qj
(aj + 1)
i(cid:54)=j vσ(cid:48)
i
+ (1 − E[Dσ
t ]) · E[Dσ
t ].
Consider the performance ranking now. At step t, the performance ranking finds a permu-
tation σ∗ ∈ Sn such that
E[Dσ∗
t
] =
i
i vσ∗
i vσ∗
aiqi
ai
i
= max
σ
i vσiaiqi
i vσiai
.
(2.1)
By optimality of the performance ranking at each step, we have that
(2.2)
where σ∗∗ is the performance ranking at step t + 1. Hence, to show that social influence is
always beneficial to the performance ranking, it suffices to show that
t+1
E[Dσ∗,σ∗∗
] ≥ E[Dσ∗,σ∗
t+1 ]
E[Dσ∗,σ∗
t+1 ] ≥ E[Dσ∗
t
].
The proof is simpler when reasoning about the playlist obtained by the performance ranking.
At step t, the performance ranking finds a playlist π∗ ∈ Sn such that
(cid:33)
(cid:80)
(cid:80)
By (2.3), π∗ satisfies(cid:88)
p
p
Consider the function
f (λ) = max
π
(cid:80)
(cid:80)
(cid:88)
p vpaπ∗
p qπ∗
p
p vpaπ∗
p
= E[Dσ∗
t
]
(cid:88)
(cid:16)
p
(cid:16)
(cid:17)
vpaπ∗
p
p − λ
qπ∗
.
(cid:88)
p
vpaπ∗
p qπ∗
p − E[Dσ∗
t
]
vpaπ∗
p =
vpaπ∗
p
p − E[Dσ∗
qπ∗
t
(2.3)
(cid:17)
]
= 0.
(2.4)
Function f is concave and strictly decreasing and hence its unique zero is the optimal value
E[Dσ∗
]. Furthermore, the optimality condition for π in function f can be characterized by
a " rearrangement inequality" which states that the maximum is reached if and only if
t
(cid:0)qπ∗
1
− λ(cid:1) ≥ ··· ≥ aπ∗
n
(cid:0)qπ∗
n − λ(cid:1)
aπ∗
1
∀λ ∈ R.
(2.5)
We are now in position to state our key theoretical result.
15
Theorem 1. The expected downloads are nondecreasing over time when the performance
ranking is used to select the playlist.
Proof. By Equation 2.2, it suffices to show that
E[Dσ∗,σ∗
t+1 ] ≥ E[Dσ∗
t
].
Let a = a1,··· , an represent the current state at time t. For ease of notation, we rename
the songs so that σ∗
i = i. We also drop the ranking superscript σ∗ from the download
notation. Denote the optimal expected number of downloads at time t as
Then, the expected number of downloads in time t + 1 conditional to time t is
(cid:80)
(cid:80)
E[Dt] =
i viaiqi
i viai
= λ∗.
i(cid:54)=j viaiqi + vj(aj + 1)qj
i(cid:54)=j viai + vj(aj + 1)
1 −
+
(cid:33)
(cid:19)
(cid:18)
1 −
+
(cid:19)
·
(cid:80)
(cid:80)
i viaiqi
i viai
(cid:18)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
i viaiqi
i viai
(cid:19)
· λ∗.
j vjajqj
i viai
i viaiqi + vjqj
i viai + vj
E[Dt+1] ≥ E[Dt]
(cid:19)
(cid:18)
1 −
+
(cid:19)
(cid:80)
(cid:80)
j vjajqj
i viai
· λ∗ ≥ λ∗.
i viaiqi + vjqj
i viai + vj
(2.6)
(2.7)
E[Dt+1] =
Proving
(cid:88)
j
=
·
(cid:32)
(cid:80)
(cid:80)
vjajqj(cid:80) viai
(cid:80)
(cid:18) vjajqj(cid:80) viai
(cid:88)
(cid:80)
·
j
(cid:88)
j
·
(cid:80)
(cid:80)
(cid:18) vjajqj(cid:80) viai
1(cid:80)
(cid:88)
(cid:0)qπ∗
i viai
aπ∗
j
1
1
(cid:88)
(cid:34)
j
amounts to showing that
which reduces to proving
or, equivalently,
(cid:34)
(cid:80)
v2
j ajqj
i viai + vj
(cid:35)
(qj − λ∗)
≥ 0
(cid:35)
≥ 0.
(qj − λ∗)
v2
j ajqj
i viai + vj
(cid:80)
− λ∗(cid:1) ≥ ··· ≥ aπ∗
n
(cid:0)qπ∗
n − λ∗(cid:1) .
16
By the rearrangement inequality (2.5), the performance ranking at step t produces a playlist
π satisfying
Therefore, if ai (qi − λ∗) ≥ aj (qj − λ∗), then vi ≥ vj for any i, j ∈ N. Hence, the perfor-
mance rank allocates all songs with a negative term after the songs with a positive term.
Define P1 = {i ∈ N (qi − λ∗) ≥ 0} and P2 = {i ∈ N (qi − λ∗) < 0}. It follows that vi ≥ vj
for all i ∈ P1, j ∈ P2. Inequality (2.7) can be rewritten as follows:
(cid:20) vjajqjvj
(cid:80)
i viai + vj
(cid:88)
j∈P1
(cid:21)
(cid:20) vjajqjvj
(cid:80)
i viai + vj
(cid:88)
j∈P2
(qj − λ∗)
+
(cid:21)
(qj − λ∗)
≥ 0.
By definition of P1 and P2, the terms in the left summation are positive and the terms in
the right summation are negative. Additionally, for any c > 0,
vi
c + vi
≥ vj
c + vj
⇔ (c + vj)vi ≥ (c + vi)vj ⇔ cvi ≥ cvj ⇔ vi ≥ vj
Hence, since v1 ≥ v2 ≥ . . . ≥ vn,
v1(cid:80)
i viai + v1
v2(cid:80)
i viai + v2
≥
≥ . . . ≥
vn(cid:80)
i viai + vn
.
Let v = mini∈P1 vi the smallest visibility assigned to an item in P1. By the rearrangement
inequality (2.5), v ≥ maxi∈P2 vi. Similarly, let q = mini∈P1 qi. By definiton of P1 and P2,
q ≥ maxi∈P2 qi. These observations leads to the following inequality:
(cid:20) vjajqjvj
(cid:80)
(cid:88)
(cid:88)
vq(cid:80)
j∈P1
i viai + v
i viai + vj
≥
j∈P1
=
vq(cid:80)
n(cid:88)
i viai + v
By Equation 2.4,
[ajvj (qj − λ∗)] +
[ajvj (qj − λ∗)]
(cid:21)
(qj − λ∗)
(qj − λ∗)
+
j∈P2
(cid:20) vjajqjvj
(cid:88)
(cid:80)
(cid:88)
vq(cid:80)
i viai + vj
i viai + v
j∈P2
[ajvj (qj − λ∗)] .
(cid:21)
n(cid:88)
j=1
[ajvj (qj − λ∗)] = 0
and the result follows.
j=1
Corollary 2. In expectation, the P-ranking policy under social influence achieves more
downloads than the any ranking policy under the independent condition.
This corollary follows directly from the fact that the expected downloads at each step do
not change under the independent condition regardless of the ranking policy in use.
17
3 Discussion
This paper presented a "measure and optimize" algorithm for the MusicLab which com-
putes, at each refresh step, an optimal ranking of the songs, given the appeal, estimated
quality, current download counts, and position visibility. This performance ranking, which
maximizes the expected number of downloads, can be computed in strongly polynomial
time, making the M&O algorithm highly scalable. Our experimental results reveal two key
findings:
1. The M&O algorithm leverages social influence to bring significant benefits in expected
downloads.
2. Social influence and position bias both bring improvements in expected downloads
and their effects are cumulative.
Our theoretical results formally validate the first finding and prove that the performance
ranking under social influence always achieves more expected downloads than any ranking
policy under the independent condition.
Our results shed an interesting and novel light about social influence and the unpre-
dictability it creates in cultural markets. The unpredictability coming from social influence
is often presented as an undesirable property of cultural markets. However, as our exper-
imental and theoretical results show, social influence, when used properly, is guaranteed to
help in maximizing the efficiency of the market. So, unless predictability is a goal per se,
social influence should be seen as an asset for optimizing cultural markets.
Our results also show that, in reasonable settings and with a performance ranking, the
unpredictability created by the use of social influence is small. Moreover, the performance
ranking identifies "blockbusters" and these "blockbusters" are the primary reason for the
increased efficiency of the market. There are conditions in the MusicLab where the best
song is not recovered and does not have the most downloads. This happens when the
best song has a "bad" appeal and is dominated by another song. Assume, for instance,
that we have only two songs and q1 > q2.
If v1A1q1 < v2A2q2 (v1 > v2), the added
visibility is not able to overcome the bad appeal: Song 1, despite its better quality, will
be dominated by song 2. Note however that these situations can be predicted, since these
dominance relationships can be identified. Equally important, social influence still helps in
these settings compared to policies not using it.
The unpredictability when using the performance ranking comes from songs that are
indistinguishable, i.e., Aiqi ≈ Ajqj.
In this case, social influence may favor one of the
songs depending on how the process unfolds. It is interesting to point out once again that
these situations can be identified a priori. Together with the dominance relationships, these
equivalence relationships helps us identify the sources of unpredictability, i.e., which songs
may emerge in the MusicLab. Moreover, it is conceivable that we can also control this
unpredictability to ensure fairness and reduce inequalities. This is a topic we are currently
investigating.
18
Our experiments have also shown that, contrary to the download ranking, the per-
formance ranking does not transform average songs into "hits". In reasonable settings, it
always identifies the "blockbusters" while, in a worst-case scenario where song appeal is neg-
atively correlated with song quality, it still promotes high-quality songs. There is a certain
robustness in the performance ranking which is not present in the download ranking.
A key insight from this research is the need to leverage social influence carefully. The
download ranking uses social influence twice: Directly through the download counts and
indirectly through the ranking. The experimental results indicate that the resulting al-
gorithm puts too much emphasis on download counts, which are not necessarily a good
indication of quality due to the sampling process. In contrast, the performance ranking
balances appeal, quality, and social influence.
Observe that, in the generative model of the MusicLab, the perceived quality of a
song is not affected by social influence: Only the probability of listening to the song is. Our
results continue to hold when the perceived quality of a song is affected by social influence
[12, 5, 15], e.g., when the perceived quality of song i is improved by i ≥ 0, when song i is
downloaded. Similarly, the results generalize to the case where the sampling probability is
given by
(cid:80)n
pi,k(σ) =
vσi(αAi + f (Di,k))
j=1 vσj (αAj + f (Dj,k))
,
where function f is positive nondecreasing.
We hope that these results will revive the debate about the consequences of social
influence. Our results show that social influence is beneficial for the efficiency of the market
and that much of its induced unpredictability can be controlled. It appears that a key issue
is to continue building our understanding of the nature of social influence and to determine
how best to use it and for which purposes: Efficiency, predictability, fairness, ...
Acknowledgments
We would like to thank Franco Berbeglia for some key contributions to the proofs and
Lorenzo Coviello for a very detailed reading of the first draft of this paper. NICTA is
funded by the Australian Government as represented by the Department of Broadband,
Communications and the Digital Economy and the Australian Research Council through
the ICT Centre of Excellence program.
References
[1] William T Bielby and Denise D Bielby. All hits are flukes: Institutionalized deci-
sion making and the rhetoric of network prime-time program development. American
Journal of Sociology, 99(5):1287 -- 1313, 1994.
19
[2] Richard E Caves. Creative industries: Contracts between art and commerce. Num-
ber 20. Harvard University Press, 2000.
[3] Arthur De Vany. Hollywood Economics: How Extreme Uncertainty Shames the Film
Industry. Routledge, 2011.
[4] Jehoshua Eliashberg and Steven M Shugan. Film critics: Influencers or predictors?
Journal of Marketing, 61(2), 1997.
[5] Simon Farrell. Social influence benefits the wisdom of individuals in the crowd. Pro-
ceedings of the National Academy of Sciences, 108(36):E625 -- E625, 2011.
[6] Peter Hedström. Experimental macro sociology: Predicting the next best seller. Sci-
ence, 311(5762):786 -- 787, 2006.
[7] Paul M Hirsch. Processing fads and fashions: An organization-set analysis of cultural
industry systems. American journal of sociology, pages 639 -- 659, 1972.
[8] Michael C Jensen. Some anomalous evidence regarding market efficiency. Journal of
financial economics, 6(2):95 -- 101, 1978.
[9] Santosh N Kabadi and Abraham P Punnen. A strongly polynomial simplex method for
the linear fractional assignment problem. Operations Research Letters, 36(4):402 -- 407,
2008.
[10] Coco Krumme, Manuel Cebrian, Galen Pickard, and Sandy Pentland. Quantifying
social influence in an online cultural market. PloS one, 7(5):e33785, 2012.
[11] Kristina Lerman and Tad Hogg. Leveraging position bias to improve peer recommen-
dation. PLOS ONE, 9(6):1 -- 8, 2014.
[12] Jan Lorenz, Heiko Rauhut, Frank Schweitzer, and Dirk Helbing. How social influence
can undermine the wisdom of crowd effect. Proceedings of the National Academy of
Sciences, 108(22):9020 -- 9025, 2011.
[13] Lev Muchnik, Sinan Aral, and Sean J. Taylor. Social influence bias: A randomized
experiment. Science, 341(6146):647 -- 651, 2013.
[14] Richard A Peterson and David G Berger. Entrepreneurship in organizations: Evidence
from the popular music industry. Administrative Science Quarterly, 16(1), 1971.
[15] Heiko Rauhut, Jan Lorenz, Frank Schweitzer, and Dirk Helbing. Reply to farrell:
Improved individual estimation success can imply collective tunnel vision. Proceedings
of the National Academy of Sciences, 108(36):E626 -- E626, 2011.
[16] Daniel Rigney. The Matthew Effect. Columbia University Press, 2010.
20
[17] Matthew J Salganik, Peter Sheridan Dodds, and Duncan J Watts. Experimental
science,
study of inequality and unpredictability in an artificial cultural market.
311(5762):854 -- 856, 2006.
[18] Matthew J Salganik and Duncan J Watts. Leading the herd astray: An experimental
study of self-fulfilling prophecies in an artificial cultural market. Social Psychology
Quarterly, 71(4):338 -- 355, 2008.
[19] Matthew J Salganik and Duncan J Watts. Web-based experiments for the study of
collective social dynamics in cultural markets. Topics in Cognitive Science, 1(3):439 --
468, 2009.
[20] Maiko Shigeno, Yasufumi Saruwatari, and Tomomi Matsui. An algorithm for fractional
assignment problems. Discrete Applied Mathematics, 56(2):333 -- 343, 1995.
[21] Robert J Shiller. Market volatility. MIT press, 1992.
[22] Andrei Shleifer. Inefficient markets: An introduction to behavioral finance. Oxford
university press, 2000.
[23] Arnout van de Rijt, Soong Moon Kang, Michael Restivo, and Akshay Patil. Field
experiments of success-breeds-success dynamics. Proceedings of the National Academy
of Sciences, 111(19):6934 -- 6939, 2014.
[24] Duncan J Watts. Everything Is Obvious: How Common Sense Fails Us. Random
House LLC, 2012.
21
|
1702.01807 | 2 | 1702 | 2018-05-30T19:06:49 | How well can machine learning predict demographics of social media users? | [
"cs.SI",
"cs.CY"
] | The wide use of social media sites and other digital technologies have resulted in an unprecedented availability of digital data that are being used to study human behavior across research domains. Although unsolicited opinions and sentiments are available on these platforms, demographic details are usually missing. Demographic information is pertinent in fields such as demography and public health, where significant differences can exist across sex, racial and socioeconomic groups. In an attempt to address this shortcoming, a number of academic studies have proposed methods for inferring the demographics of social media users using details such as names, usernames, and network characteristics. Gender is the easiest trait to accurately infer, with measures of accuracy higher than 90 percent in some studies. Race, ethnicity and age tend to be more challenging to predict for a variety of reasons including the novelty of social media to certain age groups and a lack of significant deviations in user details across racial and ethnic groups. Although the endeavor to predict user demographics is plagued with ethical questions regarding privacy and data ownership, knowing the demographics in a data sample can aid in addressing issues of bias and population representation, so that existing societal inequalities are not exacerbated. | cs.SI | cs | 1
How well can machine learning predict demographics of social media users?
Nina Cesare ([email protected]),1 Christan Grant ([email protected]),2 Quynh Nguyen
([email protected]),3 Hedwig Lee ([email protected]),4 Elaine O. Nsoesie
([email protected])1*
Institute for Health Metrics and Evaluation, University of Washington, Seattle, WA 98121, USA
1.
2. School of Computer Science, University of Oklahoma, Norman, OK 73019, USA
3. Department of Epidemiology and Biostatistics, University of Maryland College Park, MD 20742, USA
4. Department of Sociology, Washington University, St. Louis, MO 63130, USA
*Corresponding author
ABSTRACT
The wide use of social media sites and other digital technologies have resulted in an
unprecedented availability of digital data that are being used to study human behavior across
research domains. Although unsolicited opinions and sentiments are available on these
platforms, demographic details are usually missing. Demographic information is pertinent in
fields such as demography and public health, where significant differences can exist across sex,
racial and socioeconomic groups. In an attempt to address this shortcoming, a number of
academic studies have proposed methods for inferring the demographics of social media users
using details such as names, usernames, and network characteristics. Gender is the easiest trait to
accurately infer, with measures of accuracy higher than 90% in some studies. Race, ethnicity and
age tend to be more challenging to predict for a variety of reasons including the novelty of social
media to certain age groups and a lack of significant deviations in user details across
racial/ethnic groups. Although the endeavor to predict user' demographics is plagued with
ethical questions regarding privacy and data ownership, knowing the demographics in a data
sample can aid in addressing issues of bias and population representation, so that existing
societal inequalities are not exacerbated.
Keywords: social media; Twitter; blogs; demographics; population representation
2
INTRODUCTION
Social media platforms offer (near) real-time unsolicited information on users' thoughts, feelings
and experiences, allowing researchers to track attitudes and behaviors as they emerge and over
time. The individual and population scale of these data enables the study of the flow of
information through complex networks that are difficult to assess using traditional means of data
collection [1]. These data are also usually low-cost compared to traditional modes of data
collection, such as surveys.
Researchers have used these data to study social processes across a wide domain, ranging from
economics to public health. For example, posts on Facebook and Twitter have been used in real-
time tracking of public opinion on political events and social issues [2–7]. In public health, these
data have been used to study attitudes toward tobacco use [8–11], vaccinations [12, 13], sleep
disorders [14], mental health [15, 16], patient-perceived quality of care in hospitals [17], food
deserts [18], foodborne illness [19–22], trends in physical activity, weight loss, and diet [23–25],
and the dynamics of infectious diseases (such as, influenza and dengue) [26–32].
Studies have also highlighted potential limitations and biases of using these data for public health
[29, 33, 34]. One such limitation is a lack of demographic detail in many digital data sources
which limits their applicability for research and practice in fields such as demography and public
health, where it is pertinent. For example, demographic information can enable the study of
health disparities between populations from different racial or ethnic backgrounds, sex, age and
socioeconomic status to ensure that the use of these data in public health practice does not
increase existing health inequalities. Knowledge of demographics can also allow for procedures
to reweight the data to help adjust for over- and under-representation from certain groups. This
lack of demographic details has led to the proposal of a range of methods for inferring the
demographics of social media users based on their profile, postings, and network information.
Studies typically assume that a user's color preference, name, username, network characteristics,
language usage and other attributes can reveal demographic traits such as, age, gender and race
or ethnicity. While targeted advertising and recommender systems on the Internet suggest that
advanced demographic prediction methods exist, the massive troves of data used in these
processes are usually not available to academicians.
3
Here, we perform a comprehensive review on the performance of approaches proposed to detect
the gender, age, race and ethnicity of social media and blog users in the academic literature. We
characterize trends in the literature, discuss how the methods and metadata used in these studies
could impact predictions, and identify open areas for development.
Selection and Evaluation of Studies
We conducted an initial literature search in November 2016 and updated the search in November
2017. We queried combinations of the following term groups: "demographic, gender, age, or
race", "prediction, classification, or machine learning" and "social media, Twitter or blogs". We
searched three databases – Engineering Village, Microsoft Academic, and the ACM Digital
Archives – since preliminary results indicated most of the papers were published in computer
science conference proceedings. We supplemented these searches with findings from Google
Scholar, which provides access to a broad scope of academic literature, including conference
proceedings and pre-prints.
In total, we identified 160 potentially relevant articles based on their titles and summaries or
abstracts. We focused on articles that aimed to infer or predict the demographics of social media
and blog users from information provided within that context. We therefore excluded studies that
predict demographic attributes based on other sources of digital data, such as browsing history or
mobile phone messages. Additional exclusion criteria were applied to arrive at 91 articles fitting
the focus of this review (see Figure 1). We discuss the methods, group and present studies based
on the demographic trait predicted.
4
Figure 1: Literature identification and selection. We selected 160 potentially relevant articles
based on title and abstract from three academic databases and Google scholar. After excluding
irrelevant studies, we arrived at 91 articles that proposed methods for inferring the age, race,
ethnicity or gender of social media users.
RESULTS
Summary of Methods Used in Prediction
The methods presented in the selected manuscripts included machine learning approaches, raw
and adjusted data matching approaches, and human and automated facial recognition tools.
Seventy-five (82.4%) studies used supervised machine learning approaches, while two studies
used unsupervised learning. The most common machine learning methods considered included,
support vector machines [35–56], naïve Bayes [57–62], variants of the winnow algorithm [63–
66], k-nearest neighbors [37, 67, 68], naïve Bayes decision tree hybrid [69, 70], gradient boosted
decision trees [71, 72], neural networks [73–75], random forests [76], entropy models [51, 61,
77], and standard and regularized linear, logistic or log-linear regression [78–92]. These methods
were applied to data from various platforms (see Figure 2), with an overwhelming focus on
Twitter (n=57, 62.6%), likely due to the availability and wide use of Twitter data for research.
5
Five (5.5%) studies inferred demographics based on facial evaluations from either paid human
evaluators or facial recognition algorithms. The remaining 11 (12.1%) studies used raw or
adjusted data matching, meaning they directly linked user data to information from sources such
as the United States Census Bureau, or estimated the probability of belonging to particular
demographic categories given a combination of known demographic information (e.g., prior data
on the popularity of first names over time) and information from users' profiles.
Figure 2: Literature by year. Early predictions of user demographics focused on Blogs.
However, research predicting the traits of Twitter users is most common.
In addition, the measures used to assess the performance of the proposed algorithms included,
accuracy [(true positives + true negative)/total population], precision (also known as positive
predictive value = true positives/predicted positive condition), recall (also known as sensitivity=
[true positive/(true positives + false negatives)], F1-score (the harmonic mean of the precision
and recall), and area under the curve (AUC) (plot representing sensitivity/specificity pairs
corresponding to different decision thresholds; is a measure of how well a parameter can
distinguish between two diagnostic categories). Specifically, 62 (87.3%), 37 (84.1%) and 12
(60.0%) of the gender, age and race/ethnicity prediction approaches respectively, reported
6
findings using a specific performance metric. Accuracy was the most commonly reported metric;
reported in 74.6%, 47.7%, and 35.0% of studies on gender, age and race/ethnicity prediction
studies respectively. The distribution of accuracy by trait is shown in Figure 3. Studies
predicting gender had the highest median accuracy, with some studies reaching scores higher
than 0.90. Although all studies on age prediction that report accuracy used the same
methodological approach, the accuracy fell across a broad range (0.43, 0.95). In contrast, the
prediction accuracy for race/ethnicity fell within a narrower range and had a median value of
approximately 0.81.
Figure 3: Distribution of Predicted Accuracy by Trait. Methods predicting gender perform
best. The median accuracy of methods used in inferring age is similar to that of gender, but the
distribution varies widely. In contrast, the range of accuracies observed in the prediction of race
and ethnicity is narrower.
Predicting Gender
Seventy-one (78.0%) studies focused on gender prediction [35, 37–40, 42–45, 47–49, 51–53, 57–
70, 73, 74, 77–86, 91, 93–121]. Of these, 54.9% employed user postings, 25.3% used
information from user profiles (such as, profile colors), and the remainder used a combination of
postings, network and profile information.
7
Most studies proposing classification methods focused on linguistic patterns in names and posted
text. Studies that employed users' names either linked first names to offline survey data or used
these data to build models designed to improve predictive algorithms that incorporate other
characteristics. Although some sites (e.g., Twitter) do not require users to provide a valid first
and last name in their profile, this technique may still capture a large sample of users. For
example, Mislove et al. [108] found that 64.2% of Twitter users elect to include a first name in
their profile. However, approximately 71.8% of these Twitter users were male.
The performance metrics were consistently above 0.50. Of the studies that used accuracy (n=53,
74.6%) as the measure of evaluation, the average accuracy was 0.83 (range 0.65 to 0.99). Studies
that reported precision also had high values, with an average of 0.84 (range 0.58 to 0.98).
Reported F1-scores were also high, with an average value of 0.84 (range 0.68 to 0.98). We
excluded the values reported by Miller et al. [59] because they reported results for females but
not males. Overall, most of the proposed algorithms could reliably predict the gender of social
media users.
Bamman et al. [101] suggest that gender is heavily embedded within individuals' identities. It
may be equally embedded within social media users' online personas to the extent that
unsupervised clustering techniques can differentiate between users with distinct gender
identities. For example, Alowibdi et al. [69] found that profile colors alone can be effectively
used to infer the gender of Twitter users'. They predicted the gender of Twitter users with 71.4%
accuracy using only five quantized colors (i.e., compression via the reduction in the number of
colors) from users' profile fields.
Another reason why gender lends itself well to profile-based prediction techniques is because a
person's name can be a strong predictor of gender. For instance, Liu and Ruths [42] found that
adding names data to a text-dependent gender prediction improved the accuracy of their method
from 0.83 to 0.87. Both Fernandez et al. [74] and Karimi et al. [113] suggested inferring gender
by combining user names with results from facial recognition software applied to profile photos.
Mueller and Stumme [44] and Vicente et al. [105] found that even if a name is not directly
8
identifiable, it can either be cleaned in a way that renders the true name linkable with official
census data (for example, converting "Hi I'm Kaylaaaaa" to "Kayla"), or it may contain
extractable features that are gender dependent (e.g. a name that ends in a vowel is more likely to
imply a female user than a male user). Furthermore, Cesare et al. [122] also developed a
weighted ensemble model that used word and character n-grams from user names [65], the
linguistic structure of names [44], and direct matches between user names and U.S. Social
Security Administration (SSA) data to obtain an accuracy of 0.83, precision of 0.82, recall of
0.85, and F1-score of 0.84.
Studies that solely used profile metadata or combined it with users' posts had higher accuracies
than those that solely use information from users' postings. This suggests that simple, subtle cues
embedded within users' profiles can be used to accurately detect gender. In addition, studies that
used non-algorithmic methods (i.e. simple or adjusted data matching) also performed
comparably well as studies that used machine learning algorithms (i.e., average accuracy of 0.81
and 0.82 for data matching and machine learning, respectively)
Predicting Age
We identified 44 (48.4%) studies that predicted age [35, 37, 49, 52, 53, 56, 58, 62–64, 73, 75, 76,
78–80, 83, 85, 87, 89–98, 104, 106, 110, 111, 117, 118, 120, 123–130]. The proposed algorithms
were overwhelmingly text-dependent. These studies used text features and supervised learning
methods to predict exact numeric age, age category (adolescent/adult/senior) and life stage (e.g.,
student, employed, retired). Most studies capped predicted age between 30 and 40, perhaps due
to the fact that social media users tend to be younger. The reported average F1-scores of studies
that capped predicted age in the 40s was comparable to that of studies that capped predicted age
in the 30s (F1=0.74 and F1=0.71, respectively).
Several studies suggested that profile photos may also be used to assess age. McCormick et al.
[110] found that human coders are highly reliable when assessing a user's age range. When
shown 1,000 Twitter user profile images, at least two out of three workers on Amazon's
Mechanical Turk could identify the numeric age range of 96.6% and the life stage of 97.9% of
users. An and Weber [93] used Face++, a free facial recognition service that can estimate a
9
user's age within a 10-year span and observed that 49% of the 346,050 Twitter users in their
dataset had a detectable face, which is low. When comparing the predicted age of the user and
the age extracted from their profile description, they found that 25% of the predictions were
within a three-year difference and 75% of predictions were within an 11-year difference of the
user's extracted age. However, these studies did not compare estimates to ground truth data (i.e.,
true demographic information about the users sampled). Rather, estimates were compared to
population values or data not drawn from the population studied. Further work is needed to
assess the reliability, and best applications of facial recognition by human and automated coders.
Simple approaches – such as identifying mentions of age within a user's profile or matching first
names to trends in popularity over times – could be useful but not widely employed within
existing literature beyond the development of a training dataset. For the former, the available
data might be limited to teenagers or milestone years, and the latter may provide reliable
predictions for popular names only.
The reported high accuracy observed in age prediction suggests that clues from users' postings -
such as, the sociolinguistic indicators or topics discussed - provide valuable information about
the age of the user. The accuracy and F1-scores of age prediction studies that relied exclusively
on information from users' postings were slightly higher than those of studies that combined data
from users' postings with profile or network information (accuracy=0.76, 0.72, respectively; F1-
score=0.80, 0.74, respectively). Thus, using additional data besides profiles posts did not
significantly improve prediction of age.
These studies indicate that age prediction is more difficult with additional categories. In general,
the performance of prediction algorithms decreased as the number of predicted categories
increased. The average F1-scores of studies that predicted binary age categories (e.g. under/over
20 years old), three age categories (e.g. under 20, 20-40, above 40), and four age categories (e.g.
10-19, 20-29, 30-39, 40+) were 0.91, 0.73 and 0.71, respectively. The few studies that predicted
numeric age, reported an average correlation of 0.73. While prediction of numeric age is
challenging, it is most beneficial since the results can be binned into different categories as
needed.
10
Predicting Race and Ethnicity
Twenty (21.9%) of the selected studies addressed the prediction of race or ethnicity [48, 49, 54,
71, 72, 78, 79, 82, 85, 86, 91, 93, 107–111, 126, 131, 132]. The earliest studies were from 2010,
compared to 2006 for age and gender, likely due to the challenge of predicting this demographic
trait. Thirteen of these studies reported an average accuracy of 0.82 and twelve studies reported
an average F1-score of 0.70. Eight studies did not report a performance metric.
Nine (45.0%) race/ethnicity prediction studies solely relied on users' postings or combined
users' postings with profile and network metadata to infer race. The performance of these studies
suggests that posted text provides valuable insight into the race or ethnicity of the user, but that
performance improves considerably when these data are analyzed with the added context of
profile or network metadata. The average accuracy and F1-scores of text-only methods was 0.76
and 0.51, respectively, while the average accuracy and F1-scores of methods that combined text
with additional metadata was 0.81 and 0.73, respectively.
Six studies solely used profile metadata to predict race. In McCormick et al. [110], human
analysts labeled race based on profile photos. At least two out of three human coders agreed on
the race of 95.5% of 1,000 user photos. Chen et al. [49] applied n-grams and Linear Discriminant
Analysis (LDA) to data on Twitter users' neighbors' descriptions to predict race with 78.8%
accuracy (F1=0.75).
Additionally, last name was a common feature used in studies that employed profile data alone
or in conjunction with other metadata. Longley et al. [111] used the Onomap system – an online
index of names linked to geographic information developed by Mateos and colleagues [133] – to
infer the ethnicity of Twitter users in London. Similarly, Hofstra et al. [107] used last names to
determine whether Dutch Facebook users were Dutch, Moroccan or Caribbean. Neither Longley
et al. [111] or Hofstra et al. [107], however, reported performance metrics given ground truth
data. Rao et al. [109] illustrated that last names provide valuable information that can be used
solely or in conjunction with user postings to predict the ethnicity of Nigerian Facebook users
with an accuracy of approximately 0.81. Similar to the use of first names for age prediction,
ethnicity is commonly tied to last name making it traceable and useful for prediction.
11
Similarly, Chang et al. [131] and Mislove et al. [108] applied a method for inferring the race of
U.S. Twitter users based on last names and the aggregate racial distribution of the population
from which the user is drawn. However, they failed to assess their prediction against individual
ground truth data. Additional work is needed to rigorously assess the potential of this prediction
technique.
DISCUSSION
We performed a comprehensive review of studies that propose methods for inferring the age,
race, ethnicity and gender of social media and blog users. We examined trends in the literature
by trait and assessed the performance of the proposed methods.
We note that gender was the easiest trait to infer using a variety of methods and features. In
contrast, in the prediction of age, methods that relied on data from users' postings reported
higher accuracies than methods that did not use data from users' postings or combined these data
with other features. Also, the performance of age prediction algorithms generally decreased as
the number of categories predicted increased. However, trends in the literature suggest that
multi-category or numeric age predictions may improve as more advanced algorithms are used.
Likewise, in the prediction of race, user postings provided valuable information about the user
but is best if considered in conjunction with other user metadata. Also, similar to the use of first
names as a gender prediction feature, last name appears to be a strong predictor of user ethnicity.
The improvement in race and ethnicity prediction based on additional data from users' networks
and posting could be attributed to within-network segregation by race or ethnicity.
Measures of Performance
Several of the studies included in this review did not evaluate prediction performance. Nine
gender prediction approaches reviewed provided no performance metrics, and all of these used
profile metadata to generate predictive features. Similarly, only one of the seven age prediction
approaches and none of the six race/ethnicity prediction approaches that relied only on profile
metadata reported performance metrics. Furthermore, of the studies reporting a performance
metric, most used accuracy. However, accuracy is not always a good measure of performance,
12
particularly for studies with a class imbalance because the model may be over-trained to
recognize the majority group, and the overall accuracy may be excessively weighted by
classification performance specific to this group [134, 135]. To assess how well these prediction
methods are performing, accuracy needs to be reported alongside other measures of performance
such as, precision, recall, F1-score, or AUC.
Data Limitations
While further work is needed to explore the potential value of profile-based prediction features
(such as, user description and user names), findings from this review suggest that users' postings
contain valuable information for predicting users' age and race/ethnicity. This has important
implications for the scalability of this work. While sites such as Twitter and Instagram allow
researchers to collect information via public Application Programming Interfaces (APIs), these
applications are typically rate limited and may limit the scalability of a particular method [69].
For instance, when gathering timelines, the Twitter API can only make 900 calls per 15-minute
window. After reaching a rate limit, the API requires 15 minutes to reset. Although these rate
limits may not be a problem when gathering data for several thousand users, gathering data on
hundreds of thousands or millions of users would be time consuming. For researchers studying
time-sensitive topics – such as infectious disease outbreaks – this delay may be prohibitive.
Furthermore, collecting, storing and processing hundreds of tweets or ties per user for very large
samples is computationally intensive and may not be accessible to applied researchers with
limited resources.
Further, several studies did not use individual ground truth data in the evaluation of their
methods. The manner in which the test and training data used are collected may impact the
results of proposed prediction methods. The studies evaluated engaged in a variety of techniques
for measuring users' 'true' age, race and gender. Some relied on self-reported information
contained within the user's postings – for instance, by collecting recent tweets in which a user
mentions their current age [76], states that it is their birthday [56, 92, 96], or describes their
racial or ethnic identity [54, 72]. Some studies use hand-coded evaluations of users' profiles
and/or other online presences [85, 87, 89] to assess demographics, and some use facial
recognition software such as Face++ [115] or crowdsourcing platforms such as Amazon's
13
Mechanical Turk [78] to generate ground truth measures. While some studies (e.g., Bergsma and
colleagues [48]) address the effect of differently collected training datasets, further work is
needed to assess these potential biases. Ground truth data on social media and blog users, while
difficult to obtain, is important for assessing the validity of a proposed approach.
Ethical Considerations
We acknowledge that detection of user demographics poses unique ethical questions for
researchers. First, users who provide data for the development and application of classification
methods are not treated as "human subjects" in a traditional sense. According to the Belmont
Report and Common Rule, social researchers are supposed to exercise justice (i.e., ensuring fair
distribution of benefits and risks of research, as well as fair selection of research subjects),
beneficence (i.e., an obligation to maximize benefits and minimize harm toward research
subjects) and informed consent (i.e., disclosure of objectives and possible consequences of the
research toward the subject and society at large)[136]. However, implementing procedures such
as informed consent is difficult when working with tens or hundreds of thousands of
respondents. Additionally, standards regarding data ownership and publishing typically dictate
how researchers may use data from digital sources. Further work is need to determine how to
preserve human subjects research standards when working with large, digitally generated
datasets [137–139].
Ethical concerns regarding the automatic detection of user characteristics extend beyond the
individual. The term 'group privacy' refers to demographic de-anonymization, or the placement
of users within specific demographic categories [140–142]. While ethical standards regarding
the protection of research subjects typically assume that individual protection leads to group
protection [142], research suggests that disclosed affiliation with a particular demographic group
may lead to unfair targeting by marketing agencies, political campaigns, and more. For instance,
Zwitter [140] notes that political campaigns have used bot accounts to target and engage in
fabricated grassroots debates with Twitter users who fit the demographic profile of a particular
party's constituency – a practice sometimes referred to as "astroturfing". Researchers seeking to
predict user demographics not explicitly shared through their profiles should be aware of the
14
risks that disclosure of group status may generate prior to using this inferred data for analysis or
intervention.
CONCLUSIONS
We found that gender is heavily embedded within users' online presence, and that features
extracted from users' profiles – such as first names – may be used to accurately predict this trait.
Other demographic traits such as, age and race/ethnicity are more challenging to predict.
Furthermore, there is a need to assess approaches relying exclusively on profile metadata with
the same rigor as approaches that use information from networks or users' postings. Future work
in this area should be mindful of the scalability of proposed methods, as well as the implications
of these methods for individual and group privacy.
ACKNOWLDGEMENTS
This work was funded by the Robert Wood Johnson Foundation (grant #73362).
REFERENCES
1. Prier KW, Smith MS, Giraud-Carrier C, Hanson CL (2011) Identifying health-related topics
on twitter. In: International Conference on Social Computing, Behavioral-Cultural
Modeling, and Prediction. Springer, pp 18–25
2. Wang W, Rothschild D, Goel S, Gelman A (2015) Forecasting elections with non-
representative polls. Int J Forecast 31:980–991 . doi: 10.1016/j.ijforecast.2014.06.001
3. O'Connor B, Balasubramanyan R, Routledge BR, Smith NA (2010) From tweets to polls:
Linking text sentiment to public opinion time series. ICWSM 11:1–2
4. McCombs M (2014) Setting the Agenda: Mass Media and Public Opinion. John Wiley & Sons
5. Anstead N, O'Loughlin B (2015) Social Media Analysis and Public Opinion: The 2010 UK
General Election. J Comput-Mediat Commun 20:204–220 . doi: 10.1111/jcc4.12102
6. Ruths D, Pfeffer J (2014) Social media for large studies of behavior. Science 346:1063–1064
. doi: 10.1126/science.346.6213.1063
7. Rossini PGC, Hemsley J, Tanupabrungsun S, et al (2017) Social Media, U.S. Presidential
Campaigns, and Public Opinion Polls: Disentangling Effects. In: Proceedings of the 8th
International Conference on Social Media & Society. ACM, New York, NY, USA, pp 56:1–
56:5
15
8. Duke JC, Hansen H, Kim AE, et al (2014) The use of social media by state tobacco control
programs to promote smoking cessation: a cross-sectional study. J Med Internet Res 16:e169
. doi: 10.2196/jmir.3430
9. Huang J, Kornfield R, Szczypka G, Emery SL (2014) A cross-sectional examination of
. doi:
marketing of electronic cigarettes on Twitter. Tob Control 23:iii26–iii30
10.1136/tobaccocontrol-2014-051551
10. Myslín M, Zhu S-H, Chapman W, Conway M (2013) Using Twitter to Examine Smoking
Behavior and Perceptions of Emerging Tobacco Products. J Med Internet Res 15:e174 . doi:
10.2196/jmir.2534
11. Prochaska JJ, Pechmann C, Kim R, Leonhardt JM (2011) Twitter=quitter? An analysis of
. doi:
social networks. Tob Control
Twitter quit
smoking
10.1136/tc.2010.042507
tc.2010.042507
12. Bahk CY, Cumming M, Paushter L, et al (2016) Publicly Available Online Tool Facilitates
Real-Time Monitoring Of Vaccine Conversations And Sentiments. Health Aff (Millwood)
35:341–347 . doi: 10.1377/hlthaff.2015.1092
13. Salathé M, Khandelwal S (2011) Assessing Vaccination Sentiments with Online Social Media:
Implications for Infectious Disease Dynamics and Control. PLOS Comput Biol 7:e1002199
. doi: 10.1371/journal.pcbi.1002199
14. McIver DJ, Hawkins JB, Chunara R, et al (2015) Characterizing Sleep Issues Using Twitter. J
Med Internet Res 17:e140 . doi: 10.2196/jmir.4476
15. De Choudhury M, Counts S, Horvitz E (2013) Predicting postpartum changes in emotion and
behavior via social media. In: Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems. ACM, pp 3267–3276
16. Sueki H (2015) The association of suicide-related Twitter use with suicidal behaviour: A cross-
sectional study of young internet users in Japan. J Affect Disord 170:155–160 . doi:
10.1016/j.jad.2014.08.047
17. Hawkins JB, Brownstein JS, Tuli G, et al (2016) Measuring patient-perceived quality of care
in US hospitals using Twitter. BMJ Qual Saf 25:404–413 . doi: 10.1136/bmjqs-2015-
004309
18. De Choudhury M, Sharma S, Kiciman E (2016) Characterizing Dietary Choices, Nutrition,
and Language in Food Deserts via Social Media. In: Proceedings of the 19th ACM
Conference on Computer-Supported Cooperative Work & Social Computing. ACM Press,
San Francisco, California, pp 1155–1168
19. Harris JK, Mansour R, Choucair B, et al (2014) Health Department Use of Social Media to
Identify Foodborne Illness - Chicago, Illinois, 2013–2014. Center for Disease Control
16
20. Nsoesie EO, Kluberg SA, Brownstein JS (2014) Online reports of foodborne illness capture
foods implicated in official foodborne outbreak reports. Prev Med 67:264–269 . doi:
10.1016/j.ypmed.2014.08.003
21. Harrison C, Jorder M, Stern H, et al (2014) Using online reviews by restaurant patrons to
identify unreported cases of foodborne illness-New York City, 2012–2013. MMWR
63:441–445
22. Harris JK, Hawkins JB, Nguyen L, et al (2017) Using Twitter to Identify and Respond to Food
Poisoning: The Food Safety STL Project. J Public Health Manag Pract JPHMP 23:577–580
. doi: 10.1097/PHH.0000000000000516
23. Nguyen QC, Li D, Meng H-W, et al (2016) Building a National Neighborhood Dataset From
Geotagged Twitter Data for Indicators of Happiness, Diet, and Physical Activity. JMIR
Public Health Surveill 2:e158 . doi: 10.2196/publichealth.5869
24. Pagoto S, Schneider KL, Evans M, et al (2014) Tweeting it off: characteristics of adults who
tweet about a weight loss attempt. J Am Med Inform Assoc 21:1032–1037 . doi:
10.1136/amiajnl-2014-002652
25. Griffin GP, Jiao J (2015) Where does bicycling for health happen? Analysing volunteered
geographic information through place and plexus. J Transp Health 2:238–247 . doi:
10.1016/j.jth.2014.12.001
26. Corley CD, Cook DJ, Mikler AR, Singh KP (2010) Text and Structural Data Mining of
Influenza Mentions in Web and Social Media. Int J Environ Res Public Health 7:596–615 .
doi: 10.3390/ijerph7020596
27. Gomide J, Veloso A, Meira W Jr, et al (2011) Dengue Surveillance Based on a Computational
Model of Spatio-temporal Locality of Twitter. In: Proceedings of the 3rd International Web
Science Conference. ACM, New York, NY, USA, pp 3:1–3:8
28. Nsoesie EO, Brownstein JS (2015) Computational Approaches to Influenza Surveillance:
Beyond Timeliness. Cell Host Microbe 17:275–278 . doi: 10.1016/j.chom.2015.02.004
29. Nsoesie EO, Flor L, Hawkins J, et al (2016) Social Media as a Sentinel for Disease
Surveillance: What Does Sociodemographic Status Have to Do with It? PLOS Curr
Outbreaks. doi: 10.1371/currents.outbreaks.cc09a42586e16dc7dd62813b7ee5d6b6
30. Salathé M, Freifeld CC, Mekaru SR, et al (2013) Influenza A (H7N9) and the Importance of
Digital Epidemiology. N Engl J Med 369:401–404 . doi: 10.1056/NEJMp1307752
31. Chew C, Eysenbach G (2010) Pandemics in the Age of Twitter: Content Analysis of Tweets
doi:
2009 H1N1 Outbreak.
PLOS ONE
during
10.1371/journal.pone.0014118
the
5:e14118
.
17
32. Santillana M, Nguyen AT, Dredze M, et al (2015) Combining Search, Social Media, and
Traditional Data Sources to Improve Influenza Surveillance. PLOS Comput Biol
11:e1004513 . doi: 10.1371/journal.pcbi.1004513
33. Scarpino SV, Scott JG, Eggo R, et al (2016) Data Blindspots: High-Tech Disease Surveillance
Misses the Poor. Online J Public Health Inform 8: . doi: 10.5210/ojphi.v8i1.6451
34. Henly S, Tuli G, Kluberg SA, et al (2017) Disparities in digital reporting of illness: A
. doi:
assessment. Prev Med 101:18–22
socioeconomic
demographic
10.1016/j.ypmed.2017.05.009
and
35. Asoh, Hideki, Ikeda, Kazushi, Ono, Chihiro (2012) A Fast and Simple Method for Profiling a
Population of Twitter Users. In: The Third International Workshop on Mining Ubiquitous
and Social Environment. Bristol, UK, pp 19–26
36. Al Zamal F, Liu W, Ruths D Homophily and Latent Attribute Inference: Inferring Latent
Attributes of Twitter Users from Neighbors. In: Proceedings of the Sixth International
AAAI Conference on Weblogs and Social Media. Dublin, Ireland, pp 387–390
37. Santosh K, Joshi A, Gupta M, Varma V (2014) Exploiting Wikipedia Categorization for
Predicting Age and Gender of Blog Authors. In: Proceedings of the 2014 Conference on
Empirical Methods in Natural Language Processing. Aalborg, Denmark
38. Benton A, Raman A, Dredze M (2016) Learning Multiview Embeddings of Twitter Users. In:
Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics.
ACM, Berlin, Germany, pp 14–19
39. You Q, Bhatia S, Sun T, Luo J (2014) The Eyes of the Beholder: Gender prediction using
images posted in online social networks. In: 2014 IEEE International Conference on Data
Mining Workshop (ICDMW). IEEE, Shenzhen, China, pp 1026–1030
40. Nowson S, Oberlander J (2006) The Identity of Bloggers: Openness and Gender in Personal
Weblogs. In: AAAI spring symposium: Computational approaches to analyzing weblogs.
Palo Alto, CA, pp 163–167
41. Fink C, Kopecky J, Morawski M Inferring Gender from the Content of Tweets: A Region
Specific Example. In: Proceedings of the Six International Conference on Weblogs and
Social Media. pp 459–463
42. Liu W, Ruths D Using first names as features for gender inference in Twitter
43. Ludu PS (2014) Inferring gender of a Twitter user using celebrities it follows. ArXiv Prepr
ArXiv14056667
44. Mueller J, Stumme G (2016) Gender Inference using Statistical Name Characteristics in
Twitter. In: Proceedings of the The 3rd Multidisciplinary International Social Networks
Conference on SocialInformatics. ACM Press, Albany, NY, pp 1–8
18
45. Mukherjee A, Liu B (2010) Improving Gender Classification of Blog Authors. In: Proceedings
of the 2010 Conference on Empirical Methods in Natural Language Processing. Association
for Computational Linguistics, Stroudsburg, PA, USA, pp 207–217
46. Rao D, Yarowsky D, Shreevats A, Gupta M (2010) Classifying Latent User Attributes in
Twitter. In: Proceedings of the 2nd international workshop on Search and mining user-
generated contents. ACM, Toronto, Canada, pp 37–44
47. Zhang C, Zhang P (2010) Predicting gender from blog posts. Technical Report. University of
Massachusetts Amherst, USA
48. Bergsma S, Dredze M, Van Durme B, et al (2013) Broadly Improving User Classification via
Communication-Based Name and Location Clustering on Twitter. In: HLT-NAACL.
Atlanta, Georgia, pp 1010–1019
49. Chen X, Wang Y, Agichtein E, Wang F (2015) A comparative study of demographic attribute
inference in twitter. In: Proceedings of the Ninth International AAAI Conference on Web
and Social Media. Palo Alto, CA, pp 590–593
50. Bhatnagar VS, Sanket NJ, Barathi SK "Honey, We Shrunk the Weights" Gender Prediction
using Twitter Feeds and Profile Images
51. Kokkos A, Tzouramanis T (2014) A robust gender inference model for online social networks
and its application to LinkedIn and Twitter. First Monday 19: . doi: 10.5210/fm.v19i9.5216
52. Yatam SS, Reddy TR (2014) Author profiling: Predicting gender and age from blogs, reviews
& social media. Int J Eng Res Technol 3:631–633
53. Rao D, Yarowsky D (2010) Detecting Latent User Properties in Social Media. In: Proceedings
of the NIPS MLSN Workshop. Citeseer
54. Xu S, Markson C, Costello KL, et al (2016) Leveraging Social Media to Promote Public Health
Knowledge: Example of Cancer Awareness via Twitter. JMIR Public Health Surveill 2: .
doi: 10.2196/publichealth.5205
55. Siswanto E, Khodra ML (2013) Predicting latent attributes of Twitter user by employing
lexical features. In: 2013 International Conference on Information Technology and
Electrical Engineering (ICITEE). Krabi, Thailand, pp 176–180
56. Tuli, Gaurav (2015) Modeling and Twitter-based Surveillance of Smoking Contagion.
Dissertation, Virginia Polytechnic Institute and State University
57. Yan X, Yan L (2006) Gender Classification of Weblog Authors. In: Proceedings of the AAAI
Spring Symposia on Computational Approaches. AAAI, Palo Alto, CA, pp 27–29
58. Rustagi M, Prasath RR, Goswami S, Sarkar S (2009) Learning Age and Gender of Blogger
from Stylistic Variation. In: Pattern Recognition and Machine Intelligence. Springer, Berlin,
Heidelberg, pp 205–212
19
59. Miller Z, Dickinson B, Hu W (2012) Gender Prediction on Twitter Using Stream Algorithms
. doi:
Intell Sci 02:143–148
with N-Gram Character Features.
10.4236/ijis.2012.224019
Int
J
60. Wang P, Gao M, Sinnott RO (2016) Enhanced Gender Identification through Social Media
Analytics on the Cloud. In: 2016 IEEE 18th International Conference on High Performance
Computing and Communications; IEEE 14th International Conference on Smart City; IEEE
2nd International Conference on Data Science and Systems (HPCC/SmartCity/DSS). pp
1426–1433
61. Mukherjee S, Bala PK (2017) Gender classification of microblog text based on authorial style.
Inf Syst E-Bus Manag 15:117–138 . doi: 10.1007/s10257-016-0312-0
62. Goswami S, Sarkar S, Rustagi M (2009) Stylometric analysis of bloggers' age and gender. In:
Proceedings of the Third International ICWSM Conference
63. Argamon S, Koppel M, Pennebaker JW, Schler J (2007) Mining the Blogosphere: Age, gender
and the varieties of self-expression. First Monday 12:
64. Schler J, Koppel M, Argamon S, Pennebaker JW (2005) Effects of Age and Gender on
Blogging. In: AAAI spring symposium: Computational approaches to analyzing weblogs.
Palo Alto, CA, pp 199–205
65. Burger JD, Henderson J, Kim G, Zarrella G (2011) Discriminating gender on Twitter. In:
Proceedings of the Conference on Empirical Methods in Natural Language Processing.
Association for Computational Linguistics, pp 1301–1309
66. Deitrick W, Miller Z, Valyou B, et al (2012) Gender Identification on Twitter Using the
Modified Balanced Winnow. Commun Netw 04:189–195 . doi: 10.4236/cn.2012.43023
67. Cheung M, She J (2017) An Analytic System for User Gender Identification Through User
Shared Images. ACM Trans Multimed Comput Commun Appl 13:30:1–30:20 . doi:
10.1145/3095077
68. Chen J, Xiao T, Sheng J, Teredesai A (2017) Gender prediction on a real life blog data set
using LSI and KNN. In: 2017 IEEE 7th Annual Computing and Communication Workshop
and Conference (CCWC). pp 1–6
69. Alowibdi JS, Buy UA, Yu P (2013) Language independent gender classification on Twitter.
In: Procedings of the IEEE/ACM International Conference onAdvances in Social Networks
Analysis and Mining (ASONAM). IEEE, Calgary, Canada, pp 739–743
70. Alowibdi JS, Buy UA, Yu P (2013) Empirical Evaluation of Profile Characteristics for Gender
Classification on Twitter. In: 2013 12th International Conference on Machine Learning and
Applications. pp 365–369
20
71. Pennacchiotti M, Popescu A-M (2011) A Machine Learning Approach to Twitter User
Classification. In: Proceedings of the Fifth International AAAI Conference on Weblogs and
Social Media. Menlo Park, CA, pp 281–288
72. Pennacchiotti M, Popescu A-M (2011) Democrats, Republicans and Starbucks Aficionados:
User Classification in Twitter. In: Proceedings of the 17th ACM SIGKDD international
conference on Knowledge discovery and data mining. ACM, San Diego, CA, pp 430–438
73. Kim SM, Xu Q, Qu L, et al (2017) Demographic Inference on Twitter using Recursive Neural
Networks. In: Proceedings of the 55th Annual Meeting of the Association for Computational
Linguistics (Short Papers). Association for Computational Linguistics, Quebec, Canada, pp
471–477
74. Fernández D, Moctezuma D, Siordia OS (2016) Features combination for gender recognition
on Twitter users. In: 2016 IEEE International Autumn Meeting on Power, Electronics and
Computing (ROPEC). pp 1–6
75. Guimarães RG, Rosa RL, Gaetano DD, et al (2017) Age Groups Classification in Social
doi:
IEEE Access
Network Using Deep Learning.
10.1109/ACCESS.2017.2706674
5:10805–10816
.
76. Simaki V, Mporas I, Megalooikonomou V (2016) Age Identification of Twitter Users:
Classification Methods and Sociolinguistic Analysis. In: Proceedings of the 17th
International Conference on Intelligent Text Processing and Computational Linguistics
(CICLing). Montreal, Canada
77. Filippova K (2012) User demographics and language in an implicit social network. In:
Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language Learning. Association for Computational
Linguistics, Stroudsburg, PA, pp 1478–1488
78. Volkova S, Bachrach Y, Armstrong M, Sharma V (2015) Inferring Latent User Properties from
Texts Published in Social Media. In: Proceedings of the Twenty-Ninth AAAI Conference
on Artificial Intelligence. Austin, TX, pp 4296–4297
79. Kosinski M, Stillwell D, Graepel T (2013) Private traits and attributes are predictable from
. doi:
digital records of human behavior. Proc Natl Acad Sci 110:5802–5805
10.1073/pnas.1218772110
80. Sap, Maarten, Eichstaedt, Johannes, Kern, Margaret L., et al (2014) Developing Age and
Gender Predictive Lexica over Social Media. In: Proceedings of the 2014 Conference on
Empirical Methods in Natural Language Processing (EMNLP). Doha, Quatar, pp 1146–
1151
81. Reddy S, Marina del Rey CA (2016) Obfuscating gender in social media writing. In:
Proceedings of the 1st Workshop on Natural Language Processing and Computational
Social Science. Austin, TX, pp 17–26
21
82. Montasser O, Kifer D (2017) Predicting Demographics of High-Resolution Geographies with
Geotagged Tweets. In: Proceedings of the Thirty-First AAAI Conference on Artificial
Intelligence. AAAI Press, Vancouver, Canada, pp 1460–1466
83. Volkova S, Bachrach Y, Durme BV (2016) Mining User Interests to Predict Perceived Psycho-
Demographic Traits on Twitter. In: IEEE Second International Conference on Big Data
Computing Service and Applications. IEEE, pp 36–43
84. Otterbacher J (2010) Inferring gender of movie reviewers: exploiting writing style, content
and metadata. In: Proceedings of the 19th ACM international conference on Information
and knowledge management. ACM, pp 369–378
85. Culotta A, Kumar NR, Cutler J (2015) Predicting the Demographics of Twitter Users from
Website Traffic Data. In: Proceedings of the Twenty-Ninth AAAI Conference on Artificial
Intelligence. AAAI, Austin, Texas, pp 72–78
86. Culotta A, Kumar NR, Cutler J (2016) Predicting Twitter User Demographics using Distant
Supervision from Website Traffic Data. J Artif Intell ResJAIR 55:389–408
87. Nguyen D-P, Gravel R, Trieschnigg RB, Meder T (2013) " How old do you think I am?" A
study of language and age in Twitter. In: Proceedings of the Seventh International AAAI
Conference on Weblogs and Social Media. Palo Alto, CA, pp 439–448
88. Rosenthal S, McKeown K (2011) Age Prediction in Blogs: A Study of Style, Content, and
Online Behavior in Pre- and Post-social Media Generations. In: Proceedings of the 49th
Annual Meeting of the Association for Computational Linguistics: Human Language
Technologies - Volume 1. Association for Computational Linguistics, Stroudsburg, PA,
USA, pp 763–772
89. Nguyen D, Gravel R, Trieschnigg D, Meder T (2013) TweetGenie: automatic age prediction
from tweets. ACM SIGWEB Newsl 1–6 . doi: 10.1145/2528272.2528276
90. Nguyen D, Smith NA, Rosé CP (2011) Author age prediction from text using linear regression.
In: Proceedings of the 5th ACL-HLT Workshop on Language Technology for Cultural
Heritage, Social Sciences, and Humanities. Association for Computational Linguistics,
Stroudsburg, PA, pp 115–123
91. Barberá P (2016) Less is more? How demographic sample weights can improve public opinion
estimates based on Twitter data. Working Paper for NYU
92. Morgan-Lopez AA, Kim AE, Chew RF, Ruddle P (2017) Predicting age groups of Twitter
users based on language and metadata features. PLOS ONE 12:e0183537 . doi:
10.1371/journal.pone.0183537
93. An J, Weber I (2016) # greysanatomy vs.# yankees: Demographics and Hashtag Use on
Twitter. ArXiv Prepr ArXiv160301973
22
94. Al Zamal F, Liu W, Ruths D (2012) Homophily and Latent Attribute Inference: Inferring
Latent Attributes of Twitter Users from Neighbors. ICWSM 270:
95. Argamon S, Koppel M, Pennebaker JW, Schler J (2009) Automatically Profiling the Author
of an Anonymous Text. ACM Commun 52:119–123 . doi: 10.1145/1461928.1461959
96. Beretta V, Maccagnola D, Cribbin T, Messina E (2015) An Interactive Method for Inferring
Demographic Attributes in Twitter. In: Proceedings of the 26th ACM Conference on
Hypertext & Social Media. ACM Press, Guzelyurt, Northern Cyprus, pp 113–122
97. Ikeda D, Takamura H, Okumura M (2008) Semi-Supervised Learning for Blog Classification.
In: Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence. Chicago,
Illinois, pp 1156–1161
98. Zagheni E, Garimella VRK, Weber I, State B (2014) Inferring international and internal
migration patterns from Twitter data. In: roceedings of the 23rd International Conference
on World Wide Web Pages. ACM Press, Seoul, Korea, pp 439–444
99. Sloan L, Morgan J, Housley W, et al (2013) Knowing the Tweeters: Deriving sociologically
relevant demographics from Twitter. Sociol Res Online 18: doi 10.5153/sro.3001
100. Smith J (2014) Gender prediction in social media. ArXiv Prepr ArXiv14072147
101. Bamman D, Eisenstein J, Schnoebelen T (2014) Gender Identity and Lexical Variation in
Social Media. J Socioling 18:135–160
102. Fink C, Kopecky J, Morawski M (2017) Inferring Gender from the Content of Tweets: A
Region Specific Example. In: Proceedings of the Six International Conference on Weblogs
and Social Media. pp 459–463
103. Mandel B, Culotta A, Boulahanis J, et al (2012) A demographic Analysis of Online
Sentiment During Hurricaine Irene. In: NAACL-HLT. Association for Computational
Linguistics, Montreal, Canada, pp 27–36
104. Rao D, Yarowsky D, Shreevats A, Gupta M (2010) Classifying latent user attributes in
twitter. In: Proceedings of the 2nd international workshop on Search and mining user-
generated contents. ACM, pp 37–44
105. Vicente M, Batista F, Carvalho JP (2015) Twitter gender classification using user
unstructured information. In: 2015 IEEE International Conference on Fuzzy Systems
(FUZZ-IEEE). IEEE, Istanbul, Turkey, pp 1–7
106. Wang Y, Xiao Y, Ma C, Xiao Z (2016) Improving Users' Demographic Prediction via the
Videos They Talk about. In: Proc. of the Conference on Empirical Methods in Natural
Language Processing (EMNLP16). Austin, TX, pp 1359–1368
23
107. Hofstra B, Corten R, van Tubergen F, Ellison NB (2017) Sources of Segregation in Social
Networks: A Novel Approach Using Facebook. Am Sociol Rev 0003122417705656 . doi:
10.1177/0003122417705656
108. Mislove A, Lehmann S, Ahn Y-Y, et al (2011) Understanding the Demographics of Twitter
Users. In: Proceedings of the Fifth International AAAI Conference on Weblogs and Social
Media. Menlo Park, CA, pp 554–557
109. Rao D, Paul MJ, Fink C, et al (2011) Hierarchical Bayesian Models for Latent Attribute
Detection in Social Media. In: In Proceedings of the Fifth International Conference on
Weblogs and Social Media. ACM, Menlo Park, CA, pp 598–601
110. McCormick TH, Lee H, Cesare N, et al (2015) Using Twitter for Demographic and Social
Science Research: Tools for Data Collection and Processing. Sociol Methods Res. doi:
10.1177/0049124115605339
111. Longley PA, Adnan M, Lansley G (2015) The geotemporal demographics of Twitter usage.
Environ Plan A 47:465–484 . doi: 10.1068/a130122p
112. Hosseini M, Tammimy Z (2016) Recognizing users gender in social media using linguistic
features. Comput Hum Behav 56:192–197 . doi: 10.1016/j.chb.2015.11.049
113. Karimi F, Wagner C, Lemmerich F, et al (2016) Inferring Gender from Names on the Web:
A Comparative Evaluation of Gender Detection Methods. ArXiv160304322 Cs 53–54 . doi:
10.1145/2872518.2889385
114. Grbovic M, Radosavljevic, Vladan, Djuric (2015) Leveraging Blogging Activity on Tumblr
to Infer Demographics and Interests of Users for Advertising Purposes. In: Proceedings of
the 6th Workshop on Making Sense of Microposts. ACM, New York, NY
115. Merler M, Cao L, Smith JR (2015) You are what you tweet...pic! gender prediction based
on semantic analysis of social media images. In: 2015 IEEE International Conference on
Multimedia and Expo (ICME). pp 1–6
116. Ludu PS (2015) Inferring Latent Attributes of an Indian Twitter User Using Celebrities and
Class Influencers. In: Proceedings of the 1st ACM Workshop on Social Media World
Sensors. ACM, New York, NY, USA, pp 9–15
117. Zhong Y, Yuan NJ, Zhong W, et al (2015) You Are Where You Go: Inferring Demographic
Attributes from Location Check-ins. ACM Press, pp 295–304
118. Vijayaraghavan P, Vosoughi S, Roy D (2017) Twitter Demographic Classification Using
Deep Multi-modal Multi-task Learning. Association for Computational Linguistics, pp
478–483
119. Bhatnagar VS, Sanket NJ, Barathi SK (2016) "Honey, We Shrunk the Weights" Gender
Prediction using Twitter Feeds and Profile Images
24
120. Claude F, Konow R, Ladra S (2016) Fast Compressed-based Strategies for Author Profiling
of Social Media Texts. In: Proceedings of the 4th Spanish Conference on Information
Retrieval. ACM, New York, NY, USA, pp 14:1–14:6
121. Zhang Y, Dang Y, Chen H (2011) Gender Classification for Web Forums. IEEE Trans Syst
Man Cybern - Part Syst Hum 41:668–677 . doi: 10.1109/TSMCA.2010.2093886
122. Cesare N, Grant C, Hawkins JB, et al (2017) Demographics in Social Media Data for Public
Health Research: Does it matter? ArXiv Prepr ArXiv171011048
123. Rosenthal S, McKeown K (2011) Age prediction in blogs: A study of style, content, and
online behavior in pre-and post-social media generations. In: Proceedings of the 49th
Annual Meeting of the Association for Computational Linguistics: Human Language
Technologies-Volume 1. Association for Computational Linguistics, Stroudsburg, PA, pp
763–772
124. Sloan L, Morgan J, Burnap P, Williams M (2015) Who Tweets? Deriving the Demographic
Characteristics of Age, Occupation and Social Class from Twitter User Meta-Data. PLOS
ONE 10:e0115545 . doi: 10.1371/journal.pone.0115545
125. Nguyen T, Phung D, Adams B, Venkatesh S (2011) Prediction of Age, Sentiment, and
Connectivity from Social Media Text. In: Web Information System Engineering – WISE
2011. Springer Berlin Heidelberg, Sydney, Australia, pp 227–240
126. Oktay H, Firat A, Ertem Z (2014) Demographic breakdown of twitter users: An analysis
based on names. Acad Sci Eng 1–12
127. Chen J, Li S, Dai B, Zhou G (2016) Active Learning for Age Regression in Social Media.
In: China National Conference on Chinese Computational Linguistics. Springer, pp 351–
362
128. Chamberlain BP, Humby C, Deisenroth MP (2016) Probabilistic Inference of Twitter Users'
Age based on What They Follow. ArXiv Prepr ArXiv160104621
129. Wang C, Xiao B, Li X, et al (2013) Predicting Users' Age Range in Micro-blog Network.
In: Proceedings of the International Conference on Web Information Systems Engineering
(WISE 2013. Nanjing, China, pp 516–521
130. Zheng L, Yang K, Yu Y, Jin P (2013) Predicting Age Range of Users over Microblog
Dataset. Int J Database Theory Appl 6:85–94 . doi: 10.14257/ijdta.2013.6.6.08
131. Chang J, Rosenn I, Backstrom L, Marlow C (2010) ePluribus: Ethnicity on Social Networks.
In: Procedinngs of the Fourth International Conference on Weblogs and Social Media.
Washington DC, pp 18–25
132. Mohammady, Ehsan, Culotta, Aron (2014) Using County Demographics to Infer Attributes
of Twitter Users. Baltimore, Maryland, pp 7–16
25
133. Mateos P, Longley PA, O'Sullivan D (2011) Ethnicity and Population Structure in Personal
Naming Networks. PLOS ONE 6:e22943 . doi: 10.1371/journal.pone.0022943
134. Brownlee J (2014) Classification Accuracy is Not Enough: More Performance Measures
You Can Use. In: Mach. Learn. Mastery. http://machinelearningmastery.com/classification-
accuracy-is-not-enough-more-performance-measures-you-can-use/. Accessed 28 Jan 2017
135. Han J, Pei J, Kamber M (2011) Data Mining: Concepts and Techniques. Elsevier
136. (2009) Federal Policy for the Protection of Human Subjects ('Common Rule. In: HHS.gov.
https://www.hhs.gov/ohrp/regulations-and-policy/regulations/common-rule/index.html.
Accessed 26 Nov 2017
137. boyd danah, Crawford K (2012) Critical Questions for Big Data. Inf Commun Soc 15:662–
679 . doi: 10.1080/1369118X.2012.678878
138. Rivers CM, Lewis BL (2014) Ethical research standards in a world of big data.
F1000Research. doi: 10.12688/f1000research.3-38.v2
139. Vayena E, Salathé M, Madoff LC, Brownstein JS (2015) Ethical Challenges of Big Data in
Public Health. PLOS Comput Biol 11:e1003904 . doi: 10.1371/journal.pcbi.1003904
140. Zwitter A (2014) Big Data ethics. Big Data Soc 1:2053951714559253
10.1177/2053951714559253
. doi:
141. Ohm P (2012) Broken Promises of Privacy: Responding to the Surprising Failure of
Anonymization. UCLA Law Rev 57:1701–1777
142. Floridi L (2014) Open Data, Data Protection, and Group Privacy. Philos Technol 27:1–3 .
doi: 10.1007/s13347-014-0157-8
|
1902.10307 | 1 | 1902 | 2019-02-27T02:14:42 | Deep Adversarial Network Alignment | [
"cs.SI",
"cs.LG"
] | Network alignment, in general, seeks to discover the hidden underlying correspondence between nodes across two (or more) networks when given their network structure. However, most existing network alignment methods have added assumptions of additional constraints to guide the alignment, such as having a set of seed node-node correspondences across the networks or the existence of side-information. Instead, we seek to develop a general network alignment algorithm that makes no additional assumptions. Recently, network embedding has proven effective in many network analysis tasks, but embeddings of different networks are not aligned. Thus, we present our Deep Adversarial Network Alignment (DANA) framework that first uses deep adversarial learning to discover complex mappings for aligning the embedding distributions of the two networks. Then, using our learned mapping functions, DANA performs an efficient nearest neighbor node alignment. We perform experiments on real world datasets to show the effectiveness of our framework for first aligning the graph embedding distributions and then discovering node alignments that outperform existing methods. | cs.SI | cs | Deep Adversarial Network Alignment
Tyler Derr1 , Hamid Karimi1 , Xiaorui Liu1 , Jiejun Xu2 , Jiliang Tang1
1 Data Science and Engineering Lab, Michigan State University 2HRL Laboratories
{derrtyle, karimiha, xiaorui}@msu.edu, [email protected], [email protected]
9
1
0
2
b
e
F
7
2
]
I
S
.
s
c
[
1
v
7
0
3
0
1
.
2
0
9
1
:
v
i
X
r
a
Abstract
Network alignment, in general, seeks to discover
the hidden underlying correspondence between
nodes across two (or more) networks when given
their network structure. However, most existing
network alignment methods have added assump-
tions of additional constraints to guide the align-
ment, such as having a set of seed node-node cor-
respondences across the networks or the existence
of side-information. Instead, we seek to develop a
general network alignment algorithm that makes no
additional assumptions. Recently, network embed-
ding has proven effective in many network analysis
tasks, but embeddings of different networks are not
aligned. Thus, we present our Deep Adversarial
Network Alignment (DANA) framework that first
uses deep adversarial learning to discover complex
mappings for aligning the embedding distributions
of the two networks. Then, using our learned map-
ping functions, DANA performs an efficient nearest
neighbor node alignment. We perform experiments
on real world datasets to show the effectiveness of
our framework for first aligning the graph embed-
ding distributions and then discovering node align-
ments that outperform existing methods.
1 Introduction
In today's world, networks are arising almost everywhere
from social to biological networks. This has caused an in-
creased attention in the domain of network analysis. How-
ever, most efforts have primarily focused on single network
problems such as link prediction [Liben-Nowell and Klein-
berg, 2007] and community detection [Fortunato, 2010], but
many problems inherently are only defined when having mul-
tiple networks, such as the network alignment problem. In
general, network alignment aims to discover a set of node
pairs across two (or more) networks that we assume inher-
ently have a correspondence between their nodes. The ma-
jority of existing network alignment algorithms assume ad-
ditional constraints to guide the alignment process such as
a one-to-one mapping between the two networks [Zhang
and Philip, 2015], some seed node-node correspondences
(i.e., supervised) [Mu et al., 2016], sparsity in the possible
alignments [Bayati et al., 2013], and the existence of side-
information (e.g., node/edge attributes) [Zhang and Tong,
2016]. However, inherently these constraints limit the appli-
cations of these methods as in many cases these constraints
are not available due to many reasons such as data privacy.
Thus, this leaves the desire for an advanced algorithm that is
both unsupervised and assuming no side-information, which
brings in tremendous challenges.
In these formulations,
Without additional constraints, one key challenge to build
network alignment algorithms is the vast number of possi-
ble permutations of the node orderings to align nodes from
one network to another [Heimann et al., 2018].
Previ-
ous works have focused on utilizing the adjacency matrix,
or more recently, also leveraging spectral graph theory and
the Laplacian matrix representations [Nassar et al., 2018;
Hayhoe et al., 2018].
the main
idea is to discover the optimal permutation to map one net-
work's matrix representation to that of the other with mini-
mal variation between them. Various metrics have been de-
fined to measure the similarity between these matrices dur-
ing the optimization process [Guzzi and Milenkovi´c, 2017;
Aflalo et al., 2015]. Inherently the use of the adjacency ma-
trix is not scalable. Recently though, the field of network
embedding, which in general seeks to discover a low dimen-
sional representation of the nodes in a network, has seen
amazingly fast development with advanced methods provid-
ing huge improvement over purely spectral based methods
for single network tasks [Grover and Leskovec, 2016]. This
is primarily due to the condensed, space efficient, and even
richer low dimensional representations for the nodes of a net-
work. However, these network embedding methods are op-
timized separately for different graphs. In other words, em-
beddings of nodes from two networks are not aligned. Thus,
directly applying network embedding to advance network
alignment still is immensely challenging.
Meanwhile,
there have been adversarial based meth-
ods [Goodfellow et al., 2014; Isola et al., 2017; Yu et al.,
2017; Wang et al., 2017] that harness the power of deep learn-
ing for solving a variety of unsupervised problems by using
a minimax game between a generator and a discriminator.
In these adversarial based methods, the generator is trained
to attempt at "fooling" the discriminator that it is generating
"real" (and not "generated") examples while the discrimina-
tor is also trained to get better at differentiating between the
"real" and "generated" examples. This process allows for an
unsupervised way of learning a generator that can generate
examples that seemingly come from the same distribution of
the real data. These adversarial techniques have shown to
be useful in a plethora of domains including computer vi-
sion [Isola et al., 2017], natural language processing [Yu et
al., 2017], and recommendation [Wang et al., 2017].
On the one hand, network embedding algorithms have been
proven to be effective in learning representations for nodes,
but embeddings for two networks are learned separately,
which are not aligned. On the other hand, adversarial tech-
niques are powerful in learning real data distributions. Thus,
in this work, we propose to harness the power of network em-
bedding and adversarial techniques to tackle the challenging
network alignment problem without additional constraints or
knowledge outside of the network structure. The rationale is
that we can align the node representations of two networks
by taking advantage of adversarial techniques. More specif-
ically, the proposed novel Deep Adversarial Network Align-
ment (DANA) framework is composed of two stages -- one
graph distribution alignment stage and one node alignment
stage. In the graph distribution alignment stage, we utilize
deep neural networks in an adversarial framework that is able
to learn a highly complex mapping from one network's em-
bedding space to that of the other such that the mapped em-
bedding approximates the data distribution of the other net-
work's original embedding. In the node alignment stage, we
align individual nodes from two networks by using the map-
ping functions learned from the graph distribution alignment
stage. Our main contributions are as follows:
• We propose a novel unsupervised Deep Adversarial Net-
work Alignment (DANA) framework that utilizes the
power of both network embedding and adversarial train-
ing techniques to align the embedding distributions and
then perform an efficient node alignment thereafter;
• We provide an unsupervised heuristic to perform model
selection for DANA, which also simultaneously shows
the effectiveness of aligning the embedding distribu-
tions; and
• Experimental results on various datasets show the supe-
riority of DANA against numerous advanced baselines.
2 Problem Definition
In this section, we introduce the basic notations and problem
definition. First, we let N 1 = (V 1,E 1) and N 2 = (V 2,E 2)
} and
be two undirected networks with V 1 = {v1
} being sets of n1 and n2 vertices, and
V 2 = {v2
edge sets E 1 and E 2 for networks N 1 and N 2, respectively.
Now, with the aforementioned notations, we formally de-
fine the network alignment problem we want to study in this
work as follows:
2, . . . v2
n2
2, . . . v1
n1
1, v1
1, v2
Given two networks N 1 and N 2, and under the assump-
tion that there is an underlying correspondence between the
vertices V 1 and V 2, we seek to discover a set A of vertex
alignment pairs defined as:
A = {(v1, v2) v2 ∈ V 2,∀v1 ∈ V 1}
(1)
Figure 1: An illustration of our graph distribution alignment.
where for each vertex v1 in N 1 we predict a single corre-
sponding vertex v2 in N 2, such that together these pairwise
node alignments follows a global network alignment.
Actually we will solve this problem bidirectionally to align
the nodes of N 1 to N 2 and vice versa. Furthermore, we stress
that in our unsupervised setting, we do not have any known
node-node labeled correspondences nor any side-information
(such as node/edge attributes). Instead, our proposed frame-
work only requires the network structures, but could be ex-
tended to embrace such additional information (later dis-
cussed as future work).
3 Deep Adversarial Network Alignment
Framework
In this section we introduce our proposed framework, Deep
Adversarial Network Alignment (DANA), for the network
alignment problem discussed in Section 2. First, we will pro-
vide an overview of how the framework is utilized to solve
the network alignment problem by first aligning the embed-
ding distributions and then aligning the nodes. Next, we will
discuss in detail both of these key stages of our proposed
framework. Thereafter we summarize with an algorithmic
overview of our framework and also provide an analysis on
the complexity of DANA.
As previously mentioned, network embedding algorithms
have been proven to effectively learn node representations,
but embeddings for separate networks are not aligned. Thus,
we first obtain node embeddings and then use an adversarial
based method to correctly learn a complex (and even non-
linear) mapping to simultaneously align the two networks
embedding distributions. In this way the mapped embedding
from the first network approximately follows the distribution
of the other. Then, once the distributions have been aligned
the second stage uses an efficient nearest neighbor search to
match/align the individual nodes by using the mapping func-
tions obtained through the adversarial learning.
3.1 Adversarial Graph Distribution Alignment
In this section, we introduce the first stage of DANA, namely
the distribution alignment whose model is illustrated in Fig-
ure 1.
this model aligns two graphs N 1
In a nutshell,
GraphEmbeddingN1G1 2G2 1RealGeneratedN2GraphEmbeddingCycle Consistency RealGeneratedD2RealGeneratedD1RealGeneratedX1X2 and N 2 bidirectionally using two connected adversarial net-
works. The reason for connecting them is to ensure no "col-
lapse" [Zhu et al., 2017] and utilize transitivity [Zhou et al.,
2016] to regularize and prevent random alignments of the dis-
tributions which could be possible due to using complex (and
even non-linear) mappings between them.
In Figure 1, we can observe that when given two net-
works N 1 and N 2 the first step is to obtain graph embed-
dings for each of these networks. This can be obtained us-
ing one of the plethora of available methodologies, such as
node2vec [Grover and Leskovec, 2016]. Then, our goal is to
i}V 1
find a mapping between node embeddings X1 = {x1
j}V 2
for N 1 and X2 = {x2
j=1 for N 2 whose distributions are
denoted as x1 ∼ pemb(x1) and x2 ∼ pemb(x2), respec-
tively (since we are ultimately looking to align the two graph
distributions). As demonstrated in Figure 1, the model con-
tains two generators G1→2 mapping X1 → X2 and G2→1
mapping X2 → X1. Moreover, the discriminator D1 dis-
tinguishes between real embeddings of graph N 1 and those
generated from the real embedding of N 2 through G2→1 (i.e.,
{x1} and {G2→1(x2)}, respectively). Likewise, the discrim-
inator D2 distinguishes between real embeddings of N 1 and
those generated through G1→2 from the real embedding of
N 2.
Both bidirectional mappings are optimized via applying
adversarial losses [Goodfellow et al., 2014]. More precisely,
the loss function associated with aligning graph N 1 to N 2
(i.e., the mapping G1→2 : X1 → X2) is as follows:
i=1
min
G1→2
max
D2
L1→2
adv (G1→2, D2, X1, X2)
= Ex2∼pemb(x2)[log D2(x2)]
+ Ex1∼pemb(x1)[log (1 − D2(G1→2(x1)))]
(2)
where the generator G1→2 is trying to mimic the embedding
distribution of N 2 by generating embeddings G1→2(x1) (by
minimizing Eq. (2)) while simultaneously the discriminator
D2 is attempting to differentiate between x2 and G1→2(x1)
(by maximizing Eq. (2)). Similarly, the loss function of align-
ing graph N 2 to N 1 (i.e., G2→1 : X2 → X1) is as follows:
(3)
L2→1
adv (G2→1, D1, X1, X2)
= Ex1∼pemb(x1)[log D1(x1)]
+ Ex2∼pemb(x2)[log (1 − D1(G2→1(x2)))]
min
G2→1
max
D1
where in this situation the minimax game is instead between
G2→1 and D1.
By separately optimizing the loss functions in Eq. (2) and
Eq. (3) (i.e., of learning the mappings G1→2 and G2→1), we
might expect to learn an alignment between the embeddings
of N 1 to N 2, and vice versa. However, in practice, during
the training each of the separate models can map one real em-
bedding distribution to some random embeddings in the target
domain (or even collapse). More specifically, especially when
non-linearity is used in the generators, the mapping G1→2
can project embeddings of graph N 1 to some random points
in the embedding space of N 2, that although might have a
similar distribution, might also have completely distorted the
Figure 2: An illustration of our node alignment from N 1 to N 2.
proximity information between neighboring nodes that was
originally preserved in X1. In other domains, such as word-
to-word translation, adversarial techniques resorted to using
only a single directional linear mapping [Lample et al., 2018]
that could avoid these problems, but limited the complexity
and power of non-linearity in their translation/alignment.
To prevent these problems and still potentially use the
power of a non-linear generator mapping function for net-
work alignment, we introduce the cycle consistency loss sim-
ilar to [Zhu et al., 2017], which had been used for image-to-
image translation. More specifically, for a node embedding
x1 of node v1 ∈ V 1, the learned generators G1→2 and G2→1
should be able to recover and bring x1 back to the embedding
space of N 1 as follows:
x1 → G1→2(x1) → G2→1(G1→2(x1)) ≈ x1
(4)
and similarly for being able to recover the embeddings of
N 2. Intuitively, if forcing these cyclic mappings, then this
would help to prevent both the "collapse" and random align-
ment problem previously mentioned. Hence, we incorporate
the following cycle-reconstruction loss into our objective:
min
Lcyc(G1→2, G2→1, X1, X2)
(cid:104)G2→1(G1→2(x1)) − x11
(cid:105)
(cid:105)
(cid:104)G1→2(G2→1(x2)) − x21
G2→1,G1→2
= Ex1∈pemb(x1)
+ Ex2∈pemb(x2)
(5)
This leads to the graph level embedding distribution align-
(6)
max
D1,D2
adv (G2→1, D1, X1, X2)
ment to optimize the following overall loss function:
L(G1→2, G2→1, D1, D2, X1, X2)
min
G1→2,G2→1
= L1→2
adv (G1→2, D2, X1, X2) + L2→1
+ λLcyc(G1→2, G2→1, X1, X2)
where λ is a hyper-parameter controlling the balance between
ensuring a close aligning of the graph level embedding distri-
butions and the cycle consistency loss.
3.2 Nearest Neighbor Node Alignment
The second stage of DANA is to efficiently discover the node
alignments, which are based on using the discovered complex
mapping functions (i.e., the generators) from the first stage of
DANA. In this subsection, we will discuss the efficient near-
est neighbor greedy node alignment method from N 1 to N 2,
where we can then similarly perform N 2 to N 1.
As seen in Figure 2, the first step is to take the node em-
beddings X1 and map them to the embedding space of N 2
through the use of the trained generator G1→2. Next these
G1 2X1G1 2 (X1)X2 projected node embeddings of N 1 are paired with their near-
est neighbor from N 2 based on Euclidean distance. To per-
form the nearest neighbor search, we utilize a k-d tree, which
is a data structure used for performing a fast and efficient
search. [Abbasifard et al., 2014].
3.3 Algorithmic Overview and Complexity Analysis
Here we discuss an algorithmic overview of DANA along
with the computational complexity. Algorithm 1 summarizes
the entire framework including the major steps -- namely ob-
taining network embeddings (line 2), training the unsuper-
vised adversarial based graph distribution alignment (lines 4-
11), and performing nearest neighbor node alignment (lines
13-15). Next we discuss some details and the complexity of
DANA. Note that we denote nmax = max(n1, n2), and sim-
ilarly define nmin, where n1 = V 1 and n2 = V 2.
First, we obtain the embeddings for networks N 1 and N 2.
In this work, we utilize a network embedding method (more
specifically node2vec [Grover and Leskovec, 2016]) whose
complexity is O(n1) and O(n2) for networks N 1 and N 2,
respectively, resulting in overall O(nmax). Note that DANA
could use attributed embedding methods to incorporate side-
information, but we leave this as future work.
Next, we train the adversarial based graph distribution
alignment. Suppose that the algorithm is run for some con-
stant number of epochs, K, where each epoch iterates through
all the nodes (for both graphs) by randomly creating mini-
batches that perform the forward step, backpropagating error
and also updating the parameters using stochastic gradient
descent (SGD). Note that this depends on the architectures
for the generators and discriminators. However, as later dis-
cussed in Section 4, if we use small reasonable constant size
hidden layers, then the computation is also constant for each
mini-batch. Thus, due to the fact we run K epochs, the com-
plexity of the graph distribution alignment is O(nmax).
As seen in Algorithm 1 (lines 13-15), for the final step
we actually perform the node alignment bidirectionally and
select the one that has the lower average nearest neighbor
distance. First, we build the k-d tree based on the embed-
dings X2, which takes O(n2log(n2)) and then we need to
search for the nearest neighbor for all v1 ∈ V 1 using their
mapped representation G1→2(x1). The search in the worse
case for each x1 is O(n2), but O(log(n2)) in the average
case. Thus, since we perform this search for all nodes in V 1,
the total expected time is O(n2log(n2) + n1log(n2)) for the
alignment of N 1 onto N 2. Then we similarly do the align-
ment of N 2 onto N 1, resulting in the total expected time
O(nmaxlog(nmax)). This leads to the sub-quadratic total
time complexity of DANA to be O(nmaxlog(nmax)) when
ignoring the linear/constant terms.
4 Experiments
To evaluate the effectiveness of the proposed Deep Adver-
sarial Network Alignment framework, we conduct a set of
experiments for aligning real world networks. Through the
conducted experiments, we seek to answer the following two
questions: (1) Can DANA align network embeddings? and
(2) How effective is DANA at accurately discovering the true
underlying node alignment?
Algorithm 1: Deep Adversarial Network Alignment.
Input: N 1 = (V 1,E 1),N 2 = (V 2,E 2)
Output: A
1 # Utilize a Graph Embedding Method:
2 Obtain X1 and X2 from N 1 and N 2, respectively
3 # Perform Graph Distribution Alignment
4 Randomly initialize the neural network parameters (θ)
5 while Not max iterations do
6
Select a random mini-batch
B = {(x1, x2)x1 ∈ X1, x2 ∈ X2}
foreach (x1, x2) ∈ B do
7
8
Feed forward to obtain G1→2(x1), G1→2(x2),
D2(G1→2(x1)), D1(G2→1(x2)), D1(x1), D2(x2),
G2→1(G1→2(x1)), and G1→2(G2→1(x2))
end
Update parameters θ in Eq. 6 using SGD
9
10
11 end
12 # Perform Node Alignment
13 Construct A1: closest x2 ∈ k-d tree(X2) ∀x1 ∈ G1→2(X1)
14 Construct A2: closest x1 ∈ k-d tree(X1) ∀x2 ∈ G2→1(X2)
15 Set A to A1 or A2 based on mean nearest neighbor distance
4.1 Experimental Setup
Here we will discuss the datasets and how we utilize them for
our experiments, the architecture details for DANA, our pro-
posed unsupervised heuristic for model selection, and base-
line methods.
Datasets with Ground Truth Correspondence
The first two datasets we collected are Bitcoin-Alpha1 and
Bitcoin-OTC2. These networks are online marketplaces that
allow users to buy and sell things using Bitcoins. Users cre-
ate positive (or negative) links to those they trust (or distrust).
Furthermore, most users have provided their unique Bitcoin
fingerprints, thus allowing us to determine a ground-truth
mapping of users across networks, which we use to evaluate
the alignments. Note that we construct two undirected dataset
variants, the first being networks that only include positive
links (i.e., BitcoinA and BitcoinO), and the second that also
include the negative links (i.e., BitcoinAn and BitcoinOn).
Some basic statistics can be found in Table 1.
Datasets with Pseudo-Ground Truth Correspondence
Here we collected real world datasets, namely CollegeMsg3,
Hamsterster4, and Blogs4. We present the basic undirected
network statistics in Table 1. Note that we have chosen these
datasets to be used in a synthetic network alignment setting
(although themselves are real world networks), where we will
let N2 to be the original dataset, while constructing a per-
muted version as N1 and simultaneously adding some noise
in the set {5,10,20}% to N1 at random to evaluate the perfor-
mance and robustness of DANA.
1http://www.btcalpha.com
2http://www.bitcoin-otc.com
3http://snap.stanford.edu/data/
4http://konect.uni-koblenz.de/
V 1
V 2
E 1
E 2
V 1 ∩ V 2
E 1 ∩ E 2
BitcoinA
BitcoinO
(pos)
3682
3819
3591
25952
28321
24066
BitcoinAn
BitcoinOn
(pos&neg)
3783
3914
3682
28288
30691
26004
CollegeMsg Hamsterster Blogs
1899
-
-
13754
-
-
2426
-
-
16613
-
-
1224
-
-
16718
-
-
Table 1: Dataset Statistics.
DANA Architecture
First, for the graph embeddings, we utilize node2vec [Grover
and Leskovec, 2016] to obtain node embeddings of size 64.
We note that both the generator and discriminator for DANA
can be constructed in various ways. For the discriminator,
based on knowledge from other domains when using adver-
sarial based frameworks, we use a two layer fully connected
network with 512 hidden units and Leaky ReLU (Rectified
Linear Unit) [Maas et al., 2013]. For the generators, we at-
tempted using both a linear and non-linear single layer map-
ping (i.e., two possible variants). For the adversarial learn-
ing, we varied λ in {1, 10, 100}, since λ = 10 was recom-
mended in [Zhu et al., 2017]. We let η ∈ {1, 5, 25} denote
the number of times we update the generators before updat-
ing the discriminators during the alternative updating and use
the ADAM optimizer [Kingma and Ba, 2014].
In the above we mentioned three hyperparameters that we
would need to choose between, thus, we propose to use the
unsupervised heuristic of how well DANA aligns the distri-
butions as a metric to select the best parameters. We assume
the better the embedding distributions match, the better the
performance in aligning the individual nodes. Note that this
does not utilize the ground-truth alignments, but rather sim-
ply measures the average distance each mapped node is to
the nearest neighbor in the other embedded space.
Baselines
Here we introduce the set of baselines we will compare
against our proposed Deep Adversarial Network Alignment
(DANA) method. Isorank [Singh et al., 2008]/FINAL [Zhang
and Tong, 2016] is a network alignment algorithm that was
designed specifically for protein-protein interaction network
alignments and we note that FINAL is an attributed network
alignment methods that is equivalent to IsoRank when hav-
ing no edge/node attributes [Zhang and Tong, 2016]. Eige-
nAlignLR [Nassar et al., 2018] is a low rank extension of the
EigenAlign [Feizi et al., 2016] spectral based network align-
ment method. REGAL [Heimann et al., 2018] constructs
their own network embedding (while REGAL-s2v [Heimann
et al., 2018] instead uses struc2vec [Ribeiro et al., 2017]) and
then uses a nearest neighbor search for node alignments. We
also include two sparse network alignment methods, namely
SparseIsoRank [Bayati et al., 2013] (a sparse network align-
ment variation of IsoRank [Singh et al., 2008]) and NetAl-
ignBP [Bayati et al., 2013] (which uses belief propagation
to construct the alignment), where both use information for
limiting the scope of possible alignments.
We note that SparseIsoRank and NetAlignBP both assume
Figure 3: Answering question (1) where (a) shows the average near-
est neighbor distance heuristic with the accuracy for BitcoinAn and
BitcoinOn, and (b) visualizes the graph distribution alignment.
additional information to suggest to their methods which
node alignments are possible, but our problem setting does
not have such additional information. Therefore we heuris-
tically provide them information from the network structure.
More specifically, node degree similarity from N 1 and N 2
is used to provide each node in the two networks a subset
of possible nodes to pair with (as also done in [Heimann et
al., 2018]). Here we try using log2(V) and 0.01 × V for
the the set size containing the most similar nodes in terms of
absolute difference in degree. We note that for all methods
we use the default settings provided by the authors, but RE-
GAL was unable to run on networks of different sizes, thus
we only report their performance for the CollegeMsg, Blogs,
and Hamsterster datasets.
4.2 Experimental Results
Here we present the results of our experiments in network
alignment with both our known ground truth and pseudo-
ground truth datasets. Although there are multiple ways
of evaluating the performance of network alignment meth-
ods [Douglas et al., 2018], we report the accuracy, which is
by far the most commonly used.
In Figure 3(a), we have plotted the mean nearest neighbor
distances against the accuracy while taking snapshots dur-
ing DANA's optimization of the adversarial graph distribu-
tion alignment of the BitcoinAn and BitcoinOn datasets. We
can observe that as DANA better aligns the two distributions
(i.e., lower mean nearest neighbor distance) the accuracy is
also improving. In Figure 3(a), the star represents the least
mean nearest neighbor distance and we can observe it nearly
has the best accuracy. We observe the same trend across all
hyper-parameter settings and thus our unsupervised heuristic
of using the lowest mean nearest neighbor distance for model
selection works quite well in practice. To further show the
effectiveness of DANA in aligning the graph embedding dis-
tributions, we show a visualization using principle component
analysis (PCA) [Jolliffe, 2011] in Figure 3(b) where "Start"
refers to the initial random alignment and "End" shows the
final graph distribution alignment that DANA adversarially
learns. Therefore, based on Figure 3, we can conclude an an-
swer for our first question, that DANA can indeed learn to
effectively align the embeddings of two networks.
In Table 2, we ran DANA and the baselines (except for
G1 2(X1) to X2G2 1(X2) to X1StartEndX2G1 2(X1)(a)(b)Accuracy Methods
N 1: BitcoinA
N 2: BitcoinO
N 1: BitcoinAn
N 2: BitcoinOn
SparseIsoRank
NetAlignBP
IsoRank/FINAL
EigenAlignLR
REGAL-s2v
DANA
0.046
0.157
0.041
0.015
0.124
0.542
0.047
0.141
0.040
0.016
0.089
0.511
Table 2: Performance comparison with accuracy for aligning the two
variations of the Bitcoin datasets.
REGAL, since their implementation could not handle net-
works having different sizes) on the Bitcoin datasets. The
first observation is that the EigenAlignLR, SparseIsoRank,
and Isorank/FIANL all have significantly less performance
than NetAlignBP, REGAL-s2v, and DANA. It would seem
that NetAlignBP is able to effectively use the pseudo side-
information (based on node degree similarity) we provided.
Also, REGAL-s2v (which uses struc2vec embeddings) is
able to out perform EigenAlignLR (which is spectral based).
However, we also note the comparison of REGAL-s2v, which
directly performs a node alignment on embeddings from
struc2vec, against DANA, which uses adversarial learning
to correctly align network embeddings before performing the
node alignment. We can clearly see that DANA significantly
out performs REGAL-s2v and all other baseline methods.
Next, in Figure 4, we present the results for the pseudo-
ground truth datasets where we have performed the network
alignment experiments for the CollegeMsg, Hamsterster, and
Blogs datasets. In these experiments we first permuted the
nodes of the original network and then removed a portion of
the edges (i.e., level of noise) and attempted to align back to
the original network. We can observe that similar to the two
Bitcoin experiments, DANA is able to outperform all exist-
ing baseline methods for all three datasets across all levels of
noise. We also observe that as more edges are removed, it
becomes harder to align back to the original network where
all the baselines almost completely fail to align at 20%, but
yet DANA is able to still maintain a reasonable alignment.
It can also be seen that DANA and REGAL (also REGAL-
s2v in many cases) outperform the other methods, which sug-
gests again that embedding based approaches are superior.
However, as previously mentioned, REGAL-s2v does not
perform any alignment of the embeddings before perform-
ing the node alignment, and REGAL performs a similarity-
based embedding alignment through a shallow matrix factor-
ization method, but neither harness deep learning or adversar-
ial training. Furthermore, while EigenAlignLR uses spectral
based embeddings, DANA is able to harness more advanced
network embedding methods, while leading to better perfor-
mance across all levels of noise. This seems natural due to
the fact that current network embedding methods have shown
superiority over the classical spectral embeddings for a vari-
ety of network analysis tasks [Grover and Leskovec, 2016].
Thus, based on these experiments we have answered the sec-
ond question, DANA is indeed effective at aligning the cor-
responding nodes across networks, which is due to the fact it
harnesses the power of deep adversarial learning to correctly
align the embeddings of two networks.
Figure 4: Results on the three datasets with pseudo-ground truth
when varying the level of noise (i.e., percent of edges removed).
5 Related Work
Network alignment is a fundamental network analysis task
having many real world applications in user-identity link-
age [Liu et al., 2013], computer vision [Conte et al., 2004],
and bioinformatics [Singh et al., 2008]. Classical network
alignment methodologies typically were based around opti-
mizing a permutation matrix to align the matrix representa-
tions. However, some methods have introduced relaxations
such as convex or finding a doubly stochastic matrix instead
of finding a permutation matrix [Aflalo et al., 2015].
Another set of related network alignment problems are
those that are supervised. Some representative examples also
learn network embeddings, but use known node-node pairs
to align in a shared embedding space [Tan et al., 2014].
Also, there is the sparse network alignment problem [Bay-
ati et al., 2013] where the general alignment problem is sim-
plified to restrict the possible alignments between nodes. In
other words, a bipartite graph is constructed from the two net-
works being aligned, but rather than having V 1×V 2 num-
ber of possible matching pairs, they instead have a limited set
to prevent certain pairs, which could be from domain specific
knowledge or network structure. Some other specialized for-
mulations can be found for heterogeneous networks [Kong et
al., 2013] and attributed graphs [Zhang and Tong, 2016].
6 Conclusion
In this work, we proposed our Deep Adversarial Network
Alignment (DANA) framework to solve the general network
alignment problem when only provided the network struc-
ture and assuming no additional constraints. More specifi-
cally, DANA harnesses the power of adversarial learning to
align the graph embedding distributions and then thereafter
performs an efficient nearest neighbor node alignment. Fur-
thermore, we present an unsupervised heuristic to perform
model selection for DANA. Finally extensive experiments
were performed to show the effectiveness of both main stages
of DANA, while also proving DANA to be superior in per-
formance against existing network alignment methods. Our
future work consists of first extending DANA to embrace ad-
ditional constraints to aid in performing alignments, such as
node/edge attributes or assuming a seed set of known node-
node aligned pairs.
(a) CollegeMsg(b) Hamsterster(c) Blogs References
[Abbasifard et al., 2014] Mohammad Reza Abbasifard, Bi-
jan Ghahremani, and Hassan Naderi. Article: A survey
on nearest neighbor search methods. IJCA, 2014.
[Aflalo et al., 2015] Yonathan Aflalo, Alexander Bronstein,
and Ron Kimmel. On convex relaxation of graph isomor-
phism. PNAS, 2015.
[Bayati et al., 2013] Mohsen Bayati, David F Gleich, Amin
Saberi, and Ying Wang. Message-passing algorithms for
sparse network alignment. TKDD, 2013.
[Conte et al., 2004] Donatello Conte, Pasquale Foggia,
Carlo Sansone, and Mario Vento. Thirty years of graph
matching in pattern recognition. IJPRAI, 2004.
[Douglas et al., 2018] Joel Douglas, Ben Zimmerman,
Alexei Kopylov, Jiejun Xu, Daniel Sussman, , and Vince
Lyzinski. Metrics for evaluating network alignment.
In
GTA3 at WSDM, 2018.
[Feizi et al., 2016] Soheil Feizi, Gerald Quon, Mariana
Recamonde-Mendoza, Muriel M´edard, Manolis Kellis,
and Ali Jadbabaie. Spectral alignment of networks. arXiv
preprint arXiv:1602.04181, 2016.
[Fortunato, 2010] Santo Fortunato. Community detection in
graphs. Physics Reports, 2010.
[Goodfellow et al., 2014] Ian Goodfellow,
Jean Pouget-
Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen-
erative adversarial nets. In NIPS. 2014.
and
[Grover and Leskovec, 2016] Aditya Grover
Jure
Node2vec: Scalable feature learning for
Leskovec.
networks. SIGKDD, 2016.
[Guzzi and Milenkovi´c, 2017] Pietro Hiram Guzzi and Ti-
jana Milenkovi´c. Survey of local and global biological
network alignment: the need to reconcile the two sides of
the same coin. Briefings in bioinformatics, 2017.
[Hayhoe et al., 2018] Mikhail Hayhoe, Francisco Barreras,
Hamed Hassani, and Victor M Preciado. Spectre: Seedless
network alignment via spectral centralities. arXiv preprint
arXiv:1811.01056, 2018.
[Heimann et al., 2018] Mark Heimann, Haoming Shen, Tara
Safavi, and Danai Koutra. Regal: Representation learning-
based graph alignment. In CIKM, 2018.
[Isola et al., 2017] P. Isola, J. Zhu, T. Zhou, and A. A. Efros.
Image-to-image translation with conditional adversarial
networks. In CVPR, 2017.
[Jolliffe, 2011] Ian Jolliffe. Principal component analysis. In
International encyclopedia of statistical science. 2011.
[Kingma and Ba, 2014] Diederik P Kingma and Jimmy Ba.
arXiv
Adam: A method for stochastic optimization.
preprint arXiv:1412.6980, 2014.
[Kong et al., 2013] Xiangnan Kong,
Jiawei Zhang, and
Philip S. Yu. Inferring anchor links across multiple het-
erogeneous social networks. In CIKM, 2013.
[Lample et al., 2018] Guillaume Lample, Alexis Conneau,
Marc'Aurelio Ranzato, Ludovic Denoyer, and Herv´e
J´egou. Word translation without parallel data. In ICLR,
2018.
[Liben-Nowell and Kleinberg, 2007] David Liben-Nowell
and Jon Kleinberg. The link-prediction problem for social
networks. JAIST, 2007.
[Liu et al., 2013] Jing Liu, Fan Zhang, Xinying Song,
Young-In Song, Chin-Yew Lin, and Hsiao-Wuen Hon.
What's in a name?: An unsupervised approach to link
users across communities. WSDM, 2013.
[Maas et al., 2013] Andrew L Maas, Awni Y Hannun, and
Andrew Y Ng. Rectifier nonlinearities improve neural net-
work acoustic models. In DLASLP at ICML, 2013.
[Mu et al., 2016] Xin Mu, Feida Zhu, Ee-Peng Lim, Jing
Xiao, Jianzong Wang, and Zhi-Hua Zhou. User identity
linkage by latent user space modelling. In SIGKDD, 2016.
[Nassar et al., 2018] Huda Nassar, Nate Veldt, Shahin Mo-
hammadi, Ananth Grama, and David F. Gleich. Low rank
spectral network alignment. WWW, 2018.
[Ribeiro et al., 2017] Leonardo F.R. Ribeiro, Pedro H.P.
Saverese, and Daniel R. Figueiredo. Struc2vec: Learning
node representations from structural identity. SIGKDD,
2017.
[Singh et al., 2008] Rohit Singh, Jinbo Xu, and Bonnie
Berger. Global alignment of multiple protein interaction
networks with application to functional orthology detec-
tion. PNAS, 2008.
[Tan et al., 2014] Shulong Tan, Ziyu Guan, Deng Cai,
Xuzhen Qin, Jiajun Bu, and Chun Chen. Mapping users
across networks by manifold alignment on hypergraph.
AAAI, 2014.
[Wang et al., 2017] Jun Wang, Lantao Yu, Weinan Zhang,
Yu Gong, Yinghui Xu, Benyou Wang, Peng Zhang, and
Dell Zhang. Irgan: A minimax game for unifying gener-
ative and discriminative information retrieval models. SI-
GIR, 2017.
[Yu et al., 2017] Lantao Yu, Weinan Zhang, Jun Wang, and
Yong Yu. Seqgan: Sequence generative adversarial nets
with policy gradient. AAAI, 2017.
[Zhang and Philip, 2015] Jiawei Zhang and S Yu Philip.
In
Multiple anonymized social networks alignment.
ICDM, 2015.
[Zhou et al., 2016] Tinghui Zhou,
[Zhang and Tong, 2016] Si Zhang and Hanghang Tong. Fi-
nal: Fast attributed network alignment. In SIGKDD, 2016.
Philipp Krahenbuhl,
Mathieu Aubry, Qixing Huang, and Alexei A Efros.
Learning dense correspondence via 3d-guided cycle
consistency. In CVPR, 2016.
[Zhu et al., 2017] Jun-Yan Zhu, Taesung Park, Phillip Isola,
and Alexei A Efros. Unpaired image-to-image transla-
tion using cycle-consistent adversarial networks. In ICCV,
2017.
|
1608.06007 | 2 | 1608 | 2016-12-28T02:17:22 | Distributed Probabilistic Bisection Search using Social Learning | [
"cs.SI",
"cs.LG",
"cs.MA"
] | We present a novel distributed probabilistic bisection algorithm using social learning with application to target localization. Each agent in the network first constructs a query about the target based on its local information and obtains a noisy response. Agents then perform a Bayesian update of their beliefs followed by an averaging of the log beliefs over local neighborhoods. This two stage algorithm consisting of repeated querying and averaging runs until convergence. We derive bounds on the rate of convergence of the beliefs at the correct target location. Numerical simulations show that our method outperforms current state of the art methods. | cs.SI | cs | DISTRIBUTED PROBABILISTIC BISECTION SEARCH USING SOCIAL LEARNING
Athanasios Tsiligkaridis
Theodoros Tsiligkaridis
Boston University, ECE Dept.
8 St. Mary's St., Boston, MA, USA
[email protected]
MIT Lincoln Laboratory
244 Wood St., Lexington, MA, USA
[email protected]
6
1
0
2
c
e
D
8
2
]
I
S
.
s
c
[
2
v
7
0
0
6
0
.
8
0
6
1
:
v
i
X
r
a
ABSTRACT
We present a novel distributed probabilistic bisection algo-
rithm using social learning with application to target local-
ization. Each agent in the network first constructs a query
about the target based on its local information and obtains a
noisy response. Agents then perform a Bayesian update of
their beliefs followed by an averaging of the log beliefs over
local neighborhoods. This two stage algorithm consisting of
repeated querying and averaging runs until convergence. We
derive bounds on the rate of convergence of the beliefs at the
correct target location. Numerical simulations show that our
method outperforms current state of the art methods.
Index Terms -- Probabilistic bisection, consensus, decen-
tralized estimation, convergence rate, belief sharing
1. INTRODUCTION
We consider the problem of distributed bisection search us-
ing a network of agents. This problem has applications to
stochastic root finding [1], distributed group testing [2], ob-
ject tracking using cameras [3], and human-aided localization
[4]. The agents are connected through a topology and sequen-
tially search for the unknown target X ∗. Each agent forms a
query Zi,t = I(X ∗ ∈ Si,t) for some subset Si,t ⊂ X based
on its local information about the target location and obtains
a noisy response, Yi,t+1, which it uses to update its belief.
After this stage, the agents average their log-beliefs with their
neighbors. As this process is repeated after several iterations,
the agents converge to the correct consensus.
Prior work on distributed signal processing includes the
consensus literature [5, 6]. Extensions to consensus-plus-
innovations algorithms with applications to detection and es-
timation problems include nodes making noisy observations
and implementing a consensus method to spread information
around the network. Several such works on distributed es-
timation include that of [7, 8] which show that distributed
estimators are consistent and asymptotically normal. Addi-
tionally, [9, 10] studied the problem of distributed detection
and proved convergence rates on the rate of learning the
correct hypothesis.
The adaptive querying and Bayesian updating are known
in the literature as the Probabilistic Bisection Algorithm
(PBA). The PBA was first introduced by Horstein [11]. The
PBA was shown to be optimal in an information-theoretic
sense in the work by Jedynak, et al.
[12], and the conver-
gence rate of the single-agent PBA was shown to be exponen-
tial in Waeber, et al, [13]. The PBA was generalized in [4]
to multiple agents using a centralized controller strategy, and
in [14, 15] using a decentralized algorithm based on belief
consensus. In [4, 14, 15], the convergence analysis showed
that all agents reach consensus to the true target.
The contribution of this paper is to propose and analyze
a distributed bisection algorithm using a social learning ap-
proach. First, we first derive a social learning algorithm (in-
spired by the work of Lalitha, et al, [9]). Second, we derive an
asymptotic performance bound that characterizes the rate of
learning in terms of the social influence of nodes on the net-
work and the error probability of each agent. Finally, we show
using simulations, that our proposed algorithm outperforms
the case of no collaboration and improves the distributed bi-
section search algorithm presented in Tsiligkaridis, et al, [14].
2. PROBLEM FORMULATION
For concreteness, we focus on the one-dimensional case, i.e,
X = [0, 1]. In this case, the query sets Si,t take the form
of an interval [0, Xi,t], where Xi,t are the query points. The
response to the query Zi,t = I(X ∗ ∈ Si,t) is modeled as a
binary symmetric channel [12, 4, 14] and is given by:
Yi,t+1 =(Zi,t
w.p. 1 − ǫi
1 − Zi,t w.p. ǫi
In this paper, the error probabilities are constant. In [4], a
more general model was considered where the error probabil-
ity increased as the target localization error decreased.
corresponding CDF as Fi,t(x) = R x
We define pi,t(x) as the posterior distribution on the tar-
get space X = [0, 1] for agent i at time t. We also denote the
0 pi,t(u)du. The poste-
rior distributions at time t are measurable with respect to the
noisy responses {Yi,τ +1}τ ≤t up to time t. The proposed so-
cial learning algorithm consists of two stages (see Algorithm
1). In the first stage, agents perform a probabilistic bisection
and update their beliefs using the noisy response. In the sec-
ond stage, the log-beliefs are averaged among local neighbors
to spread information around the network.
The observation probability mass function used in the first
stage is given by:
li(x, yi,t+1) = f (i)
1 (y)I(x ≤ Xi,t) + f (i)
0 (y)I(x > Xi,t)
(1)
1 (y).
where f (i)
1
1 (y) = (1−ǫi)I(y=1)ǫI(y=0)
i
, f (i)
0 (y) = 1−f (i)
In the second stage, we assume that the social interaction
matrix A is a stochastic matrix corresponding to a strongly
connected, aperiodic graph G = (N , E). Here, N represents
the nodes of the network and E is the set of edges. An edge
(i, j) ∈ E exists iff Ai,j > 0.
Algorithm 1 Distributed Probabilistic Bisection Algorithm
1: Input: G = (N , E), A = {Ai,j}, {ǫi}
2: Output: { Xi,t : i ∈ N }
3: Initialize pi,0(·) to be positive everywhere.
4: repeat
5:
Stage 1: For each agent i ∈ N :
6:
7:
8:
9:
Bisect posterior density to obtain query point Xi,t =
F −1
i,t ( 1
Belief update:
2 ), and obtain noisy response yi,t+1 ∈ {0, 1}.
pi,t(x) = pi,t(x) · 2li(x, yi,t+1)
(2)
Stage 2: For each agent i ∈ N :
Average log-beliefs:
the binary symmetric channel [16].
Proof. Consider the case for a fixed node i. From Algorithm
1, we may rewrite (3) as:
log2 pi,t+1(x) =Xj
Ai,j log2 pj,t(x)
0
2PN
− log2(cid:18)Z 1
j=1 Ai,j log2 pj,t(x)dx(cid:19)
}
{z
Ai,j pj,t(x)dx
Dt
N
= log2
Xj=1
Ai,j = 0
Using Jensen's inequality, we have:
Dt ≤ log2
Z 1
0
N
Xj=1
Since Dt ≤ 0, it follows that:
log2 pi,t+1(x) ≥
N
Xj=1
Ai,j log2 pj,t(x)
(5)
Evaluating the above inequality at x = X ∗ and using
Equation (2) along with lj(X ∗, yj,t+1) = P (yj,t+1zj,t):
log2 pi,t+1(X ∗) ≥
N
Xj=1
Ai,j log2 pj,t(X ∗) +
N
Xj=1
Ai,j log2(2P (yj,t+1zj,t))
(6)
By using new variables: qi,t
def
= log2(2P (yi,t+1zi,t)), Equation (6) becomes:
def
= log2(pi,t(X ∗)) and
wj,t
pi,t+1(x) = QN
R 1
0 QN
j=1 pj,t(x)Ai,j
j=1 pj,t(x)Ai,j dx
10: until convergence
(3)
qi,t+1 ≥
N
Xj=1
Ai,jqj,t +
N
Xj=1
Using induction, it follows that:
Ai,jwj,t
(7)
3. PERFORMANCE ANALYSIS
In this section, we derive a lower bound on the posterior dis-
tribution pi,t(x) at the target X ∗. Our result shows that for
large t, we have pi,t(X ∗) ≥ 2tK for all agents i ∈ N .
Theorem 1. Assume that A corresponds to an irreducible,
aperiodic Markov chain, and pi,0(X ∗) > 0 for all i. Then,
the following asymptotic result holds:
lim inf
t→∞
log2 pi,t(X ∗)
t
≥
N
Xi=1
viC(ǫi) = K
(4)
where v is the normalized left eigenvector of A which corre-
sponds to the unit eigenvalue 2 and C(ǫi) is the capacity of
1We remark that the normalizing factor here R 1
0 pi,t(x)li(x, yi,t+1)dx
is equal to 1/2 [14].
2This is also known as the stationary distribution of the Markov chain.
qi,t ≥
t
Xτ =1
N
Xj=1
Aτ
i,jwj,t−τ +
N
Xj=1
At
i,jqj,0
(8)
Dividing both sides by t and taking the limit, it follows
that:
lim inf
t→∞
qi,t
t
≥ lim
t→∞
1
t
t
Xτ =1
N
Xj=1
Aτ
i,j wj,t−τ + lim
t→∞
1
t
N
Xj=1
At
i,jqj,0
(9)
Using 0 < pj,0(X ∗) < ∞ and At
in Equation (9) vanishes and we have:
i,j ≤ 1, the second term
lim inf
t→∞
qi,t
t
≥ lim
t→∞
1
t
t
Xτ =1
N
Xj=1
Aτ
i,j wj,t−τ =: Ki
(10)
Next, we evaluate the rate exponent Ki using the con-
vergence to a stationary distribution v for aperiodic Markov
chains. Specifically, using Theorem 7 in Sec. 2.7 from [17]:
lim
t→∞
At
i,j = vj
for any two nodes (i, j). Thus, for any arbitrarily small δ > 0,
i,j − vj ≤ δ for all t ≥ T .
there exists T = T (δ) such that At
Splitting the sum in (10):
Ki = lim
t→∞" 1
t
N
Xj=1
T −1
Xτ =1
Aτ
i,jwj,t−τ +
1
t
N
Xj=1
t
Xτ =T
Aτ
i,jwj,t−τ#
(11)
The first term in Equation (11) is negligible since it can be
upper bounded by limt→∞
τ =1 wj,t−τ . Since
wj,t−τ ≤ max{ log2(2(1 − ǫj)), log2(2ǫj)} = Bj,
1
t (T − 1)B = 0,
limt→∞
j=1 Bj = B. The second term in Equation (11)
j=1PT −1
τ =1 wj,t−τ ≤ limt→∞
j=1PT −1
t PN
1
1
dominates and is given by:
t PN
where PN
Ki = lim
t→∞
1
t
N
Xj=1
t
Xτ =T
[Aτ
i,j − vj ]wj,t−τ + lim
t→∞
1
t
N
Xj=1
t
Xτ =T
vjwj,t−τ
(12)
The first term in Equation (12) is negligible since:
lim
t→∞
1
t
≤ lim
t→∞
N
t
Xτ =T
Xj=1
Xj=1
Xτ =T
δ
t
N
t
Aτ
i,j − vj wj,t−τ
wj,t−τ ≤ lim
t→∞
δ
t
t
Xτ =T
N
Xj=1
Bj = δB
(13)
The second term in Equation (12) dominates and using the
LLN:
Ki =
=
N
Xj=1
Xj=1
N
vj" lim
t→∞
1
t
t
wj,t−τ#
Xτ =T
Xj=1
N
vj E[wj,t−τ ] =
vjC(ǫj) =: K
(14)
The proof is complete.
In Theorem 1, we derived a lower bound on the rate of
learning of each agent for the distributed bisection algorithm
(Algorithm 1). Since the rate exponent, Ki, is independent of
the agent index, i, the lower bound is the same for all agents
and depends on the channel capacities and the eigenvector
centrality, v [9]. The higher the value of vi, the larger the
contribution that node i has on the network learning rate, K.
Qt−1
Using an analogous method, we can analyze Equation
(2) for the case of no inter-agent collaboration. In this case,
A = IN , and using Equations (2) and (3), pi,t+1(X ∗) =
pi,t(X ∗) = pi,t(X ∗) + 2P (yi,t+1zi,t). Unrolling the
preceding equation over t steps, we see that pi,t(x) =
j=0 2P (yi,j+1zi,j). By taking the logarithm of both sides,
dividing by t, and applying the LLN, the convergence rate
= C(ǫi).
of the agent beliefs becomes: limt→∞
For the case of a homogeneous network, i.e.
ǫi = ǫ, ∀i,
our distributed algorithm asymptotically learns faster than
when agents do not collaborate. In general, the rate exponent
of Theorem 1 is a linear combination of agents' capacities
weighed by the eigenvector centrality. This bound is expected
to be maximized by placing nodes with low error probability
in central locations in networks where they can have a strong
social influence on other nodes.
log2 pi,t(x)
t
4. SIMULATIONS
To validate and strengthen the preceding performance analy-
sis, simulations are performed that show the proposed method
achieves better performance than the belief consensus ap-
proach of [14] and the case of no inter-agent collaboration.
We note that the computational complexity of our proposed
algorithm is on the same order as that of [14] as both al-
gorithms consist of a Bayesian bisection update and a local
averaging operation.
We randomly generate an irreducible N × N adjacency
matrix, modeling a random geometric graph [18] (see Figure
1).
In this setup, N = 20 and 18 agents have high error
probabilities (ǫi = 0.40) while the remaining 2 have low error
probabilities (ǫi = 0.05). The low error agents are the ones
with the highest two vi's, so they can positively affect the
high error nodes around them. Figure 1 displays the 2 low
error nodes and their connections in blue along with the other
nodes and their connections in black.
N PN
Figure 2 shows the average network Mean Squared Error
i=1( Xi,t − X ∗)2, and the worst-
(MSE), M SEavg(t) = 1
case network MSE, M SEmax(t) = maxi( Xi,t − X ∗)2, aver-
aged over 150 Monte Carlo trials. Our proposed method out-
performs the belief consensus approach of [14] and the case
of no collaboration. Regarding the average MSE, all three
methods converge to the consensus but our proposed method
does so much quicker than the other methods. In terms of
worst case MSE, our proposed method remains robust while
that of [14] converges slower and the method with no collab-
oration has not converged after 75 time steps. From this, we
see that the proposed method performs very well even in the
worst case and that inter-agent collaboration is advantageous
in our target localization task.
In addition to the MSE, we also plot log2 p4,t(X ∗) for
both the proposed method and that of no collaboration. The
log2 p4,t(X ∗) time series represent a base-2 logarithm of
at least as large as the lower bound K =PN
agent 4's belief, an agent with a high error probability. We
analytically showed that for the case of no collaboration, a
graph of the log2 pi,t(x∗) vs. t would be a line with slope
K = C(ǫi) that would serve as an exact bound; with our pro-
posed method, the same graph should be a line with the slope
j=1 vjC(ǫi) (14).
The analysis results are affirmed by the simulation results in
Figure 3. Here, the dotted lines represent the two preced-
ing bounds and the solid lines represent the calculations of
log2 p4,t(X ∗) for each iteration. The solid red line represents
the no collaboration method and it lines up with the exact
bound displayed in the dotted red line. The solid blue line
represents our proposed method and its slope exceeds that
displayed by the lower bound with the dotted blue line. The
simulations verify our analysis and show that the beliefs gen-
erated by our proposed method concentrate fast on the true
target location.
With our distributed algorithm, it is important to note
that we improve the network-wide performance (average and
worst case) with a small penalty in performance in the case
of no collaboration for the lowest error agent.
5. CONCLUSION
In this paper, we proposed a new distributed probabilistic
bisection algorithm for target localization and derived a lower
bound on the rate of convergence. Through analysis and sim-
ulation, we show that our proposed method attains superior
performance to other state of the art methods in terms of rate
of convergence and MSE. For future work, we can pursue an
analysis of the MSE convergence rate.
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Fig. 1. Geometric random graph. The blue nodes are the low error agents
and those in red are the high error agents.
0.15
0.1
0.05
g
v
a
E
S
M
0
0
0.4
0.3
0.2
0.1
0
0
x
a
m
E
S
M
No Collaboration
With Collaboration: Tsiligkaridis et al [2015]
With Collaboration: Proposed
10
20
30
40
Iterations
50
60
70
No Collaboration
With Collaboration: Tsiligkaridis et al [2015]
With Collaboration: Proposed
10
20
30
40
Iterations
50
60
70
Fig. 2. Average (top) and worst-case (bottom) MSE performance for our
proposed social learning algorithm, the belief consensus approach of [14],
and the case of no collaboration.
)
*
X
(
t
,
4
p
2
g
o
l
12
10
8
6
4
2
0
0
Bound - No Collaboration
Lower Bound - With Collaboration
log
(X*) - No Collaboration
(X*) - With Collaboration
4,t
p
2
log
p
2
4,t
10
20
30
40
50
60
70
Iterations
Fig. 3. Concentration of beliefs as a function of iteration. The belief of the
high error node, agent 4, of our proposed method concentrates to the target
location with a slope of 0.32 as opposed to the no collaboration case with the
slope 0.03 for the same node. Our method provides 10× improvement over
the case of no collaboration regarding the worst case scenario.
[12] B. Jedynak, P. I. Frazier, and R. Sznitman, "Twenty
Questions with noise: Bayes optimal policies for en-
tropy loss," Journal of Applied Probability, vol. 49, pp.
114 -- 136, 2012.
[13] R. Waeber, P. I. Frazier, and S. G. Henderson, "Bisec-
tion Search with Noisy Responses," Journal of Control
Optimization, vol. 53, no. 3, pp. 2261 -- 2279, 2013.
[14] T. Tsiligkaridis, B. M. Sadler, and A. O. Hero III, "On
Decentralized Estimation with Active Queries," IEEE
Transactions on Signal Processing, vol. 63, no. 10, pp.
2610 -- 2622, May 2015.
[15] T. Tsiligkaridis,
"Asynchronous Decentralized Algo-
rithms for the Noisy 20 Questions Problem," in Pro-
ceedings of the IEEE International Symposium on In-
formation Theory (ISIT), Barcelona, Spain, July 2016.
[16] T. D. Cover and J. A. Thomas, Elements of Information
Theory, Wiley, New York, NY, USA, 2006.
[17] P. G. Hoel, S. C. Port, and C. J. Stone,
Introduction
to Stochastic Processes, Houghton Mifflin Company,
Boston, MA, USA, 1972.
[18] P. Gupta and P. R. Kumar, "The Capacity of Wireless
Networks," IEEE Transactions on Information Theory,
vol. 46, no. 2, pp. 388 -- 404, March 2000.
6. REFERENCES
[1] R. Waeber, P. I. Frazier, and S. G. Henderson,
"A
Bayesian Approach to Stochastic Root Finding," in Pro-
ceedings of the Winter Simulation Conference, Phoenix,
AZ, December 2011.
[2] W. Han, P. Rajan, P. I. Frazier, and B. M. Jedynak,
"Probabilistic Group Testing Under Sum Observations:
A Parallelizable 2-Approximation for Entropy Loss,"
IEEE Transactions on Information Theory, to appear,
2016.
[3] R. Sznitman and B. Jedynak, "Active Testing for Face
Detection and Localization," IEEE Transactions on Pat-
tern Analysis and Machine Intelligence, vol. 32, no. 10,
pp. 1914 -- 1920, 2010.
[4] T. Tsiligkaridis, B. M. Sadler, and A. O. Hero III, "Col-
laborative 20 Questions for Target Localization," IEEE
Transactions on Information Theory, vol. 60, no. 4, pp.
2233 -- 2352, April 2014.
[5] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Ran-
domized Gossip Algorithms," IEEE Transactions on In-
formation Theory, vol. 52, no. 6, pp. 2508 -- 2530, June
2006.
[6] A. Dimakis, S. Kar, J. M. F. Moura, M. G. Rabbat, and
A. Scaglione, "Gossip Algorithms for Distributed Sig-
nal Processing," Proceedings of the IEEE, vol. 98, no.
11, pp. 1847 -- 1864, November 2010.
[7] S. Kar and J. M. F. Moura, "Convergence Rate Analy-
sis of Distributed Gossip (Linear Parameter) Estimation:
Fundamental Limits and Tradeoffs," IEEE Journal of
Selected Topics in Signal Processing, vol. 5, no. 4, pp.
674 -- 690, August 2011.
[8] S. Kar, J. M. F. Moura, and K. Ramanan, "Distributed
Parameter Estimation in Sensor Networks: Nonlinear
Observation Models and Imperfect Communication,"
IEEE Transactions on Information Theory, vol. 58, no.
6, pp. 3575 -- 3605, June 2012.
[9] A. Lalitha, T. Javidi, and A. Sarwate, "Social Learn-
ing and Distributed Hypothesis Testing," ArXiv e-prints,
Oct. 2014.
[10] A. Nedic, A. Olshevsky, and C. A. Uribe, "Nonasymp-
totic Convergence Rates for Cooperative Learning over
Time-Varying Directed Graphs," in Proceedings of the
American Control Conference (ACC), Chicago, IL, July
2015.
[11] M. Horstein, "Sequential Transmission Using Noiseless
Feedback," IEEE Transactions on Information Theory,
vol. 9, no. 3, pp. 136 -- 143, July 1963.
|
1510.05198 | 4 | 1510 | 2017-05-11T20:47:13 | Learning multi-faceted representations of individuals from heterogeneous evidence using neural networks | [
"cs.SI",
"cs.CL"
] | Inferring latent attributes of people online is an important social computing task, but requires integrating the many heterogeneous sources of information available on the web. We propose learning individual representations of people using neural nets to integrate rich linguistic and network evidence gathered from social media. The algorithm is able to combine diverse cues, such as the text a person writes, their attributes (e.g. gender, employer, education, location) and social relations to other people. We show that by integrating both textual and network evidence, these representations offer improved performance at four important tasks in social media inference on Twitter: predicting (1) gender, (2) occupation, (3) location, and (4) friendships for users. Our approach scales to large datasets and the learned representations can be used as general features in and have the potential to benefit a large number of downstream tasks including link prediction, community detection, or probabilistic reasoning over social networks. | cs.SI | cs | Learning Multi-faceted Representations of Individuals from
Heterogeneous Evidence using Neural Networks
Jiwei Li, Alan Ritter and Dan Jurafsky
Stanford University, Stanford, CA, USA
Ohio State University, OH, USA
jiweil,[email protected],[email protected]
7
1
0
2
y
a
M
1
1
]
I
S
.
s
c
[
4
v
8
9
1
5
0
.
0
1
5
1
:
v
i
X
r
a
Abstract
Inferring latent attributes of people on-
line is an important social computing task,
but requires integrating the many hetero-
geneous sources of information available
on the web. We propose learning individ-
ual representations of people using neural
nets to integrate rich linguistic and network
evidence gathered from social media. The
algorithm is able to combine diverse cues,
such as the text a person writes, their at-
tributes (e.g. gender, employer, education,
location) and social relations to other peo-
ple. We show that by integrating both tex-
tual and network evidence, these represen-
tations offer improved performance at four
important tasks in social media inference
on Twitter: predicting (1) gender, (2) occu-
pation, (3) location, and (4) friendships for
users. Our approach scales to large datasets
and the learned representations can be used
as general features in and have the potential
to benefit a large number of downstream
tasks including link prediction, community
detection, or probabilistic reasoning over
social networks.
Introduction
1
The recent rise of online social media presents an
unprecedented opportunity for computational so-
cial science: user generated texts provide insight
about user's attributes such as employment, educa-
tion or gender. At the same time the social network
structure sheds light on complex real-world rela-
tionships between preferences and attributes. For
instance people sharing similar attributes such as
employment background or hobbies have a higher
chance of becoming friends. User modeling based
Figure 1: Illustration for the proposed method that learns
latent representations for users, attributes and user-generated
texts based on social network information.
on information presented in social networks is an
important goal, both for applications such as prod-
uct recommendation, targeted online advertising,
friend recommendation and for helping social sci-
entists and political analysts gain insights into pub-
lic opinions and user behaviors.
Nevertheless, much important information on
social networks exists in unstructured data formats.
Important social insights are locked away, entan-
gled within a heterogenous combination of social
signals (Sun et al., 2009) - including text, networks,
attributes, relations, preferences, etc. While recent
models have attempted to link one or two aspects of
the evidence, how to develop a scalable framework
that incorporates massive, diverse social signals
including user generated texts, tens of thousands
of user attributes and network structure in an inte-
grated way, remains an open problem.
In this paper, we propose a general deep learn-
ing framework for jointly analyzing user networks,
generated context and attributes. We map users, at-
tributes, and user-generated content to latent vector
Friend(User1,User2)Friend(User2,User3)Friend(User3,User1)Friend(User4,User5)Texts:12435Hereinclass,atHarvardUniversity!612345UserNetworkStructureLike(User1,StarTrek)LivesIn(User2,Boston)StudyAt(User3,Harvard)StudyAt(User6,Harvard)16StarTrek2Boston3Harvard6UserAttributeLinkageStarTrekHarvardBoston4 representations, which are learned in a scalable way
from social network data. Figure 1 gives a brief
overview of the mechanism of the proposed model:
users are represented by similar vectors if they are
friends, share similar attributes or write similar con-
tent. Attributes are similarly clustered if associated
with similar users.1 In summary, we incorporate
diverse social signals in a unified framework, al-
lowing user embeddings to be jointly optimized
using neural networks trained on vast quantities of
rich social and linguistic context.
Based on these learned representations, our ap-
proach provides with a general paradigm on a wide
range of predictive tasks concerning individual
users as well as group behavior: user attribute in-
ference (e.g., the city the user lives in), personal
interest prediction (e.g, whether a user will like a
particular movie), and probabilistic logical reason-
ing over the social network graph. For example,
our models infer that:
• Men in California are 6.8 times more likely to
take an engineering occupation than women
in California.
• Users who work in the IT industry2 are 2.5
times more likely to like iPhones than users
working in Legal Occupations.
Our methods also have the potential to seamlessly
integrate rich textual context into many social net-
work analysis tasks including: link prediction, com-
munity detection, and so on, and the learned user
representations can be used as important input fea-
tures for downstream machine learning models, just
as how word embeddings are used in the field of
natural language processing. The major contribu-
tions of this paper can be summarized as follows:
• We propose new ways for integrating hetero-
geneous cues about people's relations or at-
tributes into a single latent representation.
• We present inference algorithms for solving
social media inference tasks related to both
individual and group behaviors based on the
learned user representations.
• Cues that we take advantage of may be noisy
and sometimes absent but by combining them
via global inference, we can learn latent facts
about people.
We evaluate the proposed model on four diverse
tasks: friend-relation prediction, gender identifica-
1This core idea is similar to collaborative filtering (Kautz
et al., 1997).
2This information comes from the Standard Occupational
Classification (SOC), as will be described later.
tion, occupation identification and user geolocation
prediction. Experimental results demonstrate im-
proved predictions by our model by incorporating
diverse evidence from many sources of social sig-
nals.
2 Social Representation Learning
Our goal is to learn latent representations from the
following three types of online information: (1)
user-generated texts (2) friend networks (3) rela-
tions and attributes.
2.1 Modeling Text
User generated texts reflect a user's interests, back-
grounds, personalities, etc. We thus propose learn-
ing user representations based on the text a user
generates.
We represent each user v by a K-dimensional
vector ev. Suppose that S denotes a sequence of
tokens S = {w1, w2, ..., wNS} generated by the
current user v. Each word w ∈ S is associated
with a K-dimensional vector ew. Let C(w) denote
the list of neighboring words for token w. w is
generated based on not only a general language
model shared across all users (namely, a model that
predicts w given C(w)), but the representation ev
of the current user :
P (wC(w), v) = p(weC)
eC =
1
Cw + 1 [
(cid:88)
w(cid:48)∈C(w)
ew(cid:48) + ev]
(1)
From Eq.1, we are predicting the current word
given the combination of its neighbors' embed-
dings and the current user embedding. This is akin
to the CBOW model (Mikolov et al., 2013) with the
only difference that the user embedding is added
into contexts. Such an idea also resembles the para-
graph vector model (Le and Mikolov, 2014) and
the multimodal language model (Kiros et al., 2014)
.
We use negative sampling, in which we ran-
domly generate negative words w∗. Let Lw denote
a binary variable indicating whether the current
word w is generated by the current user. The loss
function using negative sampling is given by:
Loss(text) = log p(Lw = 1v)+
(cid:88)
log p(Lw = 0v)
w∗
Word prediction errors are backpropogated to
user embeddings, pushing the representations of
users who generate similar texts to be similar.
2.2 Modeling User Networks
By the homophily effect,
individuals who are
friends on social networks tend to share common
characteristics. We therefore to encourage users
who are friends have similar representations.
We propose using a strategy similar to skip-gram
models, in which we consider users who are friends
on social networks analogous to neighboring words,
the representations of which we wis to be similar.
On the other hand, we want embeddings of indi-
viduals who are not friends to be distant, just as
words that do not co-appear in context. A similar
idea of transforming social graph to vector space
embeddings has been explored in the recent deep-
walk model (Perozzi et al., 2014) and Line (Tang
et al., 2015).
Suppose we have two users v and v(cid:48). The proba-
bility that the two users are and are not friends are
respectively given by:
log p(L(v, v(cid:48)) = 1) = log
log p(L(v, v(cid:48)) = 0) = log
1
1 + exp(−ev · ev(cid:48))
1 + exp(ev · ev(cid:48))
1
(2)
From Eq 2, we can see that the model favors the
cases where the dot product of friends' embeddings
is large, equivalent to their embeddings being simi-
lar. Again, we use negative sampling for optimiza-
tion. For two users v and v(cid:48) who are friends, we
sample N random users v∗, and we assume friend-
ship does not hold between them. The objective
function is therefore given by:
Similar scoring functions have been applied in a
variety of work for relation extraction (Socher et al.,
2013; Chang et al., 2014). Again we turn to neg-
ative sampling for optimization. The system ran-
domly samples none-referred-to entities and maxi-
mizes the difference between the observed relation
tuples and randomly sampled ones. Through the
model described above, users who share similar
attributes will have similar representations. At the
same time, entities that are shared by similar users
would also have similar representations and thus
occupy close positions in the vector space.
2.3 Training
The global learning objective is a linear combina-
tion of the objectives from the three categories de-
scribed above. User embeddings are shared across
these categories, and each part can communicate
with the rest: a user who publishes content about
a particular city (text modeling) can have similar
embeddings to those who live in that city (rela-
tion/attribution modeling); friends (graph model-
ing) of a basketball fan (relation/attribution model-
ing) are more likely to be basketball fans as well.
The final objective function is given as follows:
L(Θ) =Loss(text) + λ1Loss(graph)
+ λ2Loss(relation/attribute)
Θ = arg min
Θ(cid:48)
L(Θ(cid:48))
where λ1 and λ2 denote weights for different
constituents. We use Stochastic gradient decent
(Zhang, 2004) to update the parameters.
(cid:88)
v∗
Loss(graph)
= log p(L(v, v(cid:48)) = 1) +
log p(L(v, v∗) = 0)
Θt := Θt−1 − α
∂L
∂Θ
2.2.1 Modeling Relations and Attributes
Intuitively, users who share similar attributes
should also have similar representations and oc-
cupy similar position in the vector space. Suppose
that a specific relation r holds between a user v
and an entity m. We represent each user and en-
tity by a K-dimensional vector and a relation by a
K × K matrix. For any tuple (r, v, m), we map it
to a scalar within the range of [0,1], indicating the
likelihood of relation r holding between user v and
entity m:
log p(L(r, v, m) = 1)
= log
1
1 + exp(−eT
vi · Ri · em)
The system jointly learns user embeddings, word
embeddings, entity embeddings and relation matri-
ces.
Inference on Social Networks
3
In this section, we describe how to take as input
a learned user embedding for different inference
tasks on social media. We divide inference tasks
on social networks into two major categories: infer-
ence for individual behaviors and group behaviors.
The former focuses on inferring attributes of a spe-
cific user such as whether a user likes a specific
entity or whether a specific relation holds between
two users, while the latter focuses on inference over
a group of users, e.g., what is the probability of a
new yorker being a fan of Knicks.
3.1 User Attribute Inference
Given user representation evi, we wish to infer the
label for a specific attribute of a specific user. The
label can be whether a user likes an entity in a
binary classification task, or the state that a user
lives in a multi-class classification task.
Suppose that we want to predict an attribute la-
bel (denoted by tv ∈ [1, L]) for a user v. We
assume that information for this attribute is em-
bedded in user representations and therefore, and
build another neural model to expose this informa-
tion. Specifically, the model takes as input user
embedding ev and outputs the attribute label using
a softmax function as follows:
h = tanh(W · ev)
Sl = Ul · h
p(tvi = l) =
(cid:80)
exp(Sl)
l(cid:48)∈[1,L] exp(Sl(cid:48))
Parameters to learn include W and U. User rep-
resentations are kept fixed during training. The
model is optimized using AdaGrad (Zeiler, 2012).
3.2 User Relation Inference
User Relation Inference specifies whether a par-
ticular relationship holds between two users (e.g.,
whether the two users are friends). It takes em-
beddings for both users as inputs. Given a user vi
(associated with embedding evi) and a user vj (as-
sociated with embedding evj ), we wish to predict
the index of relationship label t(vi, vj) ∈ [1, L]
that holds between the two users. A neural network
prediction model is trained that takes as input the
embeddings of the two users. The model considers
the distances and angle between the two user em-
beddings. Similar strategies can be found in many
existing works, e.g., (Tai et al., 2015).
Non-linear composition is first applied to both
user representations:
hv1 = tanh(Wa · ev1)
hv2 = tanh(Wb · ev2)
Next the distance and angle between hv1 and hv2
are computed:
h+ = hv1 ⊗ hv2
h× = hv1 − hv2
The multiplicative measure h× is the elementwise
comparison of the signs of the input representations.
Finally a softmax function is used to decide the
label:
h = tanh(W × · h× + W + · h+ + W1 · hv1 + W2 · hv2)
p[t(vi, vj) = l] = softmax(U · h)
Inference of Group Behavior
Again, parameters involved are learned using
stochastic gradient decent with AdaGrad (Zeiler,
2012).
3.3
We now return to the example described in Section
1, in which we wish to estimate the probability of a
male located in California (Cal for short) having an
engineering occupation. Given a list of users, their
representations, and gold-standard labels, we first
separately train the following neural classifiers:
• whether
user
P(gender(evi)=male)
user
P(LiveIn(evi)=Cal)
• whether
a
a
is
a male,
i.e.,
lives
in Cal,
i.e.,
• whether a user takes an engineering occupa-
tion,
i.e., P(Job(evi)=engineering)
Next we estimate the general embedding (denoted
by eG) for the group of people that satisfy these
premises, namely, they are males and live in Cali-
fornia at the same time. This can be transformed to
the following optimization problem with eG being
the parameters to learn: 3
eG = arg max
e
log P (gender(e) = male)
− log P (LiveIn(e) = Cal)
(3)
Eq.3 can be thought of as an optimization prob-
lem to find a optimal value of eG. This problem
can be solved using SGD. The obtained optimal
eG is used to represent that group embedding for
users who satisfy the premises (e..g, males and
living in Cal). eG is then used as inputs to clas-
sifier P (Job(eG)) = engineering) which returns
probability of taking an engineering job.
More formally, given a list of conditions {ai} ∈
A, we want to the compute the probability that
another list of conditions (denoted by B = {bj})
hold, in which bj can be a user being an engineer.
This probability is denoted by p(BA). The al-
gorithm to compute p(BA) for group behavior
inference is summarized in Figure 2.
3Note that we assume propositions are independent, so in
the example above, being a male and living in California are
independent from each other. We leave relaxing this indepen-
dence assumption to future work.
• For ai ∈ A, bj ∈ B,
train separate classifiers
P (aie), P (bje) based on user representations e and
labeled datasets.
• Estimate group representation eG by solving the follow-
ing optimization problem using SGD:
(cid:89)
ai∈A
P (aie)
eG = arg max
• Infer the probability :
P (BA) =
e
(cid:89)
bj∈B
p(bje(group))
Figure 2: Algorithm for group behavior inference.
4 Dataset Construction
Existing commonly used social prediction datasets
(e.g., BlogCatalog and Flickr (Tang and Liu,
2009a), YouTube (Tang and Liu, 2009b)) are de-
signed with a specific task in mind: classifying
whether social links exist between pairs of users.
They contain little text or user-attribute informa-
tion, and are therefore not well suited to evaluate
our proposed model.
Social networks such as Facebook or LinkedIn
that support structured profiles would be ideal for
the purpose of this work. Unfortunately, they are
publicly inaccessible. We advert to Twitter. One
downside of relying on Twitter data is that gold-
standard information is not immediately available.
Evaluation presented in this paper therefore comes
with the flaw that it relies on downstream informa-
tion extraction algorithms or rule-based heuristics
for the attainment of "gold standards". Though not
perfect as "gold-standards" extraction algorithm
can be errorful, such a type of evaluation comes
with the advantage that it can be done automatically
to compare lots of different systems for develop-
ment or tuning in relatively large scale. Meanwhile
the way that our dataset is constructed gives impor-
tant insights about how the proposed framework
can be applied when some structured data is miss-
ing4 and how we can address these challenges by
directly from unstructured text, making this system
applicable to a much wider scenario.
4Facebook and LinkedIn do come with the ideal property
of supporting structured user profiles but hardly anyone fills
these out.
4.1 User Generated Texts and Graph
Networks
We randomly sample a set of Twitter users, discard-
ing non-English tweets and users with less than
100 tweets. For each user, we crawl their published
tweets and following / follower network using the
publicly available Twitter API. This results in a
dataset of 75 million tweets.
4.2 User Attributes
Unlike social networking websites such as Face-
book, Google+ and LinkedIn, Twitter does not sup-
port structured user profile attributes such as gender,
education and employer. We now briefly describe
how we enrich our dataset with user attributes (loca-
tion, education, gender) and user relations (friend,
spouse). Note that, the goal of this section is to
construct a relatively comprehensive dataset for the
propose of model evaluation rather than developing
user attribute extraction algorithms.
4.2.1 Location
We first associate one of the 50 US states with each
user. In this paper, we employ a rule-based ap-
proach for user-location identification.5. We select
all geo-tagged tweets from a specific user, and say
an location e corresponds to the location of the
current user i if it satisfies the following criteria,
designed to ensure high-precision: (1) user i pub-
lished more than 10 tweets from location e . (2)
user i published from location e in at least three dif-
ferent months of a year. We only consider locations
within the United States and entities are matched to
state names via Google Geocoding. In the end, we
are able to extract locations for 1.1% of the users
from our dataset.
4.2.2 Education/Job
We combine two strategies to harvest gold-standard
labels for users' occupation and educational in-
formation. We use The Standard Occupational
Classification (SOC)6 to obtain a list of occupa-
tions, a approach similar to Preot¸iuc-Pietro et al.
5 While there has been a significant work on geolocation
inference, (e.g., (Cheng et al., 2010; Conover et al., 2013;
Davis Jr et al., 2011; Onnela et al., 2011; Sadilek et al., 2012)),
the primary goals of this work are to develop user representa-
tions based on heterogenous social signals. We therefore take
a simple high-precision, low-recall approach to identifying
user locations.
6http://www.ons.gov.uk/ons/
guide-method/classifications/
current-standard-classifications/
soc2010/index.html
(2015); Preotiuc-Pietro et al. (2015). 7 A user is
assigned an occupation if one of the keywords from
the lookup table is identified in his/her profile. 12%
percent of users' occupations are identified using
this strategy.
4.2.3 Gender
Following a similar strategy as was described for
location and occupation, we take implement a sim-
ple high-precision approach for obtaining gold-
standard user gender information. We leverage the
national Social Security Gender Database (SSGD)8
to identify users' gender based on their first names.
SSGD contains first-name records annotated for
gender for every US birth since 1880 A.D9. Using
this database we assign gender to 78% of users in
our dataset.
5 Experiments
We now turn to experiments on using global infer-
ence to augment individual local detectors to infer
user's attributes, relations and preferences. All ex-
periments are based on datasets described in the
previous sections. We performed 3 iterations of
stochastic gradient descent training over the col-
lected dataset to learn embeddings. For each task,
we separate the dataset into 80% for training 10%
development and 10% for testing.
For comparison purposes, neutral models that
take into account only part of the training signals
presented naturally constitute baselines. We also
implement feature-based SVM models as baselines
for the purpose of demonstrating strength of neu-
ral models. For neural models, we set the latent
dimensionality K to 500. Pre-trained word vectors
7 SOC is a UK government system developed by the Office
of National Statistics that groups jobs into 23 major categories
(for example: or Engineering Occupations or Legal Occupa-
tions), each of which is associated with a set of specific job
titles (e.g., mechanical engineer and pediatrist for Professional
Occupations). We construct a lookup table from job occu-
pations to SOC and apply a rule-based mapping strategy to
retrieve a users' occupation information based on the free-text
user description field from their Twitter profile. Note that
this approach introduces some bias: users with high-profile
occupations are more likely to self-disclose their occupations
than users with less prestigious occupations.
8http://www.ssa.gov/oact/babynames/
names.zip
9Again, we note the large amount of related work on pre-
dicting gender of social media users (e.g., (Burger et al., 2011;
Ciot et al., 2013; Pennacchiotti and Popescu, 2011; Tang et al.,
2011).) studying whether high level tweet features (e.g., link,
mention, hashtag frequency) can help in the absence of highly-
predictive user name information. As mentioned before, we
do not adopted machine learning algorithms for attribute ex-
traction.
Model
All
Only Network
Network+Attribute
Network+Text
Accuracy
0.257
0.179
0.198
0.231
Table 1: Accuracy for different models on friend relationship
prediction from social representations.
are used based on the word2vec package.10 Em-
beddings are trained on a Twitter dataset consisting
of roughly 1 billion tokens.
graph structure and user attributes.
5.1 Friend-Relation (Graph Link) Prediction
Twitter supports two types of following patterns,
FOLLOWING and FOLLOWED. We consider two
users as friends if they both follow each other. The
friendship relation is extracted straightforwardly
from the Twitter network. Models and baselines
we employ include:
• All: The proposed model that considers text,
• Only Network: A simplified version of the pro-
posed model that only used the social graph
structure to learn user representations a Note
that by making this simplification, the model
is similar to DeepWalk (Perozzi et al., 2014)
with the exception that we adopt negative sam-
pling rather than hierarchical softmax.
• Network+Attribute: Neural models that con-
sider social graph and relation/entity informa-
tion.
• Network+text: Neural models that consider
social graph and text information.
Performance for each model is shown in Table
1. As can be seen, taking into account different
types of social signals yields progressive perfor-
mance improvement: Graph+Attribute performs
better than only graph, and All, which consider
all different types of social signals is better than
Graph+Attribute .
5.2 User Attributes: Job Occupation
We present experimental results for job classifica-
tion based on user-level representations. Evalua-
tion is performed on the subset of users whose job
labels are identified by the rule-based approach de-
scribed in the previous section. Our models are
trained to classify the top-frequent 10 categories of
job occupations
Again, all denotes the model that utilizes all
types of information. Baselines we consider in-
10https://code.google.com/p/word2vec/
Model
All
Only Network
SVM-text
Network+Text
Accuracy
0.402
0.259
0.330
0.389
Table 2: Accuracy for different models on 9-class job occupa-
tion prediction from social representations.
Model
All
Attribute+Text
Only Network
Only Text
SVM-text
Accuracy
0.840
0.575
0.804
0.785
0.828
Table 3: Accuracy for different models on 9-class job occupa-
tion prediction from social representations.
clude:
• Text-SVM: We use SVM-Light package to
train a unigram classifier that only considers
text-level information.
• Only Network: A simplified version of the pro-
posed model that trains user embedding based
on network graph and occupation information.
• Network+Text: Embeddings are trained from
user-generated texts and network information.
Experimental results are illustrated in Table 2.
As can be seen, user generated content offers infor-
mative evidence about job occupation. We also ob-
serve that considering network information yields
significant performance improvement due to the
homophily effect, which has been spotted in earlier
work (Li et al., 2014b). Again, the best performing
model is the one that considers all sorts of evidence.
5.3 User Attribute: Gender
We evaluate gender based on a dataset of 10,000
users (half male, half female). The subset is drawn
from the users whose gold standard gender la-
bels are assigned by the social-security system
described in the previous section. Baselines we
employ include: SVM-Text, in which a SVM binary
classification model is trained on unigram features;
Only-Text, in which user representations are learned
only from texts; Only-Network, in which user repre-
sentations are only learned from social graphs; and
Text+Relation, in which representations are learned
from text evidence and relation/entity information.
The proposed neural model achieves an accuracy
value of 0.840. which is very close to the best per-
formance that we are aware of described in Ciot
et al. (2013), which achieves an accuracy of 0.85-
0.86 on a different dataset. However, unlike in Ciot
et al. (2013), the proposed model does not require
Model
All
Only Network
Only Text
Network+Text
Attribute+Text
Accuracy
0.152
0.118
0.074
0.120.
0.089
Table 4: Accuracy for different models location prediction
from social representations.
massive efforts in feature engineering, which in-
volves collecting a wide variety of manual features
such as entities mentioned, links, wide range of
writing style features, psycho-lingsuitic features,
etc. This demonstrates the flexibility and scalabil-
ity of deep learning models to utilize and integrate
different types of social signals on inference tasks
over social networks,
User-generated contexts offer significant evi-
dence for gender. Again, we observe that lever-
aging all sorts of social evidence leads to the best
performance.
Experimental results are shown in Figure 3. As
can be seen, network information does significantly
help the task of gender identification, only achiev-
ing slightly better performance than random guess.
Such an argument is reinforced by the fact that
Text+Relation yield almost the same performance
as model all, which takes Text+ Relation+ network
information.
5.4 User Attribute: Location
The last task we consider is location identification.
Experiments are conducted on users whose loca-
tions have been identified using the rule-based ap-
proach described in the previous section. The task
can be thought of as a 50-class classification prob-
lem and the goal is to pick one from the 50 states
(with random-guess accuracy being 0.02). We em-
ploy baselines similar to earlier sections: only-text,
only network, text+attribute and text+network.
Results are presented in Table 3: both text
and network evidence provide informative evident
about where a user lives, leading to better perfor-
mances. Again, the best performance is obtained
when all types of social signals are jointly consid-
ered
5.5 Examples for Group Behavior Inference
Given the trained classifiers (and additionally
trained LIKE-DISLIKE classifiers with details
shown in the Appendix), we are able to perform
group behavior inference. Due to the lack gold
standard labeled dataset, we did not perform evalu-
ations, but rather list a couple of examples to give
readers a general sense of the proposed paradigm:
• P(isMale⇒isEngineer)=3.5×
P(isFemale⇒isEngineer)
• P(isMale,LiveInCalifornia⇒isEngineer)=
.
P(isFemale,LiveInCalifornia⇒isEngineer)
6.8×·
• P(LiveInColorado⇒LikeOmelet)=
1.4×P(LiveInCalifornia⇒LikeOmelet)
.
.
• P(LiveInTexas⇒LikeBarbecue)=
1.7×P(LiveInCalifornia⇒LikeBarbecue)
6 Related Work
Much work has been devoted to automatic user at-
tribute inference given social signals. For example,
(Rao et al., 2010; Ciot et al., 2013; Conover et al.,
2011; Sadilek et al., 2012; Hovy et al., 2015) fo-
cus on how to infer individual user attributes such
as age, gender, political polarity, locations, occu-
pation, educational information (e.g., major, year
of matriculation) given user-generated contents or
network information.
Taking advantage of large scale user informa-
tion, recent research has begun exploring logical
reasoning over the social network (e.g., what's the
probability that a New York City resident is a fan
of the New York Knicks). Some work (Li et al.,
2014c; Wang et al., 2013) relies on logic reason-
ing paradigms such as Markov Logic Networks
(MLNs) (Richardson and Domingos, 2006).
Social network inference usually takes advan-
tage of the fundamental propoety of homophily
(McPherson et al., 2001), which states that people
sharing similar attributes have a higher chance of
becoming friends11, and conversely friends (or cou-
ples, or people living in the same location) tend to
share more attributes. Such properties have been
harnessed for applications like community detec-
tion (Yang et al., 2013) and user-link prediction
(Perozzi et al., 2014; Tang and Liu, 2009a).
The proposed framework also focuses on at-
tribute inference, which can be reframed as rela-
tion identification problems, i.e., whether a relation
holds between a user and an entity. This work is
thus related to a great deal of recent researches on
relation inference (e.g., (Guu et al., 2015; Wang
et al., 2014; Riedel et al., 2013)).
11Summarized by the proverb "birds of a feather flock to-
gether" (Al Zamal et al., 2012).
Our work is inspired by classic work on spectral
learning for graphs e.g., (Kunegis and Lommatzsch,
2009; Estrada, 2001) and on recent research (Per-
ozzi et al., 2014; Tang et al., 2015) that learn em-
bedded representations for a graph's vertices. Our
model extends this work by modeling not only user-
user network graphs, but also incorporating diverse
social signals including unstructured text, user at-
tributes, and relations, enabling more sophisticated
inferences and offering an integrated model of ho-
mophily in social relations.
7 Conclusions
We have presented a deep learning framework for
learning social representations, inferring the latent
attributes of people online. Our model offers a new
way to jointly integrate noisy heterogeneous cues
from people's text, social relations, or attributes
into a single latent representation. The representa-
tion supports an inference algorithm that can solve
social media inference tasks related to both individ-
ual and group behavior, and can scale to the large
datasets necessary to provide practical solutions to
inferring huge numbers of latent facts about people.
Our model has the ability to incorporate various
kinds of information, and it increases in perfor-
mance as more sources of evidence are added. We
demonstrate benefits on a range of social media
inference tasks, including predicting user gender,
occupation, location and friendship relations.
Our user embeddings naturally capture the no-
tion of homophily-users who are friends, have
similar attributes, or write similar text are repre-
sented by similar vectors. These representations
could benefit a wide range of downstream appli-
cations, such as friend recommendation, targeted
online advertising, and further applications in the
computational social sciences. Due to limited pub-
licly accessible datasets, we only conduct our ex-
periments on Twitter. However, our algorithms
hold potentials to yield more benefits by combin-
ing different attributes from online social media,
such as Facebook, Twitter, LinkedIn, Flickr12,
References
Apoorv Agarwal, Boyi Xie, Ilia Vovsha, Owen Ram-
bow, and Rebecca Passonneau. 2011. Sentiment
12Images can be similarly represented as vector representa-
tions obtained from CovNet (Krizhevsky et al., 2012), which
can be immediately incorporated into the proposed framework.
analysis of twitter data. In Proceedings of the Work-
shop on Languages in Social Media. Association for
Computational Linguistics, pages 30–38.
Eugene Agichtein and Luis Gravano. 2000. Snowball:
Extracting relations from large plain-text collections.
In Proceedings of the fifth ACM conference on Digi-
tal libraries. ACM, pages 85–94.
Faiyaz Al Zamal, Wendy Liu, and Derek Ruths. 2012.
Homophily and latent attribute inference: Inferring
latent attributes of twitter users from neighbors. In
ICWSM.
John D Burger, John Henderson, George Kim, and
Guido Zarrella. 2011. Discriminating gender on
twitter. In Proceedings of the Conference on Empiri-
cal Methods in Natural Language Processing. Asso-
ciation for Computational Linguistics, pages 1301–
1309.
Kai-Wei Chang, Wen-tau Yih, Bishan Yang, and
Christopher Meek. 2014. Typed tensor decomposi-
tion of knowledge bases for relation extraction. In
Proceedings of the 2014 Conference on Empirical
Methods in Natural Language Processing (EMNLP).
pages 1568–1579.
Zhiyuan Cheng, James Caverlee, and Kyumin Lee.
2010. You are where you tweet: a content-based
approach to geo-locating twitter users. In Proceed-
ings of the 19th ACM international conference on In-
formation and knowledge management. ACM, pages
759–768.
Yejin Choi, Claire Cardie, Ellen Riloff, and Siddharth
Patwardhan. 2005. Identifying sources of opinions
with conditional random fields and extraction pat-
terns. In Proceedings of the conference on Human
Language Technology and Empirical Methods in
Natural Language Processing. Association for Com-
putational Linguistics, pages 355–362.
Morgane Ciot, Morgan Sonderegger, and Derek Ruths.
2013. Gender inference of twitter users in non-
english contexts. In EMNLP. pages 1136–1145.
Ronan Collobert and Jason Weston. 2008. A unified
architecture for natural language processing: Deep
In Pro-
neural networks with multitask learning.
ceedings of the 25th international conference on Ma-
chine learning. ACM, pages 160–167.
Mark Craven, Johan Kumlien, et al. 1999. Construct-
ing biological knowledge bases by extracting infor-
In ISMB. volume 1999,
mation from text sources.
pages 77–86.
Dmitry Davidov, Ari Rappoport, and Moshe Koppel.
Fully unsupervised discovery of concept-
2007.
specific relationships by web mining .
Clodoveu A Davis
Jr, Gisele L Pappa, Diogo
Renn´o Rocha de Oliveira, and Filipe de L Arcanjo.
2011.
Inferring the location of twitter messages
based on user relationships. Transactions in GIS
15(6):735–751.
Ernesto Estrada. 2001. Generalization of topological
indices. Chemical physics letters 336(3):248–252.
Michelle Girvan and Mark EJ Newman. 2002. Com-
munity structure in social and biological networks.
Proceedings of the national academy of sciences
99(12):7821–7826.
Alec Go, Richa Bhayani, and Lei Huang. 2009. Twit-
ter sentiment classification using distant supervision.
CS224N Project Report, Stanford pages 1–12.
Kelvin Guu, John Miller, and Percy Liang. 2015.
Traversing knowledge graphs in vector space. arXiv
preprint arXiv:1506.01094 .
Raphael Hoffmann, Congle Zhang, Xiao Ling, Luke
Zettlemoyer, and Daniel S Weld. 2011. Knowledge-
based weak supervision for information extraction
of overlapping relations. In Proceedings of the 49th
Annual Meeting of the Association for Computa-
tional Linguistics: Human Language Technologies-
Volume 1. Association for Computational Linguis-
tics, pages 541–550.
Dirk Hovy, Anders Johannsen, and Anders Søgaard.
2015. User review sites as a resource for large-scale
sociolinguistic studies. In Proceedings of the 24th
International Conference on World Wide Web.
Sean P Igo and Ellen Riloff. 2009. Corpus-based se-
mantic lexicon induction with web-based corrobora-
tion. In Proceedings of the Workshop on Unsuper-
vised and Minimally Supervised Learning of Lexical
Semantics. Association for Computational Linguis-
tics, pages 18–26.
Thorsten Joachims. 1999. Making large scale svm
learning practical .
Michael Conover, Jacob Ratkiewicz, Matthew Fran-
cisco, Bruno Gonc¸alves, Filippo Menczer, and
Alessandro Flammini. 2011. Political polarization
on twitter. In ICWSM.
Henry Kautz, Bart Selman, and Mehul Shah. 1997.
Referral web: combining social networks and col-
laborative filtering. Communications of the ACM
40(3):63–65.
Michael D Conover, Clayton Davis, Emilio Ferrara,
Karissa McKelvey, Filippo Menczer, and Alessan-
dro Flammini. 2013. The geospatial characteristics
of a social movement communication network. PloS
one 8(3):e55957.
Soo-Min Kim and Eduard Hovy. 2006. Extracting
opinions, opinion holders, and topics expressed in
online news media text. In Proceedings of the Work-
shop on Sentiment and Subjectivity in Text. Associa-
tion for Computational Linguistics, pages 1–8.
Angelika Kimmig, Stephen Bach, Matthias Broecheler,
Bert Huang, and Lise Getoor. 2012. A short intro-
duction to probabilistic soft logic. In Proceedings of
the NIPS Workshop on Probabilistic Programming:
Foundations and Applications. pages 1–4.
Ryan Kiros, Ruslan Salakhutdinov, and Rich Zemel.
2014. Multimodal neural language models. In Pro-
ceedings of the 31st International Conference on
Machine Learning (ICML-14). pages 595–603.
Efthymios Kouloumpis, Theresa Wilson, and Johanna
Moore. 2011. Twitter sentiment analysis: The good
the bad and the omg! ICWSM 11:538–541.
Zornitsa Kozareva and Eduard Hovy. 2010a. Learn-
ing arguments and supertypes of semantic relations
In Proceedings of the
using recursive patterns.
48th Annual Meeting of the Association for Compu-
tational Linguistics. Association for Computational
Linguistics, pages 1482–1491.
Zornitsa Kozareva and Eduard Hovy. 2010b. Not
all seeds are equal: Measuring the quality of text
In Human Language Technologies:
mining seeds.
The 2010 Annual Conference of the North American
Chapter of the Association for Computational Lin-
guistics. Association for Computational Linguistics,
pages 618–626.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hin-
ton. 2012.
Imagenet classification with deep con-
volutional neural networks. In Advances in neural
information processing systems. pages 1097–1105.
J´erome Kunegis and Andreas Lommatzsch. 2009.
Learning spectral graph transformations for link pre-
In Proceedings of the 26th Annual Inter-
diction.
national Conference on Machine Learning. ACM,
pages 561–568.
John Lafferty, Andrew McCallum, and Fernando CN
Pereira. 2001. Conditional random fields: Prob-
abilistic models for segmenting and labeling se-
quence data .
Quoc V Le and Tomas Mikolov. 2014. Distributed
representations of sentences and documents. arXiv
preprint arXiv:1405.4053 .
Jiwei Li, Alan Ritter, Claire Cardie, and Eduard Hovy.
2014a. Major life event extraction from twitter
based on congratulations/condolences speech acts.
In Proceedings of Empirical Methods in Natural
Language Processing.
Jiwei Li, Alan Ritter, and Eduard Hovy. 2014b.
Weakly supervised user profile extraction from twit-
ter. ACL.
Jiwei Li, Alan Ritter, and Dan Jurafsky. 2014c.
In-
ferring user preferences by probabilistic logical
arXiv preprint
reasoning over social networks.
arXiv:1411.2679 .
Miller McPherson, Lynn Smith-Lovin, and James M
Cook. 2001. Birds of a feather: Homophily in social
networks. Annual review of sociology pages 415–
444.
Tomas Mikolov, Kai Chen, Greg Corrado, and Jef-
Efficient estimation of word
arXiv preprint
frey Dean. 2013.
representations in vector space.
arXiv:1301.3781 .
Mike Mintz, Steven Bills, Rion Snow, and Dan Juraf-
sky. 2009. Distant supervision for relation extrac-
In Proceedings of the
tion without labeled data.
Joint Conference of the 47th Annual Meeting of the
ACL and the 4th International Joint Conference on
Natural Language Processing of the AFNLP: Vol-
ume 2-Volume 2. Association for Computational Lin-
guistics, pages 1003–1011.
Jukka-Pekka Onnela, Samuel Arbesman, Marta C
Gonz´alez, Albert-L´aszl´o Barab´asi, and Nicholas A
Christakis. 2011. Geographic constraints on social
network groups. PLoS one 6(4):e16939.
Olutobi Owoputi, Brendan O'Connor, Chris Dyer,
Kevin Gimpel, Nathan Schneider, and Noah A
Improved part-of-speech tagging for
Smith. 2013.
online conversational text with word clusters.
In
HLT-NAACL. pages 380–390.
Alexander Pak and Patrick Paroubek. 2010. Twitter as
a corpus for sentiment analysis and opinion mining.
In LREC.
Marco Pennacchiotti and Ana-Maria Popescu. 2011. A
machine learning approach to twitter user classifica-
tion. ICWSM 11:281–288.
Bryan Perozzi, Rami Al-Rfou, and Steven Skiena.
2014. Deepwalk: Online learning of social represen-
tations. In Proceedings of the 20th ACM SIGKDD
international conference on Knowledge discovery
and data mining. ACM, pages 701–710.
Daniel Preotiuc-Pietro, Vasileios Lampos, and Niko-
laos Aletras. 2015. An analysis of the user occupa-
tional class through twitter content .
Daniel Preot¸iuc-Pietro, Svitlana Volkova, Vasileios
Lampos, Yoram Bachrach, and Nikolaos Aletras.
2015.
Studying user income through language,
behaviour and affect in social media. PloS one
10(9):e0138717.
Delip Rao and David Yarowsky. 2010. Detecting latent
user properties in social media. In Proc. of the NIPS
MLSN Workshop.
Delip Rao, David Yarowsky, Abhishek Shreevats, and
Manaswi Gupta. 2010. Classifying latent user at-
In Proceedings of the 2nd in-
tributes in twitter.
ternational workshop on Search and mining user-
generated contents. ACM, pages 37–44.
Matthew Richardson and Pedro Domingos. 2006.
Markov logic networks. Machine learning 62(1-
2):107–136.
Sebastian Riedel, Limin Yao, Andrew McCallum, and
Benjamin M Marlin. 2013. Relation extraction with
matrix factorization and universal schemas .
Ellen Riloff, Rosie Jones, et al. 1999. Learning dictio-
naries for information extraction by multi-level boot-
strapping. In AAAI/IAAI. pages 474–479.
Alan Ritter, Luke Zettlemoyer, Mausam, and Oren Et-
zioni. 2013. Modeling missing data in distant super-
vision for information extraction. TACL 1:367–378.
Tim Rocktaschel, Sameer Singh, and Sebastian Riedel.
2015. Injecting logical background knowledge into
embeddings for relation extraction. In Proceedings
of the 2015 Human Language Technology Confer-
ence of the North American Chapter of the Associ-
ation of Computational Linguistics.
Adam Sadilek, Henry Kautz, and Jeffrey P Bigham.
2012. Finding your friends and following them to
where you are. In Proceedings of the fifth ACM inter-
national conference on Web search and data mining.
ACM, pages 723–732.
Hassan Saif, Yulan He, and Harith Alani. 2012. Seman-
In The Semantic
tic sentiment analysis of twitter.
Web–ISWC 2012, Springer, pages 508–524.
Richard Socher, Danqi Chen, Christopher D Manning,
and Andrew Ng. 2013. Reasoning with neural ten-
sor networks for knowledge base completion.
In
Advances in Neural Information Processing Systems.
pages 926–934.
Robert Speer, Catherine Havasi, and Henry Lieberman.
2008. Analogyspace: Reducing the dimensionality
In AAAI. volume 8,
of common sense knowledge.
pages 548–553.
Yizhou Sun, Yintao Yu, and Jiawei Han. 2009.
Ranking-based clustering of heterogeneous informa-
tion networks with star network schema. In Proceed-
ings of the 15th ACM SIGKDD international con-
ference on Knowledge discovery and data mining.
ACM, pages 797–806.
Kai Sheng Tai, Richard Socher, and Christopher D
Manning. 2015. Improved semantic representations
from tree-structured long short-term memory net-
works. arXiv preprint arXiv:1503.00075 .
Cong Tang, Keith Ross, Nitesh Saxena, and Ruichuan
Chen. 2011. Whats in a name: A study of
names, gender inference, and gender behavior in
facebook. In Database Systems for Adanced Appli-
cations, Springer, pages 344–356.
Jian Tang, Meng Qu, Mingzhe Wang, Ming Zhang, Jun
Yan, and Qiaozhu Mei. 2015. Line: Large-scale in-
In Proceedings of
formation network embedding.
the 24th International Conference on World Wide
Web. International World Wide Web Conferences
Steering Committee, pages 1067–1077.
Lei Tang and Huan Liu. 2009a. Relational learning via
latent social dimensions. In Proceedings of the 15th
ACM SIGKDD international conference on Knowl-
edge discovery and data mining. ACM, pages 817–
826.
Lei Tang and Huan Liu. 2009b. Scalable learning of
collective behavior based on sparse social dimen-
sions. In Proceedings of the 18th ACM conference
on Information and knowledge management. ACM,
pages 1107–1116.
Sebastian Thrun. 1996. Is learning the n-th thing any
easier than learning the first? Advances in neural
information processing systems pages 640–646.
Svitlana Volkova, Glen Coppersmith, and Benjamin
Inferring user political prefer-
Van Durme. 2014.
ences from streaming communications. In ACL.
William Yang Wang, Kathryn Mazaitis, and William W
Cohen. 2013.
Programming with personalized
pagerank: a locally groundable first-order proba-
bilistic logic. In Proceedings of the 22nd ACM in-
ternational conference on Conference on informa-
tion & knowledge management. ACM, pages 2129–
2138.
William Yang Wang and Diyi Yang. ???? Thats so an-
noying!!!: A lexical and frame-semantic embedding
based data augmentation approach to automatic cat-
egorization of annoying behaviors using# petpeeve
tweets .
Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng
Chen. 2014. Knowledge graph embedding by trans-
lating on hyperplanes. In Proceedings of the Twenty-
Eighth AAAI Conference on Artificial Intelligence.
Citeseer, pages 1112–1119.
Janyce Wiebe, Theresa Wilson, and Claire Cardie.
2005. Annotating expressions of opinions and emo-
tions in language. Language resources and evalua-
tion 39(2-3):165–210.
Bishan Yang and Claire Cardie. 2013.
Joint infer-
ence for fine-grained opinion extraction. In ACL (1).
pages 1640–1649.
Jaewon Yang and Jure Leskovec. 2013. Overlapping
community detection at scale: a nonnegative matrix
factorization approach. In Proceedings of the sixth
ACM international conference on Web search and
data mining. ACM, pages 587–596.
Jaewon Yang, Julian McAuley, and Jure Leskovec.
2013. Community detection in networks with node
In Data Mining (ICDM), 2013 IEEE
attributes.
13th international conference on. IEEE, pages 1151–
1156.
Matthew D Zeiler. 2012. Adadelta: An adaptive learn-
ing rate method. arXiv preprint arXiv:1212.5701 .
Tong Zhang. 2004. Solving large scale linear predic-
tion problems using stochastic gradient descent al-
gorithms. In Proceedings of the twenty-first interna-
tional conference on Machine learning. ACM, page
116.
Jiayu Zhou, Jianhui Chen, and Jieping Ye. 2011. Mal-
sar: Multi-task learning via structural regularization.
Arizona State University .
Appendix
Predicting Preference: Likes or Dislikes:
Here we describe how we extract user pref-
erences, namely, LIKE(USR,ENTITY) and DIS-
LIKE(USR,ENTITY).
Similar to a wide range of work on sentiment
analysis (e.g., (Choi et al., 2005; Kim and Hovy,
2006; Yang and Cardie, 2013; Agarwal et al., 2011;
Kouloumpis et al., 2011; Wang and Yang; Pak and
Paroubek, 2010; Saif et al., 2012)), our goal is to
identify sentiment and extract the target and object
that express the sentiment. Manually collecting
training data is problematic because (1) tweets talk-
ing about what the user LIKES/DISLIKES are very
sparsely distributed among the massive number of
topics people discuss on Twitter and (2) tweets ex-
pressing what the user LIKES exist in a great variety
of scenarios and forms.
To deal with data sparsity issues, we collect train-
ing data by combining semi-supervised informa-
tion harvesting techniques (Davidov et al., 2007;
Kozareva and Hovy, 2010a,b; Li et al., 2014a) and
the concept of distant supervision (Craven et al.,
1999; Go et al., 2009; Mintz et al., 2009):
information
Semi-supervised
harvesting:
We employ a seed-based information-extraction
method: the model recursively uses seed examples
to extract patterns, which are then used to harvest
new examples, which are further used as new
seeds to train new patterns. We begin with pattern
seeds including "I
like/love/enjoy (entity)",
"I
hate/dislike (entity)", "(I think) (entity)
is good/ terrific/ cool/ awesome/ fantastic", "(I
think) (entity) is bad/terrible/awful suck/sucks".
Entities extracted here should be nouns, which
is determined by a Twitter-tuned POS package
(Owoputi et al., 2013).
ation, we train 3 machine learning classifiers:
Based on the harvested examples from each iter-
• A tweet-level SVM classifier (tweet-model 1)
to distinguish between tweets that intend to
express like/dislike properties and tweets for
all other purposes.
to distinguish between like and dislike13.
• A tweet-level SVM classifier (tweet-model 2)
• A token-level CRF sequence model (entity-
model) to identify entities that are the target
of the users like/dislike.
The SVM classifiers are trained using the
SVMlight package (Joachims, 1999) with the fol-
lowing features: unigrams, bigrams, part-of-speech
tags and dictionary-derived features based on a sub-
jectivity lexicon (Wiebe et al., 2005).
The CRF model (Lafferty et al., 2001) is trained
using the CRF++ package14 using the follow-
ing features: unigrams, part-of-speech tags, NER
tags, capitalization, word shape and context words
within a window of 3 words
The trained SVM and CRF models are used to
harvest more examples, which are further used to
train updated models.
Distant Supervision: The main idea of distant
supervision is to obtain labeled data by drawing on
some external sort of evidence. The evidence may
come from a database15 or common-sense knowl-
edge16. In this work, we assume that if a relation
LIKE(USR, ENTITY) holds for a specific user, then
many of their published tweets mentioning the EN-
TITY also express the LIKE relationship and are
therefore treated as positive training data. Since
semi-supervised approaches heavily rely on seed
quality (Kozareva and Hovy, 2010b) and the pat-
terns derived by the recursive framework may be
strongly influenced by the starting seeds, adding in
examples from distant supervision helps increase
the diversity of positive training examples.
An overview of the proposed algorithm showing
how the semi-supervised approach is combined
with distant supervision is illustrated in Figure 3.
Stopping Condition: To decide the optimum
number of steps for the algorithm to stop, we man-
ually labeled a dataset which contains 200 positive
tweets (100 like and 100 dislike) with entities. se-
lected from the original raw tweet dataset rather
than the automatically harvested data. The dataset
contains 800 negative tweets . For each iteration of
data harvesting, we evaluate the performance of the
13We also investigated a 3-class classifier for like, dislike
and not-related, but found the performance constantly under-
performs using separate classifiers.
14https://code.google.com/p/crfpp/
15For example, if datasets says relation ISCAPITAL holds
between Britain and London, then all sentences with mention
of "Britain" and "London" are treated as expressing ISCAPI-
TAL relation (Mintz et al., 2009; Ritter et al., 2013).
16Tweets with happy emoticons such as :-) : ) are of positive
sentiment (Go et al., 2009).
Begin
Train tweet classification model (SVM) and entity
labeling model (CRF) based on positive/negative
data harvested from starting seeds.
While stopping condition not satisfied:
1. Run classification model and labeling model on raw
tweets. Add newly harvested positive tweets and entities
to the positive dataset.
2. For any user usr and entity entity,
if relation
LIKE(USR,ENTITY) holds, add all posts published by
usr mentioning entity to positive training data.
End
Figure 3: Algorithm for training data harvesting for extraction
user LIKE/DISLIKE preferences.
classification models and labeling model on this
human-labeled dataset, which can be viewed as a
development set for parameter tuning. Results are
reported in Table 5. As can be seen, the precision
score decreases as the algorithm iterates, but the
recall rises. The best F1 score is obtained at the
end of the third round of iteration.
iteration 1
iteration 2
iteration 3
iteration 4
tweet-model 1
tweet-model 2
entity label
tweet-model 1
tweet-model 2
entity label
tweet-model 1
tweet-model 2
entity label
tweet-model 1
tweet-model 2
entity label
Pre
0.86
0.87
0.83
0.78
0.83
0.79
0.76
0.87
0.77
0.72
0.82
0.74
Rec
0.40
0.84
0.40
0.57
0.86
0.60
0.72
0.86
0.72
0.74
0.82
0.70
F1
0.55
0.85
0.54
0.66
0.84
0.68
0.74
0.86
0.74
0.73
0.82
0.72
Table 5: Performance on the manually-labeled devset at differ-
ent iterations of data harvesting.
For evaluation, data harvesting without distant
supervision (denoted by NO-DISTANT) naturally
constitutes a baseline. Another baseline (denoted
by one-step-crf) trains a one-step CRF model,
which directly decides whether a specific token cor-
responds to a LIKE/DISLIKE entity rather than mak-
ing tweet-level decision first. Both (NO-DISTANT)
and ONE-STEP-CRF rely on the recursive frame-
work and tune the number of iterations on the afore-
mentioned gold standards. Test set consists of 100
like/dislike property related tweets (50 like and 50
dislike) with entity labels, which are then matched
with 400 negative tweets. The last baseline we
employ is a rule-based extraction approach using
the seed patterns. We report the best performance
model on the end-to-end entity extraction precision
and recall.
Model
semi+distant
no-distant
one-step (CRF)
rule
Pre
0.73
0.70
0.67
0.80
Rec
0.64
0.65
0.64
0.30
F1
0.682
0.674
0.655
0.436
Table 6: Performances of different models on extraction of
user preferences (like/dislike) toward entities.
From Table 6, we observe performance improve-
ments introduced by combining user-entity infor-
mation with distant supervision. Modeling tweet-
level and entity-level information yields better per-
formance than moldeing them in a unified model
(ONE-STEP-CRF).
We apply the model trained in this subsection
to our tweet corpora. We filter out entities that ap-
pear less than 20 times, resulting in roughly 40,000
different entities17.
17Consecutive entities with same type of NER labels are
merged.
|
1601.00139 | 1 | 1601 | 2016-01-02T05:23:48 | Group Centrality for Semantic Networks: a SWOT analysis featuring Random Walks | [
"cs.SI"
] | Group centrality is an extension of the classical notion of centrality for individuals, to make it applicable to sets of them. We perform a SWOT (strengths, weaknesses, opportunities and threats) analysis of the use of group centrality in semantic networks, for different centrality notions: degree, closeness, betweenness, giving prominence to random walks. Among our main results stand out the relevance and NP-hardness of the problem of finding the most central set in a semantic network for an specific centrality measure. | cs.SI | cs |
Group Centrality for Semantic Networks:
a SWOT analysis featuring Random Walks
Camilo Garrido, Ricardo Mora, and Claudio Gutierrez
Department of Computer Science, Universidad de Chile, Chile
{cgarrido,rmora,cgutierr}@dcc.uchile.cl
Abstract. Group centrality is an extension of the classical notion of
centrality for individuals, to make it applicable to sets of them. We per-
form a SWOT (strengths, weaknesses, opportunities and threats) anal-
ysis of the use of group centrality in semantic networks, for different
centrality notions: degree, closeness, betweenness, giving prominence to
random walks. Among our main results stand out the relevance and NP-
hardness of the problem of finding the most central set in a semantic
network for an specific centrality measure.
1
Introduction
The notion of centrality in graphs has been thoroughly studied since it was in-
troduced for social networks, being also applied in other contexts ranging from
physics to information retrieval [11], but as Freeman writes for social networks
[6], “there is certainly no unanimity on exactly what centrality is or on its con-
ceptual foundations, and there is very little agreement on the proper procedure
for its measurement.” Over the years, a great variety of measures of centrality
have been proposed, most of them pointing to the problem of finding the most
“central” (regarding the notion and/or the measure) node in a network [18].
Most of classic work on centrality studies what we could call “individual”
centrality: what is the most central node in a network for a certain measure.
This idea can be extended to a group: Given an integer k > 1, what is the
most central set S of nodes, of size k? This is known as group centrality and
was introduced by Everett and Borgatti [5] for social networks in order to make
individual measures of centrality work with groups of elements.
In this paper we study the application of the idea of group centrality to
semantic networks. To the best of our knowledge, this link has not been explored
before. It can help to shed light on the idea of a “most significative” group of
concepts in a semantic network, which in the age of big (semantic) data, seems
to be indispensable to make sense of notions like summary, table of contents,
compendium, etc. on huge semantic network datasets.
Our goal in this paper is to show the robustness of this application, develop its
properties, and present its opportunities, scope and possibilities and limitations.
We perform the study based on some classical centrality measures like de-
gree, closeness and betweenness [6].1 We present in more detail random walk
centrality which is a measure based on the notion of hitting time (informally:
the expected length of a random walk). This idea has been used with success
in modeling semantic relatedness. The intuition is that the “jump” from one
concept to another is better modeled through random walks rather than using
deterministic movements. This approach has been explored in human memory
[2], lexical relations [8] and Wikipedia corpus [13].
hooves
frog
antlers
mammal
hair
deer
animal
dog
bat
blood
robin
bird
livingthing
antlers
deer
dog
chicken
hooves
hair
animal
frog
chicken
leaves
tree
mammal
bird
livingthing
blood
feathers
cottonwood
tree
bat
red
robin
red
rose
feathers
flower
plant
color
daisy
green
cottonwood
leaves
color
plant
green
flower
rose
daisy
{ livingthing } – { mammal }
{ red, livingthing } – {mammal, plant }
k = 1
k = 2
k = 3
k = 4 { plant, red, animal, livingthing } – { mammal, plant, bird, tree }
{ bird, plant, animal } – { mammal, plant, bird }
Fig. 1: Networks built by novices (left) and experts (right) over the same set of
25 concepts. In the table, for each k it is shown the most central set of k concepts
of the novice (left) and the expert (right) networks.
Let us consider the examples shown in Fig. 1. Schvaneveldt and social sci-
entists [15] made a study of the process of construction of semantic networks.
They chose 25 natural concepts and devised a method to allow communities to
establish semantic relationships among them. The networks were built using two
different communities: on the left by novices (psychology students) and on the
right by domain experts (biologists). What are the most central concepts here?
First, let us point that the four centrality measures mentioned previously, show
remarkable similarity in their outputs (these measures do not consider labels,
only the structure of the graph). On the novices’ network, the four measures
1 Degree centrality selects the node with higher degree; closeness indicates the node
that minimizes the distance from all other nodes in the network to it; Betweenness
selects the node over which pass the higher number of shortest paths between each
pair of nodes. We will define them formally in Section 3.
indicate “livingthing”. As for the biologists’ network they indicate “mammal”.2
Also, the four of them behaved remarkable similar when searching for central
groups of different sizes, as shown in Fig. 1.
What do these examples tell? First of all, they give insights on the possi-
ble relevance and use that group centrality could have for semantic networks.
The contribution of this paper is a study of the assumptions that lie under this
optimistically/naive hypothesis, that is, an analysis of the strengths and oppor-
tunities and weaknesses and obstacles (known as SWOT analysis) of applying
this idea to semantic networks. (In the related work section we will go over the
work that has been done on ranking semantic networks, that touches some facets
of this problem).
The concrete contributions of this paper are:
1. Complexity analysis. We show that for the four previously mentioned cen-
trality measures, the problem of finding a set with optimal centrality score
is NP-hard (and its associated decision problem is NP-complete).
2. We present a small-scale study of samples of semantic networks –because
of the hardness of the computations and the difficulties to get a human
evaluation of what is “central” in a huge network, a challenge by itself– to
show tendencies and illustrate the potential and robustness of the idea of
group centrality.
3. We develop the notion of random-walk centrality, and prove several theoret-
ical properties for it. In particular, we show that testing random-walk group
centrality can be done in polynomial time.
The paper is organized as follows. In Section 2 we present the main arguments
of the SWOT analysis and related work on the subject. In Section 3 we show
the study of samples of some real semantic networks. In section 4 we include the
proofs and support for the theoretical arguments and in section 5 we present the
conclusions.
2 Group Centrality in Social Networks
Let us begin by stating formally the notions to be studied.
Basic graph notions. An undirected and simple graph is a pair G = (V, E) where
E ⊆ [V ]2, and [V ]2 is the set of all 2-elements subsets from V . The elements of
V are the vertices of G, and the ones from E are its edges. When necessary, we
will use the notation V (G) and E(G) for those sets. From now on, an element
{u, v} ∈ E will be denoted simply by uv.
A vertex u is said to be a neighbor of another vertex v, when uv ∈ E. Note
that the definition of E implies that v is also a neighbor of u. The set of neighbors
of v will be denoted by NG(v). The degree of v, dG(v) is the size of NG(v).
2 In the experiments, the biologists were asked separately to tell the most represen-
tative concept in the list, and they chose “mammal”.
A graph Pn =(cid:0){v0, v1, ..., vn},{v0v1, v1v2, ..., vn−1vn}(cid:1) with n ≥ 0, where all
vi are distinct is called a path, and the number of edges in it is its length. A cycle
is a special type of path such that v0 = vn. A path Pn in G, with n ≥ 1 such
that v0 = u and vn = v, is called a u-v path. Also G is said to be connected if
for all distinct u, v ∈ V a u-v path exists in G. A connected component of G is
a maximally connected subgraph H.
length of the shortest u-v path in G. For S ⊆ V , d(u, S) := minv∈S d(u, v).
The distance between u and v (cid:0)denoted by dG(u, v) or just d(u, v)(cid:1) is the
In what follows, all our graphs will considered to be simple, undirected and
connected.
Group Centrality The centrality measures that we study were designed to work
with single vertices [6] rather than groups of them, thus we use their natural
extension to sets as proposed in [5].
Let G = (V, E) be a graph and S ⊂ V . We denote by Cj(S) for j ∈ {d, c, bc}
the group centrality according to the centrality measure Cj.
Group degree centrality Cd(S) counts the number of vertices not in S that
are connected to some vertex in S. Multiple ties to the same vertex are counted
only once. Group closeness Cc(S) studies the value of the sum of the distances
from S to all vertices not in it. Note that for group degree, the higher the value
of Cd(S), the more central the set is. Whereas for group closeness the reciprocate
holds. Formally, they are defined as follows:
Cd(S) =
{v ∈ V \ S, N (v) ∩ S (cid:54)= φ}
V \ S
,
Cc(S) =
v∈V \S d(v, S)
V \ S
.
(cid:80)
(cid:88)
u<v
u,v /∈S
Group betweenness centrality Cbc(S) indicates the proportion of geodesics
connecting pairs of vertices (that are not in S), that pass through S
Cbc(S) =
2BC(S)
V \ S(V \ S − 1)
, with BC(S) =
σuv(S)
σuv
,
where σuv is the total number of shortest paths from u to v and σuv(S) is the
number of those paths that pass through S. As for group degree, the higher the
value of Cbc, the better.
2.1 Random Walk Centrality
Random Walks The next definitions come from the work of Lov´asz in random
walk theory [10]. Let G = (V, E) be a graph such that V = n and E = m,
where n, m ∈ N. Formally, a random walk is a sequence of vertices obtained as
follows: it starts at a vertex v0, and if at the t-th step it is at a vertex vt = u, it
moves to a neighbor v of u with probability puv = 1/d(u).
Let S ⊆ V . The hitting time for a set H(u, S) is the expected number of
steps that a random walk starting at vertex u takes to reach some vertex in S
for the first time. When S = {v} is a singleton, we will simply write H(u, v).
One of the issues that all three measures of group centrality (presented on
2.1) have in common, is that they only take into consideration shortest paths
between pairs of vertices. A semantic network represents relations between con-
cepts, motivated by an specific context. Should that context change, new and
more direct connections may arise. In that regard, as much connections as pos-
sible should be taken into consideration, which is precisely what random walk
centrality attempts to do.
Definition 1 (Random Walk Centrality). The random walk centrality of
S ⊂ V is
(cid:80)
v∈V \S H(v, S)
V \ S
h(cid:46)(S) =
.
Essentially random walk centrality is a variation of closeness that takes into
consideration all ways to reach S from a vertex v, rather than only through the
shortest path. As for group closeness, the lower the value of h(cid:46)(S), the more
central the set is.
2.2 The arguments for the SWOT analysis
Let us state them in general terms the arguments of the SWOT analysis.
Obstacles: weaknesses and threats. First of all, the problem is computationally
hard; we prove that for all four measures considered it is NP-hard:
Theorem 2. The problem of finding an optimal solution S of size k, is an NP-
hard problem for each of the four group centrality measures studied in this paper
(that is, for C ∈ {Cd, Cc, Cbc, h(cid:46)}) .
This theorem points to the difficulties of applying the notion to big data (and
in this paper for experimenting with it) and gives a first program of research on
this problem: to develop good approximate algorithms for it.
Second, there is a difficulty associated to experimenting with big data due to
the lack of good benchmarks and semantically “marked” data, that is, networks
for which we have a clear agreement on what are their most central concepts at
different granularities (i.e. the size of an optimum candidate). This is a challenge
by itself (the scope of human physical abilities) and that is the reason why we
use small (human-scale) datasets in this paper.
Third, the quality of data. It turns out that the output of the centrality and
group centrality methods depends heavily on the “quality” of the network: in
other words, in how well it represents the semantics we are looking for. The ex-
amples we discussed on section 1 are prime cases of excellent semantic networks
(in fact, developed by psychologists based on systematic experimentation). Un-
fortunately, the semantic networks that we see in real life (for example RDF
data) are usually not as good. We will show through several sample datasets
that many of our current most popular semantic data representations (e.g. RDF
graphs) have a bias as they are mostly constructed in ad-hoc manners (union
of pieces of information scattered from different sources; without the goal of
completeness, etc).3
Strengths and Opportunities. First, the notion of group centrality seems to be
worth exploring in the quest for developing summaries, small representations or
give insight about the meaning and contents of big semantic network datasets.
We show samples of different types of networks, including stable RDF datasets
like DBpedia. The results were in all cases encouraging.
Second, the notion seems to be robust and become more stable as the size
of the optimum candidate k increases. There are indications from the sample
data we worked out and partial theoretical results (Proposition 7), that in the
long term, all measures behave in the same way as k converges to V /2 (that
is, there exists a set that is optimum for all four measures). From this we have
a second line of research: to develop concluding analytical and real data studies
that can support this conjecture.
Third, the notion of random walk shows in general less variance than other
measures, i.e. gives more “unique” solutions for k > 0. This virtue and the fact
that this measure systematically gives solutions that have elements in common
with those obtained through the use of other measures, are good signals of its
significance. If we link this observation to the fact that its theoretical basis
(the notion of random walk) has been successfully used in semantic networks
(see Related Work section) there is good ground to make it a candidate for
systematic study at the theoretical and experimental level.
2.3 Related Work
Random Walks and Semantic Networks. The motivation to address random walk
centrality comes from works that have used the notion of random walks and re-
lated it with semantics. The work of Abbott et al. [2] compares the functioning
of the human mind when searching for memories with a random walk in a se-
mantic network, as both present a similar behavior. They conclude that these
results can help clarify the possible mechanisms that could account for PageR-
ank predicting the prominence of words in semantic memory. Yeh et al. [13]
use random walks to determine the semantic relatedness of two elements. In
particular, they use a personalized PageRank algorithm with a custom teleport
vector. They conclude that random walk combined with personalized PageRank
is a feasible and potentially fruitful mean of computing semantic relatedness for
words and texts. Hughes et al. [8] introduce a new measure of lexical relatedness
based on the divergence of the stationary distributions computed using random
3 We are well aware of the complexities of determining in a network of concepts which
one is more relevant. This relies on the underlying semantics of the notions involved,
on the domain from which they are taken from, on the goals that the people that
built the network had, and finally, on the quality of the network itself. Probably due
to all of the above, today we do not have standard, nor good benchmark data for
these tasks.
walks over graphs extracted from WordNet. All these works have been valuable
sources of inspiration for our method.
Relevance and quality ranking in the Semantic Web. As our motivation of
studying group centrality points to its use in the area of relevance in Semantic
data, below we discuss some related work made on it. Graves et al.[7] propose
a graph-theoretic measure called noc-order for ranking nodes in RDF graphs.
They base their ranking in a variation of closeness centrality, which measures
how costly is to reach all the nodes in the graph from one specific vertex. Note
that this idea does not use the concept of expected length (via random walks).
Zhang et al. [14] address the problem of summarizing RDF graphs using an
RDF sentence as the atomic unit. They focus in ranking the RDF sentences to
get a summary of the ontology and base their method in centrality measures
such as degree-, eigenvector-, or betweenness-centrality. Cheng et al. [3] address
the problem of ranking features of an entity in a semantic graph. Their method
relies on centrality notions enhanced with the information captured by labels.
The topics of these works is related to our work, though we focus in finding
central elements in the graph as a whole, while they do it in summarizing en-
tities/ontologies; also they do not present a conceptual notion of key elements,
rely only on experimental evaluations and do not address group centrality. Ding
et al. [4] address ranking of semantic objects on the Web. They concentrate on
documents, terms and RDF graphs as a whole. For terms, they use popularity
measures essentially based on use and population, weighted by some parameters
of the dataset (occurrences, etc.). Supekar et al. [12] address the problem of
quality of ontologies. They characterize ontological features to calculate a score
in order to determine the quality of the ontology. They characterize features like
how well the ontology reflects the domain of interest, correctness of the ontology
with respect to the domain of discourse, how easy the ontology is to understand
or the depth and width of the ontology. These last two papers deal with ranking
of ontologies as such, but do not address the problem of finding a set of relevant
concepts/nodes in RDF graphs, nor present any conceptual framework to define
them.
3 Group Centrality on Semantic Network Data
We are interested on knowing what the group centrality approach applied to
semantic networks has to offer. In order to do this, we apply group degree,
closeness, betweenness, and random walk centrality notions to semantic networks
data and study their behavior. As was stated on section 2, we restrict our analysis
to small networks for two reasons: finding a set with optimal value for an specific
centrality measure is NP-hard; and small networks (human scale) allow to get
an insight about which concepts these group centralities select and how those
selections differ from each other.
3.1 Datasets
Networks of natural concepts [15]. This is the dataset we presented in the
introduction. There we showed two networks built, respectively, by a group of
undergraduate psychology and by graduate biologist. Additionally, we show here
two other networks. A group of students from introductory psychology courses
were asked to rate the degree of relatedness of all pairwise combinations of 25
natural concepts. Using this information the authors constructed two networks,
according to two different thresholds of concept relatedness.
Perception: semantic network of common sense [16]. This network stores
knowledge about what things look and feel like. It has about 9, 000 manually
annotated relations between 4, 000 mundane concepts, e.g, cat, house, snow, etc.
DBpedia Categories [1]. This dataset covers all the Wikipedia article cate-
gories and how they are related using the SKOS vocabulary. The network covers
1, 130, 342 categories with 2, 308, 585 relations between them.
Roget’s Thesaurus This is a widely used English-language thesaurus that
contains more than 15, 000 words. This network was been used in works studying
semantic similarity and sense disambiguation, among others.
For each of the last 3 networks we extracted samples of 40 nodes using
random-walk sampling according to Leskovec and Faloutsos [?] until 40 different
nodes were visited and then we extracted the induced subgraph. Here we show
two such samples for each dataset.
{11}
closeness
degree
{11}
{4, 11}
{4, 11, 13} , {0, 11,
n
1
{2, 13} , {4, 11}
2
{0, 11, 13} , {4, 11,
3
4 {2, 4, 11, 13} , ... (23) {2, 4, 11, 13} , ... (5)]
{0, 11, 13}
{2, 4, 11, 13}
Table 1: Solutions for group centrality on Proximity Semantic Network 1.
{4, 11, 13}
{2, 4, 11, 13}
{11}
{4, 11}
{11}
{4, 11}
13} , ... (9)
13} , ... (4)
random-walk
betweenness
degree
{11}
{0, 11}
n
1
2
3
4 {2, 4, 13, 21} , ... (16)
closeness
{11}
{2, 11} , ... (2)
{2, 13, 21}
{2, 4, 13, 21}
{2, 13, 21} , ... (8)
{2, 13, 21}
{2, 4, 13, 21}
Table 2: Solutions for group centrality on Proximity Semantic Network 2.
{2, 11, 13, 21}
{2, 11, 13}
betweenness
{11}
{2, 11}
random-walk
{11}
{2, 11}
closeness
{6}
betweenness
{6} , {17}
degree
{0, 6}
{0, 6} , {10, 17}
{3, 13, 17} , {3, 10,
n
1
2
3
4 {0, 6, 10, 17} , ... (18) {0, 6, 10, 17} , ... (6)
Table 3: Solutions for group centrality on Proximity Semantic Network con-
structed by non-biologist students.
{3, 10, 17}
{0, 6, 10, 17}
{0, 13, 17}
{0, 6, 10, 17}
17}, ... (10)
{3, 10, 17}
{6}
{0, 6}
{17}
{0, 6}
random-walk
n
1
2
3 {3, 12, 19}, {3, 12, 21} {3, 12, 21}, {3, 12, 22}
closeness
{12}
{3, 12}
degree
{12}
{3, 12}
4
{3, 12, 19, 23}
{3, 12, 22}, {3, 12, 23}
{3, 12, 22}
{3, 12, 15, 19}, {3, 12, 15, 21} {3, 5, 12, 22}
Table 4: Solutions for group centrality on Proximity Semantic Network con-
structed by biologist students.
{3, 12, 19}, {3, 12, 21}
{3, 12, 19, 23}
{3, 12, 23}
betweenness
{12}
{3, 12}
random-walk
{12}
{3, 12}
degree
{4} , {20}
{4, 20}
n
1
2
{4, 18, 20}
3
4 {4, 6, 18, 20} , {0, 4, 6, 20} {0, 4, 6, 20} {4, 6, 7, 20} {4, 6, 18, 20}
Table 5: Solutions for group centrality on DBpedia sample 1.
closeness
{4, 20}
{0, 4, 20}
{20}
{4, 20}
{4, 7, 20}
{4, 6, 20} , {6, 7, 20}
random-walk
{4, 20}
betweenness
{4}
{4}
degree
{13}
n
1
2 {0, 13} , {1, 13} , ... (2)
3
4
{0, 1, 13} , ... (4)
{0, 1, 3, 13} , ... (1)
betweenness random-walk
closeness
{13}
{12, 13}
{1, 13}
{0, 1, 13}
{0, 1, 13}
{0, 1, 3, 13} {0, 1, 3, 13} {0, 1, 3, 13}
{13}
{0, 13}
{0, 1, 13}
{13}
Table 6: Solutions for group centrality on DBpedia sample 2.
degree
{25}
{14, 25}
n
1
2
3
4 {0, 4, 14, 25} , ... (10)
{4, 14, 25} , ... (5)
closeness
{14}
{5, 14} , {7, 13} , {7, 14} , {8, 13}
{4, 14, 25} , ... (4)
{1, 4, 14, 25} , ... (2)
betweenness
random-walk
{14}
{14, 25}
{4, 14, 25}
{13}
{6, 14}
{0, 6, 14}
{4, 12, 14, 25} {0, 4, 14, 25}
Table 7: Solutions for group centrality on Perception sample 1.
closeness
{29} , ... (1)
degree
{2, 29}
n
1
2
3
4 {2, 8, 29, 39} , ... (2) {2, 26, 29, 39} , ... (1) {2, 8, 10, 29} {2, 26, 29, 39}
Table 8: Solutions for group centrality on Perception sample 2.
{1}
{1, 2}
{1, 2, 8}
{0, 15, 32} , ... (7)
{2, 8, 29} , ... (3)
{29, 39}
{2, 8, 39}
{0, 15} , ... (1)
random-walk
betweenness
{29}
{15}
degree
{10}
{10, 25}
n
1
2
3
4 {10, 17, 25, 39} , ... (87) {10, 17, 25, 39} , ... (87) {10, 17, 25, 35} {10, 17, 25, 39}
Table 9: Solutions for group centrality on Roget’s Thesaurus sample 1.
closeness
{10, 25}
{10, 25, 39}
{10, 25, 35}
{10, 25, 39}
{10, 25, 39}
{25, 35}
{10, 25}
random-walk
betweenness
{35}
{10}
{35}
3.2 Discussion of results
For each semantic network, and each centrality measure, we computed the op-
timal solutions of size 1, 2, 3, and 4. The results are presented in Tables 1 to
degree
{11}
{11, 14}
{37}
{37}
n
1
2
3
4 {11, 14, 29, 37} , ... (440) {11, 14, 29, 37} , ... (440) {11, 14, 26, 37} {11, 14, 29, 37}
Table 10: Solutions for group centrality on Roget’s Thesaurus sample 2.
{11, 14, 29} , ... (6)
{1, 14, 29} , ... (6)
{37}
closeness
{14, 26}
{14, 26}
{11, 14, 29}
{26, 37}
{11, 26, 37}
betweenness
random-walk
10. It is important to note that in some cases there were multiple candidates for
optimal solution, but we only include at most two of them4 (if there are more
than two solutions, there is a mark “... (n)” where n is the number of extra
solutions) .
From the results obtained, we can advance two main observations:
Robustness of group centrality. It is noticeable that in the samples the optimal
solutions for all centrality measures become similar as the size k of the set
increases. Indeed, for groups of size 1, it can observed that there are networks
for which the solution is different, depending on the centrality notion applied
(e.g. Table 13). On the other hand, for groups of size 4, it can be observed that
in most networks there is a common set of nodes selected by all four different
measures used.
Non-ambiguity of group random-walk. As stated previously, for a given size k,
more than one optimal solution may exist for group centrality. Moreover, from
the results it can be inferred that as k increases, so does the number of optimal
candidates for certain particular measures. For example, Table 11 shows that
degree and closeness centrality have 24 (resp. 6) optimal solutions of size 4. There
is even an extreme case in Table 23, where degree and closeness have more than
400 different optimal group solutions. In a small sample of 40 nodes this variance
does not help achieve the proposed goal. However, betweenness centrality seems
to behave in general better in this matter (there is some dispersion, e.g. Table
14, but it is small). Random-walk centrality, on the other hand, is consistently
better in this regard. In our sample selections, it is the only notion for which
there is systematically a unique optimal solution for any size, and for all the
networks.
4 Theoretical aspects of Group Centrality
As we saw in previous sections, group centrality offers interesting new insights
on semantic networks. In this section we study the notion from a theoretical
point of view, focusing in the complexity of computing it.
4.1 Complexity of Group Centrality
Let us define formally the problem of computing the most central set of nodes
in a network under different centrality measures.
4 The complete results, graphs and labels of the networks can be found in the ap-
pendix.
Definition 3 (Optimumset). Given a graph G = (V, E), a positive integer k
and a function C : P(V ) → R, define optimumset(G, k, C) as the problem of
finding a set S ⊂ V , of size k such that
S ∈ arg min
U⊂V ;U=k
C(S).
For a real number α > 0, optimumset(G, k, C, α) will denote the corresponding
decision problem: find if there exists a set S ⊂ V , of size k such that C(S) = α.
Theorem 2. optimumset(G, k, C, α) is NP-complete for all four group central-
ity measures studied in this paper (that is, for C ∈ {Cd, Cc, Cbc, h(cid:46)}).
Proof. First note that it is not difficult to prove that Cd, Cc and Cbc are all com-
putable in polynomial time. Same for h(cid:46) but we will leave that demonstration
for later (see proposition 4). Thus the corresponding problems are all in NP.
Now, consider α = 1 and let Sd, Sc, Sbc and Sh(cid:46) be solutions of the corre-
sponding decision problems.
Note that Cc(Sc) = 1 (resp. Cd(Sd) = 1) if and only if Sc (resp. Sd) is a
dominating set of cardinality k in G. That is, the problem of finding a dominating
set of cardinality k in G is polynomial time reducible to optimumset(G, k, α, Cc)
and optimumset(G, k, α, Cd).
Indeed, Cc(Sc) = 1 if and only if d(v, Sc) = 1,∀v /∈ Sc, which happens if
and only if every vertex v /∈ Sc has at least one neighbor on Sc (which is also a
necessary and sufficient condition for Cd(Sd) = 1). That is, Sc (and also Sd) is
a dominating set.
On the other hand, Cbc(Sbc) = 1 (resp. Ch(cid:46) (Sh(cid:46) ) = 1) if and only if Sbc
(resp. Sh(cid:46) ) is a vertex cover of G of size k. That is, the problem of finding a vertex
cover of G of size k is polynomial time reducible to optimumset(G, k, α, Ch(cid:46) )
and optimumset(G, k, α, Cbc).
Indeed, Cbc(Sbc) = 1 if and only if σuv(Sbc) = σuv,∀u, v /∈ Sbc with u < v,
which occurs if and only if for every pair ∀u, v /∈ Sbc with u < v, all shortest
paths that connect u and v have some vertex in Sbc. This is equivalent to having
N (v) ⊆ Sbc for all v ∈ Sbc. That is, Sbc is a vertex cover of G.
Finally, Ch(cid:46) (Sh(cid:46)) = 1 if and only if h(v, Sh(cid:46)) = 1,∀v /∈ Sh(cid:46) , but h(v, Sh(cid:46)) =
1 occurs if and only if N (v) ⊂ S,∀v /∈ Sh(cid:46) . That is, Sh(cid:46) is a vertex cover of G.(cid:117)(cid:116)
4.2 Properties of Random Walk Centrality
the transition probabilities for puv = ω(uv)/ω(u), with ω(u) =(cid:80)
Given a graph G = (V, E), we can generalize the notion of random walk to the
case where there is a weight function ω : E → R on the edges of G, by changing
w∈N (u) ω(uw)
(thus the case ω ≡ 1 gives the original definition of random walk).
Note that the sequence of random vertices (vt : t = 0, 1, ...) is a Markov
chain. Let Pt will denote the distribution of vt: Pt(v) = P(vt = v). The vertex
v0 may be fixed, but may also be drawn from an initial distribution P0. This
easily proved that the distribution π(v) := ω(v)/(cid:80)
initial distribution is said to be stationary if P1 = P0 (which will imply that
Pt = P0 ∀t ≥ 0, because of the construction of the random walk). It can be
v∈V ω(w) is stationary for
every graph G and weight function ω. From now on π will be referred simply as
the stationary distribution (it is not difficult to prove that this distribution is
unique, which makes this reference valid).
Proposition 4. Given G = (V, E), a weight function ω : E → R, a subset
S ⊂ V and v ∈ V , then h(v, S) and h(cid:46)(S) can be computed in polynomial time.
For the proof, first we need the following result
Lemma 5. For a graph G = (V, E), a weight function ω and S ⊆ V define GS
as the following weighted graph:
V (GS) = (V \ S) ∪ {vS}
E(GS) = {e ∈ E/e ∩ S = φ} ∪ {e = uvS/u ∈ V, NG(u) ∩ S (cid:54)= φ}
and ωS : E(GS) −→ N defined as: ωS(uv) = ω(uv) if uv ∩ S = ∅. Otherwise,
i.e. if v = vS, then ωS(uv) = (cid:80)
ω(uw).
Then for u ∈ V \ S we have that HG(u, S) = HGS (u, vS).
w∈NG(u)∩S
Proof. For a vertex w ∈ V \ S, let Tw,S and Tw,vS be two random variables
that count the number of steps that a random walk (on G and GS respectively)
starting at vertex w takes in order to reach some vertex of S and vS (respectively)
for the first time. We will prove that Tu,S and Tu,vS have the same distribution.
Let us proceed by induction on the possible values of Tu,S and Tu,vS .
If n = 1 then
(cid:80)
P(Tu,S = 1) =
w∈NG(u)∩S
ω(u)
=
ωS(uvS)
ωS(u)
= P(Tu,vS = 1).
ω(uw)
Let n > 1 and assume that the property holds for n − 1. Then
ω(uw)
ω(u)
ω(uw)
ω(u)
P(Tw,S = n − 1)
P(Tw,vS = n − 1)
P(Tu,S = n) =
=
=
=
w∈V \S
(cid:88)
(cid:88)
(cid:88)
(cid:88)
w∈V \S
w∈V (GS )\vS
w∈V (GS )\vS
ω(uw)
ω(u)
(Tw,vS = n − 1)
ωS(uw)
ωS(u)
P(Tw,vS = n − 1) = P(Tu,vS = n).
Using this we have that HG(u, S) = E(Tu,S) = E(Tu,vS ) = HGS (u, vS).
(cid:117)(cid:116)
Proof (of proposition 5). Let P be the probability transition matrix associated
to a random walk on G, and the function ω. That is, Pij = ω(ij)
ω(i) . Define P∞ as
.
π(1) π(2) . . . π(n)
...
...
π(1) π(2) . . . π(n)
. . .
...
P∞ =
where π is the stationary distribution. Define also Z := (I − (P − P∞))−1 and
let H be the hitting time matrix of G, that is, Hij = HG(i, j). Then it can be
proved as stated in [17] that for i, j ∈ V
HG(i, j) =
Zjj − Zij
π(j)
.
(1)
Therefore, by combining equation (1) and lemma 5 we can compute h(v, S)
(cid:117)(cid:116)
by inverting a suitable matrix Z, which can be made in polynomial time.
Lemma 6. There is a parameterized family of graphs for which Random walk
centrality behaves systematically differently than degree, closeness and between-
ness group centrality.
Proof. Consider {Gm
n }, with n, m ∈ Z+ and n << m a family of graphs con-
structed as follows (see Fig. 2): for i ∈ {1, ..., m − 1}, K i
n is a subgraph of Gm
n
that is a n-clique and that is connected to v through an unique vertex ki
n. On
n of height 1 and size n with root tn
the other hand, T i
i the
only vertex in T i
n . Suppose we want to find a set
n , m + 1, C) for C ∈ {Cd, Cc, Cbc, h(cid:46)}. It is not
S that solves optimumset(Gm
difficult to prove that for group closeness and betweenness centrality the sets
n } are solutions. For group degree
SK = {v, k1
n, ..., km
n } for any 1 ≤ j ≤ m
centrality ¯SK = {tj
are possible solutions. However, one could argue the difference in connectivity
between sets SK and ST should be accounted for. Indeed, for random walk cen-
(cid:117)(cid:116)
trality, Sk (a more connected version of St) is the only solution.
n } and ST = {v, t1
n, k1
n that is neighbor of v in Gm
n } and ¯ST = {kj
n is a subtree of Gm
n, t1
n, ..., tm
n, ..., km
n, ..., tm
The following bound shows an interesting relation between random walk
centrality, degree and closeness
Proposition 7. There exists a real constant c > 0 such that for S ⊂ V
(cid:88)
v∈V \S
≤ cV \ S3.
d(u, S)
d(v)
(cid:88)
u∈V \S
h(cid:46)(S) ≤
1
V \ S
Proof. Define ∂(S) := {v ∈ S : ∃u ∈ V \ S, uv ∈ E}. Define the subgraph GS
of G as follows: V (GS) := (V \ S) ∪ ∂(S) and E(GS) := {uv ∈ E : u, v ∈
V (GS) ∧ u ∈ V \ S}.
First note that for any pair of vertices u ∈ V \ S and v ∈ ∂(S)
HG(u, S) = HGS (u, ∂(S)) ≤ HGS (u, v).
K 1
n
K 2
n
. . .
Km−1
n
Km
n
v
. . .
T 1
n
T 2
n
T m−1
n
T m
n
Fig. 2: Which set of size m + 1 is more central? SK or ST ?
The equality follows from the fact that the subgraph traversed and the target
hit in both expressions are the same, and the inequality holds because the target
of the expression on the right is more difficult to hit, thus needs more walking
around.
Using this, we have that
(cid:80)
(cid:80)
h(cid:46)(S) =
u∈V \S HG(u, S)
V \ S
≤
u∈V \S HGS (u, vu)
V \ S
,
with each vu ∈ ∂(S) chosen such that dGS (u, vu) = dG(u, S).
pair of vertices u, v ∈ V (L), HL(u, v) ≤ 2E(L)dL(u, v), we have that
On the other hand, by using the known fact that for any graph L and any
(cid:80)
u∈V \S 2E(GS)dGS (u, vu)
(cid:80)
u∈V \S HGS (u, v)
V \ S
≤
≤
1
V \ S
V \ S
(cid:88)
u∈V \S
dG(u, S)
(cid:88)
. (cid:117)(cid:116)
dG(v)
v∈V \S
5 Conclusions
We perform a SWOT analysis of the application of the notion of group centrality
to semantic networks, putting a particular focus on random walk group central-
ity. The challenges faced were mainly of two types. First, the ambiguity of the
very notion of the “most central” concept in a semantic network; and second,
the size of the networks.
In spite of these obstacles (that we discussed in the paper), we hope we
presented enough evidence that this topic deserves to be explored and researched.
We approached these obstacles by working with small samples and advancing
theoretical results. Besides stating the problem at the formal level, we proved
that the problem of computing group centrality is NP-hard, and showed that
random walk centrality comes up as a good candidate for capturing the notion
of centrality on semantic networks.
The above opens at least two lines of research on this topic: First, to find good
approximation algorithms to compute group centrality particularly for random
walk centrality. Second, the task of building a good benchmark of big seman-
tic networks together with their candidate sets to match the most important
concepts in the network.
References
1. DBpedia. http://www.dbpedia.org.
2. J. Abbott, J. Austerweil, T. Griffiths. Human memory search as a random walk
in a semantic network. Adv. Neur. Inf. Proc. Syst. 25, 3050-58, 2012.
3. G. Cheng, Th. Tran, Y. Qu. RELIN: Relatedness and Informativeness-based Cen-
trality for Entity Summarization. ISWC 2011, Springer 2011.
4. L. Ding, R. Pan, T. Finin, A. Joshi, Y. Peng, P. Kolari. Finding and ranking
knowledge on the semantic web. In ISWC 2005, 156–170. Springer, 2005.
5. M. G. Everett, S. P. Borgatti. The centrality of groups and classes. Journal of
Mathematical Sociology. 23(3): 181-201.
6. L. C. Freeman. Centrality in Social Networks. Conceptual Clarification. Social
Networks, 1 (1978/79) 215-239.
7. A. Graves, S. Adali. A method to rank nodes in an rdf graph. 7th. ISWC, 2008.
8. T. Hughes, D. Ramage. Lexical semantic relatedness with random graph walks.
2007 J. Conf. on Emp. Meth. in NLP and Comp. Nat. Lang. Learning, 581-589.
9. Jure Leskovec and Christos Faloutsos. Sampling from large graphs. Proc. 12th
ACM KDD ’06, pages 631–636.
10. L. Lov´asz Random Walks on Graphs; A Survey. Bolyai Soc., Math. Studies 2,
1993, pp. 1-46.
11. Mark EJ Newman. The structure and function of complex networks. SIAM review,
45(2):167–256, 2003.
12. K. Supekar, Ch. Patel, Y. Lee. Characterizing quality of knowledge on semantic
web. AAAI Florida AI Research Symp. (FLAIRS-2004), 17–19, 2004.
13. E. Yeh, D. Ramage, Ch. D. Manning, E. Agirre, A. Soroa. Wikiwalk: random walks
on wikipedia for semantic relatedness. 2009 Workshop on Graph-based Meth. for
NLP, 41–49, Stroudsburg, PA, USA, 2009.
14. X. Zhang, G. Cheng, Y. Qu. Ontology summarization based on rdf sentence graph.
16th Intl. Conf. on World Wide Web, 707–716. ACM, 2007.
15. R. Schvaneveldt, F. Durso, and D. Dearholt. Network structures in proximity data.
The psychology of learning and motivation, 1989, pages 249–284.
16. T. De Smedt, F. De Bleser, V. Van Asch, L. Nijs, W. Daelemans. Gravital: natural
language processing for computer graphics. Creativity and the Agile Mind: A Multi-
Disciplinary Study of a Multi-Faceted Phenomenon, 2013, pages 81
17. Howard Levinson An Eigenvalue Representation for Random Walk Hitting Times
and its Application to the Rook Graph.
18. S. Wasserman, K. Faust. Social Netowrk Analysis: Methods and Applications
(Structural Analysis in the Social Sciences). Cambridge Univ. Press, 1994.
A Appendix
Fig. 3: Proximity Semantic Network 1.
Proximity Semantic Network 1 Labels:
– 0.- animal
– 1.- antlers
– 2.- bat
– 3.- bird
– 4.- blood
– 5.- chicken
– 6.- color
– 7.- cottonwood
– 8.- daisy
– 9.- deer
– 10.- dog
– 11.- feathers
– 12.- flower
– 13.- frog
– 14.- green
– 15.- hair
– 16.- hooves
– 17.- leaves
– 18.- livingthing
– 19.- mammal
– 20.- plant
– 21.- red
– 22.- robin
– 23.- rose
– 24.- tree
degree
{11}
{4, 11}
n
1
{2, 13} , {4, 11}
2
{0, 11, 13} , {4, 11,
3
4 {2, 4, 11, 13} , ... (23) {2, 4, 11, 13} , ... (5)]
{4, 11, 13} , {0, 11,
13} , ... (9)
13} , ... (4)
closeness
{11}
betweenness
{11}
{4, 11}
{4, 11, 13}
{2, 4, 11, 13}
random-walk
{11}
{4, 11}
{0, 11, 13}
{2, 4, 11, 13}
Table 11: Solutions for group centrality on Proximity Semantic Network 1.
Fig. 4: Proximity Semantic Network 2.
Proximity Semantic Network 2 Labels:
– 0.- animal
– 1.- antlers
– 2.- bat
– 3.- bird
– 4.- blood
– 5.- chicken
– 6.- color
– 7.- cottonwood
– 8.- daisy
– 9.- deer
– 10.- dog
– 11.- feathers
– 12.- flower
– 13.- frog
– 14.- green
– 15.- hair
– 16.- hooves
– 17.- leaves
– 18.- livingthing
– 19.- mammal
– 20.- plant
– 21.- red
– 22.- robin
– 23.- rose
– 24.- tree
degree
{11}
{0, 11}
closeness
{11}
n
1
2
3
4 {2, 4, 13, 21} , ... (16)
Table 12: Solutions for group centrality on Proximity Semantic Network 2.
{2, 11} , ... (2)
{2, 13, 21}
{2, 4, 13, 21}
{2, 13, 21}
{2, 4, 13, 21}
{2, 11, 13}
{2, 11, 13, 21}
{2, 13, 21} , ... (8)
betweenness
{11}
{2, 11}
random-walk
{11}
{2, 11}
Fig. 5: Proximity network built by non-specialists.
Proximity network built by non-specialists labels:
– 0.- animal
– 1.- antlers
– 2.- bat
– 3.- bird
– 4.- blood
– 5.- chicken
– 6.- color
– 7.- cottonwood
– 8.- daisy
– 9.- deer
– 10.- dog
– 11.- feathers
– 12.- flower
– 13.- frog
– 14.- green
– 15.- hair
– 16.- hooves
– 17.- leaves
– 18.- livingthing
– 19.- mammal
– 20.- plant
– 21.- red
– 22.- robin
– 23.- rose
– 24.- tree
closeness
{6}
{3, 13, 17} , {3, 10,
17}, ... (10)
{6} , {17}
degree
{0, 6}
n
1
2
3
4 {0, 6, 10, 17} , ... (18) {0, 6, 10, 17} , ... (6)
Table 13: Solutions for group centrality on Proximity Semantic Network con-
structed by non-biologist students.
{3, 10, 17}
{0, 6, 10, 17}
{0, 13, 17}
{0, 6, 10, 17}
{0, 6} , {10, 17}
{3, 10, 17}
betweenness
{17}
{0, 6}
random-walk
{6}
{0, 6}
Fig. 6: Proximity network built by biologists.
Proximity network built by biologists labels:
– 0.- animal
– 1.- antlers
– 2.- bat
– 3.- bird
– 4.- blood
– 5.- chicken
– 6.- color
– 7.- cottonwood
– 8.- daisy
– 9.- deer
– 10.- dog
– 11.- feathers
– 12.- flower
– 13.- frog
– 14.- green
– 15.- hair
– 16.- hooves
– 17.- leaves
– 18.- livingthing
– 19.- mammal
– 20.- plant
– 21.- red
– 22.- robin
– 23.- rose
– 24.- tree
n
1
2
3 {3, 12, 19}, {3, 12, 21} {3, 12, 21}, {3, 12, 22}
closeness
{12}
{3, 12}
degree
{12}
{3, 12}
betweenness
{12}
{3, 12}
random-walk
{12}
{3, 12}
4
{3, 12, 19, 23}
{3, 12, 22}, {3, 12, 23}
{3, 12, 22}
{3, 12, 15, 19}, {3, 12, 15, 21} {3, 5, 12, 22}
Table 14: Solutions for group centrality on Proximity Semantic Network con-
structed by biologist students.
{3, 12, 19}, {3, 12, 21}
{3, 12, 19, 23}
{3, 12, 23}
Fig. 7: Sample 1 from Perception Network.
Sample 1 from Perception network labels:
– 0.- Akzidenz-Grotesk
– 1.- Christmas
– 2.- December
– 3.- Frutiger
– 4.- Helvetica
– 5.- Orient Express
– 6.- Trade Gothic
– 7.- amusement park
– 8.- carousel
– 9.- chameleon
– 10.- colorful
– 11.- consciousness
– 12.- deep
– 13.- diverse
– 14.- exotic
– 15.- grotesque
– 16.- heat
– 17.- hot
– 18.- interesting
– 19.- locomotive
– 20.- loud
– 21.- merry-go-round
– 22.- mill
– 23.- miracle
– 24.- mirror
– 25.- mysterious
– 26.- narrow
– 27.- repetitive
– 28.- roller coaster
– 29.- sans
– 30.- sans serif
– 31.- season
– 32.- sensational
– 33.- steam
– 34.- summer
– 35.- telescope
– 36.- turkey
– 37.- uniform
– 38.- waltz
– 39.- winter
degree
{25}
{14, 25}
n
1
2
3
4 {0, 4, 14, 25} , ... (10)
{4, 14, 25} , ... (5)
closeness
{14}
{5, 14} , {7, 13} , {7, 14} , {8, 13}
{4, 14, 25} , ... (4)
{1, 4, 14, 25} , ... (2)
{14}
{14, 25}
{4, 14, 25}
{13}
{6, 14}
{0, 6, 14}
betweenness
random-walk
{4, 12, 14, 25} {0, 4, 14, 25}
Table 15: Solutions for group centrality on Perception sample 1.
Fig. 8: Sample 2 from Perception Network.
Sample 2 from Perception network labels:
– 0.- New Age
– 1.- anger
– 2.- astrology
– 3.- baby
– 4.- bad
– 5.- calm
– 6.- cathedral
– 7.- child
– 8.- cirrus
– 9.- cold
– 10.- comedy
– 11.- competition
– 12.- dangerous
– 13.- dark
– 14.- deep
– 15.- evolution
– 16.- food
– 17.- fuzzy
– 18.- gene
– 19.- joke
– 20.- kid
– 21.- laugh
– 22.- light
– 23.- loiter
– 24.- negative
– 25.- old
– 26.- ominous
– 27.- park
– 28.- pillar
– 29.- playground
– 30.- reproduction
– 31.- safe
– 32.- shop
– 33.- silent
– 34.- sleep
– 35.- sly
– 36.- supermarket
– 37.- territory
– 38.- valley
– 39.- wicked
{29} , ... (1)
degree
{2, 29}
n
1
2
3
4 {2, 8, 29, 39} , ... (2) {2, 26, 29, 39} , ... (1) {2, 8, 10, 29} {2, 26, 29, 39}
Table 16: Solutions for group centrality on Perception sample 2.
{1}
{1, 2}
{1, 2, 8}
{0, 15, 32} , ... (7)
{2, 8, 29} , ... (3)
betweenness
random-walk
{29}
{29, 39}
{2, 8, 39}
closeness
{15}
{0, 15} , ... (1)
Sample 3 from Perception network labels:
– 0.- Duomo di Milano
– 1.- Middle Ages
– 2.- Milan Cathedral
– 3.- arch
– 4.- black
– 5.- boot
– 6.- calm
– 7.- cathedral
– 8.- dark
– 9.- grassland
– 10.- horse
– 11.- landscape
– 12.- marble
– 13.- mosque
– 14.- pillar
– 15.- prairie
– 16.- smooth
– 17.- soldier
– 18.- stallion
– 19.- veldt
Fig. 9: Sample 3 from Perception Network.
degree
{11}
{4, 11}
n
1
2
3 {0, 11, 13} , {4, 11, 13} , ... (9) {0, 11, 13} , {4, 11, 13} , ... (4) {4, 11, 13}
4
{2, 4, 11, 13} , ... (23)
Table 17: Solutions for group centrality on Perception sample 3.
{2, 4, 11, 13} , ... (5)
{2, 13} , {4, 11}
closeness
{11}
betweenness
random-walk
{11}
{4, 11}
{11}
{4, 11}
{0, 11, 13}
{2, 4, 11, 13} {2, 4, 11, 13}
Fig. 10: Sample 1 from DBpedia Categories network.
Sample 1 from DBpedia Categories network labels:
– 0.- 1996 films
– 1.- 1990s films
– 2.- 2000s in animation
– 3.- Decades in animation
– 4.- 1990s in animation
– 5.- 1910s in animation
– 6.- 1990s animated films
– 7.- Years in animation
– 8.- 1964 anime
– 9.- 1965 anime
– 10.- 1969 anime
– 11.- 1990 anime
– 12.- 1990 in animation
– 13.- 1993 anime
– 14.- 1993 in animation
– 15.- 1994 anime
– 16.- 1994 in animation
– 17.- 1996 anime
– 18.- 1996 in animation
– 19.- 1998 in animation
– 20.- Anime films by year
– 21.- 1990 anime films
– 22.- 1990s anime films
– 23.- 1996 anime films
– 24.- 1981 anime films
– 25.- 1960s anime films
– 26.- 1998 animated films
– 27.- 1996 animated films
– 28.- 1994 animated films
– 29.- 1993 animated films
– 30.- 1990 animated films
– 31.- 1965 anime films
– 32.- 1969 anime films
– 33.- 1945 anime films
– 34.- 1945 anime
– 35.- 1990s American animated films
– 36.- 1990s animated short films
– 37.- Tamil films of 1996
– 38.- 1964 anime films
– 39.- 1990s computer-animated films
degree
{4} , {20}
{4, 20}
n
1
2
{4, 18, 20}
3
4 {4, 6, 18, 20} , {0, 4, 6, 20} {0, 4, 6, 20} {4, 6, 7, 20} {4, 6, 18, 20}
closeness
{4, 20}
{0, 4, 20}
{20}
{4, 20}
{4, 7, 20}
{4, 6, 20} , {6, 7, 20}
{4}
betweenness
random-walk
{4}
{4, 20}
Table 18: Solutions for group centrality on DBpedia sample 1.
Fig. 11: Sample 2 from DBpedia Categories network.
Sample 2 from DBpedia Categories network labels:
– 0.- Wikipedia categories named after populated places in Italy
– 1.- Provinces of Italy
– 2.- Cities and towns in Emilia-Romagna
– 3.- Emilia-Romagna
– 4.- Parma
– 5.- Turin
– 6.- Cities and towns in Piedmont
– 7.- Communes of the Province of Parma
– 8.- Education in Italy by city
– 9.- Communes of the Province of Alessandria
– 10.- Geography of Emilia-Romagna
– 11.- Province of Pavia geography stubs
– 12.- Province of Pavia
– 13.- Bologna
– 14.- Communes of the Province of Bologna
– 15.- Buildings and structures in Bologna
– 16.- Province of Arezzo
– 17.- Province of Ancona
– 18.- Province of Lucca
– 19.- Visitor attractions in Bologna
– 20.- Tortona
– 21.- Novi Ligure
– 22.- Pavia
– 23.- Communes of the Province of Lucca
– 24.- Roman Catholic archbishops of Bologna
– 25.- History of Bologna
– 26.- Education in Turin
– 27.- Palaces in Bologna
– 28.- Education in Bologna
– 29.- University of Pavia
– 30.- Education in Emilia-Romagna
– 31.- Culture in Bologna
– 32.- Culture in Emilia-Romagna
– 33.- Buildings and structures in Pavia
– 34.- Buildings and structures in the Province of Pavia
– 35.- Companies by region of Italy
– 36.- Companies based in Bologna
– 37.- Companies based in Emilia-Romagna
– 38.- Education in Italy by region
– 39.- Media in Bologna
degree
{13}
n
1
2 {0, 13} , {1, 13} , ... (2)
3
4
{0, 1, 13} , ... (4)
{0, 1, 3, 13} , ... (1)
betweenness random-walk
{13}
closeness
{13}
{12, 13}
{1, 13}
{0, 1, 13}
{0, 1, 13}
{0, 1, 3, 13} {0, 1, 3, 13} {0, 1, 3, 13}
{13}
{0, 13}
{0, 1, 13}
Table 19: Solutions for group centrality on DBpedia sample 2.
Fig. 12: Sample 3 from DBpedia Categories network.
Sample 3 from DBpedia Categories network labels:
– 0.- Neognathae
– 1.- Birds by classification
– 2.- Procellariiformes
– 3.- Phoenicopteriformes
– 4.- Pelecaniformes
– 5.- Hornbills
– 6.- Bucerotidae
– 7.- Ardeidae
– 8.- Egretta
– 9.- Aceros
– 10.- Anthracoceros
– 11.- Buceros
– 12.- Ceratogymna
– 13.- Ocyceros
– 14.- Penelopides
– 15.- Bucerotinae
– 16.- Tockus
– 17.- Phalacrocoracidae
– 18.- Fregatidae
– 19.- Falconiformes
– 20.- Accipitridae
– 21.- Oceanodroma
– 22.- Hydrobatinae
– 23.- Hydrobatidae
– 24.- Phalacrocorax
– 25.- Plotopteridae
– 26.- Bucerotiformes
– 27.- Limnofregata
– 28.- Fregata
– 29.- Oceanitinae
– 30.- Aviceda
– 31.- Dryotriorchis
– 32.- Erythrotriorchis
– 33.- Anorrhinus
– 34.- Fregetta
– 35.- Oceanites
– 36.- Botaurus
– 37.- Anhingidae
– 38.- Rhyticeros
– 39.- Pelecanus
degree
{5}
{4, 5}
n
1
2
3
4 {4, 5, 20, 23} {4, 5, 20, 23} {4, 5, 20, 23} {4, 5, 20, 23}
closeness
{1}
{1, 5}
betweenness
random-walk
{1}
{1, 5}
{4, 5, 23}
{1}
{4, 5}
{4, 5, 23}
{4, 5, 23}
{4, 5, 23}
Table 20: Solutions for group centrality on DBpedia sample 3.
Fig. 13: Sample 1 from Roget’s Thesaurus network.
Sample 1 from Roget’s Thesaurus network labels:
– 0.- Jack
– 1.- ardor
– 2.- chips
– 3.- cob
– 4.- cordiality
– 5.- dress down
– 6.- empressement
– 7.- fervency
– 8.- fervidness
– 9.- fervor
– 10.- fire
– 11.- foment
– 12.- give a whipping
– 13.- go
– 14.- gusto
– 15.- heartiness
– 16.- hearty
– 17.- heat
– 18.- jacktar
– 19.- kale
– 20.- lather
– 21.- mariner
– 22.- navigator
– 23.- passionateness
– 24.- sailorman
– 25.- salt
– 26.- sea dog
– 27.- seafarer
– 28.- seafaring man
– 29.- spondulix
– 30.- sugar
– 31.- trim
– 32.- unction
– 33.- vehemence
– 34.- verve
– 35.- warm
– 36.- warm the blood
– 37.- warmth
– 38.- warmth of feeling
– 39.- whale
{10, 25, 39}
{35}
{10}
degree
{10}
{10, 25}
n
1
2
3
4 {10, 17, 25, 39} , ... (87) {10, 17, 25, 39} , ... (87) {10, 17, 25, 35} {10, 17, 25, 39}
Table 21: Solutions for group centrality on Roget’s Thesaurus sample 1.
closeness
{10, 25}
{25, 35}
{10, 25, 35}
{10, 25}
{10, 25, 39}
betweenness
random-walk
{35}
{10, 25, 39}
Fig. 14: Sample 2 from Roget’s Thesaurus network.
Sample 2 from Roget’s Thesaurus network labels:
– 0.- alliterate
– 1.- assonate
– 2.- bedpan
– 3.- blench
– 4.- blink
– 5.- chevy
– 6.- chuck
– 7.- chunk
– 8.- cracked
– 9.- cut
– 10.- dash
– 11.- dodge
– 12.- duck
– 13.- fight shy of
– 14.- fling
– 15.- hoar
– 16.- lay
– 17.- pelt
– 18.- poesy
– 19.- poetry
– 20.- pound
– 21.- pull back
– 22.- pun
– 23.- recoil
– 24.- reft
– 25.- rhyme
– 26.- rime
– 27.- scour
– 28.- shrink
– 29.- shy
– 30.- shy at
– 31.- side step
– 32.- slit
– 33.- song
– 34.- stab
– 35.- submarine
– 36.- swerve
– 37.- verse
– 38.- whack
– 39.- whop
degree
{29}
{26, 29}
n
1
2
3
4 {10, 12, 26, 32} , ... (4) {10, 12, 24, 26} , ... (4) {10, 12, 26, 29} {10, 12, 26, 29}
closeness
{26, 29}
{26, 29}
{10, 12, 26}
{9, 29}
{9, 26, 29}
betweenness
random-walk
{10, 12, 26}
{10, 12, 26}
{9}
{9}
{29}
Table 22: Solutions for group centrality on Roget’s Thesaurus sample 2.
Fig. 15: Sample 3 from Roget’s Thesaurus network.
Sample 3 from Roget’s Thesaurus network labels:
– 0.- ass
– 1.- bring about
– 2.- bring to effect
– 3.- bring to pass
– 4.- burro
– 5.- chucklehead
– 6.- create
– 7.- cuddy
– 8.- dickey
– 9.- do
– 10.- doit
– 11.- donkey
– 12.- dub
– 13.- duffer
– 14.- effect
– 15.- gawk
– 16.- gawkhammer
– 17.- gawky
– 18.- generate
– 19.- get
– 20.- lobster
– 21.- longear
– 22.- looby
– 23.- loon
– 24.- lubber
– 25.- lummox
– 26.- lump
– 27.- lunkhead
– 28.- majority
– 29.- mass
– 30.- moke
– 31.- most
– 32.- plurality
– 33.- preponderance
– 34.- preponderancy
– 35.- put up a tree
– 36.- slouch
– 37.- stick
– 38.- swab
– 39.- the greater number
degree
{11}
{11, 14}
{37}
{37}
n
1
2
3
4 {11, 14, 29, 37} , ... (440) {11, 14, 29, 37} , ... (440) {11, 14, 26, 37} {11, 14, 29, 37}
Table 23: Solutions for group centrality on Roget’s Thesaurus sample 3.
{11, 14, 29} , ... (6)
{1, 14, 29} , ... (6)
{37}
closeness
{14, 26}
{14, 26}
{11, 14, 29}
{26, 37}
{11, 26, 37}
betweenness
random-walk
|
1708.07609 | 2 | 1708 | 2018-05-09T22:22:32 | Clique-based Method for Social Network Clustering | [
"cs.SI",
"physics.soc-ph",
"stat.AP"
] | In this article, we develop a clique-based method for social network clustering. We introduce a new index to evaluate the quality of clustering results, and propose an efficient algorithm based on recursive bipartition to maximize an objective function of the proposed index. The optimization problem is NP-hard, so we approximate the semi-optimal solution via an implicitly restarted Lanczos method. One of the advantages of our algorithm is that the proposed index of each community in the clustering result is guaranteed to be higher than some predetermined threshold, $p$, which is completely controlled by users. We also account for the situation that $p$ is unknown. A statistical procedure of controlling both under-clustering and over-clustering errors simultaneously is carried out to select localized threshold for each subnetwork, such that the community detection accuracy is optimized. Accordingly, we propose a localized clustering algorithm based on binary tree structure. Finally, we exploit the stochastic blockmodels to conduct simulation studies and demonstrate the accuracy and efficiency of our algorithms, both numerically and graphically. | cs.SI | cs |
Clique-based Method for Social Network
Clustering
Guang Ouyang, Dipak K. Dey, Panpan Zhang
May 11, 2018
Abstract
In this article, we develop a clique-based method for social network clus-
tering. We introduce a new index to evaluate the quality of clustering results,
and propose an efficient algorithm based on recursive bipartition to maxi-
mize an objective function of the proposed index. The optimization problem
is NP-hard, so we approximate the semi-optimal solution via an implicitly
restarted Lanczos method. One of the advantages of our algorithm is that the
proposed index of each community in the clustering result is guaranteed to be
higher than some predetermined threshold, p, which is completely controlled
by users. We also account for the situation that p is unknown. A statistical
procedure of controlling both under-clustering and over-clustering errors si-
multaneously is carried out to select localized threshold for each subnetwork,
such that the community detection accuracy is optimized. Accordingly, we
propose a localized clustering algorithm based on binary tree structure. Fi-
nally, we exploit the stochastic blockmodels to conduct simulation studies
and demonstrate the accuracy and efficiency of our algorithms, both numeri-
cally and graphically.
Keywords: Clique-score index, localized clustering algorithm, modularity, so-
cial network, spectral analysis,stochastic block model
1
Introduction
Networks are proliferating all around us, and they appear in different forms, such
as (hardwired) electrical grids or (virtual) social relationships. Networked systems
spread in various scientific and applied disciplines, for example, the Internet, the
World Wide Web, metabolic networks, neural networks, food webs and social net-
works, etc. In this paper, we place our focus on social networks. Social network
1
analysis is a branch of the social science which is an academic discipline studying
a society and the behavior of entities therein. In sociometric or other quantitative
studies, social networks are usually modeled by graph structures consisting of a set
of nodes or vertices connected by directed arcs or undirected edges. More specif-
ically, the actors in a social network are represented by nodes. For each pair of
nodes, if there appears some pattern of ties, they will be connected by an edge.
Edges can be directed or undirected, depending on the feature and interpretation
of the network. There are a variety of social networks, such as the Facebook [29],
marriage networks [25] and smart business networks [24].
Due to solid theoretical foundation of statistics, statistical methodologies have
been a significant input to the studies of social networks. A plethora of statistical
methods have been established and developed to uncover relational structure of
social networks, and dyadic ties between actors therein. In this paper, we focus on a
significant research topic in social networks-clustering. More precisely, our goal
is to develop some statistical methods for accurately clustering actors in a social
network into mutually exclusive communities. This process, in many literatures, is
called "community detection."
Roughly speaking, the fundamental principle underlying social network clus-
tering is that a group of actors who are excessively connected are more likely to
form a community. Technically speaking, the formation of a community requires
that the connections of actors within the community are significantly higher than
the connections between actors from different communities. From the sociological
point of view, the occurrence of high connection density between actors in a com-
munity is usually due to some kind of homology or homogeneity of actors. For
instance, consider a friendship network on Facebook. Intrinsically, students from
the same department of a college are more likely form a friendship community, as
they have a very high probability to know and friend each other. The homogeneities
are reflected in the location parameter (i.e., college) and the academic parameter
(i.e., department). On the other hand, students with different education, social or
geographic background are much less likely to be connected.
Past research on social network clustering can be summarized into two cate-
gories. One approach is to propose a (parametric or nonparametric) probabilistic
graphical model (PGM) which characterizes the community structures of a social
network. Pioneering work of such model-based approach were the p1 model [13]
and the stochastic blockmodel (SBM) [14]. Several successful models were pro-
posed for community detection in the last two decades, including an extension of
the SBM [22], a latent position model [12], a latent position cluster model [11],
and a mixed membership SBM in [1], etc. We refer the interested readers to [10]
for a complete and comprehensive review of PGMs for social network clustering.
Another approach is to consider a metric that can be used to quantitatively evaluate
2
the quality of social network clustering. The task is to specify an objective func-
tion based on the proposed metric, and the ultimate goal is to design an efficient
algorithm to optimize the objective function over all potential clustering strategies.
In this article, we call this class of approaches metric-based methods. Precursory
work in this direction traced back to [28], in which a measure called cluster co-
efficient was proposed to evaluate mutual acquaintance between actors in a social
network. We refer the interested readers to [18] and [19] for extensive discussions
about cluster coefficient. While cluster coefficient is the metric that is often used
for random graphs or dynamic network models, we place our focus on cluster-
ing problems of static networks in this manuscript. Representative and popular
methods for static network clustering are the spectral clustering method [21], the
normalized cut approach [27], and the modularity maximization method [20], etc.
The contribution of this paper is as follows:
1. We propose a novel clustering method based on graph cliques such that the
clique score (a new index defined to evaluate clustering outcomes) of each
identified community is higher than some threshold p.
2. When threshold p is unknown, we develop a systematic strategy to select
localized thresholds p (for each subnetwork which requires further subdivi-
sion) by controlling over-clustering and under-clustering simultaneously.
3. The two algorithms proposed in this manuscript are efficient, easy to imple-
ment, and provide consistently reliable clustering results.
We would like to point out that even though undirected binary social networks
are considered in this manuscript for the sake of interpretation, our method can
be extended to directed or weighted networks in a similar manner, done mutatis
mutandis. The rest of this manuscript is organized as follows. In Section 2, we
introduce some notations that will be used throughout the paper. In Section 3, we
briefly review two classical clustering methods. Section 4 is divided into two sub-
sections. In Section 4.1, we propose a new metric to quantitatively measure the
quality of clustering results, and then establish an objective function based on it,
followed by an algorithm which is designed upon the idea of recursive bipartition
to optimize the objective function. In Section 4.2, we consider the situation that
the global parameter p is not predetermined, and develop a method to compute
localized parameter p so as to maximize the detection accuracy of our algorithm.
We further propose a localized algorithm (corresponding to localized parameter p),
with some modifications on the algorithm developed in Section 4.1. We present
several simulation examples in Section 5 to demonstrate the efficiency of our al-
gorithms and accuracy and consistency of clustering results. In addition, we nu-
merically compare the performance of the proposed algorithm and the traditional
3
modularity maximization algorithm. Lastly, we address some concluding remarks
and potential future studies in Section 6.
2 Notations
In this section, we introduce the notations that will be used throughout the manuscript.
Let G = G(V, A) be an undirected graph to model a social network of V = n
actors, where A is an n × n adjacency matrix such that
in which
aij =
for i, j ∈ V . Let D be an n × n diagonal matrix such that
1,
0,
(cid:40)
deg (1)
0
...
0
D =
A = (aij)n×n,
Nodes i and j are connected,
otherwise,
,
0
deg (2)
...
0
···
···
...
··· deg (n)
0
0
...
where deg (i) represents the degree of node i ∈ V . Consider the matrix L defined
as
L = D − A.
The normalized Laplacian matrix of G [5, page 2] is then given by
L = D−1/2LD−1/2 = I − D−1/2AD−1/2,
where I denotes the identity matrix of rank n.
Two graph invariants on which many classical network clustering methods de-
pend on are volume and cut. The volume of a graph G, denoted Vol(G), is the total
number of degrees of the nodes in G, i.e.,
n(cid:88)
i=1
Vol(G) =
deg (i).
The cut, on the other hand, is defined on subgraphs of G. Let G1 and G2 be two
disjoint subgraphs of G, the cut of G1 and G2, denoted Cut(G1, G2), is the number
of edges linking G1 and G2, i.e.,
(cid:88)
aij.
i∈G1,j∈G2
Cut(G1, G2) =
4
3 Prerequisites
In this section, we give brief reviews of the spectral clustering method with a con-
centration on the normalized cut and the modularity maximization approach as
prerequisites.
3.1 Spectral network clustering
We first look at the spectral network clustering algorithm developed in [27] and [21],
inspired from the spectral graph theory [5]. Generally speaking, a measure called
the normalized cut was adopted as an objective function to quantify the number of
edges across different communities. Suppose that a social network G is split into
h communities G1, G2, . . . , Gh. The normalized cut is given by
h(cid:88)
k=1
Ncut(G1, G2, . . . , Gh) =
Cut(Gk, Gk)
Vol(Gk)
,
(1)
where Gk denotes the complement of Gk in G, for k = 1, 2, . . . , h. Consider
an n × h community indicating matrix P = (pik)n×h, in which each entry is
pik = 1/(cid:112)Vol(Gk) for node i in community Gk. The optimization problem cor-
responding to Equation (1) is, in fact, a discrete trace minimization problem, which
can be approximated by a standard trace minimization problem with a relaxation
of the discreteness condition as follows:
min
G1,G2,...,Gh
Tr(P(cid:62)LP) ≈ min
P∈Rn×h
Tr(P(cid:62)LP),
(2)
subject to P(cid:62)DP = I. Let C = D1/2P. The optimization problem on the right
hand side of Equation (2) is equivalent to
Tr(C(cid:62)LC),
min
C∈Rn×h
subject to C(cid:62)C = I. This is a standard Rayleigh quotient problem, and the solu-
tion of C is composed of the first h eigenvectors of L; see [5] and [15]. Thus, the
solution to the right hand side of Equation (2) can be obtained by solving the first
h eigenvectors in a generalized eigenvalue system Lv = λDv, which can be done
via an adaptive algorithm based on bipartition proposed in [27].
3.2 Modularity maximization
Another popular algorithm for social network clustering, known as the modularity
maximization, was first introduced in [20]. In essence, the underlying principle is to
5
partition a social network into mutually exclusive communities such that the num-
ber of edges across different communities is significantly less than the expectation,
whereas the number of edges within each community is significantly greater than
the expectation. In [20], a bipartition situation was considered, and the clustering
outcome was evaluated by a measure-modularity-defined as
Q =
1
2Vol(G)
s(cid:62)Bs,
where s is an n × 1 column indication vector such that
(cid:40)
si = 1,
si = −1,
if node i belongs to community 1;
if node i belongs to community 2,
and B = (bij)n×n is an n × n matrix with entires
bij = aij − deg (i) deg (i)
Vol(G)
,
for i = 1, 2, . . . , n. Subdivisions for existing communities are available by repeat-
edly implementing the proposed modularity algorithm. The decision of whether or
not subdividing an existing community Gk of size nGk depends on an associated
modularity matrix BGk =
with entries
(cid:16)
(cid:17)
bGk
ij
nGk×nGk
ij = bij − δ(i, j)
bGk
bil,
(cid:88)
l∈Gk
where δ(·,·) denotes the Kronecker delta function. Subdivision for Gk is termi-
nated if the largest eigenvalue of BGk is zero.
4 Clique-based clustering algorithm
In this section, we propose a novel algorithm for social network clustering. More
specifically, we define a new measure called the p-clique index to quantitatively
evaluate the quality of clustering outcome. The section is divided into two subsec-
tions. We introduce a clique-based clustering algorithm in Section 4.1, and then
extend it to a localized clique-based clustering algorithm in Section 4.2. The lo-
calized algorithm enables us to update the threshold p for each subnetwork so as
to well control the potential over-clustering or under-clustering problems. We will
discuss the details in the sequel.
6
4.1 Clique-based clustering algorithm
In this section, We give a clique-based clustering algorithm, inspired from the mod-
ularity maximization algorithm proposed in [20]. In graph theory, a clique is de-
fined as a complete graph on a set of nodes, i.e., each pair of nodes is connected
by an edge. A clique is called maximal if it cannot be extended to any larger-size
clique by including any adjacent node. The ideal clustering outcome is that each
community in a network is a maximal clique. In reality, this is hard to achieve. For
many real world social networks, it is even difficult to guarantee that each com-
munity forms a clique. Therefore, an appropriate measure to assess the degree of
connectivity of a community is needed. The next graph invariant measure can be
used to gauge the internal link density of each community in a network.
Definition 1 (clique score). The clique score of a community (cluster, subnetwork)
is the ratio of the number of observed ties to the number of edges in the clique over
the same number of nodes.
For example, a community consisting of 10 nodes and 18 internal links has
(cid:1) = 0.4. It is obvious that a clique always has clique score one.
clique score 18/(cid:0)10
2
Our goal is to develop an adaptive algorithm to maximize the overall clique scores
(e.g., weighted average) of all communities in a network, subject to the clique
score of each community exceeding some predetermined threshold 0 ≤ p ≤ 1.
The condition of "exceedance" in our algorithm is essential as it guarantees that
none of the communities in our clustering outcome performs extremely bad or does
not achieve the minimum standard. In addition to this, the choice of p is flexible,
depending on the users' needs or the realistic features of communities in a social
network. In Section 4.2, we will discuss how to choose an appropriate value of p to
optimize the performance of the proposed algorithm when no prior information of
p is available. In the next definition, we bridge the gap between clique graph and
p.
Definition 2 (p-clique). A p-clique is a random graph of a set of nodes, of which
each pair is connected by an edge independently with probability p, for 0 ≤ p ≤ 1.
The p-clique defined in this manuscript is not novel, and structurally it is equiv-
alent to the Erdos-R´enyi model proposed by [6]. The definition of p-clique (c.f.
Definition 2) adopts an alternative interpretation of the Erdos-R´enyi model given
by [9]. Since the nodes are connected independently, there is no structure of com-
munities or clusters in p-cliques theoretically. Hence, p-cliques appear to be a
proper benchmark model in our study. The expected number of edges in a p-clique
on n nodes is np(1 − p).
7
G1, G2, . . . , Gh respectively with community sizes n1, n2, . . . , nh such that(cid:80)h
Suppose that a social network G of size n is clustered into h communities,
n. Our task is to search for a clustering rule such that the total degree of nodes in
each community is significantly larger than the expected number of edges of a p-
clique of the same size, whereas the total number of links across different commu-
nities is minimized. Based on this idea, we propose a measure called the p-clique
index as follows, and our ultimate goal is to design an adaptive algorithm for an
optimal clustering rule where the p-clique index is maximized.
k=1 nk =
Definition 3 (p-clique index). Let G be a social network that consists of n nodes,
and C = [G1, G2, . . . , Gh] be a clustering rule which divides G into h communi-
ties. The p-clique index is given by
1
n(n − 1)
(cid:32) h(cid:88)
(cid:0)Vol(Gk) − pnk(nk − 1)(cid:1)
(cid:1)
− (cid:88)
(cid:0)Cut(Gk, Gl) − pnknl
(cid:88)
(cid:0)(aij − p)δ(ci, cj) + (p − aij)(1 − δ(ci, cj))(cid:1), (3)
1≤k(cid:54)=l≤h
1
k=1
D(c, p) =
=
n(n − 1)
1≤i(cid:54)=j≤n
where c = (c1, c2, . . . , cn)(cid:62) is the membership indication vector for nodes.
The essence of p-clique index is to reward connected nodes in the same com-
munity (with 1 − p) and disconnected nodes from different communities (with p),
but penalize connected nodes from different communities (with −(1− p)) and dis-
connected nodes in the same community (with −p). Different from model-based
methods, this approach is not to fit data (i.e., observation) to any type of generative
model or p-clique structures. Rather, our goal is to determine a clustering rule, in
which each community has a higher clique score than a predetermined threshold p.
The algorithm for our clique-based clustering approach is based on the hi-
erarchical clustering algorithm developed for modularity maximization in [20].
To begin with, we consider bipartition, i.e., clustering a social network into two
communities 1 and 2. Define an alternative membership indication vector, s =
(s1, s2, . . . , sn)(cid:62), as follows:
(cid:40)
si = 1,
si = −1,
if node i belongs to community 1,
if node i belongs to community 2,
8
for i = 1, 2, . . . , n. Then, the p-clique index (c.f. Equation (3)) is equivalent to
D(s, p) =
1
n(n − 1)
(aij − p)sisj =
1
n(n − 1)
sT C(p)s,
(cid:88)
1≤i(cid:54)=j≤n
(cid:110)
s(cid:62)
(cid:17)
where the p-clique matrix C(p) is given by
C(p) = A − p(1n×n − In×n),
in which 1n×n is an n × n matrix of all ones. In what follows, the network clus-
tering is converted to an optimization problem
s(cid:62)C(p)s
(cid:111)
(cid:110)
max
(4)
.
s
This is equivalent to
max
s
{D(s, p) − D(1n×1, p)}
1
s(cid:62)C(p)s − 1(cid:62)
max
n(n − 1)
s
(cid:111)
=
=
1
n(n − 1)
max
s
n×1C(p)1n×1
C(p) −
1
n
n(cid:88)
i,j=1
C(p)ij
In×n
,
s
(5)
where C(p)ij denotes the (i, j)th entry of p-clique matrix C(p). Since s = {−1, 1}n
is dyadic, the optimization problem is NP-hard. To relax the problem, our strategy
is to allow s to be any normalized real-valued vector. One solution to Equation (5)
is the eigenvector v = (v1, v2, . . . , vn)(cid:62) corresponding to the largest eigenvalue
In×n, denoted by λmax. We thus obtain a natural
approximation of the solution; that is, we cluster the nodes with respect to the signs
i,j=1 C(p)ij
n
(cid:80)n
of C(p) −(cid:16) 1
of the components of v: (cid:40)
sapprox
i
sapprox
i
= 1,
= −1,
vi ≥ 0,
vi < 0,
for i = 1, 2, . . . , n.
We remark that we would rather work on the optimization problem in Equa-
tion (5) than that in Equation (4) analytically, albeit they are mathematically iden-
tical. The underlying reason is that Equation (5) provides an instinctive insight
of the decision rule of partitioning an existing community. Our rule is that we
do not further subdivide the existing cluster if the eigenvalue λmax ≤ 0 and
9
D(1n×1, p) ≥ 0. Both of the conditions are needed since λmax ≤ 0 indicates
that D(1n×1, p) ≥ max{maxs{D(s, p)}, 0} and D(1n×1, p) ≥ 0 implies that the
link density of the existing community is already higher than p.
We next propose an algorithm based on recursive bipartition for our p-clique
approach. In practice, we first determine the p-clique matrix, C(G)(p) for a given
network G of size n. We partition G into two clusters G1 and G2 with respect to
the signs of the entries of the eigenvector corresponding to the largest eigenvalue
of C(G)(p), and then compute the additional contribution to the p-clique index due
to the division, denoted by
∆D(p) =
1
n(n − 1)
siC(G)(p)ijsj − C(G)(p)ij
.
(cid:88)
(cid:16)
i,j∈G
(cid:17)
We continue to apply the bipartition algorithm respectively to G1 (according to
the associated p-clique matrix C(G1)(p), where C(G1)(p) is a submatrix extracted
directly from C(G)(p) only for the nodes contained in G1) and G2 (according to
the associated p-clique matrix C(G2)(p) obtained in a similar manner), if at least
one of the two regularity criteria is met. We terminate the algorithm until no further
partition is needed. The algorithmic procedure is presented in Algorithm 1.
Algorithm 1: The bipartition-based algorithm for p-clique index maxi-
mization
Input: p-clique matrix C(G)(p) of network G, empty community list L
1 Procedure BiPartition(G, C(G)(p));
2 Compute the eigenvector s corresponding to the largest eigenvalue of
3 Partition G into G1 and G2 according to the signs of entries of s;
4 Compute the additional contribution to p-clique index, ∆D(p);
C(G)(p);
5 if ∆D(p) > 0 or(cid:80)
i,j∈G C(G)
BiPartition(G1, C(G1)(p));
BiPartition(G2, C(G2)(p));
ij (p) < 0 then
6
7
8 else
9
10 end
Add G to community list L
Output: Community list L
To conclude this section, we address some remarks on the proposed algorithm.
The essence of Algorithm 1 is to maximize the additional contributions to the over-
all p-clique index, whenever a division of an existing cluster is implemented (ver-
sus its current state) . As mentioned, the optimization problem is NP-hard, and our
10
algorithm will continue to be executed if(cid:80)
strategy is to relax the problem by allowing s to be any real vector with (cid:96)2 norm
equal to one. Our goal is to determine the eigenvector associated with the largest
eigenvalue of the p-clique matrix, and specifically our strategy is to exploit an im-
plicitly restarted Lanczos method developed in [4]. It is worthy of noting that our
(p) < 0 for some G(cid:48) even
when ∆D(p) ≤ 0, since, under such circumstance, the clique index of G(cid:48) is less
than p. We take the quality (i.e., having clique score higher than predetermined
threshold p) of each resulting community as a higher priority, bearing with non-
optimal grade for the overall p-clique index. We view this as a major feature, as
well as an advantage, of our algorithm, since the choice of threshold p is absolutely
flexible, depending on the users' needs.
i,j∈G(cid:48) C(G(cid:48))
ij
On the other hand, a natural question arising from our algorithm is " what if
there is no specific requirement or prior information for p?" An arbitrary choice of
p may lead to over-clustering or under-clustering problems. Having this concern
in mind, we propose a modified algorithm based on Algorithm 1 in Section 4.2.In
addition, we carry out a statistical procedure for "reasonable" selection of unspec-
ified p so as to simultaneously minimize the errors of over-clustering and under-
clustering.
4.2 Localized clustering algorithm
We have demonstrated that the algorithm proposed in Section 4.1 is advantageous
for social network clustering, as it allows users to control the quality of each re-
sulting community. In many cases, the threshold parameter p is predetermined or
preselected, based on users' needs or past research experience. For the cases that
p is not specified, it is needed to develop a systematic method to choose an opti-
mal assignment of p. Probably a question that needs to be answered in advance
is whether or not there exists such optimum p. An analogous issue occurs for the
modularity maximization algorithm, and was discussed in [26] and [17], where the
conclusion from both was negative for the existence of such overall optimal value
of p. In this section, we consider a localized clustering strategy, the core idea of
which is to select different values of p for different subnetworks.
In essence, our strategy for the selection of p is a process of balancing over-
clustering and under-clustering. Intrinsically, a large value of p usually results in
small sizes of clusters (i.e., over-clustering), whereas a small value of p sometimes
fails to guarantee the quality of clusters (i.e., under-clustering). The procedure that
balances between over-clustering and under-clustering is analogous to dealing with
type I errors and type II errors in statistical hypothesis testings. We borrow these
two terminologies in our study. Let us call the error of over-clustering as type I
error and the error of under-clustering as type II error. Our goal is determine a
11
distinct value of p for each existing cluster such that both types of errors are well
controlled.
Recall the p-clique (i.e., the Erdos-R´enyi model) introduced in Definition 2.
The model is constructed completely at random without any cluster structure, sug-
gesting that we can exploit it to control type I error, and accordingly determine an
upper bound of p. Consider an Erdos-R´enyi graph, ER(n, p0), where n and p0 are
given. Intuitively, we need to set the threshold parameter p significantly less than
p0 so that there is a small probability to divide ER(n, p0) into two subgraphs or
more by our algorithm. Assume that α is the maximal percentage of the number
of nodes to be split from the Erdos-R´enyi graph under tolerance. Our goal turns to
find the largest value of p such that at most αn nodes are split off of the majority.
Let X denote the observed inter-across link density between the split group (i.e.,
αn nodes) and the remainder (i.e., (1− α)n nodes), and FX be its distribution. An
instinctive and reasonable choice of p is thus the αth percentile of FX. Although
the expected number of inter-across edges is (1 − α)np0 in theory, the inter-across
link density is anticipated to be much smaller than the internal density within the
large group of (1 − α)n nodes. In this regard, we approximate FX by a truncated
normal distribution, where the top α is curtailed; that is, X is normally distributed
with mean
(cid:115)
E[X] = p0 − φ(zα)
1 − α
p0(1 − p0)
(1 − α)n
,
and variance
Var[X] =
p0(1 − p0)
(1 − α)n
(cid:34)
(cid:18) φ(zα)
1 − α
(cid:19)2(cid:35)
,
1 − zαφ(zα)
1 − α
−
where φ(·) denote the density function of the standard normal distribution, and zα
is the (1− α)th percentile of standard normal. Thus, we obtain an upper bound for
p as a function of α, which is given by
(cid:111)
(cid:112)Var[X]
(cid:115)
p0(1 − p0)
(1 − α)n
(cid:41)
,
(6)
(cid:110)
(cid:40)
pU(α) = max
= max
0, E[X] − zα
0, p0 − ξ(α)
where ξ(α) is a constant depending on α only.
The next task is to determine the lower bound for p. Suppose that we have two
independent Erdos-R´enyi graphs which are respectively denoted by ER(n1, p1)
and ER(n2, p2), and p12 is the link density between the two graphs. Let β be
the probability that the two graphs are merged. We once again consider normal
12
approximation, and conclude that the two graphs are both statistically significant if
we have
(cid:115)
min{p1, p2} > p12 + zβ
p12(1 − p12)
n1n2
,
(7)
the right-hand side of which in fact forms the lower bound for the parameter p.
However, this lower bound is intractable in general as it requires the information of
p12, which is usually unknown for most real-world networks. Therefore, we place
focus on the upper bound for p, and our strategy is to set the parameter p as large
as possible under tolerance so as to minimize the probability of merging significant
subnetworks.
Finally, we propose a new algorithm for unspecified p based on Algorithm 1
proposed in Section 4.1. The new algorithm is designed for controlling type I
and II errors simultaneously. As mentioned, however, it seems to be unreasonable
to have a global threshold p for our algorithm. We elaborate the reasonings and
demonstrate our solution via the following illustrative example.
Suppose that we cluster a network G of size n into two subnetworks G1 of size
n1 and G2 of size n2. Initially, we adopt a threshold p(G) which depends on the
observed link density of G. However, if we continue to use p(G) as the threshold
when G1 or G2 or both need to be further clustered, we may have to bear with over-
clustering or under-clustering risks in the subdivision processes. Alternatively, we
suggest to reset the threshold(s) before subdividing G1 or G2 or both. In other
words, we choose
(cid:40)
(cid:40)
and
p(G1) = max
p(G2) = max
0, p1 − ξ(α)
0, p2 − ξ(α)
(cid:115)
(cid:115)
(cid:41)
(cid:41)
p1(1 − p1)
(1 − α)n1
p2(1 − p2)
(1 − α)n2
as updated thresholds for G1 and G2, respectively. From then on, the threshold
parameters are refreshed in this manner for all subnetworks that need to be fur-
ther subdivided. As for each subnetwork to which a fresh threshold parameter is
assigned, we call the new algorithm localized clustering algorithm.
One of the most effective ways to illustrate this heirachical clustering process
is probably to exploit binary tree. We start with a root node that represents the
original network which needs to be clustered. At the first level, two child nodes
are attached to the root node. The child nodes represent either subnetworks or
communities. If a child node is a subnetwork which requires further subdivision, it
will carry over two higher-level child nodes and itself turns to an internal node. If
13
a child node is a community which does not need further subdivision, it becomes
a terminal node in the tree. We use rectangle for internal nodes and circle for
terminal nodes. When the algorithm is terminated, the clustering result is reflected
in all the terminal nodes in the binary tree. An example of binary tree is given in
Figure 1.
Network G
Subnetwork G1
Subnetwork G2
Community 1
Subnetwork G3
Community 2
Community 3
Community 4
Community 5
Figure 1: An example of binary tree that represents a result of five communities
We are now in the position of defining a new clique index, the local clique
index, modified from the p-clique index given in Definition 3.
Definition 4 (local clique index). Consider a network G consisting of n nodes. Let
T be the binary tree that describes a hierarchical clustering procedure on G, and
{v} be the collection of all internal nodes of T . The local clique index is given by
(cid:88)
(cid:88)
(cid:16)
LD(T ) =
1
n(n − 1)
v∈T
i(cid:54)=j∈v
(aij − pv)δ(v)
ij + (pv − aij)
1 − δ(v)
ij
,
(8)
(cid:16)
(cid:17)(cid:17)
where pv is the threshold of parameter p for internal node (subnetwork) v.
As shown in Figure 1, our new algorithm is again based on recursive bipartition
procedures, and the goal is to maximize the score function given in Equation (8).
Analogous to Algorithm 1, we need to determine the local clique matrix for each v
in T , i.e.,
C(v) = A(v) − pv (1nv×nv − Inv×nv ) ,
14
where A(v) is the updated adjacency matrix for subnetwork v and nv is the size of
v. Besides, we need to compute the additional contribution from each bipartition
of v to the overall score LD(T ), i.e.,
(cid:16)
(cid:17)(cid:17)
.
(cid:88)
(cid:16)
i(cid:54)=j∈v
∆LD(v) =
1
n(n − 1)
(aij − pv)δ(v)
ij + (pv − aij)
1 − δ(v)
ij
(cid:80)
The algorithm for hierarchical clustering process is terminated as ∆LD(v) ≤
i(cid:54)=j∈v(aij − pv) for all existing v in T , and then all remaining inter-
n(n−1)
nal nodes v turn out to be terminal nodes, i.e., added to the community list. We
summarize our strategy in Algorithm 2, slightly modified from Algorithm 1.
1
Algorithm 2: The algorithm of network clustering based on local clique
index maximization
Input: Network G, binary tree T with a single root node representing
G, tolerance of type I error α
(cid:114)
1 Procedure LocalBiPartition(G);
2 Determine network size nG and the clique score pobs
G − ξ(α)
(cid:41)
3 Compute the threshold pG = max
;
4 Create local clique matrix C(G) = A(G) − pG (1nG×nG − InG×nG);
5 Compute the eigenvector s corresponding to the largest eigenvalue of
G of G;
G (1−pobs
pobs
G )
(1−α)n
(cid:40)
0, pobs
C(G);
6 Partition G into G1 and G2 with respect to the signs of s;
7 Compute additional contribution, ∆LD(G), to overall local clique index
by partitioning G;
(cid:80)
i(cid:54)=j∈v(aij − pG) then
Add G1 to the left child node position of G;
LocalBiPartition(G1);
Add G2 to the right child node position of G;
LocalBiPartition(G2);
8 if ∆LD(G) >
9
10
11
12
nG(nG−1)
1
Output: Binary tree T
5 Simulations
In this section, we conduct some simulation studies to evaluate the performance
of two clique-based clustering algorithms proposed in the previous sections. We
also show that the proposed algorithm outperforms the modularity maximization
approach [20] for network clustering. The SBM in [22] is adopted to generate
15
networks, as they allow us to predetermine the structure of simulated networks,
which can be used as the ground truth for comparisons.
To begin with, we briefly review the SBM. Given a network of n nodes which
belong to h nonempty communities, let n1, n2, . . . , nh be respectively the size of
each community. For i = 1, 2, . . . , n, a mapping c : {1, 2, . . . , n} (cid:55)→ {1, 2, . . . , h}
preserves the membership information for each node labeled with i. An h × h
probability matrix B describes the link densities within every community, as well
as those between different communities, i.e., B =
. Notice that we
Bc(i),c(j)
have to require Bc(i),c(j) ≤ min
for all i, j, when predefining
matrix B; otherwise, the simulated network would be against the clustering struc-
ture obtained by our algorithm(s). Suppose that B (the probabilistic structure of
network) is specified, we are able to simulate the entries in the adjacency matrix
Bc(i),c(i), Bc(j),c(j)
(cid:110)
(cid:111)
(cid:16)
(cid:17)
(cid:16)
(cid:17)
aij = Bernoulli
Bc(i),c(j)
for i < j. Suppose that the network is undirected, we have aji equal aij by sym-
metry. We also assume that all the entries on the diagonal equal 0, as loops are not
considered in our study.
To quantitatively evaluate the performance of algorithms, we adopt two well-
defined robust measures: the normalized mutual information (NMI) and the Ad-
justed Rand Index (ARI), respectively proposed in [8] and in [16]. Suppose that
T is the ground truth of community structure, and S is the clustering result of an
algorithm, the NMI of T and S is given by
−2(cid:80)CT
k=1
k=1 Nk· log
(cid:80)CT
(cid:80)CS
(cid:16) Nk·
(cid:17)
N··
l=1 Nkl log
+(cid:80)CS
(cid:16) NklN··
Nk·N·l
l=1 N·l log
(cid:17)
(cid:16) N·l
N··
(cid:17) ,
NMI(T, S) =
where CT and CS are the number of communities for T and S, respectively; (Nkl)
is a CT × CL confusion matrix, in which Nkl denotes the number of nodes that
should be in community k according to the truth, but are mis-clustered into com-
munity l according to algorithm S; Nk·, Nl·, and N·· are standard definitions of the
sum of the kth row, the sum of the lth column, and the overall sum of the confusion
matrix, respectively. We borrow the same notations and give the definition of ARI
as follows:
(cid:0)Nkl
(cid:80)CS
(cid:80)CT
(cid:104)(cid:80)CT
(cid:1) +(cid:80)CS
(cid:0)Nk·
k=1
l=1
2
k=1
2
(cid:1) −(cid:104)(cid:80)CT
(cid:0)Nk·
(cid:1)(cid:105) −(cid:104)(cid:80)CT
(cid:0)Nl·
k=1
2
(cid:1)(cid:80)CS
(cid:0)Nk·
(cid:0)Nl·
(cid:1)(cid:80)CS
(cid:1)(cid:105)(cid:14)(cid:0)N··
(cid:0)Nl·
(cid:1)
(cid:1)(cid:105)(cid:14)(cid:0)N··
2
2
l=1
(cid:1) .
l=1
2
k=1
2
l=1
2
2
ARI(T, S) =
1
2
16
(cid:16)
(cid:40)
(cid:40)
1,
0,
M (T )
ij =
1,
0,
and
M (S)
ij =
(cid:16)
In addition, we propose a measure analogous to NMI. Once again, assuming
that T is the true structure of the simulated network based on SBMs, and S is the
analogy from our clustering algorithm, we consider two n × n binary matrices,
M (T )
, defined respectively as follows:
(cid:17)
(cid:17)
and
M (S)
ij
ij
Nodes i and j are in the same community in model T ,
otherwise;
Nodes i and j are in the same community by algorithm S,
otherwise.
Our measure is defined on a cluster level, and in the form of an h × h matrix. The
fundamental principle of our measure is simple; that is, we compute the proportion
of mis-clustered nodes, from cluster to cluster. More precisely, for all nodes in
different communities k and l (k (cid:54)= l) under T , the error measure is given by
kl =
i,j s.t.
c(T ) (i)=k,c(T )(j)=l
(cid:88)
(cid:88)
(cid:12)(cid:12)(cid:12)M (T )
ij − M (S)
nknl
ij
(cid:12)(cid:12)(cid:12)M (T )
ij − M (S)
nk(nk − 1)
ij
(cid:12)(cid:12)(cid:12)
,
(cid:12)(cid:12)(cid:12)
.
where nk and nl are the sizes of communities k and l, respectively. The error
measure for a single community k is defined analogously, i.e.,
kk =
i,j s.t.
c(T )(i)=c(T )(j)=k
Ultimately, the error matrix is given by
= (kl)h×h.
We would like to point out that our error matrix measure can be more appealing
in some comprehensive analysis, as it shows where (i.e., in which community or
communities) exactly a large amount of mis-clustering occurs when the clustering
result far deviates from ground truth.
5.1 Effect of threshold p
In the first experiment, we show that a correct choice of threshold p has a significant
impact on the p-clique index clustering algorithm (c.f. Algorithm 1). We simulate
17
Cluster (size: 100) Cluster 2 (size: 10) Cluster 3 (size: 10)
Cluster 1 (size: 100)
Cluster 2 (size: 10)
Cluster 3 (size: 10)
0.2
0.05
0.05
0.05
0.5
0.05
0.05
0.05
0.5
Table 1: Link densities of clusters of a simulated SBM1
an SBM consisting of 120 nodes which are clustered into three communities of
sizes 100, 10 and 10, respectively. The parameters of in-group and cross-group
link densities are summarized in Table 1.
The expected clique score of the simulated SBM1 is approximately 0.1597,
which can be used as an estimate of the overall link density of the network. Let
the error parameter α be equal to 0.025. According to Equation (6), the upper
bound for the threshold p is 0.09541. On the other hand, the lower bound for p
(c.f. Equation (7)) is max{0.0635, 0.0927} = 0.0927. Suppose that we choose
p = 0.09, simulate 100 independent SBM1s, and compute the distance-based error
matrix . Albeit the threshold p is just slightly less than the lower bound, we have
a large under-clustering error for clusters 2 and 3, i.e., 23 = 32 = 0.2218 with
standard error SE(23) = SE(32) = 0.0376, indicating that the probability of
misclassifying nodes from these two clusters is about 22% in average. Suppose
that the value of p is increased to 0.11 (even though p = 0.11 is greater than the
upper bound), the errors 23 = 32 drop dramatically to 0.0614 with standard error
SE(23) = SE(32) = 0.0172. As 23 = 32 ≈ 6% is the largest entry in the
error matrix for p = 0.11, it seems that over-clustering does not bring too much
trouble in this experiment; the reason is that all predefined in-group densities are
significantly larger than cross-group densities. Due to the limit of space, we refer
the interested readers to [23, Sections 4.2 and 4.3] for more analogous examples.
Nevertheless, as long as the parameter p is fairly close to the proposed threshold
selection interval from the above, the clustering outcomes for this experiment (c.f.
Table 1) are under satisfactory. We choose the optimal value of p = 0.9527, and
depict the result in Figure 2, where the three communities are clearly identified and
colored by blue, red and green.
18
Figure 2: Clustering result of SBM1 by Algorithm 2; p = 0.9527
5.2 Global threshold v.s. localized thresholds
The next experiment is designed to compare the performance of the algorithms
proposed in this manuscript. We show that the localized algorithm provides more
reliable outcomes when the threshold p is unknown. We consider SBM2 with 140
nodes which are clustered into three communities of sizes 100, 20 and 20, respec-
tively. The parameters of in-group and cross-group link densities are summarized
in Table 2.
Cluster 1 (size: 120) Cluster 2 (size: 20) Cluster 3 (size:20)
Cluster 1 (size: 120)
Cluster 2 (size: 20)
Cluster 3 (size: 20)
0.2
0.05
0.05
0.05
0.6
0.12
0.05
0.12
0.8
Table 2: Link densities of clusters of simulated SBM2
The expected clique score of the simulated SBM2 is approximately 0.1546.
We again set the error parameter α at 0.025. The associated threshold parameter p
19
according to Equation (6) is 0.0959. As the parameter p is less than the cross-group
link density between communities 2 and 3, it seems to be difficult to separate the
two clusters with the overall threshold p via Algorithm 1. To verify our conjecture,
we simulate 100 independent SBM2s, compute the NMI for each replication, and
take the average as an estimate. We obtain (cid:91)NMI = 0.8488 with standard error
0.0025. In addition, we compute the proposed block-wise distance-based measure,
summarized in Table 3.
Cluster 1
Cluster 2
Cluster 3
11 = 0.0006
12 = 0.0.0012
13 = 0.0012
SE(11) = 0.0011 SE(12) = 0.0003 SE(13) = 0.0003
21 = 0.0012
22 < 10−4
23 = 0.9890
SE(21) = 0.0030
SE(22) < 10−4
SE(23) = 0.0010
31 = 0.0012
32 = 0.9890
33 = 0.0010
SE(31) = 0.0030 SE(32) = 0.0010 SE(23) = 0.0009
Cluster 1
Cluster 2
Cluster 3
Table 3: The block-wise error measure of clustering SBM2 via Algorithm 1
Although community 1 is successfully identified, we observe that the estimate
of error rate between communities 2 and 3 is 23 = 0.9890 = 98.90%, which sug-
gests that the BiPartition algorithm with the global threshold p = 0.0959 fails to
separate these two communities almost surely. This is also reflected in the cluster-
ing result (via Algorithm 1) of the simulated SBM (c.f. Table 2) given in Figure 3.
The entire network is divided into two cluster (rather than three), which are colored
by red and green.
We remedy this problem by applying the localized clustering algorithm, i.e.,
Algorithm 2. We start with the root node (the original simulated SBM) in the bi-
nary tree structure. We use threshold p = 0.0959 to bipartition the root node,
and obtain community 1, and subnetwork 2, where the latter requires for further
clustering. Subnetwork 2 consists of 40 nodes and the expected clique score is
approximately 0.4026. With the same α = 0.025, the updated threshold param-
eter for subnetwork 2 is 0.2536. We would like to mention that in practice we
also compute the threshold parameter for subnetwork 1 (community 1), and obtain
20
Figure 3: Clustering result of SBM2 by Algorithm 1
0.1231 for implementing Algorithm 2, where we find that no further subdivision is
needed. After subdividing subnetwork 2, we continue to compute new thresholds
for both subsequent subnetworks, and ultimately find that they form the other two
communities in the network.
Another 100 independent SBMs are simulated, and the NMI for each simulated
network is computed. The mean estimate is (cid:91)NMIlocal = 0.9677 with standard
error 0.0034. We also compute the block-wise error rates, and present them in
Table 4, where we find a significant improvement of 23 = 32 = 0.0019 versus
0.9890. The small error rate implies that the communities 2 and 3 are significantly
identifiable according to the updated threshold for subnetwork 2. Analogously,
we depict the clustering result of a simulated SBM via Algorithm 2, shown in
Figure 4. The network is successfully divided into three communities as predefined
in Table 2. The communities are colored by blue, red and green.
5.3 The propose algorithm v.s. modularity maximization
Notice that the algorithm proposed in this manuscript is inspired from the mod-
ularity maximization algorithm.
In this section, we would like to compare the
proposed algorithm with the modularity approach.
In the literature, there is an
extensive discussion about the limitations and drawbacks of the modularity-based
algorithms. One of the most significant problems of the modularity method is
21
Cluster 1
11 = 0.0231
Cluster 2
12 < 10−4
SE(11) = 0.0033
SE(12) < 10−4
Cluster 3
13 = 0.0001
SE(13) < 10−4
21 < 10−4
SE(21) < 10−4
31 < 10−4
SE(31) < 10−4
22 = 0.0334
23 = 0.0019
SE(22) = 0.0057 SE(23) = 0.0019
32 = 0.0019
33 = 0.0357
SE(32) = 0.0019 SE(23) = 0.0067
Cluster 1
Cluster 2
Cluster 3
Table 4: The block-wise error measure of clustering SBM2 via Algorithm 2
that the maximization of modularity score is not always consistent with the opti-
mized clustering outcome, which was addressed by [7]. This limitation is related
to under-clustering, as it seems that the modularity maximization algorithm in-
tends to merge small clusters, especially in large-size networks. It is believed that
the modularity-based methods undergo over-clustering problems as well. Both of
these issues were stressed and discussed in [26] and [17]. To the best of our knowl-
edge, neither of them has been systematically solved up to date. We refer the read-
ers to [3] for extensive discussions about inconsistency of clustering results from
the modularity maximization algorithm. On the other hand, both over-clustering
and under-clustering concerns are considered and well controlled in our algorithm,
which empirically ensures consistency. We present the following three simulation
examples to show resistance of the proposed algorithm in this manuscript.
We reconsider SBM1, a simulated network contains two extremely small com-
munities which are extremely loosely linked; SBM2, a simulated network contains
two relatively small communities which are fairly loosely linked; and SBM3, a
simulated network which is partitioned into communities of approximately even
size. The structural parameters are summarized in Table 5.
We respectively apply the modularity maximization algorithm and the localized
clique-based algorithm to these three networks, and evaluate the clustering results
by using the mean estimates of both NMI and ARI, based on 100 independent
copies of each simulated network. The results are presented in Table 6.
Based on the experimental results, we observe that the modularity maximiza-
22
Figure 4: Clustering result of SBM2 by Algorithm 2
tion algorithm performs extremely poorly for SBM1. As a matter of fact, the
modularity algorithm even fails to distinguish the two smaller communities (c.f.
clusters 2 and 3). As the sizes of small communities and the link density there-
between increase, the modularity maximization algorithm recovers (as reflected in
the results for SBM2), but apparently still does not appear as well-performed as
our algorithm. For networks comprising similar-size communities (e.g., SBM3),
the modularity maximization algorithm seems to outperform our algorithm, but
not significantly. Having the randomness of simulated networks in mind, we thus
conclude that the proposed algorithm in this manuscript is preferred due to its con-
sistency and reliability.
5.4 Time complexity
In the era of big data, researchers always concern about the efficiency of algo-
rithms, especially the newly proposed algorithm, when the network size is large.
In the last part of this section, we look into this issue numerically by implementing
the localized algorithm in Python for several networks of different sizes and struc-
ture. Both mean estimates of NMI and ARI are computed. Our testing results are
summarized in Table 7. We are convinced that the proposed algorithm is able to
provide accurate clustering results in a relatively short amount of time.
23
Cluster 1 (size: 60) Cluster 2 (size: 40) Cluster 3 (size:40)
Cluster 1 (size: 60)
Cluster 2 (size: 40)
Cluster 3 (size: 40)
0.2
0.05
0.05
0.05
0.6
0.12
0.05
0.12
0.8
Table 5: Link densities of clusters of a simulated SBM3
Network
(cid:91)NMI
Modularity maximization
SE((cid:91)NMI) (cid:100)ARI
SE((cid:100)ARI) (cid:91)NMI
Localized clique-based method
SE((cid:91)NMI) (cid:100)ARI
SE((cid:100)ARI)
SBM1
0.1681
0.0057
0.0514
0.0056
0.9309
0.0047
0.8556
0.0074
SBM2
0.7525
0.0092
0.5159
0.0106
0.9677
0.0030
0.9525
0.0035
SBM3
0.9568
0.0048
0.9055
0.0054
0.9477
0.0039
0.8719
0.0064
Table 6: Comparison of clustering results respectively from the modularity maxi-
mization algorithm and the localized clique-based algorithm
6 Concluding remarks
In this section, we address several remarks and discuss some possible future work.
To conclude, we propose a new clique-based measure in this manuscript to eval-
uate the quality of network clustering, and design an algorithm to maximize an
objective function based on the proposed measure. The clique score of each com-
munity in our clustering result is guaranteed to be higher than some predetermined
threshold p. We also consider the situation at which the threshold p is unspecified.
We develop an approach which accounts for over-clustering and under-clustering
problems simultaneously to select localized p for the network and its subsequent
networks. An associated localized algorithm is proposed and discussed.
Studies of networks or network models usually involve big data problems.
When network size or parameter space or both are large, it is always challeng-
ing to use model-based clustering methods, as many of them depend on accurate
but slow Bayesian MCMC algorithms, for example [22]. However, the methods
24
Number of clusters Number of simulations (cid:91)NMI (cid:100)ARI
Time
3
4
10
25
100
100
20
20
0.8596
0.9309
92.7 ms
0.9687
0.9779
94.5 ms
0.9895
0.9901
1.66 s
0.8960
0.8869
12.6 s
Size
120
1270
7000
20000
Table 7: A summary of community detection for several simulated SBMs via Al-
gorithm 2
proposed in this manuscript attempt to convert clustering analysis to optimization
problems. Therefore, many sophisticated machine learning techniques and well-
developed approximation methods are ready to use to deal with big data issues.
It is worthy of mentioning that the methods considered in this manuscript also
applicable to sparse social network, since the computation of p-clique index is
primarily based on the p-clique matrix C(p) = A − p(1n×n − In×n). This matrix
is not sparse even if the adjacency matrix A is sparse. Numerical methods, such
as the Lanczos method in [4], promise that the eigenvector corresponding to the
largest eigenvalue of C(p) always can be determined very fast.
The development of the localized clustering algorithm in this paper depends on
a strong assumption that the null network is an Erdos-R´enyi model. This may not
be true for many real-world networks. It is suggested in [2] that many networks
around us follow power law, which is not the case for the Erdos-R´enyi graphs. We
conjecture that it may be more accurate to consider a scale-free network as null for
the studies of social networks possessing power-law property, and we will consider
future research in this direction.
References
[1] AIROLDI, E.M., BLEI, D.M., FIENBERG, S.E. and XING, E.P. Mixed mem-
bership stochastic blockmodels. Journal of Machine Learning Research, 9,
1981–2014. (2008)
[2] BARAB ´ASI, A.-L. and ALBERT, R. Emergence of scaling in random net-
works. Science, 286, 509–512. (1999)
25
[3] BICKEL, P.J. and CHEN, A. A nonparametric view of network models and
Newman-Girvan and other modularities Proceedings of the National Academy
of Sciences of the United States of America, 106, 21068–21073. (2009)
[4] CALVETTI, D., REICHEL, L. and SORENSEN, D. An implicitly restarted
Lanczos method for large symmetric eigenvalue problems. Electronic Trans-
actions on Numerical Analysis, 2, 1–21. (1994)
[5] CHUNG, F.R.K. Spectral Graph Theory. American Mathematical Society,
Providence, RI. (1997)
[6] ERD OS, P. and R ´ENYI, A. On random graphs I. Publicationes Mathematicae,
6, 290–297. (1959)
[7] FORTUNATO, S. and BARTH ´ELEMY, M. Resolution limit in community de-
tection. Proceedings of the National Academy of Sciences of the United States
of America, 104, 36–41. (2007)
[8] FRED, A. and JAIN, A. Robust data clustering, in the Proceedings of IEEE
Computer Society Conference on Computer Vision and Pattern Recognition, 2,
128–133. (2003)
[9] GILBERT, E. N. Random graphs. Annals of Mathematical Statistics, 30, 1141–
1144. (1959)
[10] GOLDENBERG, A., ZHENG, A.X., FIENBERG, S.E. and AIROLDI, E.M.
A survey of statistical network models. Foundations and Trends in Machine
Learning, 2, 129–233. (2010)
[11] HANDCOCK, M.S., RAFTERY, A.E. and TANTRUM, J.M. Model-based
clustering for social networks. Journal of the Royal Statistical Society, Series
A, 170, 301–354. (2007)
[12] HOFF, P.D., RAFTERY, A.E. and HANDCOCK, M.S. Latent space ap-
proaches to social network analysis. Journal of the American Statistical Associ-
ation, 97, 1090–1098. (2002)
[13] HOLLAND, P. W. and LEINHARDT, S. An exponential family of probability
distributions for directed graphs. Journal of the American Statistical Associa-
tion, 76, 33–50. (1981)
[14] HOLLAND, P.W., LASKEY, K.B. and LEINHARDT, S. Stochastic blockmod-
els: First steps. Social Networks, 5, 109–137. (1983)
26
[15] HORN, R.A. and JOHNSON, C.R. Matrix Analysis. Cambridge University
Press, New York. (1985)
[16] HUBERT, L. and ABRABIE, P. Comparing partitions. Journal of Classifica-
tion, 2, 193–218. (1985
[17] LANCICHINETTI, A. and FORTUNATO, S. Limits of modularity maximiza-
tion in community detection. Physical Review E, 84, 066122. (2011)
[18] NEWMAN, M.E.J. The structure of scientific collaboration networks. Pro-
ceedings of the National Academy of Sciences of the United States of America,
98, 404–409. (2001)
[19] NEWMAN, M.E.J., STROGATZ, S.H. and WATTS, D.J.. Random graphs
with arbitrary degree distributions and their applications. Physical Review E,
64, 026118. (2001)
[20] NEWMAN, M.E.J. Modularity and community structure in networks. Pro-
ceedings of the National Academy of Sciences of the United States of America,
103, 8577–8582. (2006)
[21] NG, A.Y., JORDAN, M.I. and WEISS, Y. On spectral clustering: Analysis
and an algorithm. Advances in Neural Information Processing Systems, 14, 849–
856. (2001)
[22] SNIJDERS, T.A.B. and NOWICKI, K. Estimation and prediction for stochas-
tic blockmodels for graphs with latent block structure. Journal of Classification,
14, 75–100. (1997)
[23] OUYANG, G. Social Network Community Detection. Ph.D. dissertation. Uni-
versity of Connecticut. (2015)
[24] PAO, L.-F. Discovering the dynamics of smart business networks. Computa-
tional Management Science, 1, 445–458. (2014)
[25] PEI, X., ZHAN, X.-X. and JIN, Z. Application of pair approximation method
to modeling and analysis of a marriage network. Applied Mathematics and Com-
putation, 294, 280–293. (2017)
[26] REICHARDT, J. and BORNHOLDT, S. Statistical mechanics of community
detection. Physical Review E, 74, 016110. (2006)
[27] SHI, J. and MALIK, J. Normalized cuts and image segmentation. IEEE
Transaction on Pattern Analysis and Machine Intelligence, 22, 888–905. (2000)
27
[28] WATTS, D.J. and STROGATZ, S.H. Collective dynamics of "small-world"
networks. Nature, 440–442. (1998)
[29] WOHLGEMUTH, J. and MATACHE, M.T. Small-wold properties of Facebook
group networks. Complex Systems, 23, 197–225. (2014)
28
|
1902.03093 | 1 | 1902 | 2019-01-31T16:59:01 | A large-scale crowdsourced analysis of abuse against women journalists and politicians on Twitter | [
"cs.SI",
"cs.LG",
"stat.AP",
"stat.ML"
] | We report the first, to the best of our knowledge, hand-in-hand collaboration between human rights activists and machine learners, leveraging crowd-sourcing to study online abuse against women on Twitter. On a technical front, we carefully curate an unbiased yet low-variance dataset of labeled tweets, analyze it to account for the variability of abuse perception, and establish baselines, preparing it for release to community research efforts. On a social impact front, this study provides the technical backbone for a media campaign aimed at raising public and deciders' awareness and elevating the standards expected from social media companies. | cs.SI | cs |
A large-scale crowd-sourced analysis of abuse against
women journalists and politicians on Twitter
Laure Delisle∗†
Element AI
Alfredo Kalaitzis∗†
Element AI
Krzysztof Majewski†
Element AI
Archy de Berker†
Element AI
Milena Marin‡
Amnesty International
Julien Cornebise†
Element AI
Abstract
We report the first, to the best of our knowledge, hand-in-hand collaboration
between human rights activists and machine learners, leveraging crowd-sourcing
to study online abuse against women on Twitter. On a technical front, we carefully
curate an unbiased yet low-variance dataset of labeled tweets, analyze it to account
for the variability of abuse perception, and establish baselines, preparing it for
release to community research efforts. On a social impact front, this study provides
the technical backbone for a media campaign aimed at raising public and deciders'
awareness and elevating the standards expected from social media companies.
1
Introduction
Social media platforms have become a critical space for women and marginalized groups to express
themselves at an unprecedented scale. Yet a stream of research by Amnesty International [1, 2]
showed that many women are subject to targeted online violence and abuse, which denies them the
right to use social media platforms equally, freely, and without fear. Being confronted with toxicity at
a massive scale leaves a long-lasting effect on mental health, sometimes even resulting in withdrawal
from public life altogether [3]. A first smaller-scale analysis of online abuse against women UK
Members of Parliament (MPs) on Twitter [4, 5] proved the impact such targeted campaigns can have:
it contributed to British Prime Minister Theresa May publicly calling out the impact of online abuse
on democracy [6].
This laid the groundwork for the larger-scale Troll Patrol project that we present here: a joint effort
by human rights researchers and technical experts to analyze millions of tweets through the help
of online volunteers. Our main research result is the development of a dataset that could help in
developing tools to aid online moderators. To that end, we i) Designed a large, enriched, yet unbiased
dataset of hundreds of thousands of tweets; ii) Crowd-sourced its labeling to online volunteers; iii)
Analyzed its quality via a thorough agreement analysis, to account for the personal variability of
abuse perception; iv) Compared multiple baselines with the aim of classifying a larger dataset of
millions of tweets. Beyond this collaboration, this should allow researchers worldwide to push the
envelope on this very challenging task -- one of many in natural language understanding [7].
The social impact Amnesty International is aiming for is ultimately to influence social media compa-
nies like Twitter into increasing investment and resources -- under any form -- dedicated to tackling
online abuse against women. With this study, we contribute to this social impact by providing the
research backbone for a planned media campaign in November 2018.
2 Crowd-sourcing an importance-sampled enriched set
Core to this study is the careful crafting of a large set of tweets followed by a massive crowd-sourced
data labeling effort.
∗Equal contribution. †{laure,freddie,km,archy,julien}@elementai.com; ‡[email protected]
32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
Figure 1: Left to right: Distribution of annotations-per-tweet: To analyze agreement, we used only tweets
annotated more than twice (∼73k); Values of Contain Abuse are ordinal; Type and Medium conditioned on
Contain Abuse (cid:54)= No: the majority of abuse is not easily classified, and the vast majority of abuse is textual.
Studied population: We selected 778 women politicians and journalists with an active, non-protected
Twitter account, with fewer than 1 million followers, including most women British MPs and all
US Congresswomen, and journalists from a range of news organizations representing a diversity of
media bias. Full details are in Appendix A.
Tweet collection: 14.5M tweets mentioned at least one woman of interest during 2017. We obtained
a subset of 10K per day sampled uniformly from Twitter's Firehose, minus tweets deleted since
publication, totaling 2.2M tweets.
Pre-labeling selection: Taking into account the average labeling time per tweet from a pilot study,
the expected duration of the campaign, and the expected graders' engagement, we targeted labeling
at most 275K tweets in triplicate. We first selected 215K tweets, correcting the 10K daily cap using
per-day stratified sampling proportional to each day's actual volume. While this sample is statistically
representative of the actual tweet distributions, its class imbalance would induce high variance into
any estimator, and waste the graders' engagement. We therefore enriched the dataset with 60K tweets
pre-filtered through the Naive-Bayes classifier pre-trained in [5]. To maintain statistical non-bias, we
keep track of the importance sampling weights.
Volunteers labeling via crowd-sourcing: Finally, these tweets, properly randomized, were deployed
through Amnesty Decoders, the micro-tasking platform based on Hive [8] and Discourse [9] where
Amnesty International engages digital volunteers (mostly existing members and supporters) in
human rights research. Great effort was put into designing a user-friendly, interactive interface,
accessible at [10] -- see Appendix D for screenshots. After a video tutorial, volunteers were shown
an anonymized tweet from the randomized sample, then were asked multiple-choice questions: 1)
"Does the tweet contain problematic or abusive content?" (No, Problematic, Abusive). Unless their
answer was No, the follow-up questions were "What type of problematic or abusive content does it
contain?" (at least one of six) and (optional question) "What is the medium of abuse?" (one of four).
See Fig. 1 for details and summary statistics. At all times they had access to definitions and examples
of abusive and problematic content, and the typologies thereof -- see Appendix E.
By August 2018, 157K unique tweets containing 167K mentions of the studied individuals had been
categorized at least once, totalling 337K labels, thanks to the contribution of 4, 537 online volunteers.
Experts labeling: In addition to engaging digital volunteers, Amnesty also asked three experts
(Amnesty's researcher on online abuse against women, Amnesty's manager of the Troll Patrol project
and an external expert in online abuse) to label a sub-set of 568 tweets. Those tweets were sampled
from tweets labeled by exactly three volunteers as of June 8, 2018. To ensure low variance in the
estimates, we once again used importance sampling, inflating the proportion of potentially abusive
tweets by sampling 500 tweets uniformly from those labeled as "Abusive" by the Naive-Bayes
classifier mentioned in [5] and "Basic Negative" by Crimson Hexagon's sentiment analysis (our
Firehose access provider), and 500 tweets uniformly sampled on the remainder.
Re-weighting after importance sampling: To ensure that any inference or training based on the
enriched sample is representative of the Twitter distribution, we use importance sampling to re-weight
the tweets in the empirical distribution. The weights are defined as the ratio of the target distribution
(as estimated by the daily counts) and the enriched distribution -- see Appendix C for the full derivation
of the weights.
2
123456789times annotated25k50kno. tweetsAnnotation countsNoProblematicAbusive0255075%300k28k7kIs the content problematic or abusive?10k20kNASexual threatsHomophobia or transphobiaEthnic or religious slurPhysical threatsRacismSexism or misogynyOther0%2%2%5%5%8%22%55%Type of abusive content103104VideoOther attachmentnaImageText0.5%0.5%0.7%4.3%94.1%Medium of abuse 3 Analysis and generalization
3.1 Agreement analysis
We quantified the agreement among raters -- within crowd and within experts -- using Fleiss' kappa (κ),
a statistical measure of inter-rater agreement [11]. κ is designed for nominal (non-ordinal categorical)
variables, e.g. Fig 1(c), whereas in ordinal variables κ tends to underestimate the agreement because
it treats the disagreement between Problematic <> Abusive the same as No <> Abusive. We also use
the intra-class correlation (ICC) [12] for ordinal categorical annotations, like Contains Abuse: No
< Problematic < Abusive. We define κ and icc below, and further explain in Appendix B.
Figure 2: Visualizing the distribution of annotations a(t) (+ jitter for clarity) in the multinomial 2-simplex.
The corners are events of complete agreement. The center is no agreement with the non-ordinal assumption, but
c = 3, c ∼ P (C); Agreement
partial agreement with ordinality. Left to right: Simulated perfect agreement, a(t)
among 3 experts on 1000 tweets: empirical probabilities are visually amplified by over-sampling a(t) to 20k;
Agreement among N = 3 Decoders per tweet: if N > 3, raters are chosen randomly; Simulated agreement-by-
chance only: a(t) ∼ Multinomial (N = 3, p = P (C)). The multinomial assumes independence between trials.
A hierarchical modeling approach can capture inter-rater dependence [13, 14].
Fleiss' kappa:A rater can annotate a tweet as class c ∈ C = {No, Problematic, Abusive}. The
annotation a = (aN o, aP r, aAb), ac ∈ {0, 1, 2, 3}, Σcac = N, contains the class counts for a tweet
annotated by N raters. The overall agreement for a set of tweets T is κ = 1T Σtκ(t), where
N (N−1) ∈ [0, 1] is the fraction of
κ(t) = Σcr(t)
pairs of raters that agree on c. P (C = c) = pc is the empirical probability of c, hence Σcp2
c is the
probability of agreement-by-chance.
is the within-tweet agreement, and rc = ac(ac−1)
c −Σcp2
1−Σcp2
c
c
ICC (intra-class correlation): Let A ∈ RT×N be the matrix of annotations, Ai,j ∈ {0, 1, 2}
(ordinal values raters can assign), and each row is a tweet annotated by N random raters. The
N ΣjAi,j and the overall mean is µ = 1T Σiµi. The within-tweet
tweet-specific mean is µi = 1
(N−1) Σj (Ai,j − µi)2, and its average Vw = 1T ΣiV (i) is the
disagreement for tweet i is V (i) = 1
overall within-tweet variance. Similarly, the between-tweet variance is Vb = NT−1 Σi(µi − µ)2. The
icc can now be expressed in terms of a one-way ANOVA [12]: icc =
, the fraction of
variation in annotations that is not explained by between-tweet disagreements.
Results: Table 1 and Figure 2 (mid left / mid right) show more agreement among the experts than
among the volunteers -- higher κ and ICC among the former. There is also more agreement when
assessing the presence of abuse than when assessing the type of abuse.
Vb+(N−1)Vw
Vb−Vw
Figure 3: Left & center: performance of Perspective API and fine-tuned BERT classifiers, with respect to the
experts' and crowd's labels. Right: performance of the crowd-as-a-classifier against the experts' labels. Note:
Recall is equivalent to TPR, hence the y-axes (TPR<>Recall) of the two plots are aligned.
3
Problematic0123Abusive0123No0123Simulated perfect agreementProblematic0123Abusive0123No0123Agreement amongst expertsProblematic0123Abusive0123No0123Agreement amongst DecodersProblematic0123Abusive0123No0123Simulated agreement by chance12log10 probability0.00.51.0FPR0.00.20.40.60.81.0TPR0.00.51.0precision(0.29, 0.52)(0.45, 0.41)weighted ROC / Precision-Recall (Perspective)vs Experts (auc=0.78, F1-opt=0.38)vs Crowd (auc=0.89, F1-opt=0.43)F1-optimal0.00.51.0FPR0.00.20.40.60.81.0TPR0.51.0precision(0.50, 0.44)(0.35, 0.57)weighted ROC / Precision-Recall (fine-tuned BERT)vs Experts (auc=0.89, F1-opt=0.47)vs Crowd (auc=0.91, F1-opt=0.43)F1-optimal0.00.51.0FPR0.00.20.40.60.81.0TPR0.51.0precision(0.41, 0.53)weighted ROC / Precision-Recall (Crowd)vs Experts (auc=0.90, F1-opt=0.46)F1-optimal Table 1: Agreement per variable and
per labeling cohort.
Experts
Labels from
Contain Abuse
Type of Abuse
Crowd
κ
.26
.16
ICC κ
.35
-
.54
.74
ICC
.70
-
Table 2: Classifier performance vs. crowd vs. expert labels
Labels from
Naive Bayes
Crimson Hexagon
Davidson et al.
Perspective API
Fine-tuned BERT
Crowd
Crowd
P recision Recall F ∗
.13
.17
.20
.14
.36
.53
.43
.45
.43
.35
-
-
.25
.40
.27
.41
.57
-
1 AP
.11
-
.25
.34
.40
-
Experts
P recision Recall F ∗
.40
.32
.04
.05
.39
.35
.38
.29
.47
.50
.41
.46
.27
.04
.46
.52
.44
.53
1 AP
.21
-
.25
.25
.36
.39
3.2 Comparison of baseline classifiers
The core focus in this study is to build and analyze the dataset, with a view to extend that analysis to
the remaining 2M unlabeled tweets using state of the art models. We prepare this follow-up research
community effort by establishing baselines on various classification models.
Classifiers: In Table 2, Naive Bayes refers to the classifier from [5]. Crimson Hexagon refers to
sentiment labels -- Category and Emotion -- from Crimson Hexagon. We also benchmarked the
pre-trained classifier from Davidson et al. [15]. Perspective API refers to the public toxicity scoring
API provided by Jigsaw [16, 17]. We also trained our own model, which combined a pre-trained
BERT embedder [18] and an abuse-specific embedding trained from scratch. For details see F.
Methodology: For this analysis, we conflate the labels Problematic and Abusive into one positive
(Abusive) class. The crowd labels are the majority votes over labels on tweets labeled by exactly
three volunteers. The expert labels are majority votes over labels from the three domain experts
mentioned in Section 2. For Crimson Hexagon, we define Abusive as the intersection of Category
= Basic Negative and Emotion = Anger Disgust.
Results: Table 2 shows the F ∗
1 (optimal F1 score), corresponding precision and recall, and the
Average Precision (AP ), to evaluate several abuse detection classifiers with respect to labels from the
crowd and from the experts.
4 Discussion
Dataset availability and reproducibility: Amnesty International intends to publish as much of the
dataset as possible to encourage replication and further research on the topic. At the very least the
URLs of the tweets and the grades will be made public. Publishing the actual tweets is more delicate
due to Twitter Terms and Conditions. Publishing the meta-data on the graders (gender, location)
would be of great interest, but is still under discussion from an ethical point of view.
Future work: We aim to eventually apply different classifiers to the whole unlabeled dataset, so as
to scale up the human rights researchers' work by sifting through the huge amount of tweets. As
shown in Section 3.2, this will require a careful tuning of models to increase the precision beyond its
current performance. In parallel, we also want to analyze the scale, typology and intersectionality of
abuse, either on the labeled set or on the classified extra 2M tweets, for the planned media campaign.
Social impact: The sheer volume of hateful speech on social media has recently prompted govern-
ments to put strong pressure on social media companies to remove such speech [19]. The moderation
of abusive messages at scale requires some form of automated assistance. Our results highlight the
double challenge of automatic abuse classification: the subjectivity in the labels and the limited ability
of state-of-the-art classifiers to generalize beyond training data. This all points toward the need for
systems where human subtlety and context awareness are empowered by automatic pre-screening.
Whether the companies themselves should be trusted with (or required to implement) such moderation,
or whether they should fund or be supervised by a third-party neutral watchdog, goes far beyond a
purely technical conversation. This is why collaboration between technical experts (machine learners,
data scientists) and domain experts (human rights researchers, anti-censorship activist, etc.), as well
as society in a broader sense, is so important for genuinely impactful AI for Social Good efforts.
Acknowledgements
We are extremely grateful to all the volunteers from Amnesty Decoders for their hard work. We would like to
thank Nasrin Baratalipour, Francis Duplessis, Rusheel Shahani and Andrei Ungur for modelling support. We
also thank Jerome Pasquero for his support, and the Perspective team for access to their API.
4
References
[1] Azmina Dhrodia. Unsocial media: The real toll of online abuse against women, November
2017. URL www.medium.com/amnesty-insights/unsocial-media-the-real-toll-
of-online-abuse-against-women-37134ddab3f4. [Online; posted 20-November-2017].
[2] Amnesty International. Toxic Twitter, 2018. URL https://www.amnesty.org/en/latest/
research/2018/03/online-violence-against-women-chapter-1.
[3] Committee on Standards in Public Life.
Intimidation in public life.
2017. URL
https://assets.publishing.service.gov.uk/government/uploads/system/
uploads/attachment_data/file/666927/6.3637_CO_v6_061217_Web3.1__2_.pdf.
[4] Azmina Dhrodia. Unsocial media: Tracking Twitter abuse against women MPs, Septem-
ber 2017.
URL www.medium.com/@AmnestyInsights/unsocial-media-tracking-
{T}witter-abuse-against-women-mps-fc28aeca498a. [Online; posted 04-September-
2017].
[5] Ekaterina Stambolieva. Technical report, Amnesty International, 2017. URL https://
drive.google.com/file/d/0B3bg_SJKE9GOenpaekZ4eXRBWk0/view.
[6] The Guardian.
to democracy',
URL https://www.theguardian.com/society/2018/feb/05/theresa-may-
2018.
calls-abuse-in-public-life-a-threat-to-democracy-online-social-media.
Theresa may calls abuse in public life 'a threat
[7] Roberto González-Ibáñez, Smaranda Muresan, and Nina Wacholder. Identifying sarcasm in
Twitter: A closer look. pages 581 -- 586, 2011. URL http://dl.acm.org/citation.cfm?id=
2002736.2002850.
[8] New York Times Labs. Hive: Open-source crowdsourcing framework, 2014. URL http:
//nytlabs.com/blog/2014/12/09/hive-open-source-crowdsourcing-framework/.
[9] Discourse. Discussion forum. URL https://www.discourse.org/.
[10] Troll patrol, 2018. URL https://decoders.amnesty.org/projects/troll-patrol.
[11] Joseph L Fleiss. Measuring nominal scale agreement among many raters. Psychological bulletin,
76(5):378, 1971.
[12] Patrick E Shrout and Joseph L Fleiss. Intra-class correlations: uses in assessing rater reliability.
Psychological bulletin, 86(2):420, 1979.
[13] Alfredo Kalaitzis and Ricardo Silva. Flexible sampling of discrete data correlations without
In Advances in Neural Information Processing Systems, pages
the marginal distributions.
2517 -- 2525, 2013.
[14] Ricardo Silva and Alfredo Kalaitzis. Bayesian inference via projections. Statistics and Comput-
ing, 25(4):739 -- 753, 2015.
[15] Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. Automated hate
speech detection and the problem of offensive language, 2017. URL https://aaai.org/ocs/
index.php/ICWSM/ICWSM17/paper/view/15665.
[16] Jigsaw. Perspective API. URL https://www.perspectiveapi.com.
[17] Hossein Hosseini, Sreeram Kannan, Baosen Zhang, and Radha Poovendran. Deceiving google's
perspective api built for detecting toxic comments. arXiv preprint arXiv:1702.08138, 2017.
[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of
deep bidirectional transformers for language understanding, 2018.
[19] Björn Gambäck and Utpal Kumar Sikdar. Using convolutional neural networks to classify
hate-speech. 2017.
5
[20] Randal Douc and Eric Moulines. Limit theorems for weighted samples with applications to
sequential Monte Carlo methods. Annals of Statistics, 36(5):2344 -- 2376, 2008.
[21] Hugging Face Development Team. PyTorch pretrained BERT. URL https://github.com/
huggingface/pytorch-pretrained-BERT.
[22] Luis von Ahn.
resources/.
List of abusive words.
URL https://www.cs.cmu.edu/~biglou/
6
A Population definition
We selected politicians and journalists with an active, non protected Twitter account, with fewer than
1 million followers. The group included:
• All women members of the British Parliament (220, including 22 women who left parliament
during the June 2017 elections and excluding one politician with over 1 million followers);
• All women in the United States Congress (107, excluding 3 politicians with more than 1
• And women journalists working at the following news organizations, selected to represent a
million followers);
diversity of media bias:
-- Breitbart (16),
-- Daily Mail (78),
-- The Sun (54),
-- The Guardian (124),
-- The New York Times (278),
-- Gal-Dem (23),
-- PinkNews (9).
B Agreement analysis
B.1 Fleiss' Kappa
Notation A rater can annotate a tweet as class c ∈ C = {No, Problematic, Abusive}. The
annotation tuple a = (aN o, aP r, aAb), ac ∈ {0, 1, 2, . . .}, Σcac = N, contains the class-specific
counts for a tweet annotated by N raters.
Estimation of κ The within-class agreement-ratio rc = ac (ac−1)
N (N−1) ∈ [0, 1] is the ratio of pairs of
raters that agree on c, over the total of pairs of N raters. P (C = c) = pc is the empirical marginal
probability of class c. Hence, Σc p2
c is the overall probability of agreement by chance across a dataset
of tweets. For a specific tweet t we can compute the within-class agreement κ(t)
, where
the numerator is the agreement-above-chance attained on c, and the denominator is the best-case-
scenario (maximal) agreement-above-chance attainable across classes. Hence κ(t)
is the fraction that
c
the attained agreement in c contributes to the best-case scenario, while accounting for agreement-by-
chance. Finally, the within-tweet agreement for a tweet t is the sum across classes, κ(t) = Σcκ(t)
c ,
and the overall agreement across a set of tweets T is the expectation
c −p2
1 − Σcp2
c = r(t)
c
c
κ = ET [κ(.)] ≈ 1T Σtκ(t).
(1)
B.2 ICC (intra-class correlation)
Notation We denote the matrix of annotation as A ∈ RT × N . In this work, Ai, j ∈ {0, 1, 2}
(ordinal values raters can assign), where each row represents a tweet annotated by N random raters.
Algorithm The tweet-specific mean is µi = 1
N ΣjAi,j and the overall mean is µ = 1T Σiµi. We
(N−1) Σj (Ai,j−µi)2.
can express the within-tweet disagreement as the within-tweet variance V (i) = 1
Then the average of within-tweet disagreements expresses the overall within-tweet variance, Vw =
1T ΣiV (i). Similarly, the between-tweet variance is Vb = NT−1 Σi(µi − µ)2. Note that the i-th tweet
is polarized when V (i) is maximized, i.e. half of the raters choose 0 and the other half choose 2. In
the extreme scenario that all tweets are maximally polarizing, Vb = 0. Therefore Vb expresses the
overall tendency for disagreement-by-chance. All classes of ICC are equivalent to a type of ANOVA
(ANalysis Of VAriance) in linear mixed-effects model of annotations [12]. In our case,
(2)
Vb − Vw
Vb + (N − 1)Vw
icc(1, k) =
7
Intuitively, the ANOVA framework defines agreement as the fraction of variation in annotations that
is not explained by between-tweet disagreements.
Systemic disagreement As mentioned above, in extreme scenarios where Vb is small, the ICC can
be negative. Negative κ and icc values might seem like an artifact of degenerate or extreme data, only
to be dismissed as no agreement in the downstream analysis. At closer inspection, the numerator
shows that subtracting the agreement-by-chance yields a measure of systemic disagreement: e.g.
expecting P (agreement-by-chance) = 0.9 but observing agreement only 20% of the time, implies a
systemic cause for polarizing opinions (e.g. controversial content, raters annotating with different
rules).
C Importance sampling analysis
Population and Crimson sets W and C: We denote the population (World set) as W , and the
sample obtained from the Twitter firehose (Crimson set) as C. Members of the sets W and C are
observation tuples (t, k, d), where t is the text content of a tweet, k ∈ {0, 1} is the output of a Naive
Bayes Classifier NBC : t (cid:55)→ k, and d is the day in 2017 that a tweet was published:
C ⊂ W = {(t, k, d)}
(3)
Distributions pW and pC: We define pW and pC as the probability mass over sets W and C,
respectively, and any marginals and conditionals thereof:
(4)
(5)
The density pW (d) is directly available from the daily total volumes nd of tweets matching the query,
total that is provided by Crimson Hexagon alongside the smaller sampled set C:
pW (t, k, d) = pW (t, kd) pW (d)
pC(t, k, d) = pC(t, kd) pC(d)
nd = {(t, k, d(cid:48)) ∈ W : d(cid:48) = d} provided as metadata,
pW (d) =
.
(6)
nd(cid:80)
d(cid:48) nd(cid:48)
The Crimson set C is constructed by uniform sampling over tweets in W , such that for any day d, the
conditional probabilities over both sets are equal:
pC(t, kd) = pW (t, kd)
Then, using eq. (7) in (5):
pC(t, k, d) = pW (t, kd) pC(d)
Constructed set A: The final set A is defined as the union
A = B ∪ F,
where
B = {(t, k, d) ∼ pB(t, k, d) (cid:39) pW (t, kd) pW (d) (cid:39) pW (t, k, d)}
approximates the world joint distribution through stratified sampling per day, and
F = {(t, k, d) ∈ C\B : k = 1}
(11)
is an enriched sample resulting from pre-filtering by a simple Naive Bayes classifier. The cardinalities
of these sets are: C = 2.2M, B = 215k, F = 60k and A = 275k.
With β =
B+F, and z(d) a normalizing constant depending on d:
F
pA(t, kd) =
where I(.) is the indicator function.
The conditional probabilities of a tweet are identical in W and A:
z(d)
βI(k = 1) pA(t, kd) + (1 − β) pA(t, kd)
pW (tk, d) = pA(tk, d) .
Combining equations (13) and (12) leads to:
pA(t, kd) ∝ βI(k = 1) pW (tk, d) pW (kd) + (1 − β) pW (tk, d) pW (kd) .
8
(7)
(8)
(9)
(10)
(12)
(13)
(14)
Importance weights wi: Estimating statistics on the world set W using the samples in set A can
be achieved using importance sampling, i.e. assigning a specific weight pW (t, k, d)/pA(t, k, d) to
each triplet (t, k, d) in A.
For each tweet (t, k, d) ∈ A, we define the weighting function w
w(t, k, d) =
=
pW (t, k, d)
pA(t, k, d)
pW (tk, d) pW (k, d)
pA(tk, d) pA(k, d)
.
Injecting equation (13) in equation (15), we can simplify by pW (tk, d):
w(t, k, d) =
=
pW (k, d)
pA(k, d)
pW (kd) pW (d)
pA(kd) pA(d)
.
(15)
(16)
Since A is a finite set, the probability mass functions pA(kd) and pA(d) in equation (16) are directly
accessible by simple counting.
The probability mass functions pW (d) is known from (6). The term pW (kd) is not available in closed
form, but can be estimated straightforwardly. Indeed from equation (7) we have pW (kd) = pC(kd),
and the latter can be estimated by simple counting on C, leading to empirical estimate:
{(t, k(cid:48), d(cid:48)) ∈ C : k(cid:48) = k}
This leads to the final plug-in estimator of the importance weights:
pW (kd) =
{(t, k(cid:48), d(cid:48)) ∈ C : k(cid:48) = k, d(cid:48) = d}
w(t, k, d) =
pW (kd)pW (d)
pA(kd)pA(d)
.
.
(17)
(18)
For any given function f (t, k, d), we therefore estimate its expectation in the whole population W
using the self-normalized importance estimator:
EW [f (t, k, d)] =
f (ti, ki, di) .
(19)
(cid:88)
(ti,ki,di)∈A
wi(cid:80)
j wj
where for any tweet (ti, ki, di) with GUID (Globally Unique Identifier) i we use the estimated
unnormalized importance weight wi := w(ti, ki, di).
Note that for full mathematical rigour, the asymptotic consistency of the importance sampling
estimator EW could be proven by showing that the replacement of the density estimator pW in the
plug-in estimator w is asymptotically valid. Such a proof could proceed along the lines of [20], but is
outside of the scope of this article.
9
D Labeling tool screenshots
The workflow presented to each grader by the labelling tool is illustrated in Figure 4 and Figure 5.
Figure 4: First stage of labeling: Initial screen showing an anonymized tweet, with anonymized handles and
first question.
(a) Second stage of
labeling:
Identification of the type of abuse.
(b) Third, optional stage of label-
ing: Identification of the part of
the tweet that carries the abuse.
(c) Warning displayed after clas-
sifying a tweet as abusive, to min-
imize the impact on the labelers'
mental health.
Figure 5: Follow-up stages, conditional on the first stage of labeling.
E Definitions and examples used in Troll Patrol - Trigger Warning
Abusive content Abusive content violates Twitter's own rules and includes tweets that promote
violence against or threaten people based on their race, ethnicity, national origin, sexual orientation,
gender, gender identity, religious affiliation, age, disability, or serious disease.
Examples include physical or sexual threats, wishes for the physical harm or death, reference to
violent events, behaviour that incites fear or repeated slurs, epithets, racist and sexist tropes, or other
content that degrades someone. For more information, see Twitter's hateful conduct policy.
In examples shown below, tweets were anonymized and only show a standard template (incl. the
author handle, the author profile picture, the tweet date and time, likes and retweets).
10
Problematic content Hurtful or hostile content, especially if it were repeated to an individual on
multiple or cumulative occasions, but not as intense as an abusive tweet. It can reinforce negative or
harmful stereotypes against a group of individuals (e.g. negative stereotypes about a race or people
who follow a certain religion). Such tweets may have the effect of silencing an individual or groups
of individuals.
Sexism or misogyny
Insulting or abusive content directed at women based on their gender, includ-
ing content intended to shame, intimidate or degrade women. It can include profanity, threats, slurs
and insulting epithets.
Racism Discriminatory, offensive or insulting content directed at a woman based on her race,
including content that aims to attack, harm, belittle, humiliate or undermine her.
Homophobia or transphobia Discriminatory, offensive or insulting content directed at a woman
based on her sexual orientation, gender identity or gender expression. This includes negative
comments towards bisexual, homosexual and transgender people.
Ethnic or religious slur Discriminatory, offensive or insulting content directed at a woman based
on her ethnic or religious identities.
11
Physical threats Direct or indirect threats of physical violence or wishes for serious physical harm,
death, or disease.
Sexual threats Direct or indirect threats of sexual violence or wishes for rape or other forms of
sexual assault.
Other There will be some tweets that fall under the 'other category' that are problematic and/or
abusive. For example, statements that target a user's disability, be it physical or mental, or content
that attacks a woman's nationality, health status, legal status, employment, etc.
F Abuse Classification Model
Model architecture We used a pretrained BERT model [18] (12 layers, 768 units per layer) as
the basis for our classification model. We took the final-layer representation of the first token in the
sequence as a fixed-length tweet embedding (see Figure 3 of [18]). The model was implemented
in Pytorch, and made use of the BERT implementation provided at [21], which in turn utilizes a
pre-trained model provided by Google.
To account for out-of-vocabulary abusive words, we added a second single-layer word embedding
(128 units), which we trained from scratch with a limited abusive vocabulary. This vocabulary
included a list of 1300 'possibly abusive' words available online [22], and the 1000 words which
occurred most disproportionately in the abusive class of the training data. To obtain a fixed-length
representation from this embedder, we took the mean across words in each tweet.
We concatenated these two representations to obtain a fixed-length tweet representation (of length
896), and passed througha fully-connected layer of 64 units before returning a decision via a binary
softmax layer.
Data We split the crowd-sourced data into train, validation, and test sets (90 : 5 : 5). We adjust all
reported performance metrics for the original importance sampling (see C).
Training We trained the model end-to-end with stochastic gradient descent (learning rate = 0.0001,
momentum=0.9) for 11 epochs, minimizing a cross-entropy loss.
12
Figure 6: We combined a pre-trained BERT model with a word embedding exclusively including abusive words.
The two embeddings were concatenated and passed through a fully-connected layer before a softmax layer
returned the prediction. Numbers in orange are layer widths.
13
|
1912.05123 | 1 | 1912 | 2019-12-11T05:35:36 | Collective Identity Formation on Instagram -- Investigating the Social Movement Fridays for Future | [
"cs.SI"
] | In recent years, social media changed the way individuals participate in social movements. While activists demonstrate on the street to fight for a public goal, members of specific movements can also act collective online. Thus, different aspects might influence the formation of collective identity and therefore drive collective action on social media. This study combines the perspectives of social identity- and identity theory in order to examine how members of an opinion-based group contribute to the collective group/social identity formation and therefore, to collective action. To this end, we applied automated text classification techniques to Instagram communication related to the social movement Fridays for Future. Analysing 1,137 comments showed that individuals mainly express Group Cohesion and Emotional Attachment rather than Solidarity by commenting on Instagram. This study further presents a proposed model of collective group/social identity of collective action. Succeeding research aims at enhancing the classification and testing the model. | cs.SI | cs | Australasian Conference on Information Systems
2019, Perth Western Australia
Brünker et al.
Collective Identity Formation on Instagram
Collective Identity Formation on Instagram -- Investigating
the Social Movement Fridays for Future
Research in Progress
Felix Brünker
Professional Communication in Electronic Media/Social Media
University of Duisburg-Essen
Duisburg, Germany
Email: [email protected]
Fabian Deitelhoff
Institute for the Digital Transformation of Application and Living Domains
University of Applied Science and Arts
Dortmund, Germany
Email: [email protected]
Milad Mirbabaie
Faculty of Business Studies and Economics
University of Bremen
Bremen, Germany
Email: [email protected]
Abstract
In recent years, social media changed the way individuals participate in social movements. While
activists demonstrate on the street to fight for a public goal, members of specific movements can also
act collective online. Thus, different aspects might influence the formation of collective identity and
therefore drive collective action on social media. This study combines the perspectives of social identity-
and identity theory in order to examine how members of an opinion-based group contribute to the
collective group/social identity formation and therefore, to collective action. To this end, we applied
automated text classification techniques to Instagram communication related to the social movement
Fridays for Future. Analysing 1,137 comments showed that individuals mainly express Group Cohesion
and Emotional Attachment rather than Solidarity by commenting on Instagram. This study further
presents a proposed model of collective group/social identity of collective action. Succeeding research
aims at enhancing the classification and testing the model.
Keywords: Collective Identity, Identity Theory, Collective Action, Social Movement, Instagram
304
Brünker et al.
Collective Identity Formation on Instagram
Australasian Conference on Information Systems
2019, Perth Western Australia
1 INTRODUCTION
The way individuals participate in social movements changed over the last decades due to the
development of social media platforms (Tye et al. 2018). Individuals now have opportunity structures
(Mirbabaie et al. 2016) to be part of social movements in a virtual way. This step breaks the boundaries
of physical participation in social movements on the streets. However, individuals can not only be part
of a movement by a low-cost effort via social media but rather activist may use social media to catalyse
the participation and popularity of certain social movement which they support. In this way, the
relevance of social movements may increase for society and political agenda such as the #metoo debate
did (Manikonda et al. 2018). However, one recent movement which tries to put the topic of climate
change on the local, as well as global political agenda, is Fridays for Future (Neimark 2019). To this end,
individuals organise and communicate via social media (Stieglitz et al. 2016) and therefore establish
online opinion-based groups (Stieglitz et al. 2018; Thomas et al. 2012) by dedicated social media
accounts such as the profile of "Greta Thunberg", the public face of Fridays for Future. Further, being
part of such an opinion-based group might support common convictions and therefore show up
collective actions.
Regarding this context through the lens of identity theory reveals that not only a group membership
influences the formation of identity within an opinion-based group. Likewise, information systems such
as social media might play a discrete role in identity formation (Carter et al. 2017; Carter and Grover
2015; Liu and Chan 2010). According to identity theory, individuals form their identities by internalised
meanings connected to the Self (Person), an individualised Role, or as a member of a Group (McCall
and Simmons 1966; Stets and Serpe 2013). However, identity theory does not comprehensively include
inter- and intragroup dynamics (Hogg 2018; Turner et al. 1987). Likewise, regarding identity formation
during social movements reveals a multitude of group dynamics which identity theories miss to explain
in-depth (Postmes et al. 2006). Therefore, we consider social identity theory with its subtheory
collective identity to fulfil this gap (Davis et al. 2019).
Therefore, we plan to examine how collective action is influenced by the collective group and social
identity, i.e. collective group/social identity. As this is research in progress, we aim to answer as a first
step how collective group/social identity is formed on social media within a social movement. Thus, we
derived the following research question:
RQ1: How is the collective group/social identity formed on Instagram within an opinion-based
community?
In order to answer the above research question, we analysed comments from two posts on Instagram by
Greta Thunberg, the public face of the social movement Fridays for Future. Overall, we manually
collected 1,137 comments from February 2019 until July 2019 for further analysis. To better understand
collective identity formation and therefore collective action on social media, we provide first insights
how members of an opinion-based community contribute to collective identity formation. As most of
the research on social media focus on Twitter or Facebook (Bunker et al. 2017; Mirbabaie et al. 2017; Oh
et al. 2013; Stieglitz et al. 2017), we decided to examine the case Fridays for Future on Instagram as the
social media platform of choice. In this way, we aim to broaden the understanding of social movement
research on social media by providing findings according to a platform which differs in the way of usage
to Twitter or Facebook. Therefore, we expect to provide new preliminary findings of social movements
on social media in the area of information systems research.
The paper is structured as follows. First, we provide an overview of the collective identity formation on
social media as well as its role in collective action. Second, we outline our research design by describing
the examined case, the data collection, and the data analysis. Furthermore, we present our preliminary
findings based on the data analysis in the subsequent section. Last, we provide a conclusion of our
outcome and an outlook for a follow-up study and further research.
2 BACKGROUND
2.1 Collective Identity Formation on Social Media
In order to understand the formation of (social) collective identity and its impact on collective action on
social media, it is crucial to investigate the role of the individual and (social) group influences (Thomas
et al. 2012). To this end, identity theory describes the individual identity by three overlapping
dimensions: Person, Role and Group/Social (Davis et al. 2019; Stets and Serpe 2013). The dimensions
Person and Role refer to "identities which are internalized meanings attached to the self as a unique
person [as well as] an occupant of a role" (Davis et al. 2019, p. 3). Likewise, group identity describes the
305
Australasian Conference on Information Systems
2019, Perth Western Australia
Brünker et al.
Collective Identity Formation on Instagram
membership in an (activist) community whereas social identity is untied from a network and an
expression of one's status or lifestyle (Davis et al. 2019).
However, taking only identity theory into account is not sophisticating to describe the complex process
of collective identity formation during a social movement. Thus, Davis et al. (2019) integrated the theory
of collective identity into the dimension group/social of identity theory and called it collective
group/social identity (see Figure 1). The scholars described collective identity as a subtheory of social
identity theory that pertains to activist identification on social media (Davis et al. 2019). In this context,
collective group/social identity closes the gap between identity theory and social identity theory by
providing a theoretical foundation which considers the interrelation between individual, interpersonal
and group/social in an information system such as social media (Burke and Stryker 2016; Davis et al.
2019; Miller et al. 2016).
Figure 1: Conceptual integration of collective identity into identity theory (Davis et al. 2019)
Furthermore, collective identity combines the bases group cohesion, emotional attachment, and
solidarity (Polletta and Jasper 2001). These three bases might show up on social media in several ways.
However, even in heterogeneous networks which are loosely connected are these three bases important
for identity formation and maintenance during a social movement (Melucci 1988; Miller et al. 2016). In
the context of social movements, social media provides individuals self-verifying capabilities by various
participation patterns such as liking, commenting or connecting which others (Miller et al. 2016).
Subsequently, evolving personal networks could drive individuals to support the collective. This can be
motivated by emotional attachments on social media (Melucci 1988; White 2010). However, several
influencing factors might form collective action and drive participation in social movements as the social
identity model of collective action (SIMCA) suggests (van Zomeren et al. 2008).
2.2 Collective Identity and Collective Action
Collective action can be described as an action which is exclusively based on the voluntary cooperation
of individuals (Marwell and Oliver 1993). Furthermore, collective action strives to establish a public or
semi-public good (Heckathorn 1996). In this context, existing research often explains group interactions
as a central factor for social identity formation (Postmes et al. 2006; Postmes and Spears 2008; van
Zomeren et al. 2008) as well as opinion-based group identification (Thomas et al. 2012). According to
the integrative meta-analysis of collective action research by van Zomeren et al. (2008), the three
subjective variables perceived injustice, perceived efficacy, and social identification could influence
collective action. The scholars derived the SIMCA (Figure 2) which emphasises a central role of social
identity in promoting collective action. It shows how social identification is formed by group
interactions.
Figure 2: The social identity model of collective action (SIMCA) (van Zomeren et al. 2008)
However, Thomas et al. (2012, 2016) proposed an approach focussing on opinion-based group identity
as well as group-based effects on the formation of collective action. But, the studies neglected the link
306
Australasian Conference on Information Systems
2019, Perth Western Australia
Brünker et al.
Collective Identity Formation on Instagram
between identity theory and the formation of a collective identity towards collective action (Davis et al.
2019). Especially in social movements, individuals forming a collective identity which could emerge on
the basis of group members experiencing a common set of interests or goals (Stets and Serpe 2013).
However, people differ in the level of participation during a social movement which could be related to
individuals' multiple identities or the extent of commitment to the group (Stets and Serpe 2013). Thus,
examining the formation of collective identity is crucial to understand the evolving of collective action
in social movements. We, therefore, aim to examine collective group/social identity (Davis et al. 2019)
to investigate the role of identity formation on collective action.
3 RESEARCH DESIGN
3.1 Case Description
The social movement Fridays for Future grows in global attention due to the large devotion of the head
of the movement, Greta Thunberg. Thunberg is a young Swedish climate activist who started the global
phenomenon Fridays for Future by striking school each Friday. Moreover, this social movement is
mainly driven by young students. Fridays for Future appeals to the citizen in several countries around
the world to stand together and fight against climate change. The members of the movement utilising
social media to organise new strikes and demonstrations. Nowadays, the social movement reached
global attention. The participation in Fridays for Future scores about 4717 strikes in more than 3800
cities across every continent ("Strike List Countries" 2019).
3.2 Data Collection
In order to investigate how collective identity is formed on Instagram by the three bases of collective
identity: solidarity, emotional attachment, and group cohesion, we manually collected 1,137 comments
of two public available post by Greta Thunberg on Instagram. The two posts were chosen due to their
call for global participation on the first (post one) and second (post two) global strike for climate.
Furthermore, the social media platform Instagram is chosen as the object of research due to its
popularity, especially in the generation of current students. Likewise, public communication about a
certain topic builds up under the umbrella of a distinct post on Instagram. Thus, examining the
communication related to two major calls for participation in global climate strikes might give relevant
insights into the research objective.
3.3 Data Analysis
To analyse the collected comments for both posts, we saved the comment data in an SQLite database.
For every comment, we collected the publicly available data, consisting of the comment content (text),
date of creation, and the username of the poster. Additionally, we connected every comment with the
shortcode, Instagram's unique identifier for a post.
For the first analysis step, we annotated every comment with the language used within the comment
content. We could not annotate 372 out of the 1,137 comments, because the content contains only
mentions of other Instagram users, hashtags, or Emojis. We did not use Emojis, hashtags, or the
usernames for deducing the language, because this is not reliable and might lead to false conclusions.
We found sixteen different languages used in the comment contents. The dominant language is English
(584), followed by Italian (65), Swedish (53), Spanish (25), and French (10). The remaining eleven of
the sixteen languages are below ten occurrences each. Therefore, we focused on English comments in
our further analysis, to focus on the majority of comments containing one language and its specific
characteristics for categorising the comments.
In order to categorise every comment by the three bases of collective identity, we created three lists of
words, Emojis, and hashtags. To this end, we randomly selected 200 comments out of the 1,137 and
manually categorised words, Emojis, and hashtags to the fitting collective identity base. The subsequent
step for categorising the comments into the three collective identity bases is automated through a SQL
statement. During this step, a single comment can be assigned to multiple collective identity bases. We
searched for the previously defined words, Emojis, and hashtags within the comment contents with a
like operator. We do not count the number of occurrences for every word, Emoji or hashtag, we are
tracking if one is present or not. Therefore, a comment can be between none or all categories. This is
important, because a comment can, for example, be an emotional attachment and enhancing the group
cohesion.
307
Brünker et al.
Collective Identity Formation on Instagram
Australasian Conference on Information Systems
2019, Perth Western Australia
4 PRELIMINARY FINDINGS
Out of the 584 comments, the above-mentioned SQL statement could not classify 145 English comments
in at least one category. This is because of not explainable words (e.g. "periotd"), maybe due to
misspelling, very short questions (e.g. "Only ??") or comments in general (e.g. "Me") with just one or
two words, or because of text that does not make any sense in the context of Fridays for Future in general
or the two posts specifically. Some comments could be classified as troll comments, with content like
"lol" or "nah". Such comments can also be an Emotional Attachment, but without more context, we
decided to not classify them because the content is ambiguous. For this first analysis, we do not take
troll comments or short comments/questions into account. Due to the fact, that these comments are not
classified to one of the three bases, they are automatically ignored in our subsequent analysis. The
following table shows the words, Emojis, and hashtags for our classification:
Collective Identity
Bases
Group Cohesion
Emotional
Attachment
Words
Hashtags
we, us, together, strong,
can't wait, don't stop,
let's do, there, with you,
see you, @
cool, super, yes, stupid,
no, fuck, dumb, cunt,
asshole, bitch, shit,
useless, rock, best
#cleanair, #cleanenergy,
#FridaysForFuture, #ClimateStrike,
#climatestrikes, #climatecrisis,
#SchoolStrike4Climate,
#climatechange
-
Emojis
-
Solidarity
thank, thanks, legend,
hero
#GreatGreta, #DropoutGreta,
#fightthesystem, #govegan,
#endanimalabuse, #vegan,
#savetheplanet, #yesshecan
Table 1. Words, Emojis, and hashtags used for classification.
The corresponding SQL statement categorized the remaining 439 comments as Group Cohesion (299),
Emotional Attachment (279), and Solidarity (86). It is evident, that some comments were, at least,
categorized as two bases. 24 English comments are in all of the three categories. This is the case for
comments with, on average, more text than the other comments (more than 10 words), or due to many
used Emojis and/or hashtags. If only comments in exclusively one category are considered, our analysis
reveals 120 comments according to Group Cohesion, 100 to Emotional Attachment, and 18 to Solidarity.
The third table shows, that it is uncommon for a comment to belong to just one of our three categories.
The Fridays for Future movement arouses strong emotions among their followers. Therefore, and
because of the amount of Emojis and hashtags, a comment in the category Group Cohesion is likely to
be in the Emotional Attachment category, too, and vice versa. The data and our findings suggest, that
the two bases Group Cohesion and Emotional Attachment are the dominant factors expressing collective
identity within an opinion-based group in the context of Fridays for Future.
5 CONCLUSION AND NEXT STEPS
As this research in progress, we provide preliminary results on collective identity formation on social
media. Our findings show to what extent the three bases of collective social/group identity arise on
Instagram within an opinion-based community (Friday for Futures). In order to conceptualise the
concept of collective identity and collective action, we derived the proposed model of collective
group/social identity of collective action (Figure 3). We plan to test this model in a succeeding study to
examine the impact of each dimension on the collective group/social identity as well as the indirect effect
on collective action. To this end, we aim to conduct a laboratory study based on the preliminary findings
of this research in progress. The findings of this study provide first insights to how each base (Solidarity,
Emotional Attachment, and Group Cohesion) occur on Instagram within an opinion-based group.
308
Australasian Conference on Information Systems
2019, Perth Western Australia
Brünker et al.
Collective Identity Formation on Instagram
Figure 3: Proposed model of collective group/social identity of collective action
However, at this point of research, we need to expand the preliminary findings to observe a broader
dataset and review the findings against another sample of data. First, we need further classification on
the given dataset, including the phased-out languages. Therefore, we aim to enhance the classification
of the comments in order to get more precise results according to the new data. Likewise, we aim to
improve the list for the automated classification according to the new data. Second, we plan to gather
data from several official Fridays for Future Instagram accounts as well as Twitter communication
related to the Fridays for Future movement in order to validate the findings.
6 REFERENCES
Bunker, D., Mirbabaie, M., and Stieglitz, S. 2017. "Convergence Behaviour of Bystanders : An Analysis
of 2016 Munich Shooting Twitter Crisis Communication," in Proceedings of the 28th
Australasian Conference on Information Systems (ACIS), pp. 1 -- 11.
Burke, P. J., and Stryker, S. 2016. "Identity Theory: Progress in Relating the Two Strands," in New
Directions in Identity Theory and Research, New York, NY: Oxford University Press, pp. 657 --
810.
Carter, M., Compeau, D., Kennedy, M. I. L., and Schmalz, M. 2017. "The Content and Context of Identity
in a Digital Society," in Proceedings of the 25th European Conference on Information Systems
(ECIS), Guimarães, Portugal, pp. 3245 -- 3254.
Carter, M., and Grover, V. 2015. "Me, My Self, and I(T): Conceptualizing Information Technology
Identity and Its Implications," MIS Quarterly (39:4), pp. 931 -- 957.
Davis, J. L., Love, T. P., and Fares, P. 2019. "Collective Social Identity: Synthesizing Identity Theory and
Social Identity Theory Using Digital Data," Social Psychology Quarterly.
Heckathorn, D. D. 1996. "The Dynamics and Dilemmas of Collective Action," American Sociological
Review (61:2), pp. 250-277.
Hogg, M. A. 2018. "Social Identity Theory," in Contemporary Social Psychological Theories, Standford
CA: Stanford University Press.
Liu, N., and Chan, H. C. 2010. "Understanding the Influence of Social Identity on Social Support Seeking
Behaviors in Virtual Healthcare Community," in ICIS 2010 Proceedings.
Manikonda, L., Beigi, G., Liu, H., and Kambhampati, S. 2018. "Twitter for Sparking a Movement , Reddit
for Sharing the Moment: # Metoo through the Lens of Social Media," in 11th SBP-BRiMS 2018,
DC, USA.
Marwell, G., and Oliver, P. 1993. The Critical Mass in Collective Action: A Micro-Social Theory, New
York: Cambridge University Press.
McCall, G. J., and Simmons, J. L. 1966. Identities and Interactions, New York: Free Press.
Melucci, A. 1988. "Getting Involved: Identity and Mobilization in Social Move- Ments," International
Social Movement Research (1:26), pp. 329 -- 348.
Miller, S. D., Taylor, V., and Rupp, L. J. 2016. "Social Movements and the Construction of Queer
Identity," in New Directions in Identity Theory and Research, New York, NY: Oxford University
Press, pp. 443 -- 69.
Mirbabaie, M., Stieglitz, S., and Eiro, M. R. 2017. "#IronyOff -- Understanding the Usage of Irony on
Twitter during a Corporate Crisis," in PACIS 2017 Proceedings.
309
Australasian Conference on Information Systems
2019, Perth Western Australia
Brünker et al.
Collective Identity Formation on Instagram
Mirbabaie, M., Stieglitz, S., and Volkeri, S. 2016. "Volunteered Geographic Information and Its
Implications for Disaster Management," in Proceedings of the 49th Hawaii International
Conference on System Sciences (HICSS), pp. 207 -- 216.
Neimark, B. 2019. "Address the Roots of Environmental Crime," Science (364:6436), (J. Sills, ed.), pp.
138 -- 138.
Oh, O., Agrawal, M., and Rao, H. R. 2013. "Community Intelligence and Social Media Services: A Rumor
Theoretic Analysis of Tweets During Social Crises," MIS Quarterly (37:2), pp. 407 -- 426.
Polletta, F., and Jasper, J. M. 2001. "Collective Identity and Social Movements," Annual Review of
Sociology (27:1), pp. 283 -- 305.
Postmes, T., Baray, G., Haslam, S. A., Morton, T. A., and Swaab, R. I. 2006. "The Dynamics of Personal
and Social Identity Formation," in Individuality and the Group: Advances in Social Identity,
London, United Kingdom: SAGE Publications Ltd, pp. 215 -- 236.
Postmes, T., and Spears, R. 2008. Toward an Integrative Social Identity Model of Collective Action : A
Quantitative Research Synthesis of Three Socio-Psychological Perspectives, (134:4), pp. 504 -- 535.
Stets, J. E., and Serpe, R. T. 2013. "Identity Theory," in Handbook of Social Psychology, J. DeLamater
and A. Ward (eds.), Dordrecht: Springer Netherlands, pp. 31 -- 60.
Stieglitz, S., Meske, C., Ross, B., and Mirbabaie, M. 2016. "Going Back in Time to Predict the Future -
The Complex Role of the Data Collection Period in Social Media Analytics," Information Systems
Frontiers.
Stieglitz, S., Mirbabaie, M., and Potthoff, T. 2018. "Crisis Communication on Twitter during a Global
Crisis of Volkswagen -- The Case of 'Dieselgate,'" in Proceedings of the 51st Hawaii International
Conference on System Sciences (HICSS), pp. 513-522.
Stieglitz, S., Mirbabaie, M., Schwenner, L., Marx, J., Lehr, J., and Brünker, F. 2017. "Sensemaking and
Communication Roles in Social Media Crisis Communication," in Proceedings Der 13.
Internationalen Tagung Wirtschaftsinformatik (WI 2017), St. Gallen, pp. 1333 -- 1347.
"Strike List Countries." 2019. Fridays For Future, May 30. (https://fridaysforfuture.org/events/list,
accessed July 30, 2019).
Thomas, E. F., Mavor, K. I., and McGarty, C. 2012. "Social Identities Facilitate and Encapsulate Action-
Relevant Constructs: A Test of the Social Identity Model of Collective Action," Group Processes &
Intergroup Relations (15:1), pp. 75 -- 88.
Thomas, E. F., McGarty, C., and Mavor, K. 2016. "Group Interaction as the Crucible of Social Identity
Formation: A Glimpse at the Foundations of Social Identities for Collective Action," Group
Processes & Intergroup Relations (19:2), pp. 137 -- 151.
Turner, J. C., Hogg, M. A., Oakes, P. J., Reicher, S. D., and Wetherell, M. S. 1987. Rediscovering the
Social Group: A Self-Categorization Theory, Cambridge, MA: Basil Blackwell.
Tye, M., Leong, C., Tan, F. T. C., Tan, B., and Khoo, Y. H. 2018. "Social Media for Empowerment in
Social Movements: The Case of Malaysia's Grassroots Activism," Communications of the
Association for Information Systems (42:1), pp. 408 -- 430.
White, R. W. 2010. "Structural Identity Theory and the Post-Recruitment Activism of Irish Republicans:
Persistence, Disen- Gagement, Splits, and Dissidents in Social Movements," Social Problems
(57:3), pp. 341 -- 370.
van Zomeren, M., Postmes, T., and Spears, R. 2008. "Toward an Integrative Social Identity Model of
Collective Action: A Quantitative Research Synthesis of Three Socio-Psychological Perspectives,"
Psychological Bulletin (134:4), pp. 504 -- 535.
Acknowledgments
This work was supported by the Deutsche Forschungsgemeinschaft (DFG) under grant No. GRK 2167,
Research Training Group "User-Centred Social Media". Furthermore, this project has received funding
from the European Union's Horizon 2020 research and innovation programme under the Marie
Skłodowska-Curie grant agreement No 823866.
Copyright: © 2019 Brünker, Deitelhoff & Mirbabaie. This is an open-access article distributed under the terms
of the Creative Commons Attribution-NonCommercial 3.0 Australia License, which permits non-commercial
use, distribution, and reproduction in any medium, provided the original author and ACIS are credited.
310
|
1503.03137 | 1 | 1503 | 2015-03-11T00:52:20 | Comparing Network Structures of Commercial and Non-commercial Biohacking Online-communities | [
"cs.SI"
] | This paper compares two biohacking groups, Bulletproof Executive and DIYbio, whose distinct goals result in differences in social network structures, activities and entry points. | cs.SI | cs | COMPARING NETWORK STRUCTURES OF COMMERCIAL AND NON-COMMERCIAL
BIOHACKING ONLINE-COMMUNITIES
Sipra Bihani, Michael Hartman, Florian Sobiegalla, Amanda Rosenberg
IIT Stuart School of Business, University of Bamberg
10 West 35th Street, 18th Floor
Chicago USA, Bamberg Germany
[email protected], [email protected], [email protected],
[email protected]
PROCEEDINGS COINs15
ABSTRACT
This paper compares
two biohacking groups,
Bulletproof Executive and DIYbio, whose distinct
goals result in differences in social network structures,
activities and entry points.
INTRODUCTION
The term biohacking was first used online in
December 2008 and now encompasses a wide range
of topics and activities (Google Trends). We define
biohacking as the amateur practice of biological
experimentation for a self-defined purpose using a
variety of DIY devices and techniques in a non-
traditional setting (Bennett 2009; Delfanti 2013;
personal communication, October 29 2014).
The two groups this paper analyzes are DIYbio and
Bulletproof Executive (BE.) DIYbio was founded in
2008 by Mackenzie Cowell and Jason Bobe with the
mission of “establishing a vibrant, productive and
safe community of DIY biologists” (DIYbio). BE is
led by founder Dave Asprey, who has a personal and
financial interest in discovering methods that help
people “reach
the state of high performance”
(Bulletproof). These two groups have grown their
networks online in different ways.
COMPARISON OF BIOHACKING GROUPS
Both BE and DIYbio use online social media to share
knowledge, expand their network, and meet their
respective goals. Both groups have a website, a blog,
a Facebook page, and a Twitter account; DIYbio also
has a newsletter and Google group while BE has
multiple Twitter accounts, a YouTube channel, an
Instagram account, a Google+ page, a podcast, and an
online store. The five methods used to develop the
comparison are secondary research for context on
biohacking and
interview an active
the 2014 Bulletproof
biohacker who attended
Conference to learn the diversity of motivations of
biohackers; a survey to DIYbio local chapters to
understand their definition of biohacking and current
activities; and social network analysis with Condor,
trends; an
Gephi, and Wordle to create a network of actors; and
content analysis with TagCrowd of the two group’s
blogs to analyze activities.
Figure 1 (left) : Twitter fetch for “DIYbio” on Dec 6,
2014; Figure 2 (right): Twitter fetch for “LondonBioHack”
on Dec 6, 2014
DIYbio has the mission of democratizing biological
research, which has implications for its network
structure and entry points. As Figure 1 shows, the
network has no central node. The network
is
composed of 55 regional COINs from 21 countries
that set their own protocols (DIYbio). Figure 2 shows
the tweets of “LondonBioHack”, which is a local
chapter. Actors tweeting about this group are more
connected with one another than those tweeting about
DIYbio. Since actors create new technology and
ideas, LondonBioHack and DIYbio both coolfarm
with projects such as OpenerPCR and an outline of
biohacking ethics, respectively (survey, Nov 19, 2014,
DIYbio, Gloor 2010).
In regards to entry points, DIYbio has no form of
public outreach, such as advertising, for people
outside of the network to learn about it. The local
groups get public visibility by disseminating their
findings on their chapter websites and through
members’ personal blogs. In accordance with Gloor,
DIYbio can be classified as a self-selecting
collaborative innovation network (COIN) because
members aim to contribute to current scientific
knowledge in an open source environment (2007,
2010). People who are or want to become a part of
the virtual network are intrinsically motivated to
reach a shared goal and develop their own set of rules.
with a CIN in which actors share a common interest
and spot new trends that will become cool (Gloor
2007).
accounts:
@bulletproofexec
Figure 3: Twitter Fetch for “Bulletproof Coffee” in blue,
“Bulletproofexec” in red, “Dave Asprey” in green on Nov.
19, 2014. The central yellow node is (@bulletproofexec.)
To examine the structure of BE, we looked at its two
Twitter
and
@BPNutrition. As Figure 3 shows, Asprey, as
represented by the handle @bulletproofexec, is a
central node in the BE network; @BPNutrition is a
bridge to other accounts. There are tweets about
Bulletproof Coffee by accounts that are not directly
connected to either Asprey or @BPNutrition. The
commercial product has spurred attention to form a
Collaborative Interest Network (CIN) with members
who share a common interest and practice without
necessarily identifying as biohackers (Gloor, 2010).
Asprey continues to be a coolhunter by brokering a
network with different biohacking circles and experts
on self-improvement topics that he brings together on
his social media and at the annual Bulletproof
Conference (Gloor 2007). In an interview with a
participant at the conference, the biohacker said that
he felt like it was the first real biohacking conference
because it covered a variety of topics such as electron
flow, gut ecology, habit formation, and nutrient
injections (personal communication, October 29
2014.) This brokering enables BE to be influential in
many biohacking topic areas and provides channels
for commercialized products.
its different social media outlets, BE
Through
promotes the company’s central product and entry
point, called Bulletproof Coffee, which is marketed
as more effective at improving productivity and
efficiency than the average cup of coffee. People
without a specific interest in biohacking can learn
about Bulletproof Coffee by stumbling upon it at a
local coffee shop or by reading about it in media
outlets, such as The New York Times (Bulletproof).
As Figure 4 shows, the easily identifiable Bulletproof
Coffee had growth in Google searches that have far
surpassed searches for related terms. Based on the
commercial entry point and the mix of biohackers
and non-biohackers, BE can be described as a COIN
Figure 4: Google Trend visualization of interest in search
terms from January 2008 to December 2014.
CONCLUSION
The research finds a significant interrelation among
organization goals, structure, activity, and entry
points as shown in Table 1. DIYbio is an
interconnected network of deeply engaged biohackers
that coolfarms, whereas, by coolhunting and forming
CINs, BE exposes many people to biohacking
methods and products for self-improvement.
BE
DIYbio
coolhunting
commercial
no leader
coolfarming
open-source
Network structure a COIN with a CIN a COIN of COINs
Network influence leader
Activity
Entry points
Table 1: Comparison of BE and DIYbio based on influence,
structure, focus, and entry points.
Since biohacking is a young and growing practice,
some questions remain about the social network
structures of the two organizations under examination.
As the communities mature, their structure may
evolve. For instance, will the CIN members drinking
the COIN and begin
Bulletproof Coffee enter
biohacking? Will
the umbrella organization of
DIYbio continue to connect networks or will they
become more and more autonomous?
REFERENCES
Bennett, Gaymon & Nils Gilman, Anthony Stavrianakis & Paul
Rabinow (2009, Dec). From synthetic biology to biohacking: are
we prepared? Nature Biotechnology, 27, 12, 1109-1111.
Bulletproof (2014) Website. Retrieved from: http://bit.ly/1FMNiXp.
Visited on 2015, Feb 9.
Delfanti, Alessandro. (2013). Biohackers: The Politics of Open
Science. New York: Pluto Press.
DIYbio.
http://DIYbio.org.
Gloor, P., & Cooper, S. (2007). Coolhunting: Chasing Down the
Next Big Thing. New York: AMACOM.
Gloor, P. (2010). Coolfarming: Turn Your Great Idea into the Next
Big Thing. AMACOM, NY 2010.
Google Trends. (2014, Dec 6). ‘Bulletproof Coffee’, ‘DIYbio’,
‘Biohacking’ Topics. Retrieved from: http://bit.ly/1Bnqy0y
10). DIYbio. Retrieved
(2014, Nov
from:
|
1712.02433 | 1 | 1712 | 2017-12-06T22:59:51 | Identifying Data Noises, User Biases, and System Errors in Geo-tagged Twitter Messages (Tweets) | [
"cs.SI"
] | Many social media researchers and data scientists collected geo-tagged tweets to conduct spatial analysis or identify spatiotemporal patterns of filtered messages for specific topics or events. This paper provides a systematic view to illustrate the characteristics (data noises, user biases, and system errors) of geo-tagged tweets from the Twitter Streaming API. First, we found that a small percentage (1%) of active Twitter users can create a large portion (16%) of geo-tagged tweets. Second, there is a significant amount (57.3%) of geo-tagged tweets located outside the Twitter Streaming API's bounding box in San Diego. Third, we can detect spam, bot, cyborg tweets (data noises) by examining the "source" metadata field. The portion of data noises in geo-tagged tweets is significant (29.42% in San Diego, CA and 53.47% in Columbus, OH) in our case study. Finally, the majority of geo-tagged tweets are not created by the generic Twitter apps in Android or iPhone devices, but by other platforms, such as Instagram and Foursquare. We recommend a multi-step procedure to remove these noises for the future research projects utilizing geo-tagged tweets. | cs.SI | cs | Identifying Data Noises, User Biases, and System Errors in
Geo-tagged Twitter Messages (Tweets)
Ming-Hsiang Tsou
Hao Zhang
Chin-Te Jung
Center for Human Dynamics in the
Center for Human Dynamics in the
Esri (Beijing) R&D Center,
Mobile Age, San Diego State
University, 5500 Campanile Drive,
San Diego, CA 92182, USA.
01-6195940205
[email protected]
Mobile Age, San Diego State
University, 5500 Campanile Drive,
San Diego, CA 92182, USA.
[email protected]
Beijing, China
[email protected]
ABSTRACT
Many social media researchers and data scientists collected geo-
tagged tweets to conduct spatial analysis or identify spatiotemporal
patterns of filtered messages for specific topics or events. This
paper provides a systematic view to illustrate the characteristics
(data noises, user biases, and system errors) of geo-tagged tweets
from the Twitter Streaming API. First, we found that a small
percentage (1%) of active Twitter users can create a large portion
(16%) of geo-tagged tweets. Second, there is a significant amount
(57.3%) of geo-tagged tweets located outside the Twitter Streaming
API's bounding box in San Diego. Third, we can detect spam, bot,
cyborg tweets (data noises) by examining the "source" metadata
field. The portion of data noises in geo-tagged tweets is significant
(29.42% in San Diego, CA and 53.47% in Columbus, OH) in our
case study. Finally, the majority of geo-tagged tweets are not
created by the generic Twitter apps in Android or iPhone devices,
but by other platforms, such as Instagram and Foursquare. We
recommend a multi-step procedure to remove these noises for the
future research projects utilizing geo-tagged tweets.
Keywords
Social media, Geo-tagged, Data noises, Twitter, Tweets.
1. INTRODUCTION
Twitter is one of the most popular social media platforms used in
academic research works due to its large number of users,
comprehensive metadata, openness of messages, and public
available application programming interfaces (APIs). Tweets are
the actual Twitter messages created by users to express their
feelings, events, or activities within 140 characters including
spaces. Geocoding tasks for tweets are important for social media
analytics because data scientists and researchers need to aggregate
social media messages or users into a city, a region, or nearby
points of interests (POI) for location-based analysis and regional
trend analysis. Currently,
the public Twitter Application
Programming Interfaces (APIs) can provide five
types of
geocoding sources: 1. Geo-tagged coordinates, 2. Place check-in
location (bounding box), 3. User Profile Location, 4. Time Zones,
5. Texts containing locational information (explicit or implicit
information). Among the five geocoding methods, tweets with
geo-tagged coordinates is the most popular data source used in
location-based social media research. Geo-tagged tweets have
precise latitude and longitude coordinates (decimal degrees) stored
in a metadata field of tweets, called "geo" (a deprecated field name
in APIs) or "coordinates" (the current field name in APIs). When
users turn on the precise location tag function on their Twitter
accounts (which is off by default), their tweets will be geo-tagged
using GPS or Wi-Fi signals in their mobile devices. Since many
users do not enable precise location tags, there are only around 1%
of tweets containing geo-tagged information. The percentage of
geo-tagged tweets may vary among different topics or keywords.
For example, during a wildfire event, the percentage of geo-tagged
tweets can become 4% or higher collected by using wildfire related
keywords [1].
researchers used geo-tagged
Geo-tagged tweets are valuable data for social media researchers
and geographers to study geographic context and spatial association
within social media data. One popular method of collecting geo-
tagged tweets is to utilize Twitter's Streaming API with a
predefined bounding box or multiple predefined keywords.
Previous works have demonstrated that social media data collected
by the Twitter Streaming API (free version with 1% sampling rate)
are a good sample of Twitter's Firehose API (very expensive
version providing100% tweet data in Twitter servers) [2]. In
academics, many
for
conducting spatial analysis and GIS operations for their research
projects. For example, the 2013 special issue of "Mapping
Cyberspace and Social Media" in Cartography and Geographic
Information Science [3] includes seven refereed research papers
and four out of seven paper are using geo-tagged tweets as their
main data sources. There are two types of Twitter APIs in general,
Streaming API for collecting real-time feeds of Twitter messages
and Search API for collecting historical tweets (up to 7 or 9 days
before the search date) with specific keywords or user names with
database query methods. This paper will only focus on the
characteristics of geo-tagged tweets collected from the Twitter
Streaming API.
tweets
2. TWITTER SPAMS, BOTS, AND
CYBORGS
Previous research has identified three major types of data noises in
Twitter data, spams, bots, and cyborgs. Twitter spams, similar to
email spams, vary in different forms. Some spams are transparent
and easy to be identified while some spams are sophisticated [4].
Spams are usually created for reaching more users and increase the
financial gain for spammers. Spammers now use multiple platforms
to disseminate spams which also include social media services [5].
Many researchers have addressed the importance of eliminating
spams for a clearer and safer online networking environment [6].
The existence of spams in Twitter was first noticed in 2008 by
studying over 100,000 Twitter users and grouping them base on
follower-to-following ratios [7]. The groups contain broadcasters
(users with large amount of followers), acquaintances (users who
have a follower-to-following ratio close to 1), and miscreants (users
who follow a lot and less followed) [7]. Many Twitter spams
include URLs. According to Grier et al. [8], 8% of the URLs shared
on Twitter are considered as the linkage to malicious websites,
which contains malware or scams [8]. Clicking on those embedded
URL may cause serious problems to users' privacy and local
computers [9]. The harmfulness of spams does not only limit to
malware and scams. Spam message also waste the storage space on
server side [10].
to defeat
the definition of spam accounts
Twitter released its rules on eliminating and controlling spams and
abuse which include the regulation on banning the spam accounts
permanently, and
[11].
Unfortunately, some messages from spam accounts have been
changed correspondingly
these rules. Previously,
researchers tried to detect spam account by building a URL
blacklist. However, 90% of users have already clicked on a spam
URL before it is put into blacklist [8]. Furthermore, the URL
shorten service also lead to the uncertainty on detecting spam URL
[12]. Currently, three major trends for detecting spam accounts
include data compression algorithms [13], machine learning [14],
and statistics [15]. In recent years, more and more new spam
detection methods are created based on these three bases. Lin and
Huang [10] evaluated the tweets by examining the URL rate and
the interaction rate. Lumezanu and Feamster [5] categorized the
tweets by characterizing the publishing behavior and analyzing the
effectiveness of spam. Bayesian statistic was utilized for
classifying the tweets [16]. It is worth to learn that most of the spam
tweets are not sent by human but bots. However, bots does not only
send out spam tweets, but also send out useful information
sometimes, such as weather information, traffic update, and
earthquake events.
The existence of bots may not be widely recognized by all the
Twitter users. According to Japan Times, the bot is defined as
follow: "Twitter-bots are small software programs that are
designed to mimic human tweets. Anyone can create bots, though it
usually requires programming knowledge. Some bots reply to other
users when they detect specific keywords. Others may randomly
tweet preset phrases such as proverbs. Or if the bot is designed to
emulate a popular person (celebrity, historic icon, anime character
etc.) their popular phrases will be tweeted. Not all bots are fully
machine-generated, however, and interestingly the term "bot" has
also come to refer to Twitter accounts that are simply "fake"
accounts. [17]."
Meanwhile, cyborgs, a mix of humans and bots, refer to either bot-
assisted humans or human-assisted bots [18]. A bot can send tweets
automatically by called Twitter APIs. Sometimes, after a bot
receives audience, the creator (human) may tweet through bots
which led to a merged version of humans and bots – cyborgs. The
emergence of bots and cyborgs should be attributed to the growing
users population and open nature of Twitter [18]. Bots and cyborgs
generate many tweets everyday by providing various information,
including news updates, advertisements, emergency information, et
al. [18]. Some researchers study how to increase bots' influence in
social media [19]. On the other hand, identifying bot and cyborgs
are not easy [20]. According to Chu and his team [8], they identified
that 10.5% of Twitter accounts are bots while 36.2% are classified
as cyborgs. The existence of bots and cyborgs bring both pros and
cons. The information being tweeted, such as news, job posting,
SAMPLE: 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, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
SMSociety'17, July 28–30, 2017, Toronto, ON, Canada.
Copyright 2010 ACM 1-58113-000-0/00/0010 …$15.00.
DOI: http://dx.doi.org/...
allow people with the access to latest updates. However, the spam
tweets sent by bots are harmful to social media users.
Azmandian [21] introduce a two-steps procedure for eliminating
the bots from the Twitter data. The two-steps are: (1) All the Twitter
users whose tweets contain URLs more than 70% of the time will
be identified as bots. (2) All the users who traveled for more than
120km/h will be identified as a bot. However, these previous
research did not focus on the elimination of spams, bots and
cyborgs in geo-tagged tweets. This paper will focus on how to
identify and remove data noises and errors in geo-tagged tweets.
3. SYSTEM ERRORS IN THE
STREAMING APIS FOR GEO-TAGGED
TWEETS
Since the Twitter Streaming API provides the bounding box
option to enable users to search and collect geo-tagged tweets
within the bounding box area. Many people think that all collected
geo-tagged tweets by the Streaming API will be restricted within
the user-defined bounding box. However, we compared two case
studies of Streaming API with bounding boxes for the County of
San Diego and the City of Columbus during one month
(November, 2015). Both cases illustrated that there are only
42.7% of tweets are contained within San Diego County and
83.8% of tweets are within the City of Columbus (Table 1).
Table 1. The percentage of geo-tagged tweets within the original
bounding box or within the State boundary in the County of San
Diego and the City of Columbus collected in one month
(November, 2015).
Figure 1 illustrates the spatial distribution of geotagged tweets
collected by the Twitter Streaming API using the bounding box of
San Diego County during two months (October and November,
2015). Around 42.7% of the geo-tagged tweets are within the
boundary of San Diego County and 57.3% of tweets are outside the
original defined bounding box. We also conducted kernel density
maps using geotagged tweets and found out that the major hot spots
(clustered areas) are located around San Diego Downtown and two
other hot spots are around the cities of Los Angeles and San
Francisco. Figure 1 illustrates that the distribution of outside-
bounding box tweets are mainly located in California and a few
tweets are located in Mexico. There are a small amount of tweets
located in South America, Europe, and Southeast Asia.
Figure 2 illustrates the spatial distribution of geotagged tweets
using the bounding box of the City of Columbus in Ohio, USA
during October and November, 2015. 83.8% of tweets are within
the actual boundary of the City of Columbus. 15.8% of tweets are
within the Ohio State. Only 0.4% of geotagged tweets are outside
the Ohio State.
Comparing the two case studies, the San Diego case has much
higher ratio of outside bounding box tweets. The Columbus case
Percentage to total TweetsTweets in San Diego County (SDC)9794442.7%131464Tweets in Cali. State (excluding SDC)5638224.6%Other Regions7508232.7%Total Tweets229408Percentage to total TweetsTweets in Columbus City (CC)5329183.8%10279Tweets in Ohio State (excluding CC)1004315.8%Other Regions2360.4%Total Tweets63570GeoViewer@SanDiegoNovemberOutside BoundaryGeoViewer@ColumbusNovemberOutside Boundary
has better results of using bounding box to collect the mid-size city
level tweets.
The Twitter rule #2 may be the key reason for collecting tweets
outside the bounding box. For example, if a tweet selects
"California" as the place to check-in, the place box of California
will overlap with the bounding box of San Diego. Therefore, the
tweet outside San Diego will be selected. However, there are still
several global geo-tagged tweets which cannot be explained by the
heuristic rules. We will recommend that the future users of Twitter
Search APIs should add the geo-filter procedure after collecting
geo-tagged tweets using bounding box methods to remove the
system errors.
4. DATA NOISES FROM COMMERCIAL
BOT AND CYBORG TWEETS
In this paper, we defined the noises in social media as the messages
created by bots or cyborgs with commercial purposes (such as
advertisements and marketing events) since most social media
researchers focus on self-expression messages by actual human
beings. After manually reviewing hundreds of geo-tagged tweets in
our collection, we found out that one easy way to detect the bot or
cyborg tweets for advertisement is to examine a metadata field in
tweets, called "source". If a tweet was created on an iPhone device,
the source field will be "Twitter for iPhone". If a tweet was created
by specific bots or web programs, the source field could be
"TweetMyJOBS", "dlvr.it", or "AutoCarSale". Therefore, by
classifying different types of "source" values in the metadata, we
can remove these noises created by bots or cyborgs for commercial
purposes (Table 2).
We collected one month of geotagged tweets in San Diego County
in November 2015. After removing the tweets outside the San
Diego County, 97,944 tweets are within San Diego County. After
reviewing hundreds of different source names in tweets, we created
a black list for data noises and advertisement tweets based on the
source names in San Diego (Table 2). Among the total tweets in
San Diego, there are 29.42% tweets being recognized as noise data
by our list. Table 3 illustrates the top six categories of noise sources,
including Jobs (21.17%), Advertisement (3.49%), Weather
(2.18%), Earthquake information (1.06%), News (0.97%), and
Traffic (0.53%).
Similar to the previous comparison studies, we also collect one
month of geotagged tweets in the City of Columbus in November
2015. Table 3 illustrated the black list for data noises and
advertisement tweets in Columbus. There are 53.47% tweets being
recognized as noise data in Columbus based on our black list which
is much higher than the San Diego case. The top five categories of
noise sources in Columbus are Jobs (43.23%), Advertisement
(3.63%), Traffic (2.79%), News (2.10%), and Weather (1.72%).
We cannot find the Earthquake information bots in the City of
Columbus.
Table 2. The Proportion of data noises with different "source"
names in one month of geo-tagged tweets (November, 2015) in
San Diego County
Figure 1. The spatial distribution of geo-tagged tweets using the
Streaming API with the bounding box of San Diego County
(October and November 2015) at county, state, and world scales.
Figure 2. The spatial distribution of geo-tagged tweets using the
streaming API with the bounding box of Columbus City (October
and November 2015) at county, state, and world scales.
There is one possible explanation about this system error generated
by the Twitter APIs. Based on the Twitter API documents
(https://dev.twitter.com/streaming/overview/request-
parameters#locations), the streaming API uses the following
heuristic rules to determine whether a given Tweet falls within a
bounding box:
If the coordinates field is populated, the values there will be
tested against the bounding box. Note that this field uses
geoJSON order (longitude, latitude).
If coordinates is empty but place is populated, the region
defined in place is checked for intersection against the
locations bounding box. Any overlap will match.
If none of the rules listed above match, the Tweet does not
match the location query. Note that the geo field is deprecated,
and ignored by the streaming API.
1.
2.
3.
services with personalized recommendations and tips. The generic
Twitter apps are the fifth for Android and the seventh for iPhone
devices (Figure 3).
Figure 4 illustrates the proportion of "source" categories within the
City of Columbus. The most popular platform in Columbus is the
"TweetMyJOB" bots. The second popular platform is Instagram.
The ranking of the rest categories are similar to the San Diego case
study.
Table 3. The Proportion of data noises with different "source"
names in one month of geo-tagged tweets (November, 2015) in the
City of Columbus
Figure 3. The numbers of Tweets produced by different platforms
inside the San Diego bounding box during the month of November,
2015.
Figure 3 illustrates the proportion of "source" categories including
both noise and non-noise tweets within one month of geo-tagged
tweets (November 2015) from the Streaming API within the San
Diego County boundary. The red color indicates the tweets labelled
as bots or cyborgs based on our black list. The green color indicates
the tweets created by generic Twitter platform (such the Android or
iOS Twitter Apps). The blue color indicates the tweets from other
third-party apps or services, such as Instagram or Foursquare. The
most popular source (platform) in San Diego geo-tagged tweets is
Instagram (46,484 tweets, 47.46%), which is a very popular social
media platform for sharing photos and videos either publicly or
privately on mobile devices. Instagram users can extend their photo
sharing to other platforms, including Twitter and Facebook. Since
the default setting of Instagram is geotag-enabled. Therefore, most
Instagram messages sharing on Twitter include detail geolocations.
The second popular platform is "TweetMyJOB", which contains
job market advertising tweets using the geolocations of recruiting
companies. Foursquare is a location-based search-and-discovery
Figure 4. The numbers of Tweets produced by different platforms
inside the City of Columbus bounding box during the month of
November, 2015.
5. WHERE ARE THESE BOTS?
This section will focus on the visualization of the actual locations
of these bot and cyborg tweets in our collected data. From a spatial
analysis perspective, if we cannot remove these noises, the spatial
distribution patterns of these noises may cause a significant
problem or biases in the outcomes of spatiotemporal analysis.
We started to map the most popular bots in our black list:
TweetMyJOBS, which is a Twitter-based recruiting service
embedded with the job advertisement tweets. We found out that
most of TweetMyJOBS tweets are around the center of cities (San
Diego downtown or Columbus downtown) (Figure 5) due to the
higher density of business buildings and addresses in the center of
Source categorySource nameHashtagTweet numberPercentageJobTweetMyJOBS16005SafeTweet by TweetMyJOBS4726CareerCenter6Total2073721.17%Advertisementdlvr.it2837Golfstar269dine here182Simply Best Coupons77Auto City Sales56sp_californiaCoupon41Total34213.49%WeatherCities2105iembot24Sandaysoft Cumulus7Total21362.18%EarthquakeEarthquake762everyEarthquake203EarthquakeTrack.com69QuakeSOS9Total10431.06%NewsSan Diego Trends843WordPress.com111Total9540.97%TrafficTTN SD traffic512TTN LA traffic11Total5230.53%Percentage of Noise:29.42%Source categorySource nameHashtagTweet numberPercentageJobTweetMyJOBS16789SafeTweet by TweetMyJOBS6250Total2303943.23%Advertisementdlvr.it1642circlepix147dine here77Beer Menus53sp_ohioCoupon4DanceDeets4sp_oregonCoupon3SmartSearch2JCScoop1LeadingCourses.com1Total19343.63%TrafficTTN CMH traffic1486Total14862.79%NewsColumbus Trends1021eLobbyist80WordPress.com10twitterfeed8stolen_bike_alerter1Total11202.10%WeatherCities578iembot337Total9151.72%Percentage of Noise:53.47%
cities. The red color areas in Figure 5 are the high kernel density
areas created by the clusters of tweets.
should remove these bot or cyborg tweets before conducting any
spatiotemporal analysis using geotagged tweets.
Figure 5. The spatial distribution of the TweetMyJOBS tweets in
San Diego (top) and Columbus (bottom) during the month of
November, 2015. Red color areas are the clustered tweets in these
locations.
There are other spatial distribution patterns of bots or cyborgs based
on their source names:
• Cities is one of the twitter bots for weather forecasting. Their
spatial distribution is following the locations of major weather
stations or the center of local neighborhood throughout the
whole cities.
•
• Dlvr.it is a new service for attracting and engaging audiences
across the web with powerful content sharing tools. It can help
users to distribute their posted social media messages to other
platforms automatically. All dlvr.it tweets are located in a
single point (at the center of San Diego downtown and at the
center of Columbus downtown).
San Diego Trends and Columbus Trends is used for sending
out the local news. There is only one location tweeting through
San Diego Trends and Columbus Trends which is the
locations of San Diego City Hall and Columbus City Hall.
TNN CMH traffic is the platform of posting traffic accidents.
The whole platform generated 1486 tweets in Columbus City
during the month of November, 2015. According to the spatial
distribution and kernel density map below, the spatial
distribution of their tweets are highly correlated to major roads
and road intersections (Figure 6).
•
Based on these finding, we can conclude that the spatial
distributions of bots and cyborgs are not random. Different types of
bots have their unique spatial distribution patterns. Researchers
Figure 6. The spatial distribution of the TNN CMH traffic tweets
in Columbus during the month of November, 2015. Red dots are
the clustered tweets created by the traffic bots.
6. TWEETING FREQUENCY IN GEO-
TAGGED TWEETS (USER BIASES)
To explore potential user biases in geo-tagged tweets, this study
first calculated the frequency of geo-tagged tweets for individual
users. Geo-tagged tweets were collected within the bounding boxes
of San Diego County, CA and the City of Columbus, OH in U.S.
throughout the month of November, 2015. After removing the bots
and cyborgs, 69,317 human tweets are within San Diego bounding
box and 15,916 unique Twitter users (accounts) were identified
within the collected tweets. Figure 7 (left) reveals the number of
users along with their geo-tagged rates throughout the whole month
of November, 2015 in San Diego. Over 7,900 users only had one
tweet during the whole month, which consists up to 49% of total
users. More than 80% of Twitter users created less than 5 tweets in
the whole month. But 1% of Twitter users created 16% of total
Tweets. This finding is very similar to other literatures in Twitter
message analysis [22][23].
Tweets per user (San Diego)
Tweets per user (Columbus)
Figure 7. The number of users along with their geo-tagged tweeting
rates within November 2015 in San Diego County (left) and in
Columbus City (right).
Using the bounding box of the Columbus City, 29,902 human
tweets were identified for the month of November, 2015. 8,758
unique Twitter users are identified. Over 5,000 users in Columbus
only tweeted once during the whole month, which consists up to
58% of total users (Figure 7, right). Over 86% of Twitter users
created less than 5 tweets in the whole month. Meanwhile, 1% of
Twitter users created 19% of total Tweets.
The patterns of tweets per user in both San Diego and Columbus
cases are similar (Figure 1). The San Diego case has larger number
of human tweets and unique users. The Columbus case has higher
portion of users with only one tweet within one month and the
higher percentage of tweets created by 1% top users. Table 4 shows
a side by side comparison between the two case studies.
Table 4. User biases comparison between San Diego and
Columbus cases.
To avoid the user biases (small percentage of users creating large
amounts of tweets), we can calculate the number of unique users
rather than the number of tweets for statistic analysis [24]. We can
also remove top 1% or 5% of active users if we need to analyze the
common messages from the general population. Another possible
method is to select one tweet per user for sentiment or statistic
analysis.
7. DISCUSSION
When social media researchers and geographers utilized geo-
tagged tweets for their research projects, they need to identify and
remove data noises, user biases, and system errors from their
collected geo-tagged tweets. This paper provides a systematic view
to illustrate the characteristics of geo-tagged tweets from the
Streaming API focusing their noises, user biases, and system errors.
We recommend the following data process procedure for removing
the bot and cyborg tweets:
1. Using the Twitter Streaming with a bounding box to collect
geotagged tweets.
2. Use the same bounding box to filter out (remove) tweets
outside the bounding box or the actual boundary of cities using
GIS software.
necessarily reflect the views of the National Science Foundation.
The authors thank other HDMA team members' contribution to
this research.
REFERENCES
[1] Wang, Zheye, Xinyue Ye, and Ming-Hsiang Tsou.
(2016). Spatial, temporal, and content analysis of
Twitter for wildfire hazards. Natural Hazards: 1-18.
DOI: 10.1007/s11069-016-2329-6.
[2] Morstatter, Fred, Jürgen Pfeffer, Huan Liu, and
Kathleen M. Carley. "Is the sample good enough?
comparing data from twitter's streaming api with
twitter's firehose." arXiv preprint arXiv:1306.5204
(2013).
[3] Tsou, Ming-Hsiang, and Michael Leitner.
"Visualization of social media: seeing a mirage or a
message?." Cartography and Geographic Information
Science 40, no. 2 (2013): 55-60.
[4] Yardi, S., Romero, D., Schoenebeck, G., & boyd, d.
(2009). Detecting spam in a Twitter network. First
Monday, 15(1). doi:10.5210/fm.v15i1.2793.
[5] Lumezanu, C. & Feamster, N. (2012). Observing
Common Spam in Tweets and Email. IMC'12 Boston,
MA, USA.
[6] Chu, Z., Widjaja, I., & Wang, H. (2012). Detecting Social
Spam Campaigns on Twitter. Proceedings of the Applied
Cryptography and Network Security Conference (ACNS'12).
455–472. DOI:http://dx.doi.org/10.1007/978-3-642-31284-
7_27
[7] Krishnamurthy, B., Gill, P., & Arlitt, M. (2008, August). A
few chirps about twitter. In Proceedings of the first workshop
on Online social networks (pp. 19-24). ACM.
3. Manually review the source fields in the collected tweets to
identify top 10 or more bot or cyborg tweets and create a black
list for each city.
[8] Grier, C., Thomas, K., Paxson, V., & Zhang, M. (2010).
@spam: The underground on 140 characters or less. ACM
CCS.
4. Remove bot or cyborg tweets using the black list.
[9] Thosmas, K., & Nicol, D. M. (2010). The koobface botnet
5.
If the research needs to analyze common messages from the
general population, we should use the numbers of unique users
rather than the numbers of tweets in spatial analysis. Another
possible solution is to remove the top 1% or 5% active users
from the database.
Hopefully, by identifying these errors, biases, and noises,
researchers can remove these bots and cyborgs from their raw data
before conducting actual spatial analysis and improve their research
findings and outcomes.
8. ACKNOWLEDGMENTS
This material is based upon work supported by the National
Science Foundation under Grant No. 1416509, project titled
"Spatiotemporal Modeling of Human Dynamics Across Social
Media and Social Networks" and Grant No. 1634641, IMEE
project titled "Integrated Stage-Based Evacuation with Social
Perception Analysis and Dynamic Population Estimation". Any
opinions, findings, and conclusions or recommendations
expressed in this material are those of the author and do not
and the rise of social malware. In Proceedings of the 6rh
ACM Symposium on Information, Computer and
Communications Security.
[10] Lin, PC., & Huang, PM. A study of effective features for
detecting long-surviving Twitter spam accounts. IEEE
International Conference on Advanced Communication
Technology (ICACT), 2013 15th, 2013.
[11] Twitter (2016). The Twitter Rules. Archived from
https://support.twitter.com/articles/20170467.
[12] Uribl. (2013). Realtime URL Blacklist. Archived from
http://uribl.com/about.shtml.
[13] Bratko, A., Filipic, B., Cormack, G. V., Lynam, T. R., &
Zupan, B. (2006). Spam filtering using statistical data
compression methods. J. Mach. Learn. Res. 7:2673-2698.
[14] Goodman, J., Heckerman, D., & Rounthwaite, R. (2005).
Stopping spam. Scientific American 292(4):42–88.
[15] Fetterly, D., Manasse, M., & Najork, M. (2004). Spam, damn
spam, and statistics: using statistical analysis to locate spam
web pages. In Workshop on the Web and Databases.
Human TweetsHuman Users 1 time1-5 timesMost active userSan Diego69,31715,91649%84.20%903Columbus29,9028,75858%89.50%964Ratio of User who tweeted
[16] Wang, A. (2010). Don't Follow Me: Spam Detecting in
Twitter. Proc. Int'l Conf. Security and Cryptography
(SECRYPT).
tweets. Agents and Data Mining Interaction, vol.7607 of
Lecture Notes in Computer Science, pp. 139-149, Springer
Berlin Heidelberg
[17] Akimoto, A. (2011). Japan the Twitter nation. The Japan
[22] Li, Linna, Michael F. Goodchild, and Bo Xu. "Spatial,
Times. Archived from
http://www.japantimes.co.jp/life/2011/05/18/digital/japan-
the-twitter-nation/
[18] Chu, Z., Gianvecchio, S., Wang H., & Jajodia, S. (2010).
Who is Tweeting on Twitter, Human, Bot, or Cyborg.
ACSAC, Austin, Texas, USA.
[19] Messias, J., Schmidt, L., Oliveira, R., & Benevenuto, F.
(2013). You followed my bot! Transforming robots into
influential users in Twitter. First Monday, 18(7).
doi:10.5210/fm.v18i7.4217
[20] Boshmaf, Y., Muslukhov, I., Beznosov, K., & Ripeanu, M.
(2011). The socialbot network. In Proceedings of the 27th
Annual Computer Security Applications Conference (p. 93).
[21] Azmandian, M., Singh, K., Gelsey, B., Chang, Y.-H. &
Maheswaran, R. (2013). Following human mobility using
temporal, and socioeconomic patterns in the use of
Twitter and Flickr." Cartography and Geographic
Information Science 40, no. 2 (2013): 61-77.
[23] Sloan, L., & Morgan, J. (2015). Who Tweets with Their
Location? Understanding the Relationship between
Demographic Characteristics and the Use of Geoservices and
Geotagging on Twitter. PloS one, 10(11), e0142209.
[24] Tsou, Ming-Hsiang. "Research challenges and
opportunities in mapping social media and Big Data."
Cartography and Geographic Information Science 42,
no. sup1 (2015): 70-74.
|
1901.03629 | 1 | 1901 | 2019-01-11T16:11:45 | A Community-aware Network Growth Model for Synthetic Social Network Generation | [
"cs.SI",
"physics.soc-ph"
] | This study proposes a novel network growth model named ComAwareNetGrowth which aims to mimic evolution of real-world social networks. The model works in discrete time. At each timestep, a new link (I) within-community or (II) anywhere in the network is created (a) between existing nodes or (b) between an existing node and a newcoming node, based on (i) random graph model, (ii) preferential attachment model, (iii) a triangle-closing model, or (iv) a quadrangle-closing model. Parameters control the probability of employing a particular mechanism in link creation. Experimental results on Karate and Caltech social networks shows that the model is able to mimic real-word social networks in terms of clustering coefficient, modularity, average path length, diameter, and power law exponent. Further experiments indicate that ComAwareNetGrowth model is able to generate variety of synthetic networks with different statistics. | cs.SI | cs |
A Community-aware Network Growth Model for Synthetic Social
Network Generation
Furkan Gursoy
Bertan Badur
[email protected]
[email protected]
Dept. of Management Information Systems
Bogaziçi University
Istanbul, Turkey
Abstract
This study proposes a novel network growth model named ComAwareNetGrowth which aims to mimic
evolution of real-world social networks. The model works in discrete time. At each timestep, a new
link (I) within-community or (II) anywhere in the network is created (a) between existing nodes or (b)
between an existing node and a newcoming node, based on (i) random graph model, (ii) preferential
attachment model, (iii) a triangle-closing model, or (iv) a quadrangle-closing model. Parameters control
the probability of employing a particular mechanism in link creation. Experimental results on Karate
and Caltech social networks shows that the model is able to mimic real-word social networks in terms
of clustering coefficient, modularity, average path length, diameter, and power law exponent. Further
experiments indicate that ComAwareNetGrowth model is able to generate variety of synthetic networks
with different statistics.
Keywords Networks and graphs · Network growth models · Graph generators · Communities
1 Introduction
Purpose of this study is to build a network growth model that is able to generate synthetic networks which
resemble real-world social networks. Existing network growth models in the literature often aim to reproduce
only a very small set of network statistics. Our work, on the other hand, attempts to reproduce a larger list of
various statistics to evaluate the proposed model's accuracy.
The contribution of this study includes using combination of different growth mechanisms under a community-
aware framework, providing a growth model to understand evolution of many social networks, and providing a
generative mechanism which can be used in generating synthetic networks.
The rest of this study is structured as follows. In Section 2, the relevant literature is briefly reviewed. In
Section 3, the proposed model is formally described. In section 4, experimental results and findings are presented.
The conclusion and final remarks are given in Section 5.
Citation: Gursoy, F., & Badur, B. (2018). A Community-aware Network Growth Model for Synthetic Social Network Generation.
Proceedings of the 5th International Management Information Systems Conference
1
2 Background
Generally speaking, random graph models assume that the probabilities of links (edges) between nodes (vertices)
are independent of each other and equal. There are closely related versions of random graph models. For instance,
in original Erdos-Renyi model (Erdos & Renyi, 1959), all graphs on a fixed node set with a fixed number of links
are equally likely. On the other hand, in the model introduced by Gilbert (1959), each link has a fixed probability
of being present or absent, independently of the other links.
Random graph models and many other popular models in the literature are originally static models which
enable the study of structural features of networks. Another category of network models is growth models
(also called as generative network models). Preferential attachment model is an example of a model in the
latter category. (Simon, 1955), (Price, 1976), (Barabasi & Albert, 1999) are the seminal papers on preferential
attachment model. The main idea behind this model is as follows. At each time, a new node arrives and makes a
certain number of links with existing nodes using the node degrees as probability of making a link with the node.
This mechanism is able to generate a degree distribution which follows a power law which is observed in most
real-world networks.
Triadic closure is a concept in social network theory, first suggested by German sociologist Simmel (1908). It
follows the idea that two nodes are more likely to be connected if they have common neighbors. Generalizing
this, we can consider n-th order neighbors instead of only first-order neighbors. In this way, for instance, a
quadrangle-closure implies that two nodes are more likely to be connected if they have neighbors who are
neighbors with each other. Once a quadrangle is closed, a four-cycle is created. Lazega and Pattison (1999)
examine whether cycles larger than tri-cycles could be observed in an empirical setting to a greater extent than
could be accounted for by parameters for configurations involving at most three nodes. Snijders et al. (2006)
describe this as four-cycle partial conditional dependence and proposes it as a new configuration.
Kimura et al. (2004) present a model where links are created by a mixture of preferential attachment, uniform
attachment, and community-based attachment model. Leskovec et al. (2008) propose a network evolution model
considering the micromechanisms in social networks. Their model utilizes preferential attachment and triangle-
closure models in creating links. Lim et al. (2016) provide a comprehensive review of different approaches in
generating realistic synthetic graphs.
3 Proposed Model
The proposed model works in discrete time steps. It assumes that each node belongs to exactly one community.
At each time step, a new link is to be created. Nodes are assumed to arrive in uniform time intervals. The time
interval is found by dividing number of links (m) to number of nodes (n), thus a new node arrives at each m/n-th
step.
When a new node arrives, it is assigned a community label with respective probabilities given as input.
Then, the arriving node makes a link either within its community or in the whole network based on respective
probabilities specified by the input parameter. The link can be made with a uniformly selected random node or
with a node selected by using node degrees as the proxy probabilities (i.e., preferential attachment). If the link is
to be made within community, node degrees for preferential attachment process is calculated only considering the
links in the community.
At the time steps where no new node arrives, a new link is created between the existing nodes. One end of
the link is selected randomly among all existing nodes. The other end is selected within community/in whole
network, randomly or based on preferential attachment model; similar to the procedure explained previously for
a newly arriving node. However, another mechanism is available for links between existing nodes that is not
available for newly arriving nodes: triangle or quadrangle closing models. For a given node, a triangle is closed
by making a link to its second-order neighbor. A quadrangle is closed by making a link to its third-order neighbor.
2
More mechanisms can be created based on the proposed n-th order specification but considering that diameter
and average path length of observed networks are relatively small, including such mechanisms with higher n
value would mean making a link specifically to a distant node rather than a node in the neighborhood.
The selection of mechanisms (e.g., within community/in whole network, random/preferential attachment/triangle-
closure/quadrangle-closure) are controlled by input parameters. In some cases, the selected mechanism is not
able to generate a new link for reasons such as lack of possible triangles or quadrangles, or because all possible
links for the given node and mechanism already exist in the network. In those cases, at that time step, no action is
performed and model continues with the next time step. Although the property of uniform arrival of nodes is
distorted, the resulting network still has n nodes and m links.
In addition, potential target nodes are selected before checking whether they already have a link with the
source node. If a selected node is already connected, no action is performed at that time step too. This also results
in violation of uniform arrival of nodes. However, we chose to keep it this way to prevent overdensification of
relatively small communities where a randomly selected node is more likely to be an immediate neighbor.
Each mechanism employed in ComAwareNetGrowth model corresponds to the simplified versions of possible
link formations in real-world networks. The stronger the communities, the more probable for a node to make
connection within its community. Random links are usually less frequent in most social networks but actually
an existing mechanism. Preferential attachment reflects the 'rich get richer' phenomena. Triangle closures and
quadrangle closures are justified by the observation that nodes are usually more likely to make links within their
neighborhood rather than distant nodes.
Algorithm 1 presents our proposed model in general with omitting some details. The complete source code
in R software language and some network datasets generated using the proposed methodology are available at
furkangursoy.github.io.
4 Experimental Analysis
In this section, several experiments are conducted and results are analyzed. In the first part, we attempt to generate
two real-world undirected simple social networks: Zachary's Karate Club (Zachary, 1977) and Caltech Facebook
Network (Traud et al., 2011). The first network, called as Karate in the rest of the study, is a well-known social
network of a university karate club where links exist between members who interact outside the club. The latter
network, called as Caltech in the rest of the study, consists of the complete set of users from the Facebook network
of California Institute of Technology.
Table 1 displays various statistics of the two networks. The statistics are calculated using R's igraph package.
In calculation of modularity; first, communities are detected by utilizing fast greedy community detection
algorithm available in R. Estimated communities, then, are used in calculation of modularity. The sizes of the
found communities are also utilized as input probabilities for belonging to communities in the network growth
experiments.
Table 1: Network Statistics
n
34
769
m
78
16656
Clustering
Coefficient
0.26
0.29
Karate
Caltech
Avg. Path
Length Modularity Diameter
5.00
6.00
2.41
2.34
0.38
0.33
Power Law
Exponent
2.55
1.50
3
Algorithm 1 ComAwareNetGrowth: The Proposed Community-aware Network Growth Model
4.1 Experiments on Karate Network
The following parameter values are used to generate a social network that mimics Karate network: n = 34,
m = 78, k = 3, C = {0.24, 0.5, 0.26}, comp = 0.8, rpn = 0.22, ppn = 0.78, rpe = 0.04, ppe = 0.14,
c3pe = 0.41, and c4pe = 0.41. The experiments are repeated 10 times, and statistics for the generated and
observed networks are given in Table 2.
On average, the proposed growth model with the given parameter settings is able to mimic the Karate network
in general. However, modularity and and power law exponent1 (Newman, 2005) statistics are not closely matched.
Although most of the statistics can be generated with small standard deviations, modularity and clustering
coefficient show relatively larger variations in these experiments. This might be due to the very small size of the
network.
1Most networks follow power-law in the tail. Accordingly, in fitting the power law, lower bound is set as 2 for Karate and 5 for all
other experiments.
4
Table 2: Experimental Results for Karate Network
Avg. Path
Clustering
Coefficient
0.18
0.24
0.22
0.25
0.19
0.30
0.20
0.27
0.30
0.27
0.24
0.04
0.26
-0.02
Run#1
Run#2
Run#3
Run#4
Run#5
Run#6
Run#7
Run#8
Run#9
Run#10
Mean
StdDev
Observed
Diff.
Length Modularity Diameter
4.00
5.00
4.00
5.00
5.00
5.00
5.00
6.00
6.00
5.00
5.00
0.63
5.00
0.00
0.26
0.26
0.20
0.27
0.34
0.27
0.21
0.22
0.22
0.34
0.26
0.05
0.38
-0.12
2.32
2.42
2.35
2.47
2.50
2.38
2.31
2.52
2.55
2.56
2.44
0.09
2.41
0.03
Power Law
Exponent
2.20
2.13
2.10
2.31
2.22
2.20
2.24
2.02
2.14
2.19
2.18
0.08
2.55
-0.37
Figure 1 visualizes the communities found in the original Karate network and a synthetic network generated
by our model. In the synthetic network, communities are less separated than the original network. Especially, two
communities (displayed in white and black) do not come as separate communities. Note that, this is just one of
the generated networks. Other networks might more closely resemble the original network.
Figure 1: (a) Observed Network, (b) Generated Network
4.2 Experiments on Caltech Network
The following parameter values are used to generate a network that mimics Caltech network: n = 769, m =
16656, k = 8, C = {0.375, 0.341, 0.254, 0.017, 0.005, 0.004, 0.003, 0.003}, comp = 0.85, rpn = 0.333,
ppn = 0.666, rpe = 0.091, ppe = 0.182, c3pe = 0.363, and c4pe = 0.363. The experiments are repeated 10
times, and statistics for the generated and observed networks are given in Table 3.
5
Experimental results show that average path length, modularity, and power law exponent are closely matched
with negligible standard deviations. However, clustering coefficient and diameter values are not replicated
successfully in terms of difference between the generated and observed statistic. Caltech network is a larger
network than the Karate network. Therefore, as desired, variances in the generated statistics are almost zero.
Consequently, it can be concluded that given a sufficient number of nodes and links, our proposed growth model
is able to produce stable results.
Table 3: Experimental Results for Caltech Network
Clustering
Coefficient
0.17
0.17
0.17
0.17
0.17
0.17
0.17
0.17
0.17
0.17
0.17
0.00
0.29
-0.12
Run#1
Run#2
Run#3
Run#4
Run#5
Run#6
Run#7
Run#8
Run#9
Run#10
Mean
StdDev
Observed
Diff.
Avg. Path
Length Modularity Diameter Power Law
Exponent
1.52
2.19
2.20
1.52
1.52
2.20
1.52
2.19
1.52
2.19
1.52
2.19
2.20
1.52
1.52
2.20
1.52
2.19
1.52
2.20
2.20
1.52
0.00
0.00
1.50
2.34
-0.14
0.02
4.00
4.00
4.00
4.00
4.00
4.00
4.00
4.00
4.00
4.00
4.00
0.00
6.00
-2.00
0.32
0.32
0.33
0.34
0.32
0.33
0.32
0.33
0.33
0.32
0.33
0.01
0.33
0.00
4.3 Other Experiments
Another set of experiments are performed without the purpose of mimicking any real network but to explore
the variety of networks which can be generated with our proposed growth model. Parameter settings for eight
experiments are presented in Table 4. The first and last four experiments are the same except that m = 2000
for former experiments whereas m = 5000 for latter experiments, hence a denser network. Each experiment is
repeated for 10 times. The mean and standard deviations of generated statistics are presented in Table 5.
Table 4: Settings for Other Experiments
n m k C
{.2, 0.2, 0.2, 0.2, 0.2}
Exp#1 500 2000 5
Exp#2 500 2000 5
{.2, 0.2, 0.2, 0.2, 0.2}
Exp#3 500 2000 5 {.005, 0.055, 0.11, 0.28, 0.55}
Exp#4 500 2000 5 {.005, 0.055, 0.11, 0.28, 0.55}
{.2, 0.2, 0.2, 0.2, 0.2}
Exp#5 500 5000 5
Exp#6 500 5000 5
{.2, 0.2, 0.2, 0.2, 0.2}
Exp#7 500 5000 5 {.005, 0.055, 0.11, 0.28, 0.55}
Exp#8 500 5000 5 {.005, 0.055, 0.11, 0.28, 0.55}
comp rpn ppn rpe ppe c3pe c4pe
0.50 0.50 0.5 0.25 0.25 0.25 0.25
0.75 0.33 0.66 0.10 0.20 0.30 0.40
0.75 0.33 0.66 0.10 0.20 0.30 0.40
0.25 0.17 0.83 0.12 0.63 0.12 0.12
0.50 0.50 0.50 0.25 0.25 0.25 0.25
0.75 0.33 0.66 0.10 0.20 0.30 0.40
0.75 0.33 0.66 0.10 0.20 0.30 0.40
0.25 0.17 0.83 0.12 0.63 0.12 0.12
The experiments confirm that standard deviations are indeed very low. Diameter might be seen as an exception
to this but it is mostly due to the nature of that statistic. Since it is an integer value, even if the model generates
only two adjacent integer values for this statistic, standard deviation might come as large at the first sight.
As the number of links increase from 2000 to 5000, values for average path Length and diameter get smaller
as expected. On the other hand, clustering coefficent increases. There does not seem to be a meaningful change
in the values of modularity.
6
Table 5: Other Experimental Results
Avg. Path
Clustering
Coefficient
0.10
0.00
0.11
0.00
0.10
0.00
0.11
0.01
0.15
0.00
0.16
0.00
0.15
0.00
0.17
0.00
Exp#1 Mean
Std.Dev.
Exp#2 Mean
Std.Dev.
Exp#3 Mean
Std.Dev.
Exp#4 Mean
Std.Dev.
Exp#5 Mean
Std.Dev.
Exp#6 Mean
Std.Dev.
Exp#7 Mean
Std.Dev.
Exp#8 Mean
Std.Dev.
Length Modularity Diameter
7.50
0.71
7.60
0.52
7.60
0.70
7.60
0.52
5.10
0.32
4.90
0.32
5.20
0.42
5.00
0.00
0.38
0.01
0.47
0.01
0.34
0.02
0.42
0.01
0.36
0.01
0.44
0.00
0.31
0.01
0.40
0.02
3.31
0.02
3.33
0.03
3.29
0.02
3.39
0.02
2.51
0.00
2.52
0.01
2.51
0.01
2.55
0.01
Power Law
Exponent
2.41
0.03
2.41
0.04
2.39
0.03
2.47
0.04
1.82
0.01
1.81
0.01
1.82
0.01
1.82
0.01
5 Conclusion
In this work, we have developed a network growth model with the aim of exploring the mechanisms behind most
real-world networks, and being able to mimic real-world networks through these mechanisms. The proposed
growth model serves as an initial step to build a realistic growth model which can generate graphs with any set of
given statistics.
When there exist a sufficient number of nodes and links, our model generates networks which do not vary
between themselves in terms of the network statistics employed in this study. Such stability is highly desired.
However, some network statistics are not being successfully replicated. This might be because one or combination
of the two things. First, parameter values are determined based on a few manual experiments rather than estimating
them from the real network. Second, the mechanisms employed in the growth model might not be sufficient to
generate any social network easily, which suggest further development of the model.
Given the limitations of the current work, more effort in future should be directed towards developing more
accurate mechanisms, exploring relationships between the mechanisms, and estimating parameter values via
more systematic calibration experiments.
7
References
Barabasi, A. L., & Albert, R. (1999). Emergence of scaling in random networks. Science, 286(5439), 509-512.
doi:10.1126/science.286.5439.509
Erdos, P., & Renyi, A. (1959). On random graphs I. Publicationes Mathematicae (Debrecen), 6, 290-297.
Gilbert, E. N. (1959). Random graphs. The Annals of Mathematical Statistics, 30(4), 1141-1144.
Kimura, M., Saito, K., & Ueda, N. (2004). Modeling of growing networks with directional attachment and
communities. Neural Networks, 17(7), 975-988. doi:10.1016/j.neunet.2004.01.005
Lazega, E., & Pattison, P. E. (1999). Multiplexity, generalized exchange and cooperation in organizations: A case
study. Social Networks, 21(1), 67-90. doi:10.1016/s0378-8733(99)00002-7
Leskovec, J., Backstrom, L., Kumar, R., & Tomkins, A. (2008). Microscopic evolution of social networks.
Proceeding of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining -
KDD 08. doi:10.1145/1401890.1401948
Lim, S., Lee, S., Powers, S. S., Shankar, M., & Imam, N. (2016). Survey of Approaches to Generate Realistic
Synthetic Graphs. doi:10.2172/1339361
Newman, M. (2005). Power laws, Pareto distributions and Zipfs law. Contemporary Physics, 46(5), 323-351.
doi:10.1080/00107510500052444
Price, D. D. (1976). A general theory of bibliometric and other cumulative advantage processes. Journal of the
American Society for Information Science, 27(5), 292-306. doi:10.1002/asi.4630270505
Simmel, G. (1908). Sociologie. Untersuchungen über die Formen der Vergesellschaftung. Duncker & Humblot.
Simon, H. A. (1955). On a Class of Skew Distribution Functions. Biometrika, 42(3/4), 425. doi:10.2307/2333389
Snijders, T. A., Pattison, P. E., Robins, G. L., & Handcock, M. S. (2006). New Specifications for Exponential
Random Graph Models. Sociological Methodology, 36(1), 99-153. doi:10.1111/j.1467-9531.2006.00176.x
Traud, A. L., Kelsic, E. D., Mucha, P. J., & Porter, M. A. (2011). Comparing Community Structure to Character-
istics in Online Collegiate Social Networks. SIAM Review, 53(3), 526-543. doi:10.1137/080734315
Zachary, W. W. (1977). An Information Flow Model for Conflict and Fission in Small Groups. Journal of
Anthropological Research, 33(4), 452-473. doi:10.1086/jar.33.4.3629752
8
|
1905.10351 | 3 | 1905 | 2019-08-28T15:36:05 | Tempus Volat, Hora Fugit -- A Survey of Tie-Oriented Dynamic Network Models in Discrete and Continuous Time | [
"cs.SI",
"stat.AP"
] | Given the growing number of available tools for modeling dynamic networks, the choice of a suitable model becomes central. The goal of this survey is to provide an overview of tie-oriented dynamic network models. The survey is focused on introducing binary network models with their corresponding assumptions, advantages, and shortfalls. The models are divided according to generating processes, operating in discrete and continuous time. First, we introduce the Temporal Exponential Random Graph Model (TERGM) and the Separable TERGM (STERGM), both being time-discrete models. These models are then contrasted with continuous process models, focusing on the Relational Event Model (REM). We additionally show how the REM can handle time-clustered observations, i.e., continuous time data observed at discrete time points. Besides the discussion of theoretical properties and fitting procedures, we specifically focus on the application of the models on two networks that represent international arms transfers and email exchange. The data allow to demonstrate the applicability and interpretation of the network models. | cs.SI | cs | Tempus Volat, Hora Fugit - A Survey of
Tie-Oriented Dynamic Network Models in
Discrete and Continuous Time
Cornelius Fritz∗, Michael Lebacher and Goran Kauermann
Department of Statistics, Ludwig-Maximilians-Universitat Munchen
Abstract
Given the growing number of available tools for modeling dynamic networks, the
choice of a suitable model becomes central. The goal of this survey is to provide an
overview of tie-oriented dynamic network models. The survey is focused on introducing
binary network models with their corresponding assumptions, advantages, and short-
falls. The models are divided according to generating processes, operating in discrete
and continuous time. First, we introduce the Temporal Exponential Random Graph
Model (TERGM) and the Separable TERGM (STERGM), both being
time-discrete models. These models are then contrasted with continuous process mod-
els, focusing on the Relational Event Model (REM). We additionally show how the
REM can handle time-clustered observations, i.e., continuous time data observed at
discrete time points. Besides the discussion of theoretical properties and fitting pro-
cedures, we specifically focus on the application of the models on two networks that
represent international arms transfers and email exchange. The data allow to demon-
strate the applicability and interpretation of the network models.
Keywords -- Continuous-Time, Discrete-Time, Event Modeling, ERGM, Random
Graphs, REM, STERGM, TERGM
∗[email protected]
1
9
1
0
2
g
u
A
8
2
]
I
S
.
s
c
[
3
v
1
5
3
0
1
.
5
0
9
1
:
v
i
X
r
a
1
Introduction
The conceptualization of systems within a network framework has become popular within
the last decades, see Kolaczyk (2009) for a broad overview. This is mostly because network
models provide useful tools for describing complex dependence structures and are applicable
to a wide variety of research fields. In the network approach, the mathematical structure of
a graph is utilized to model network data. A graph is defined as a set of nodes and relational
information (ties) between them. Within this concept, nodes can represent individuals,
countries or general entities, while ties are connections between those nodes. Dependent on
the context, these connections can represent friendships in a school (Raabe et al., 2019),
transfers of goods between countries (Ward et al., 2013), sexual relations between people
(Bearman et al., 2004) or hyperlinks between websites (Leskovec et al., 2009) to name just
a few. Given a suitable data structure for the system of interest, the conceptualization as
a network enables analyzing dependencies between ties. A central statistical model that
allows this is the Exponential Random Graph Model (ERGM, Robins and Pattison, 2001).
This model permits the inclusion of monadic, dyadic and hyperdyadic features within a
regression-like framework.
Although the model allows for an insightful investigation of within-network dependencies,
most real-world systems are typically more complex. This is especially true if a temporal
dimension is added, which is relevant, as most systems commonly described as networks
evolve dynamically over time. It can even be argued that most static networks are de facto
not static but snapshots of a dynamic process. A friendship network, e.g., typically evolves
over time and influences like reciprocity often follow a natural chronological order.
Of course, this is not the first paper concerned with reviewing temporal network models.
Goldenberg et al. (2010) wrote a general survey covering a wide range of models. The authors
laid the foundation for further articles and postulated a soft division of statistical network
models into latent space (Hoff et al., 2002) and p1 models (Holland and Leinhardt, 1981), all
originating in the Eds-Rnyi-Gilbert random graph models (Erdos and R´enyi, 1959; Gilbert,
1959). Kim et al. (2018) give a contemporary update on the field of dynamic models building
on latent variables. Snijders (2005) discusses continuous time models and reframes the
independence and reciprocity model as a Stochastic Actor oriented Model (SAOM, Snijders,
1996). Block et al. (2018) provide an in-depth comparison of the Temporal Exponential
Random Graph Model (TERGM, Hanneke et al., 2010) and the SAOM with special focus
on the treatment of time. Further, the ERGM and SAOM for networks which are observed
at single time points are contrasted by Block et al. (2019), deriving theoretical guidelines for
model selection based on the differing mechanics implied by each model.
In the context of this compendium of articles, the scope is to give an update on the
dynamic variant of the second strand of models relating to p1 models. We therefore extend
2
Figure 1: Tree diagram summarizing the dependencies between models originating in the
Erds-Rnyi-Gilbert graph model, models situated in a box with a grey background are dis-
cussed in this article. This graph is an update of Figure 6.1 in Goldenberg et al. (2010).
the summarizing diagram of Goldenberg et al. (2010) as depicted in Figure 1. Generally, we
divide temporal models into two sections, by differentiating between discrete and continuous
time network models. In this endeavor, we emphasize reviewing tie-oriented models. Tie-
oriented models are concerned with formulating a stochastic model for the existence of a tie
contrasting the actor-oriented approach by Snijders (2002), which specifies the model from
the actors point of view (Block et al., 2018). The Dynamical Actor-Oriented model (DyNAM,
Stadtfeld and Block, 2017) adopts this actor-oriented paradigm to event data. This type of
model was formulated with a focus on social networks (Snijders, 1996). Contrasting this, tie-
oriented models can be viewed as more general, since they are also applicable to non-social
networks.
Statistical models for time discrete data rely on an autoregressive structure and condition
the state of the network at time point t on previous states. This includes the TERGM and
the Separable TERGM (STERGM, Krivitsky and Handcock, 2014). There exists a variety of
recent applications of the TERGM. White et al. (2018) use a TERGM for modeling epidemic
disease outcomes and Blank et al. (2017) investigate interstate conflicts. In He et al. (2019)
Chinese patent trade networks are inspected and Benton and You (2017) use a TERGM
for analyzing shareholder activism. Applications of STERGMs are given for example by
Stansfield et al. (2019) that model sexual relationships and Broekel and Bednarz (2019) that
study the network of research and development cooperation between German firms.
3
Erdös-Rényi-Gilbdert(Gilbert 1959, Erdös and Rényi 1959)Latent Space ModelsStatic Variant (Hoff, Raftery, Handcock2002)Dynamic Variant (Sarkar, Moore 2005)p1 models(Holland and Leinhardt1981)Static Variant Dynamic Variant ERGM(Frank and Strauss1986)TERGM(Hanneke and Xing2006)STERGM(Krivitski and Handcock2012)Markov Chain-BasedDiscrete TimeMarkov Process-BasedContinuous TimeSAOM (Snijders 1996)DyNAM(Stadtfeld et al. 2017)REM(Butts 2008)Longitudinal DataTime Stamped Data In case of time-continuous data, the model regards the network as a continuously evolving
system. Although this evolution is not necessarily observed in continuous time, the process
is taken to be latent and explicitly models the evolution from the state of the network at
time point t− 1 to t (Block et al., 2018). In this paper we discuss the relational event model
(REM, Butts, 2008) for the analysis of event data. Applications of the REM are manifold
and range from explaining the dynamics of health behavior sentiments via Twitter (Salath´e
et al., 2013), inter-hospital patient transfers (Vu et al., 2017), online learning platforms (Vu
et al., 2015), and animal behavior (Tranmer et al., 2015) to structures of project teams
(Quintane et al., 2013). Eventually, the REM is adapted to time-discrete observations of
networks. That is, we observe the time-continuous developments of the network at discrete
observation times only. Henceforth, we use the term time-clustered for this special data
structure.
In reviewing dynamic network models, we assume a temporal first-order Markov depen-
dency. To be more specific, this implies that the network at time point t only depends on
the previous observation of the network. This characteristic is widely used in the analysis of
longitudinal networks (Hanneke et al., 2010; Krivitsky and Handcock, 2014) and the result-
ing conditional independence among states of the network facilitates the estimation with an
arbitrary number of time points. In that respect, it suffices to only include two observational
moments for illustrative purposes, since the interpretation and estimation with a longer se-
ries of networks is unchanged. Lastly, the comparison of the methods at hand in a clear-cut
manner is hence enabled.
The paper is structured as follows. In Section 2 we give basic definitions that are used
throughout the paper and present the two data examples that will be analyzed as illustrative
examples. After that, Section 3 introduces time-discrete and Section 4 time-continuous
network models. They are applied in Section 5 on two data sets and Section 6 concludes.
Additional results relating to the applications can be found in the Supplementary Material.
2 Definitions and Data Description
2.1 Definitions
This article regards directed binary networks, with ties representing directed relations be-
tween two nodes at a time point. The respective information can be represented in an
adjacency matrix Yt = (Yij,t)i,j=1,..,n ∈ Y, where Y = {Y : Y ∈ {0, 1}n×n} represents the
set of all possible networks with n nodes. The entry (i, j) of Yt is "1" if a tie is outgo-
ing from node i to j in year t and "0" otherwise. Further, the discrete time points of
the observations of Yt are denoted as t = 1, . . . , T . We restrict our analysis to two time
points in both exemplary networks, which suffices for comparison. Hence, we set T = 2. In
4
t
Time
Number of events
Number of nodes
Number of possible ties n(n − 1)
Density
Transitivity
Reciprocity
Repetition
n
Arms Trade Network
2016
−
180
2017
−
180
32 220
0.021
0.195
0.081
−
32 220
0.020
0.202
0.083
0.641
Email Network
Period 1 Period 2
4 957
88
7 656
0.123
0.407
0.7
−
2 537
88
7 656
0.087
0.345
0.687
0.574
Table 1: Descriptive statistics for the international arms trade network (left) and the Euro-
pean research institutions email correspondence (right).
many networks, including our running examples self-loops are meaningless. We therefore fix
Yii,t ≡ 0 ∀ i ∈ {1, . . . , n} throughout the article. Further, all sub-scripted temporal indices
(Yt) are assumed to take discrete and all indices in brackets (Y (t)) continuous values. The
temporal indicator t denotes the observation times of the network and to notationally differ
this from time-continuous model we write t for continuous time.
To sufficiently compare different models, we use two application cases. The first one
represents the international trade of major weapons, which is given by discrete snapshots
of networks that are yearly aggregated over time-continuous trade instances, i.e., the time-
stamped information is not observed. Whereas the second application, a network of email
traffic, comes in time-stamped format, that can be aggregated to discrete-time observations.
2.2 Data Set 1: International Arms Trade
The data on international arms trading for the years 2016 and 2017 are provided by the
Stockholm International Peace Research Institute (SIPRI, 2019). To be more specific, infor-
mation on the exchange of major conventional weapons (MCW) together with the volume
of each transfer is included. In order to have a binary network representation, we discretize
the data and set edges Yij,t to "1" if country i sent arms to country j in t.
The left side of Table 1 gives some descriptive measures (Csardi and Nepusz, 2006) and
Figure 2 visualizes the arms trade network using the software Gephi (Bastian et al., 2009).
The density of a network is the proportion of realized edges out of all possible edges and
is similar in both years, indicating the sparsity of the modeled network. Clustering can
be expressed by the transitivity measure, providing the percentage of triangles out of all
connected triplets. Reciprocity in a graph is the ratio of reciprocated ties and is similar in
both years. As expressed by the high percentage of repeated ties, most countries seem to
5
Figure 2: The International Arms Trade as a binary network in 2016 (left) and 2017 (right).
Nodes that are isolated in both years are not depicted for clarity and the node size relates
to the sum of the out- and in-degree. The labels of the nodes are the ISO3 codes of the
respective countries.
continue trading with the same partners.
Additionally, different kinds of exogenous covariates may be controlled for in statistical
network models. In the given example we use the logarithmic Gross Domestic Product (GDP)
(World Bank, 2017) as monadic covariates concerning the sender and receiver of weapons.
We also include the absolute difference of the so called polity IV index (Center for systemic
Peace, 2017), ranging from zero (no ideological distance) to 20 (highest ideological distance),
as a dyadic exemplary covariate. These covariates are assumed to be non-stochastic and we
denote them by xt. See the Supplementary Material for a list of all included countries and
their ISO3 code.
6
BRAAFGCANINDUSACHNDZAFRADEUITANLDRUSZAFSWEAREGBRBLRAGOISRAUTARGARMAUSAZETURBHSBELBHRBGDYUGUKRBTNBOLCZEBWAESPCHEBFABDICMRSVKNORCHLUZBCOLZARCIVHRVDNKDOMECUEGYFINSLVESTETHGABGHAGRCHUNKGZIDNPOLIRNIRQBGRKORSVNIRLJAMJPNJORKAZKENKWTLVALBNLTULUXMKDMWIBRNMYSTHAMLIMLTMRTMUSMEXYUGMARMOZMYMNAMNPLNICNGAPAKOMNSGPPANPERNZLPHLPRTQATROMSAUGEOSENSLESDNTJKTZATTOTUNTKMUGAVNMZMBALBTCDCYPFJIGINHNDLAOLSONERPNGRWALKABRAAFGCANINDUSACHNDZAFRADEUITANLDRUSZAFSWEAREGBRBLRAGOISRAUTARGARMAUSAZETURBHSBELBHRBGDYUGUKRBTNBOLCZEBWAESPCHEBFABDICMRSVKNORCHLUZBCOLZARCIVHRVDNKDOMECUEGYFINSLVESTETHGABGHAGRCHUNKGZIDNPOLIRNIRQBGRKORSVNIRLJAMJPNJORKAZKENKWTLVALBNLTULUXMKDMWIBRNMYSTHAMLIMLTMRTMUSMEXYUGMARMOZMYMNAMNPLNICNGAPAKOMNSGPPANPERNZLPHLPRTQATROMSAUGEOSENSLESDNTJKTZATTOTUNTKMUGAVNMZMBALBTCDCYPFJIGINHNDLAOLSONERPNGRWALKA 2.3 Data Set 2: European Research Institution Email Correspon-
dence
The second network under study represents anonymized email exchange data between insti-
tution members of a department in a European research institution (see Paranjape et al.,
2017, Email EU Core, 2019). In this data set, we observe events ω = (i, j, t) that represent
emails sent from department member i to department member j at a specific time point t.
The data contains n = 89 persons and is recorded over 802 days. For this paper, we
select the first two years and split them again into two years, labeled Period 1 and Period
2. Within the first period, 8 068 events are recorded and 4 031 in the second period. We
only regard one-to-one email correspondences, therefore we exclude all group mails from the
analysis. In the right column of Table 1 the descriptive measures for the two aggregated
networks are given and in Figure 3 they are visualized. All descriptive statistics are higher
in the email exchange network as compared to the arms trade network. In comparison to
the arms trade network, the aggregated network is more dense with more than 10% of all
possible ties being realized. In both years the transitivity measure is relatively higher in
both time periods. The high share of reciprocated ties is intuitive given that the network
represents email exchange between institution members that may collaborate. No covariates
are available for this network. See Annex A for the visualization of the degree distributions
of both applications.
3 Dynamic Exponential Random Graph Models
3.1 Temporal Exponential Random Graph Model
The Exponential Random Graph Model (ERGM) is among the most popular models for the
analysis of static network data. Holland and Leinhardt (1981) introduced the model class,
which was subsequently extended with respect to fitting algorithms and network statistics
(see Lusher et al., 2012, Robins et al., 2007). Spurred by the popularity of ERGMs, dynamic
extensions of this model class emerged, pioneered by Robins and Pattison (2001) who de-
veloped time-discrete models for temporally evolving social networks. Before we start with
a description of the model, we want to highlight that the TERGM as well as the STERGM
are most appropriate for equidistant time points. That is, we observe the networks Yt at
discrete and equidistant time points t = 1, ..., T . Only in this setting, the parameters allow
for a meaningful interpretation. See Block et al. (2018) for a deeper discussion.
Hanneke et al. (2010) is the main reference for the TERGM, a model class that utilizes
the Markov structure and, thereby, assumes that the transition of a network from time
point t − 1 to time point t can be explained by exogenous covariates as well as structural
7
Figure 3: The European research institution email correspondence aggregated to a binary
network divided into Period 1 (day 1-365, left) and Period 2 (day 1-365, right). The node
size relates to the sum of the out- and in-degree.
components of the present and preceding networks. Using a first-order Markov dependence
structure and conditioning on the first network, the resulting dependence structure of the
model can be factorized into
Pθ(YT , ..., Y2Y1, x1, ..., xT ) = Pθ(YTYT−1, xT )··· Pθ(Y3Y2, x3)Pθ(Y2Y1, x2).
(1)
In the formulation above, it is assumed that the joint distribution can be decomposed into
yearly transitions from Yt−1 to Yt. Further, it is assumed that the same parameter vector
θ governs all transitions. Often, this is an unrealistic assumption for networks observed
at many time-points because the generative process may change other time. Therefore, it
can be useful to allow for different parameter vectors for each transition probability (i.e.
θT , θT−1, ...). In such a setting, the parameters for each transition can either be estimated
sequentially (e.g. Thurner et al., 2018) or by using smooth time-varying effects (e.g. Lebacher
et al., 2019).
Given the dependence structure (1), the TERGM assumes that the transition from Yt−1
to Yt is generated according to an exponential random graph distribution with the parameter
θ:
.
(2)
Pθ(Yt = ytYt−1 = yt−1, xt) =
exp{θT s(yt, yt−1, xt)}
κ(θ, yt−1, xt)
8
Generally, s(yt, yt−1, xt) specifies a p-dimensional function of sufficient network statistics
which may depend on the present and previous network as well as on covariates. These
network statistics can include static components, designed for cross-sectional dependence
structures (see Morris et al., 2008 for more examples). However, the statistics s(yt, yt−1, xt)
explicitly allow temporal interactions, e.g. delayed reciprocity
sdelrecip(yt, yt−1) ∝(cid:88)
i(cid:54)=j
yji,tyij,t−1.
(3)
This statistic governs the tendency whether a tie (i, j) in t − 1 will be reciprocated in t.
Another important temporal statistic is stability
(yij,tyij,t−1 + (1 − yij,t)(1 − yij,t−1)) .
(4)
sstability(yt, yt−1) ∝(cid:88)
i(cid:54)=j
In this case, the first product in the sum measures whether existing ties in t − 1 persist
in t and the second term is one if non-existent ties in t − 1 remain non-existent in t. The
proportionality sign is used since in many cases the network statistics are scaled into a specific
interval (e.g. [0, n] or [0, 1]). Such a standardization is especially sensible for networks where
the actor set changes with time. Additionally, exogenous covariates can be included, e.g.,
time-varying covariates xij,t
sdyadic(yt, xt) =
yij,txij,t.
(5)
(cid:88)
i(cid:54)=j
There exists an abundance of possibilities for defining interactions between ties in t − 1
and t. From this discussion and equation (2) it also becomes evident, that in a situation
where the interest lies in the transition between two periods, a TERGM can be modeled
simply as an ERGM, including lagged network statistics. This can be done for example by
incorporating yij,t−1 as explanatory variable in (5) which is mathematically equivalent to the
stability statistic (4). In the application we call this statistic repetition (Block et al., 2018).
Concerning the estimation of the model, maximum likelihood estimation appears to be
a natural candidate due to the simple exponential family form (2). However, the nor-
malization constant in the denominator of model (2) often poses an inhibiting obstacle
when estimating (T)ERGMs. This can be seen by inspecting the normalization constant
y∈Y exp{θT s(yt, yt−1, xt)}, that requires summation over all possible net-
works y ∈ Y. This task is virtually infeasible, except for very small networks. Therefore,
Markov Chain Monte Carlo (MCMC) methods have been proposed in order to approximate
the logarithmic likelihood function (see Geyer and Thompson (1992) for Monte Carlo max-
κ(θ, yt−1, xt) = (cid:80)
9
imum likelihood and Hummel et al. (2012) for its adaption to ERGMs). The article by
Caimo and Friel (2011) provides an alternative algorithm that uses MCMC-based inference
in a Bayesian model framework. Another approach is to employ maximum pseudolikelihood
estimation (MPLE, Strauss and Ikeda, 1990) that can be viewed as a local alternative to
the likelihood (van Duijn et al., 2009) but is often regarded as unreliable and poorly under-
stood in the literature (Hunter et al., 2008, Handcock et al., 2003). However, the MPLE is
asymptotically consistent (Desmarais and Cranmer, 2012) and the often suspect standard
errors can be corrected via bootstrap (Leifeld et al., 2018). A notable special case arises if
the network statistics are restricted such that they decompose to
s(yt, yt−1, xt) =
yij,tsij(yt−1, xt),
(6)
(cid:88)
i(cid:54)=j
with sij being a function that is evaluated only at the lagged network yt−1 and covariates xt
for tie (i, j). With this restriction, we impose that the ties in t are independent, conditional
on the network structures in t − 1. This greatly simplifies the estimation procedure and
allows to fit the model as a logistic regression model (see for example Almquist and Butts,
2014) without the issues related to the MPLE.
A problem, that is very often encountered when fitting (T)ERGMs with endogenous
network statistics is called degeneracy (Schweinberger, 2011) and occurs if most of the prob-
ability mass is attributed to network realizations that provide either full or empty networks.
One way to circumvent this problems is the inclusion of modified statistics, called geometri-
cally weighted statistics (Snijders et al., 2006). Using the definitions of Hunter (2007), the
geometrically weighted out-degree distribution (GWOD) controls for the out-degree distri-
bution with one statistic, via
sGW OD(yt) = exp{αO} n−1(cid:88)
(cid:0)1 − (1 − exp{−αO})k(cid:1)Ok(yt),
(7)
k=1
with Ok(yt) being the number of nodes with out-degree k in t and αO as the weighting
parameter. Correspondingly, the in-degree distribution is captured by the geometrically
weighted in-degree distribution (GWID) statistic by exchanging Ok(yt) with Ik(yt), which
counts the number of nodes with in-degree k, and αO with αI. While on the one hand, the
weighting often effectively counteracts the problem of degeneracy, the statistics become more
complicated to interpret. Negative values of the associated parameter typically indicate a
centralized network structure.
Regarding statistics capturing clustering, the most common geometrically weighted tri-
angular structure is called geometrically weighted edge-wise shared partners (GWESP) and
10
builds on the number of two-paths that indirectly connect two nodes i and j given the
presence of an edge (i, j):
sGW ESP (yt) = exp{αS} n−2(cid:88)
(cid:0)1 − (1 − exp{−αS})k(cid:1)Sk(yt),
(8)
k=1
where αS is a weighting parameter. The number of edges with k shared partners (Sk(yt))
is uniquely defined in undirected networks.
If the edges are directed it must be decided,
which combination should form a triangle, see Lusher et al. (2012) for a discussion. As
a default, the number of directed two-paths is chosen (Goodreau et al., 2009). Generally,
a positive coefficient for GWESP indicates that triadic closure increases the probability of
edge occurrence and globally a positive value for the associated parameter means more triadic
closure as compared to a regime with a negative value (Morris et al., 2008).
3.2 Separable Temporal Exponential Random Graph Model
An useful improvement of the TERGM (2) is the STERGM proposed by Krivitsky and
Handcock (2014). This model can be motivated by the fact that the stability term leads to
an ambiguous interpretation of its corresponding parameter. Given that we include (4) in
a TERGM and obtain a positive coefficient after fitting the model it is not clear whether
the network can be regarded as "stable" because existing ties are not dissolved (i.e. yij,t =
yij,t−1 = 1) or because no new ties are formed (i.e. yij,t = yij,t−1 = 0). To disentangle this,
the authors propose a model that allows for the separation of formation and dissolution.
Krivitsky and Handcock (2014) define the formation network as Y + = Yt ∪ Yt−1, being
the network that consists of the initial network Yt−1 together with all ties that are newly
added in t. The dissolution network is given by Y − = Yt ∩ Yt−1 and contains exclusively
ties that are present in t and t − 1. Given the network in t − 1 together with the formation
and the dissolution network we can then uniquely reconstruct the network in t, since Yt =
Y +\(Yt−1\Y −) = Y − ∪ (Y +\Yt−1). Define θ = (θ+, θ−) as the joint parameter vector that
contains the parameters of the formation and the dissolution model. Building on that,
Krivitsky and Handcock (2014) define their model to be separable in the sense that the
parameter space of θ is the product of the parameter spaces of θ+ and θ− together with
conditional independence of formation and dissolution given the network in t − 1:
Pθ(Yt = ytYt−1 = yt−1, xt) =
Pθ+(Y + = y+Yt−1 = yt−1, xt)
(cid:123)(cid:122)
(cid:124)
Formation Model
Dissolution Model
(cid:125)
(cid:124)
11
Pθ−(Y − = y−Yt−1 = yt−1, xt)
.
(9)
(cid:123)(cid:122)
(cid:125)
Figure 4: Conceptual representation, illustrating formation and dissolution in the STERGM.
The structure of the model is visualized in Figure 4. On the left hand side the state of the
network Yt−1 is given, consisting of two ties (i, h) and (h, j). In the top network all ties that
could possibly be formed are shown in dashed and the actual formation in this example (i, j)
is shown in solid. On the bottom, the two ties that could possibly be dissolved are shown
and in this example (h, j) persists while (i, j) is dissolved. On the right hand side of Figure 4
the resulting network at time point t is displayed. Given this structure and the separability
assumption (9), it is assumed that the formation model is given by
Pθ+(Y + = y+Yt−1 = yt−1, xt) =
exp{(θ+)T s(y+, yt−1, xt)}
κ(θ+, yt−1, xt)
,
(10)
with κ(θ+, yt−1, xt) being the normalization constant. Accordingly, the dissolution model
can be defined. Inserting the separable models in (9) makes it apparent, that the STERGM
is a subclass of the TERGM since
Pθ(Yt = ytYt−1 = yt−1, xt) =
exp{(θ+)T s(y+, yt−1, xt)}
exp{(θ−)T s(y−, yt−1, xt)}
κ(θ−, yt−1, xt)
κ(θ+, yt−1, xt)
exp{θT s(yt, yt−1, xt)}
=
κ(θ, yt−1, xt)
12
ihjYt−1DissolutionFormationihjihjihj=∖(∖)YtY+Yt−1Y− with θ = (θ+, θ−)T , s(yt, yt−1, xt) = (s(y+, yt−1, xt), s(y−, yt−1, xt))T and the normalization
constant set accordingly.
For practical reasons it is important to understand that the term dissolution model is
somewhat misleading since a positive coefficient in the dissolution model implies that nodes
(or dyads) with high values for this statistic are less likely to dissolve. This is also the
standard implementation in software packages, but can simply be changed by switching the
signs of the parameters in the dissolution model.
The network statistics are used similarly as in a cross-sectional ERGM. In Krivitsky and
Handcock (2014) they are called implicitly dynamic because they are evaluated either at the
formation network y+ or the dissolution network y−, which are both formed from yt−1 and
yt. For example, the number of edges is separately computed now for the formation and the
dissolution network, giving either the number of edges that newly formed or the number of
edges that persisted. For example, reciprocity in the formation network is defined as
y+
jiy+
ij
(11)
srecip(y+, yt−1) = srecip(y+) ∝(cid:88)
i(cid:54)=j
and in case of the dissolution model, y+ is simply exchanged with y−. Similarly, and edge
covariates or the geometrically weighted statistics shown in equations (5), (7) and (8) are
now functions of y+ or y− not yt.
3.3 Model Assessment
In analogy to binary regressions models, the (S)TERGM can be evaluated in terms of their
receiver-operator-
characteristic (ROC) curve or precision-recall (PR), where the latter puts more emphasis
on finding true positives (e.g. Grau et al., 2015). A comparison between different models is
possible using, for example, the Akaike Information Criterion (AIC, Claeskens and Hjort,
2008). Here we want to highlight, that the AIC fundamentally builds on the log likelihood,
which in most realistic applications is only available as an approximation, see Hunter et al.
(2008) for further discussion.
However, in statistical network analysis it is often argued that suitable network models
should not exclusively provide good predictions for individual edges, but also be able to repre-
sent topologies of the observed network. The dominant approach to asses the goodness-of-fit
of (S)TERGMs is based on sampling networks from their distribution under the estimated
parameters and then comparing network characteristics of these sampled networks with the
same ones from the observed network (Hunter et al., 2008). For this approach, it is recom-
mendable to utilize network characteristics that are not used for specifying the model. For
instance, models that include the GWOD statistic (7) may not be compared to the simulated
13
values of it but against the out-degree distribution.
Hanneke et al. (2010) point out that for networks with more than one transition from
t−1 to t available it is possible to employ a "cross-validation-type" assessment of the fit. The
parameters can be fit repetitively to all observed transitions except one hold-out transition.
It is then checked, how well the network statistics from the hold-out transition period are
represented by the ones sampled from the coefficients obtained from all other transitions.
4 Relational Event Model
4.1 Time-Continuous Event Processes
The second type of dynamic network models results by comprehending network changes
as a continuously evolving process (see Girardin and Limnios, 2018 as a basic reference
for stochastic processes). The idea was originally introduced by Holland and Leinhardt
(1977). According to their view, changes in the network are not occurring at discrete time
points but as a continuously evolving process, where only one tie can be toggled at a time.
This framework was extended by Butts (2008) to model behavior, which is understood as a
directed event at a specific time, that potentially depends on the past. Correspondingly, the
observations in this section are behaviors which are given as triplets ω = (i, j, t) and encode
the sender i, receiver j, and exact time point t. This fine-grained temporal information
is often called time-stamped or time-continuous, we adopt the latter name. Furthermore,
we only regard dyadic events in this article, i.e., a behavior only includes one sender and
receiver.
The concept of behavior, hereinafter called event, generalizes the classical concept of
binary relationships based on graph theory as promoted by Wasserman and Faust (1994).
This event framework does not intrinsically assume that ties are enduring over a specific time
frame (Butts, 2009; Butts and Marcum, 2017). For example in an email exchange network,
sending one email at a specific time point is merely a brief event, which does not convey the
same information as a durable relationship. Therefore, the time-stamped information cannot
adequately be represented in a binary adjacency matrix without having to aggregate the
relational data at the cost of information loss (Stadtfeld, 2012). Nevertheless, a friendship
between actor i and j at a given time point can still be viewed as an event that has an
one-to-one analogy to a tie in the classical framework.
The overall aim of Relational Event Models (REM, Butts, 2008) is to understand the
dynamic structure of events conditional on the history of events (Lerner et al., 2013). This
dynamic structure, in turn, controls how past interactions shape the propensity of future
events. To make this model feasible, we leverage results from the field of time-to-event
analysis, or survival analysis respectively (see, e.g., Kalbfleisch and Prentice, 2002 for an
14
overview). The central concept of the REM can be motivated by the introduction of a
multivariate time-continuous Poisson counting process
N (t) = (Nij(t) i, j ∈ {1, . . . , n}),
(12)
where Nij(t) counts how often actors i and j interacted in [0, t). Note that we indicate
continuous time t with a tilde to distinguish from the discrete time setting with t = 1, 2, . . . , T
assumed in the previous section. Process (12) is characterized by an intensity function λij(t)
for i (cid:54)= j, which is defined as:
λij(t) = lim
dt↓0
P(Nij(t + dt) = Nij(t) + 1)
dt
.
This is the instantaneous probability of observing a jump of size "1" in Nij(t), which indicates
observing the event (i, j, t). Since we assume that there are no self-loops λii(t) ≡ 0 ∀ i =
1, . . . , n holds.
4.2 Time-Continuous Observations
Butts (2008) introduced the REM to analyze the intensity λij(t) of process (12) when time-
continuous data on the events are available. He assumed that the intensity is constant
over time but depends on time-varying relational information of past events and exogenous
covariates. Vu et al. (2011) extended the model by postulating a semi-parametric intensity
similar to Cox (1972):
λij(t N (t), x(t), θ) = λ0(t)exp(cid:8)θT sij
(cid:0)N (t), x(t)(cid:1)(cid:9),
(13)
where λ0(t) is an arbitrary baseline intensity, θ ∈ Rp the parameter vector and sij
a statistic that depends on the (possibly time-continuous) covariate process x(t) and the
counting process just prior to t.
(cid:0)N (t), x(t)(cid:1)
(cid:0)N (t), x(t)(cid:1). Solely the differing level of the model needs to be accounted for, since model
(cid:0)N (t), x(t)(cid:1) from the position
sij
(13) takes a local time-continuous point of view to understand the relational nature of the
observed events. This necessitates defining the statistics sij
of specific ties, in contrast to the globally defined statistics s(yt, yt−1, xt) in (2). To give an
example, the tie-level version of reciprocity for the event (i, j) is defined as
Generally, similar statistics as already introduced in Section 3 can be included in
(cid:0)N (t), x(t)(cid:1) = I(cid:0)Nji(t) > 0(cid:1),
sij,reciprocity
15
where I(·) is the indicator function. It only regards, whether already having observed the
event (j, i) prior to t has an effect on λij(t N (t), x(t), θ), in comparison to the network level
version (3) of delayed reciprocity that counted all reciprocated ties between the networks yt
and yt−1.
Degree statistics can be specified as either sender- or receiver-specific. If we, e.g., want
to control for the out-degree of the sender the corresponding tie-oriented statistic is:
(cid:0)N (t), x(t)(cid:1) =
n(cid:88)
I(cid:0)Nih(t) > 0(cid:1).
sij,SOD
The in-degree of the receiver can be formulated accordingly.
h=1
Clustering in event sequences may be captured by different types of nested two-path
configurations. For instance, the tie-oriented version of directed two-paths, henceforth called
transitivity, is given by:
(cid:0)N (t), x(t)(cid:1) =
n(cid:88)
I(cid:0)Nih(t) > 0(cid:1)I(cid:0)Nhj(t) > 0(cid:1).
sij,T RA
h=1
(cid:0)N (t), x(t)(cid:1) equal to the covariate values of interest. Since the effect of a past event at
The inclusion of monadic and dyadic exogenous covariates becomes straightforward by setting
sij,dyadic
time δ, say, on a present event at time t may vary according to the elapsed time t−δ, Stadtfeld
and Block (2017) introduced windowed effects, which only regard events that occurred in a
pre-specified time window, e.g. a year. We will come back to this point in the next section.
If time-continuous observations are available each dimension of the observed counting
process is conditional on the past independent. This, in turn, enables the construction of a
likelihood, which can subsequently be maximized. Assuming that Ω is the set of all observed
events and T the interval of observation, the likelihood can be written as:
L(θ) =
λij(t N (t), x(t), θ)exp
λkh(u N (u), x(u), θ)du
(14)
(cid:111)
.
(cid:89)
(i,j,t)∈Ω
(cid:90)
(cid:110) −
n(cid:88)
T
k,h=1
This likelihood is straightforward to maximize in the case of a parametric baseline intensity
λ0(t), for example Butts (2008) assumes λ0(t) = γ0. Alternatively, Butts (2008) analyzed
events with ordinal temporal information. In this setting, the likelihood is equal to the partial
likelihood introduced by Cox (1972) for estimating parameters of semi-parametric intensities
as in (13). Letting Ut denote the set of all possible events that could have occurred at time
16
point t but did not, the partial likelihood for continuous event data is defined as:
(cid:89)
(cid:80)
PLcont(θ) =
λij(t N (t), x(t), θ)
λkh(t N (t), x(t), θ)
(i,j,t)∈Ω
(k,h)∈Ut
.
(15)
Consecutively, Λ0(t) = (cid:82) t
0 λ0(u)du can be estimated with a Nelson Aalen estimator (see
Kalbfleisch and Prentice, 2002 for further details on the estimation).
When dealing with large amounts of event data the main obstacle is evaluating the sum
over the intensities of all possible ties in (15) (Butts, 2008). One exact option is to trade a
longer running time for a slimmer memory footprint by means of a coaching data structure.
Vu et al. (2011) exploit this by saving prior values of the sum and subsequently changing
it event-wise by elements of Ut whose covariates changed. Alternatively, Vu et al. (2015)
proposes approximate routines that utilize case-control sampling and stratification for the
Cox model (Langholz and Borgan, 1995). More precisely, the sum is only calculated over a
sampled subset of possible events in addition to stratification. Lerner and Lomi (2019) go
one step further and sample events out of Ω for the calculation of PL(θ) in (15).
Extensions of this model building on already well-established methods in social network
and time-to-event analysis were numerously proposed. Perry and Wolfe (2013) used a strat-
ified Cox model in (13). Stadtfeld et al. (2017) adopted the Stochastic Actor oriented Model
(SAOM) to events. DuBois and Smyth (2010) and DuBois et al. (2013) extended the Stochas-
tic Block Model (SBM) for time-stamped relational events. Further, DuBois et al. (2013)
adopted a Bayesian hierarchical model to event data when information is only available in
smaller groups.
4.3 Time-Clustered Observations
Generally, the approach discussed above requires time-continuous network data, meaning
that we observe the precise time points of all events. To give an instance, in the first data
example, this means that we need the exact time point t of an arms trade between country i
and j. Often, such exact time-stamped data are not available and, in fact, trading between
states can hardly be stamped with a single time point t. Indeed, we often only observe the
time-continuous network process at discrete time points t = 1, . . . , T . In such setting, we may
assume a Markov structure in that we do not look at the entire history of the process N (t)
but just condition the intensity (13) on the history of events from the previous observation
t − 1 to t. Technically this means that N (t) is adapted to Y (t) := N (t) − N (t − 1) and x(t)
for t ∈ [t − 1, t]. We then reframe (13) as:
λij(t Y (t), x(t), θ) = λ0(t)exp(cid:8)θT sij
(cid:0) Y (t), x(t)(cid:1)(cid:9).
(16)
17
In other words, we assume that the intensity of events between t−1 and t does not depend on
states of the multivariate counting process (12) prior to t− 1. For this reason, all endogenous
statistics introduced in Section 4.2 are now evaluated on Y (t) instead of N (t). This is a
reasonable assumption, if one is primarily interested in short-term dependencies between
the individual counting processes. It enables a meaningful comparison to the models from
Section 3 that assume an analog discrete Markov property. However, we want to emphasize
that this dependence structure is not vital to inferential results.
If we observe the continuous process at discrete time points it is inevitable that we observe
time clustered observations, meaning that two or more events happen at the same time point.
Under the term tied observations this phenomenon is well known in time-to-event analysis
and treated with several approximations. One option is the so-called Breslow approximation
(see Peto, 1972; Breslow, 1974). Let therefore
Ot = {(i, j) Nij(t) − Nij(t − 1) > 0}
where element (i, j) is replicated Nij(t) − Nij(t − 1) times in Ot, that is if an event between
i and j occurred multiple times in the interval from t − 1 to t then (i, j) appears respective
times in Ot. Given that we have not observed the exact time point of an event we also get
no information on the baseline intensity λ0(t) in (13) for t ∈ [t − 1, t] so that the model
simplifies to a discrete choice model structure (see, e.g., Train, 2009) which resembles the
partial likelihood (15) and is defined as:
(cid:0) Y (t), x(t)(cid:1)}
(cid:0) Y (t), x(t)(cid:1)}(cid:1)nt ,
(17)
(cid:81)
(cid:0)(cid:80)
T(cid:89)
t=1
(i,j)∈Ot
(k,h)∈Ut
exp{θT sij
exp{θT skh
PLclust(θ) =
where nt = Ot. Alternatively, one can replace the denominator in (17) by considering all
possible orders of the unobserved events in Ot giving the average likelihood as introduced
by Kalbfleisch and Prentice (2002). Since this can be a combinatorial and hence numerical
challenge, random sampling of time point orders among the time-clustered observations can
be used with subsequent averaging, which we call Kalbfleisch-Prentice approximation (see
Kalbfleisch and Prentice, 2002). Further techniques to deal with unknown time ordering
are augmenting the clustered events into possible paths of ordered events and adapting the
maximum likelihood estimation proposed for the SAOM by Snijders et al. (2010) or using
random sampling of the ordering. This can be legitimized in cases where we may assume
independence among events happening in one year, since the events take a long time to
materialize (Snijders, 2017).
18
4.4 Model Assessment
In comparison to the assessment for models operating in discrete time, widely accepted
methods dealing with relational event data are scarce. The proposals either stem from
time-to-event analysis or regard link prediction, which is the task of predicting the most
likely next event given the history of past events (Liben-Nowell and Kleinberg, 2007). One
example of the former option is the usage of Schnfeld residuals by Vu et al. (2017) to check
the assumption of proportional intensities, which is central to semi-parametric models as
the one proposed by Cox (1972). For the latter approach, we need to define a predictive
measure that quantifies how well the next event is predicted. Vu et al. (2011) proposed the
recall measure that estimates the percentage of test events which are in the list of K most
likely next events according to a given model. Evaluating this percentage for different values
of K permits a visualization of the predictive capabilities of the model. The strength of
the predicted intensity allows the ordering of events according to the probability of being
observed next. If we model the propensity of time-clustered events that represent binary
adjacency matrices one can alternatively adopt the analysis of the ROC and PR curve
introduced in Section 3.3.
5 Application
When it comes to software, there exist essentially three main R packages that are designed
for fitting TERGMs and STERGMs. Most important is the extensive statnet library
(Goodreau et al., 2008) that allows for simulation-based fitting of ERGMs. The library
contains the package tergm with implemented methods for fitting STERGMs using MCMC
approximations of the likelihood. However, currently the package tergm (version 3.5.2) does
not allow for fitting STERGMs with time-varying dyadic covariates for more than two time
periods jointly. The package btergm (Leifeld et al., 2018) is designed for fitting TERGMs us-
ing either maximum pseudo-likelihood or MCMC maximum likelihood estimation routines.
In order to obtain Bayesian Inference in ERGMs, the package bergm by Caimo and Friel
(2014) can be used. Besides implementations in R, the stand-alone program PNet (Wang
et al., 2006) allows for simulating, fitting and evaluating (T)ERGMs.
In order to ensure
comparable estimates we estimate the TERGM, as well as the STERGM, with the statnet
library, using MCMC-based likelihood estimation techniques. We use the package ergm and
include delayed reciprocity and the repetition of previous ties as dyadic covariates. The
STERGM is fitted using the tergm package.
Marcum and Butts (2015) implemented the R package relevent (version 1.0-4) to es-
timate the REM for time-stamped data. It was followed by the package goldfish (version
1.2) by Stadtfeld and Hollway (2018) for modeling event data with precise and ordinal tem-
19
poral information with an actor- and tie-oriented variant of the REM. Furthermore, it is
highly customizable in terms of endogenous and exogenous user terms and will be used in
the following applications.
We want to remark that the STERGM coefficients are implicitly dynamic, while in the
TERGM all network statistics except the lagged network and delayed reciprocity terms are
evaluated on the network in t. All covariates of the REM are continuously updated and the
intensity at time point t ∈ [t − 1, t] only depends on events observed in [t − 1, t). Like the
building period proposed by Vu et al. (2011), the events in t− 1 are only used for building up
the covariates and not directly modeled. Due to no compositional changes, we did not scale
any statistics. Moreover, we refer to the Supplementary Material for the model assessment.
5.1 Data Set 1: International Arms Trade
The results obtained for the arms trading data section are displayed in Table 2. For a detailed
interpretation of effects focusing on political, social, and economic aspects we refer to the
relevant literature (e.g. Thurner et al., 2018). Here we want to comment on a few aspects
only. While we do not have timestamps for the arms trades, the longitudinal networks can
still be viewed as time-clustered observations enabling the techniques from Section 4.3.
Both, the TERGM (column 1) and the REM (column 4) identify the repetition of previous
ties as a driving force in the dynamic structure of the network. Degree-related covariates,
which are GWID and GWOD in the (S)TERGM and in- and out-degree in the REM, capture
centrality in the network. The coefficients of the GWID and GWOD are negative and have
low p-values in the TERGM. This stands in contrast to the STERGM, where these effects
are only pronounced in the formation model (column 2), while they are insignificant effect
in the dissolution model (column 3). Hence, these effects suggest a centralized pattern in
the formation network, which is also captured by the TERGM. In the REM an analogous
pattern can be detected, since a higher in-degree of the receiver increases the respective
intensity, thus spurs trade relations. Similar interpretations hold for the out-degree of the
sender. Overall, countries that have a high out-degree are more likely to send weapons
and countries with a high in-degree to receive weapons, which again results in a centralized
network structure as indicated by the estimates in the TERGM and STERGM.
Lastly, consistent effects among the models were also found for the exogenous covariates.
Consider, for instance, the coefficient of the logarithmic GDP of the importing country. The
TERGM assigns a significantly higher probability to observe in-going ties to countries with
a high GDP just like the REM. However, disentangling the model towards formation and
dissolution we see strongly significant coefficients in the dissolution model while the effect
for the formation model is weakly significant.
Based on the independence assumption in (9) we can sum up the two AIC values and
20
TERGM
STERGM
REM
Formation Dissolution
−
−
−
−
Repetition
Edges
Reciprocity
3.671∗∗∗
(0.132)
−15.632∗∗∗ −17.186∗∗∗ −16.987∗∗∗
(3.587)
(1.809)
−0.058
−0.258
(0.619)
(0.306)
−1.823∗∗∗
−0.412
In-Degree (GWID)
(0.442)
(0.278)
−0.326
Out-Degree (GWOD) −3.220∗∗∗
(0.533)
(0.304)
0.150
0.050
(0.066)
(0.126)
−0.016
−0.024∗
(0.017)
(0.010)
0.323∗∗∗
0.313∗∗∗
(0.088)
(0.048)
0.327∗∗∗
0.165∗∗∗
(0.087)
(0.043)
−949.833
−258.425
1917.666
532.849
1917.666
(2.168)
−0.620
(0.436)
−2.106∗∗∗
(0.379)
−4.126∗∗∗
(0.462)
0.076
(0.071)
−0.028∗
(0.014)
0.394∗∗∗
(0.054)
0.135∗
(0.054)
−675.327
1366.654
Log Likelihood
AIC
(cid:80) AIC
log(GDP) Receiver
log(GDP) Sender
GWESP
Polity Score
1899.503
2.661∗∗∗
(0.143)
−
−
−0.109
(0.181)
0.060∗∗
In-Degree Receiver
(0.015)
0.010∗∗ Out-Degree Sender
(0.004)
0.010
(0.029)
−0.016
(0.009)
0.395∗∗∗
(0.039)
0.192∗∗∗
(0.032)
Transitivity
Table 2: Arms trade network: Comparison of parameters obtained from the TERGM (first
column), STERGM (Formation in the second column, Dissolution in the third column) and
REM (fourth column). Standard errors in brackets and stars according to p-values smaller
than 0.001 (∗∗∗), 0.05 (∗∗) and 0.1 (∗). Decay parameter of the geometrically weighted
statistics is set to log(2) and the Kalbfleisch-Prentice approximation was used with 100
random orderings of the events to find the estimates of the REM.
see that the AIC value of the STERGM is smaller than of the TERGM.
5.2 Data Set 2: European Research Institution Email Correspon-
dence
As already indicated by the descriptive statistics in Table 1, the email network seems to be
driven by three major structural influences: repetition, reciprocity, and transitive clustering.
21
TERGM
STERGM
REM
Formation Dissolution
−
−
2.27∗∗∗
(0.084)
−
−
1.655∗∗∗
(0.075)
−0.004
In-Degree Receiver
(0.003)
−0.0001 Out-Degree Sender
(0.003)
0.070∗∗∗
(0.008)
Transitivity
Repetition
Edges
Reciprocity
In-degree (GWID)
Out-degree (GWOD)
GWESP
Log Likelihood
AIC
(cid:80) AIC
−
−
1.367∗∗∗
(0.107)
−2.237∗∗∗
−5.755∗∗∗ −4.853∗∗∗
(0.224)
(0.247)
(0.237)
2.586∗∗∗
2.498∗∗∗
0.398∗∗∗
(0.226)
(0.157)
(0.112)
1.060∗∗
1.349∗
0.709
(0.415)
(0.648)
(0.333)
−0.369
−0.411
0.031
(0.397)
(0.431)
(0.312)
0.429∗∗∗
0.655∗∗∗
1.560∗∗∗
(0.086)
(0.111)
(0.110)
−1723.732 −1000.506 −505.431
3459.464
1020.862
3459.464
3031.874
2011.012
Table 3: Email exchange network: Comparison of parameters obtained from the TERGM
(first column), STERGM (Formation in the second column, Dissolution in the third column)
and REM (fourth column). Standard errors in brackets and stars according to p-values
smaller than 0.001 (∗∗∗), 0.05 (∗∗) and 0.1 (∗). Decay parameter of the geometrically weighted
statistics is set to log(2).
The estimates from Table 3 demonstrate, that all models were able to identify these forces.
According to the REM (column 4), the event network of email traffic in the research
institution is not centralized and primarily based on collaboration between coworkers. We
can draw those conclusions from insignificant estimates of degree-related statistics and highly
significant estimates regarding reciprocity and repetition. In the TERGM (column 1) we
find a positive and significant effect of GWID, while no effect can be found in the STERGM
(columns 2 and 3). The estimates of repetition and reciprocity in the REM and TERGM are
very pronounced. For instance, the estimates of the REM imply that a reciprocated event is
19.6 times more likely than an event with the same covariates only not being reciprocated.
Interestingly, the STERGM detects a lower effect of GWESP in the formation and dissolution
than the TERGM. The effect of the delayed reciprocity in the TERGM is less relevant than
reciprocity in the formation and dissolution model. This strongly differing effect size results
from the mathematical formulation of the statistics given in equations (3) and (11).
Contrasting the AIC values of the TERGM and STERGM shows that the dynamic struc-
22
ture of the email network is again better explained by the STERGM. In the Supplementary
Material we fit the TERGM and STERGM to multiple time points.
6 Conclusion
6.1 Further Models
Snijders (1996) formulated a two-stage process model operating in a continuous-time frame-
work. The dynamics are considered to evolve according to unobserved micro-steps. At
first, a sender out of all eligible actors gets the opportunity to change the state of all his
outgoing ties. Consecutively, the actor needs to evaluate the probability of changing the
present configuration with each possible receiver, which entails each actors knowledge of the
complete graph whenever he has the possibility to toggle one of his ties. Lastly, the deci-
sion is randomly drawn relative to the probabilities of all possible actions. In general, the
SAOM is a well-established model for the analysis of social networks, that was successfully
applied to a wide array of network data, e.g., in Sociology (Agneessens and Wittek, 2012;
de Nooy, 2002), Political Science (Kinne, 2016; Bichler and Franquez, 2014), Economics
(Castro et al., 2014), and Psychology (Jason et al., 2014). Estimation of this model variant
is predominantly carried out with the R package RSiena (Ripley et al., 2013).
Another notable model that can be regarded as a bridge between the ERGM and continuous-
time models is the Longitudinal ERGM (LERGM, Snijders and Koskinen, 2013; Koskinen
et al., 2015). In contrast to the TERGM, the LERGM assumes that the network evolves in
micro-steps as a continuous time Markov process with an ERGM being its limiting distri-
bution. Similar to the SAOM, the model builds on randomly assigning the opportunity to
change, followed by a function that governs the probability of a tie change. This model is
still tie-oriented, meaning that dyadic ties instead of actors are chosen and then have the
option to change the current network.
6.2 Summary
In this article, we put emphasis on tie-oriented dynamic network models. Comparisons
between these models can be drawn on the level at which each implied generating mechanism
works and how time is perceived. The overall aim in the TERGM is to find an adequate
distribution of the adjacency matrix Yt conditioning on information of previous realizations
of the network. In the separable extension, the aim remains unchanged, only splitting Yt
into two smaller sub-networks that include all possible ties that were and were not present in
Yt−1 separately. While the (S)TERGM proceeds in discrete time, the REM tackles modeling
the intensity on the tie level in continuous time conditional on past events. Therefore,
23
the TERGM and STERGM take a global and REM a local point-of-view, which results in
substantially different interpretations of the estimates.
Furthermore, we analyzed two data sets that represent two types of network data that
are traditionally either modeled by the TERGM and REM. By extending the REM to time-
clustered observations and aggregating events to binary adjacency matrices a meaningful
comparison between the STERGM, TERGM, and REM is enabled.
Acknowledgement
We thank the anonymous reviewers for their careful reading and constructive comments.
The project was supported by the European Cooperation in Science and Technology [COST
Action CA15109 (COSTNET)]. We also gratefully acknowledge funding provided by the
German Research Foundation (DFG) for the project KA 1188/10-1: International Trade of
Arms: A Network Approach. Furthermore we like to thank the Munich Center for Machine
Learning (MCML) for funding.
References
Agneessens, F. and R. Wittek (2012). Where do intra-organizational advice relations come
the role of informal status and social capital in social exchange. Social Net-
from?
works 34 (3), 333 -- 345.
Almquist, Z. W. and C. T. Butts (2014). Logistic Network Regression for Scalable Analysis
of Networks with joint Edge/Vertex Dynamics. Sociological methodology 44 (1), 273 -- 321.
Bastian, M., S. Heymann, and M. Jacomy (2009). Gephi: An open source software for
exploring and manipulating networks.
Bearman, P., J. Moody, and K. Stovel (2004). Chains of Affection: The Structure of Ado-
lescent Romantic and Sexual Networks. American Journal of Sociology 110 (1), 44 -- 91.
Benton, R. A. and J. You (2017). Endogenous dynamics in contentious fields: Evidence from
the shareholder activism network, 2006 -- 2013. Socius 3, 2378023117705231.
Bichler, G. and J. Franquez (2014). Conflict Cessation and the Emergence of Weapons
Supermarkets, pp. 189 -- 215. Cham: Springer International Publishing.
Blank, M., M. Dincecco, and Y. M. Zhukov (2017). Political regime type and warfare:
evidence from 600 years of european history. Available at SSRN 2830066 .
24
Block, P., J. Koskinen, J. Hollway, C. Steglich, and C. Stadtfeld (2018). Change we can
believe in: Comparing longitudinal network models on consistency, interpretability and
predictive power. Social Networks 52, 180 -- 191.
Block, P., C. Stadtfeld, and T. Snijders (2019). Forms of Dependence: Comparing SAOMs
and ERGMs From Basic Principles. Sociological Methods & Research 48 (1).
Breslow, N. (1974). Covariance Analysis of Censored Survival Data. Biometrics 30 (1),
89 -- 99.
Broekel, T. and M. Bednarz (2019, Jan). Disentangling link formation and dissolution in
spatial networks: An application of a two-mode stergm to a project-based r&d network in
the german biotechnology industry. Networks and Spatial Economics.
Butts, C. (2008). A Relational Event Framework for Social Action. Sociological Methodol-
ogy 38 (1), 155 -- 200.
Butts, C. (2009). Revisiting the foundations of network analysis. Science 325 (5939), 414 --
416.
Butts, C. and C. Marcum (2017). A Relational Event Approach to Modeling Behavioral
Dynamics. ArXiv e-prints.
Caimo, A. and N. Friel (2011). Bayesian inference for exponential random graph models.
Social Networks 33 (1), 41 -- 55.
Caimo, A. and N. Friel (2014). Bergm: Bayesian exponential random graphs in r. Journal
of Statistical Software 61 (2), 1 -- 25.
Castro, I., C. Casanueva, and J. L. Galn (2014). Dynamic evolution of alliance portfolios.
European Management Journal 32 (3), 423 -- 433.
Center for systemic Peace (2017). Polity IV Annual Time-Series, 1800-2015, Version 3.1.
http://www.systemicpeace.org. Accessed: 2017-06-02.
Claeskens, G. and N. L. Hjort (2008). Model selection and model averaging. Cambridge:
Cambridge University Press.
Cox, D. (1972). Regression Models and Life-Tables. Journal of the Royal Statistical Society.
Series B (Methodological) 34 (2), 187 -- 220.
Csardi, G. and T. Nepusz (2006). The igraph software package for complex network research.
InterJournal, Complex Systems 1695 (5), 1 -- 9.
25
de Nooy, W. (2002). The dynamics of artistic prestige. Poetics 30 (3), 147 -- 167.
Desmarais, B. A. and S. J. Cranmer (2012). Statistical mechanics of networks: Estimation
and uncertainty. Physica A: Statistical Mechanics and its Applications 391 (4), 1865 -- 1876.
DuBois, C., C. Butts, D. McFarland, and P. Smyth (2013). Hierarchical models for relational
event sequences. Journal of Mathematical Psychology 57 (6), 297 -- 309. Social Networks.
DuBois, C., C. Butts, and P. Smyth (2013). Stochastic blockmodeling of relational event
In C. M. Carvalho and P. Ravikumar (Eds.), Proceedings of the Sixteenth
dynamics.
International Conference on Artificial Intelligence and Statistics, Volume 31 of Proceedings
of Machine Learning Research, Scottsdale, Arizona, USA, pp. 238 -- 246. PMLR.
DuBois, C. and P. Smyth (2010). Modeling Relational Events via Latent Classes. In Pro-
ceedings of the 16th ACM SIGKDD International Conference on Knowledge Discovery and
Data Mining, KDD '10, New York, NY, USA, pp. 803 -- 812. ACM.
Email EU Core (2019). Email EU core temporal, Department 3. https://snap.stanford.
edu/data/email-Eu-core-temporal-Dept3.txt.gz. Accessed: 2019-08-05.
Erdos, P. and A. R´enyi (1959). On Random Graphs i. Publicationes Mathematicae Debre-
cen 6, 290.
Geyer, C. J. and E. A. Thompson (1992). Constrained Monte Carlo maximum likelihood for
dependent data. J. R. Statist. Soc. B , 657 -- 699.
Gilbert, E. N. (1959, 12). Random graphs. Ann. Math. Statist. 30 (4), 1141 -- 1144.
Girardin, V. and N. Limnios (2018). Applied Probability (2 ed.). Heidelberg: Springer.
Goldenberg, A., A. X. Zheng, S. E. Fienberg, and E. M. Airoldi (2010). A Survey of Statistical
Network Models. Foundations and Trends in Machine Learning 2 (2), 129 -- 233.
Goodreau, S. M., M. S. Handcock, D. R. Hunter, C. T. Butts, and M. Morris (2008). A
statnet Tutorial. Journal of Statistical Software 24 (9), 1.
Goodreau, S. M., J. A. Kitts, and M. Morris (2009). Birds of a feather, or friend of a
friend? Using exponential random graph models to investigate adolescent social networks.
Demography 46 (1), 103 -- 125.
Grau, J., I. Grosse, and J. Keilwagen (2015). PRROC: computing and visualizing precision-
recall and receiver operating characteristic curves in R. Bioinformatics 31 (15), 2595 -- 2597.
26
Handcock, M. S., G. Robins, T. Snijders, J. Moody, and J. Besag (2003). Assessing degen-
eracy in statistical models of social networks. Technical report, Citeseer.
Hanneke, S., W. Fu, E. P. Xing, et al. (2010). Discrete temporal models of social networks.
Electronic Journal of Statistics 4, 585 -- 605.
He, X., Y. bo Dong, Y. ying Wu, G. rui Jiang, and Y. Zheng (2019). Factors affecting evolu-
tion of the interprovincial technology patent trade networks in china based on exponential
random graph models. Physica A: Statistical Mechanics and its Applications 514, 443 --
457.
Hoff, P. D., A. E. Raftery, and M. S. Handcock (2002). Latent Space Approaches to Social
Network Analysis. Journal of the American Statistical Association 97 (460), 1090 -- 1098.
Holland, P. and S. Leinhardt (1977). A dynamic model for social networks. The Journal of
Mathematical Sociology 5 (1), 5 -- 20.
Holland, P. W. and S. Leinhardt (1981). An exponential family of probability distributions
for directed graphs. J. Am. Statist. Ass. 76 (373), 33 -- 50.
Hummel, R. M., D. R. Hunter, and M. S. Handcock (2012). Improving simulation-based
algorithms for fitting ERGMs. Journal of Computational and Graphical Statistics 21 (4),
920 -- 939.
Hunter, D. R. (2007). Curved exponential family models for social networks. Social net-
works 29 (2), 216 -- 230.
Hunter, D. R., S. M. Goodreau, and M. S. Handcock (2008). Goodness of fit of social network
models. J. Am. Statist. Ass. 103 (481), 248 -- 258.
Hunter, D. R. and M. S. Handcock (2006). Inference in curved exponential family models
for networks. Journal of Computational and Graphical Statistics 15 (3), 565 -- 583.
Jason, L. A., J. M. Light, E. B. Stevens, and K. Beers (2014). Dynamic Social Networks in
Recovery Homes. American Journal of Community Psychology 53 (3-4), 324 -- 334.
Kalbfleisch, J. and R. Prentice (2002). The Statistical Analysis of Failure Time Data. Wiley-
Blackwell.
Kim, B., K. H. Lee, L. Xue, and X. Niu (2018). A review of dynamic network models with
latent variables. Statist. Surv. 12, 105 -- 135.
27
Kinne, B. J. (2016). Agreeing to arm: Bilateral weapons agreements and the global arms
trade. Journal of Peace Research 53 (3), 359 -- 377.
Kolaczyk, E. D. (2009). Statistical analysis of network data. Methods and Models. New York:
Springer Science & Business Media.
Koskinen, J., A. Caimo, and A. Lomi (2015). Simultaneous modeling of initial conditions and
time heterogeneity in dynamic networks: An application to Foreign Direct Investments.
Network Science 3 (1), 58 -- 77.
Krivitsky, P. N. and M. S. Handcock (2014). A separable model for dynamic networks. J.
R. Statist. Soc. B 76 (1), 29 -- 46.
Langholz, B. and O. Borgan (1995). Counter-matching: A stratified nested case-control
sampling method. Biometrika 82 (1), 69 -- 79.
Lebacher, M., P. W. Thurner, and G. Kauermann (2019, Mar). A Dynamic Separable
Network Model with Actor Heterogeneity: An Application to Global Weapons Transfers.
arXiv e-prints.
Leifeld, P., S. J. Cranmer, and B. A. Desmarais (2018). Temporal exponential random graph
models with btergm: estimation and bootstrap confidence intervals. Journal of Statistical
Software 83 (6), doi: 10.18637/jss.v083.i06.
Lerner, J., M. Bussmann, T. Snijders, and U. Brandes (2013). Modeling frequency and type
of interaction in event networks. Corvinus journal of sociology and social policy 4 (1),
3 -- 32.
Lerner, J. and A. Lomi (2019). Reliability of relational event model estimates under sam-
pling: how to fit a relational event model to 360 million dyadic events. arXiv preprint
arXiv:1905.00630 .
Leskovec, J., K. J. Lang, A. Dasgupta, and M. W. Mahoney (2009). Community Structure
in Large Networks: Natural Cluster Sizes and the Absence of Large Well-Defined Clusters.
Internet Mathematics 6 (1), 29 -- 123.
Liben-Nowell, D. and J. Kleinberg (2007). The link-prediction problem for social networks.
J. Am. Soc. Inf. Sci. Technol. 58 (7), 1019 -- 1031.
Lusher, D., J. Koskinen, and G. Robins (2012). Exponential random graph models for social
networks: Theory, methods, and applications. Cambridge: Cambridge University Press.
28
Marcum, C. and C. Butts (2015). Constructing and Modifying Sequence Statistics for
relevent Using informR in R. Journal of Statistical Software 64 (5), 1 -- 36.
Morris, M., M. S. Handcock, and D. R. Hunter (2008). Specification of exponential-family
random graph models: terms and computational aspects. Journal of Statistical Soft-
ware 24 (4), 1548.
Paranjape, A., A. R. Benson, and J. Leskovec (2017). Motifs in temporal networks.
In
Proceedings of the Tenth ACM International Conference on Web Search and Data Mining,
pp. 601 -- 610. ACM.
Perry, P. and P. Wolfe (2013). Point process modelling for directed interaction networks.
Journal of the Royal Statistical Society: Series B (Statistical Methodology) 75 (5), 821 -- 849.
Peto, R. (1972). Contribution to the discussion of the paper by Dr. Cox. Journal of the
Royal Statistical Society. Series B (Methodological) 34 (2), 187 -- 220.
Quintane, E., P. Pattison, G. Robins, and J. Mol (2013). Short- and long-term stability in
organizational networks: Temporal structures of project teams. Social Networks 35 (4),
528 -- 540.
Raabe, I. J., Z. Boda, and C. Stadtfeld (2019). The Social Pipeline: How Friend Influence
and Peer Exposure Widen the STEM Gender Gap. Sociology of Education 92 (2), 105 -- 123.
Ripley, R., K. Boitmanis, and T. Snijders (2013). RSiena: Siena - Simulation Investigation
for Empirical Network Analysis. https://CRAN.R-project.org/package=RSiena. R
package version 1.2-12.
Robins, G. and P. Pattison (2001). Random graph models for temporal processes in social
networks. Journal of Mathematical Sociology 25 (1), 5 -- 41.
Robins, G., P. Pattison, Y. Kalish, and D. Lusher (2007). An introduction to exponential
random graph (p*) models for social networks. Social Networks 29 (2), 173 -- 191.
Salath´e, M., D. Q. Vu, S. Khandelwal, and D. Hunter (2013). The dynamics of health
behavior sentiments on a large online social network. EPJ Data Science 2 (1), 4.
Sarkar, P. and A. W. Moore (2006). Dynamic Social Network Analysis using Latent Space
Models. In Y. Weiss, B. Scholkopf, and J. C. Platt (Eds.), Advances in Neural Information
Processing Systems 18, pp. 1145 -- 1152. MIT Press.
Schweinberger, M. (2011).
Instability, sensitivity, and degeneracy of discrete exponential
families. Journal of the American Statistical Association 106 (496), 1361 -- 1370.
29
SIPRI
(2019).
Arms Transfers Database.
https://www.sipri.org/databases/
armstransfers. Accessed: 2019-03-01.
Snijders, T. (1996). Stochastic actororiented models for network change. The Journal of
Mathematical Sociology 21 (1-2), 149 -- 172.
Snijders, T. (2002). The statistical evaluation of social network dynamics. Sociological
Methodology 31 (1), 361 -- 395.
Snijders, T. (2005). Models for Longitudinal Network Data, pp. 215247. Structural Analysis
in the Social Sciences. Cambridge University Press.
Snijders, T. (2017). Comment: Modeling of coordination, rate functions, and missing order-
ing information. Sociological Methodology 47 (1), 41 -- 47.
Snijders, T. and J. Koskinen (2013). Longitudinal Models, pp. 130140. Structural Analysis
in the Social Sciences. Cambridge University Press.
Snijders, T., J. Koskinen, and M. Schweinberger (2010). Maximum likelihood estimation for
social network dynamics. The Annals of Applied Statistics 4 (2), 567.
Snijders, T., P. E. Pattison, G. L. Robins, and M. S. Handcock (2006). New specifications
for exponential random graph models. Sociological Methodology 36 (1), 99 -- 153.
Stadtfeld, C. (2012). Events in social networks : a stochastic actor-oriented framework for
dynamic event processes in social networks. Ph. D. thesis.
Stadtfeld, C. and P. Block (2017). Interactions, Actors, and Time: Dynamic Network Actor
Models for Relational Events. Sociological Science 4 (14), 318 -- 352.
Stadtfeld, C. and J. Hollway (2018). goldfish: Goldfish -- Statistical network models for
dynamic network data. R package version 1.2.
Stadtfeld, C., J. Hollway, and P. Block (2017). Dynamic Network Actor Models: Investigat-
ing Coordination Ties through Time. Sociological Methodology 47 (1), 1 -- 40.
Stansfield, S. E., J. E. Mittler, G. S. Gottlieb, J. T. Murphy, D. T. Hamilton, R. Detels,
S. M. Wolinsky, L. P. Jacobson, J. B. Margolick, C. R. Rinaldo, et al. (2019). Sexual role
and hiv-1 set point viral load among men who have sex with men. Epidemics 26, 68 -- 76.
Strauss, D. and M. Ikeda (1990). Pseudolikelihood estimation for social networks. Journal
of the American statistical association 85 (409), 204 -- 212.
30
Thurner, P. W., C. S. Schmid, S. J. Cranmer, and G. Kauermann (2018). Network In-
terdependencies and the Evolution of the International Arms Trade. Journal of Conflict
Resolution.
Train, K. (2009). Discrete Choice Methods with Simulation. Cambridge University Press.
Tranmer, M., C. S. Marcum, B. Morton, D. Croft, and S. de Kort (2015). Using the relational
event model (rem) to investigate the temporal dynamics of animal social networks. Animal
Behaviour 101, 99 -- 105.
van Duijn, M. A., K. J. Gile, and M. S. Handcock (2009). A framework for the comparison
of maximum pseudo-likelihood and maximum likelihood estimation of exponential family
random graph models. Social Networks 31 (1), 52 -- 62.
Vu, D., L. Alessandro, M. Daniele, and P. Francesca (2017). Relational event models for
longitudinal network data with an application to interhospital patient transfers. Statistics
in Medicine 36 (14), 2265 -- 2287.
Vu, D., A. Asuncion, D. Hunter, and P. Smyth (2011). Dynamic egocentric models for cita-
tion networks. In Proceedings of the 28th International Conference on Machine Learning,
ICML 2011, Bellevue, Washington, USA, June 28 - July 2, 2011, pp. 857 -- 864.
Vu, D., D. Hunter, P. Smyth, and A. Asuncion (2011). Continuous-Time Regression Models
for Longitudinal Networks. In J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F. Pereira,
and K. Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 24, pp.
2492 -- 2500. Curran Associates, Inc.
Vu, D., P. Pattison, and G. Robins (2015). Relational event models for social learning in
MOOCs. Social Networks 43, 121 -- 135.
Wang, P., G. Robins, and P. Pattison (2006). Pnet: A program for the simulation and
estimation of exponential random graph models. University of Melbourne.
Ward, M., J. S. Ahlquist, and A. Rozenas (2013). Gravity's Rainbow: A Dynamic Latent
Space Model for the World Trade Network. Network Science 1.
Wasserman, S. and K. Faust (1994). Social Network Analysis: Methods and Applications.
Structural Analysis in the Social Sciences. Cambridge University Press.
White, L. A., J. D. Forester, and M. E. Craft (2018). Covariation between the physiologi-
cal and behavioral components of pathogen transmission: Host heterogeneity determines
epidemic outcomes. Oikos 127 (4), 538 -- 552.
31
World Bank (2017). World Bank Open Data, Real GDP. http://data.worldbank.org/.
Accessed: 2017-04-01.
32
A Annex: Additional Descriptives
Figures 5 and 6 depict the distributions of in- and out-degrees in the two networks. Building
on in- and out-degree of all nodes, these distributions represent the relative frequency of all
possible in- and out-degrees in the observed networks, which is calculated with the igraph
package in R (Csardi and Nepusz, 2006).
In the arms trade network, a strongly asymmetric relation is revealed, indicating that
about 70% of the countries do not export any weapons, while a small percentage of countries
accounts for the major share of trade relations. The distribution of the in-degree is not that
extreme but still we have roughly one third of all countries not importing at all.
The email exchange network shows a different structure. Here, many medium-sized in-
degrees can be found and only roughly 10% of all nodes have receiver no emails. For the out-
degree, this number doubles (roughly 20% have not sent emails). Further, the distribution
of the out-degree is more skewed then the one for the in-degree.
Figure 5: Arms trade network: Barplots indicating the distribution of the in- and out-degrees.
Black bars indicate the values of year 2016 and grey bars of 2017.
I
0.00.20.40.60.8051015DegreeProportionIn−Degree0.00.20.40.60.80204060DegreeProportionOut−Degree Figure 6: Email exchange network: Barplots indicating the distribution of the in- and out-
degrees. Black bars indicate the values of period 1 and grey bars of period 2.
II
0.000.050.100.150.20010203040DegreeProportionIn−Degree0.000.050.100.150.20010203040DegreeProportionOut−Degree B Supplementary Material
B.1 Countries in the International Arms Trade Network
Country Name
Niger
Nigeria
Norway
Oman
Country Name
Germany
Ghana
Greece
Grenada
Guatemala
Guinea
Guinea-Bissau
Palau
Panama
Papua New Guinea
Paraguay
Peru
Philippines
Poland
Portugal
Qatar
Romania
Russia
Rwanda
Saint Kitts and Nevis
Saint Lucia
ISO3
ISO3
ISO3
Country Name
NER
DEU
AFG
Afghanistan
NGA
GHA
ALB
Albania
NOR
GRC
DZA
Algeria
OMN
GRD
AND
Andorra
PAK
GTM Pakistan
AGO
Angola
PLW
GIN
ATG
Antigua and Barbuda
PAN
GNB
ARG
Argentina
PNG
GUY
ARM Guyana
Armenia
PRY
HTI
Haiti
AUS
Australia
PER
HND
Honduras
AUT
Austria
PHL
HUN
Hungary
AZE
Azerbaijan
POL
ISL
Iceland
BHS
Bahamas
PRT
IND
India
BHR
Bahrain
QAT
IDN
Indonesia
BGD
Bangladesh
ROM
IRN
Iran
BRB
Barbados
RUS
IRQ
Iraq
BLR
Belarus
RWA
IRL
Ireland
BEL
Belgium
KNA
ISR
Israel
BLZ
Belize
ITA
LCA
Italy
BEN
Benin
JAM Saint Vincent and the Grenadines VCT
Jamaica
BTN
Bhutan
WSM
JPN
BOL
Japan
Bolivia
JOR
SMR
BWA Jordan
Botswana
STP
KAZ
BRA
Brazil
SAU
KEN
BRN
Brunei Darussalam
SEN
KOR
BGR
Bulgaria
YUG
KOS
BFA
Burkina Faso
SYC
KWT
BDI
Burundi
SLE
KGZ
KHM Kyrgyzstan
Cambodia
SGP
LAO
CMR
Cameroon
LVA
SVK
CAN
Canada
SVN
LBN
Cape Verde
CPV
SLB
LSO
Central African Republic CAF
ZAF
LBR
TCD
Chad
ESP
LBY
CHL
Chile
LKA
LTU
CHN
China
SDN
COL
Colombia
LUX
COM Macedonia (FYROM) MKD
Comoros
SUR
SWZ
ZAR Madagascar
DR Congo
SWE
COG Malawi
Congo
CHE
CRI
Costa Rica
TJK
CIV
Cote dIvoire
TZA
HRV Mali
Croatia
THA
CUB Malta
Cuba
Cyprus
CYP Marshall Islands
TMP
TGO
CZE
Czech Republic
TTO
DNK Mauritius
Denmark
TUN
DMA Mexico
Dominica
TUR
DOM Micronesia
Dominican Republic
TKM
ECU Moldova
Ecuador
UGA
EGY Mongolia
Egypt
El Salvador
SLV
UKR
ARE
GNQ Morocco
Equatorial Guinea
GBR
EST
Estonia
USA
ETH Myanmar
Ethiopia
URY
FJI
Fiji
Namibia
UZB
FIN
Finland
Nauru
VUT
FRA
France
Nepal
Gabon
GAB
VNM
Netherlands
ZMB
GMB New Zealand
Gambia
Georgia
GEO
ZWE
Samoa
San Marino
Sao Tome and Principe
Saudi Arabia
Senegal
Serbia
Seychelles
Sierra Leone
Singapore
Slovakia
Slovenia
Solomon Islands
South Africa
Spain
Sri Lanka
Sudan
Suriname
MDG Swaziland
MWI
MYS
MDV
MLI
MLT
MHL
MRT
MUS
MEX
FSM
MDA
MNG Uganda
YUG
Ukraine
United Arab Emirates
MAR
MOZ
United Kingdom
MYM United States
NAM Uruguay
NRU
NPL
NLD
NZL
NIC
Sweden
Switzerland
Tajikistan
Tanzania
Thailand
Timor-Leste
Togo
Trinidad and Tobago
Tunisia
Turkey
Turkmenistan
Kazakhstan
Kenya
South Korea
Kosovo
Kuwait
Laos
Latvia
Lebanon
Lesotho
Liberia
Libya
Lithuania
Luxembourg
Malaysia
Maldives
Mauritania
Montenegro
Mozambique
Nicaragua
Uzbekistan
Vanuatu
Viet Nam
Zambia
Zimbabwe
Table 4: Countries included in the analysis of the international trade network with the ISO3
codes, that are used in the graphical representations of the network.
III
B.2 Simulation-based Goodness-of-fit in (S)TERGMs
In Figures 7 and 10 we show simulation-based godness-of-fit (GOF) diagnostics for the the
TERGM model and in Figures 8, 9, 11 and 12 for the STERGM in the formation and
dissolution model, respectively. The figures are created by the R package ergm (version
3.10.4) and follow the approach of Hunter et al. (2008).
In all three models, the fitted
model is used in order to simulate 100 new networks. Based on these, different network
characteristics are computed and visualized in boxplots.
The standard characteristics used are the complete distributions of the in-degree, out-
degree, edge-wise shared partners and minimum geodesic distance (i.e. number of node pairs
with shortest path of length k between them). The solid black line indicates the measure-
ments of these characteristic in the observed network. These statistics show whether mea-
sures like GWID, GWOD and GWESP are sufficient to reproduce global network patterns.
Because many shares are rather small, we visualize the simulated and observed measures on
a log-odds scale.
On the bottom of the figures it is shown how well the actual network statistics are
reproduced. Note, that both models compare different things as the TERGM is evaluated
at yt while the STERGM regards y+ and y−. Overall, all plots indicate a satisfying fit of
the respective models.
IV
B.2.1 Data Set 1: International Arms Trade
Figure 7: Arms trade network: Simulation-based goodnes-of-fit diagnostics in the TERGM.
Boxplots give the evaluations of the respective network characteristics at the simulated net-
works and the solid line gives the actual values from the observed network. First four
panels give the log-odds of a node for different in-degrees (top left), out-degrees (top right),
edge-wise shared partners (middle, left) and minimum geodesic distance (middle right). All
included rescaled network statistics on the bottom panel.
V
012345678910121416−5−4−3−2−10in degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllll0481217222732374247525762−6−4−201out degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll012345678910111213−6−5−4−3−2−1edge−wise shared partnerslog−odds for an edgellllllllllllllllllllllllllllllllllllllllll12345678910NR−10−6−202minimum geodesic distancelog−odds for a dyadllllllllllllllllllllllllllllllllledgecov.lnetedgesedgecov.lrecipgwodeg.fixed.0.693147180559945edgecov.poldiffnodeocov.lgdpnodeicov.lgdp−4−2024model statisticslog−odds for the statisticlllllllllllllllllllllllllllGoodness−of−fit diagnostics Figure 8: Arms trade network: Simulation-based goodnes-of-fit diagnostics in the STERGM
for the formation model. Boxplots give the evaluations of the respective network character-
istics at the simulated networks and the solid line gives the actual values from the observed
network. First four panels give the log-odds of a node for different in-degrees (top left), out-
degrees (top right), edge-wise shared partners (middle, left) and minimum geodesic distance
(middle right). All included rescaled network statistics on the bottom panel.
VI
0246810121416182022−6−5−4−3−2−1in degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll04813182328333843485358636873−6−4−201out degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll012345678911131517−7−5−3−1edge−wise shared partnerslog−odds for an edgelllllllllllllllllllllllllllllllllllllllllllllllllllllllll123456789NR−10−6−4−202minimum geodesic distancelog−odds for a dyadlllllllllllllllllllllllllllllledgesmutualgwodeg.fixed.0.693147180559945edgecov.poldiffnodeocov.lgdpnodeicov.lgdp−4−2024model statisticslog−odds for the statisticllllllllllllllllllllllllFormation: Goodness−of−fit diagnostics Figure 9: Arms trade network: Simulation-based Goodnes-of-fit diagnostics in the STERGM
for the dissolution model. Boxplots give the evaluations of the respective network character-
istics at the simulated networks and the solid line gives the actual values from the observed
network. First four panels give the log-odds of a node for different in-degrees (top left), out-
degrees (top right), edge-wise shared partners (middle, left) and minimum geodesic distance
(between them, middle right). All included rescaled network statistics on the bottom panel.
VII
0123456789101214−5−4−3−2−10in degreelog−odds for a nodellllllllllllllllllllllllllllllllllllllllllllllll03691317212529333741454953−6−4−20out degreelog−odds for a nodellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll012345678910−5−4−3−2−10edge−wise shared partnerslog−odds for an edgelllllllllllllllllllllllllllllllll12345678NR−10−6−202minimum geodesic distancelog−odds for a dyadllllllllllllllllllllllllllledgesmutualgwodeg.fixed.0.693147180559945edgecov.poldiffnodeocov.lgdpnodeicov.lgdp−4−2024model statisticslog−odds for the statisticllllllllllllllllllllllllDissolution: Goodness−of−fit diagnostics B.2.2 Data Set 2: European Research Institution Email Correspondence
Figure 10: Email exchange network: Simulation-based Goodnes-of-fit diagnostics in the
TERGM. Boxplots give the evaluations of the respective network characteristics at the sim-
ulated networks and the solid line gives the actual values from the observed network. First
four panels give the log-odds of a node for different in-degrees (top left), out-degrees (top
right), edge-wise shared partners (middle, left) and minimum geodesic distance (middle
right). All included rescaled network statistics on the bottom panel.
VIII
02468111417202326293235−5−4−3−2In−DegreeLog−Odds for a Nodellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll024681013161922252831−5−4−3−2−1Out−DegreeLog−Odds for a Nodellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll02468101214161820−7−6−5−4−3−2Edge−Wise Shared Partnerslog−odds for an edgellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll123456789NR−8−6−4−20Minimum Deodesic Distancelog−odds for a dyadllllllllllllllllllllllllllllllRepetitionEdgesReciprocityGWIDGWODGWESP−4−2024Model Statisticslog−odds for the statisticllllllllllllllllllGoodness−of−fit diagnostics Figure 11: Email exchange network: Simulation-based Goodnes-of-fit diagnostics in the
STERGM for the formation model. Boxplots give the evaluations of the respective network
characteristics at the simulated networks and the solid line gives the actual values from the
observed network. First four panels give the log-odds of a node for different in-degrees (top
left), out-degrees (top right), edge-wise shared partners (middle, left) and minimum geodesic
distance (middle right). All included rescaled network statistics on the bottom panel.
IX
036912162024283236404448−5.0−4.0−3.0in degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll036912162024283236404448−5.0−4.0−3.0−2.0out degreelog−odds for a nodellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll02468111417202326293235−7−6−5−4−3−2edge−wise shared partnerslog−odds for an edgelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll12345678NR−8−6−4−20minimum geodesic distancelog−odds for a dyadllllllllllllllllllllllllllledgesmutualgwideg.fixed.0.693147180559945gwesp.fixed.0.693147180559945−4−2024model statisticslog−odds for the statisticlllllllllllllllFormation: Goodness−of−fit diagnostics Figure 12: Email exchange network: Simulation-based Goodnes-of-fit diagnostics in the
STERGM for the dissolution model. Boxplots give the evaluations of the respective network
characteristics at the simulated networks and the solid line gives the actual values from
the observed network. First four panels give the log-odds of a node for different in-degrees
(top left), out-degrees (top right), edge-wise shared partners (middle, left) and minimum
geodesic distance (between them, middle right). All included rescaled network statistics on
the bottom panel.
X
024681013161922252831−5−4−3−2−1in degreelog−odds for a nodellllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll02468101214161820222426−4−3−2−1out degreelog−odds for a nodelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll012345678911131517−6−5−4−3−2−1edge−wise shared partnerslog−odds for an edgellllllllllllllllllllllllllllllllllllllllllllllllllllll123456789NR−8−6−4−20minimum geodesic distancelog−odds for a dyadlllllllllllllllllllllllllllllledgesmutualgwideg.fixed.0.693147180559945gwesp.fixed.0.693147180559945−4−2024model statisticslog−odds for the statisticlllllllllllllllDissolution: Goodness−of−fit diagnostics B.3 ROC-based Goodness-of-fit
B.3.1 Data Set 1: International Arms Trade
As already stated in Section 3.3 of the main paper techniques for assessing the fit of a
probabilistic classification can be used when working with binary network data. In the case
of observations at discrete time points this allows an informal comparison of the models
proposed in Section 3 and 4.3.
In the case of the TERGM and STERGM the application of the ROC- and PR-curve
follows from the conditional probability of observing a specific tie (see equation (2.5) of
Hunter and Handcock, 2006). For the REM we predict the intensities of all possible events
given the information of t − 1 and use this value as a score in the calculation of the ROC
curve. While the latter approach is non-standard and can only be applied to REMs that
regard durable ties, it enables a direct comparison between the models as shown in Figure 13.
The results of the ROC curve indicate a generally good fit of all models. In the STERGM
more parameters are estimated, which seems to lead to a slightly bigger area under curve
(AUC) values as compared to the REM and TERGM. Similar to the conclusions from the
ROC curve, the PR curve favors the TERGM and STERGM over the REM.
Figure 13: Arms trade network: ROC and PR curves from the TERGM (dotted line),
STERGM (dotdashed line), and REM (solid line).The AUC values of the respective curves
are indicated in brackets.
XI
0.000.250.500.751.000.000.250.500.751.00SpecificitySensitivityTERGM (0.975)STERGM (0.985)REM (0.969) 0.000.250.500.751.000.000.250.500.751.00RecallPrecisionTERGM (0.633)STERGM (0.686)REM (0.611) B.3.2 Data Set 2: European Research Institution Email Correspondence
The second data set regards reoccurring events in the REM, which are aggregated for the
analysis of the TERGM and STERGM. Therefore, the ROC and PR curve are only available
for the TERGM and STERGM. The results are depicted in Figure 14. For this data set, the
ROC curves favor the TERGM. Yet, when emphasis is put on finding the true positives, the
PR curve detects a better model fit of the STERGM.
Figure 14: Email exchange network: ROC and PR curves from the TERGM (dotted line),
STERGM (dotdashed line), and REM (solid line). The AUC values of the respective curves
are indicated in brackets.
As explained in Section 4.4 one option to asses the goodness-of-fit of REMs is the recall
measure as proposed by Vu et al. (2011). We apply the measure in three different situations
that may be of interest when measuring the predictive performance of relational event models:
predict the next tie, next sender, and next receiver. The worst case scenario in terms of
predictions of a model would be random guessing of the next sender, receiver, or event, the
resulting recall rates are indicated by the dotted lines. The results in Figure 15 exhibit a
good predictive performance of the REM, i.e.
in about 75% of the events the right sender
and receiver is among the 25 most likely senders and receivers.
XII
0.000.250.500.751.000.000.250.500.751.00SpecificitySensitivityTERGM (0.92)STERGM (0.91)0.250.500.751.000.000.250.500.751.00RecallPrecisionTERGM (0.63)STERGM (0.658) (a)
(b)
(c)
Figure 15: Email exchange network: Recall Curves of REM regarding the next sender (a),
receiver (b), and event (c). The dotted line indicates the measure under random guessing of
the next sender, receiver, or event.
XIII
0.000.250.500.751.000255075CutoffRecall0.000.250.500.751.000255075CutoffRecall0.00.10.20.30.40.50100200300400CutoffRecall B.4 Application with multiple time points
In the main article, we fitted a TERGM as well as a STERGM to two time points, called
period 1 and period 2. However, it is possible to fit these models to multiple transitions. In
order to do so, we took the first two years of the email exchange network and aggregated
the deciles into 10 binary networks. Using again a first order Markov assumption and
conditioning on the first network, this allows to fit a TERGM as well as a STERGM to
the remaining nine networks. For comparison we additionally fit a REM to the data set.
The estimation is done using the function mtergm from the package btergm (version 3.6.1)
(Leifeld et al., 2018) that implements MCMC-based maximum likelihood. The STERGM is
fitted again using the package tergm (version 3.5.2).
The corresponding results can be found in Table 5 in column 1 to 3. Note that the
parameter estimates still refer to the transition from t − 1 to t and can interpreted in the
same way as in the main article.
In that regard note, that it is now assumed that the
coefficients stay constant with time. Possible approaches to relax this assumption were
given in the main article. The estimates of the REM (column 4) are consistent with the
(S)TERGM but slightly differ to the main article, since now we condition only on the first
out of 10 periods and hence model more events.
XIV
TERGM
STERGM
REM
Formation Dissolution
−
−
−
−
Edges
Repetition
Reciprocity
1.986∗∗∗
(0.058)
−3.435∗∗∗ −4.485∗∗∗ −0.769∗∗∗
(0.107)
(0.073)
(0.059)
2.755∗∗∗
1.090∗∗∗
1.761∗∗∗
(0.113)
(0.069)
(0.079)
−0.307∗
−1.239∗∗∗ −0.870∗∗∗
In-Degree (GWID)
(0.152)
(0.156)
(0.099)
Out-Degree (GWOD) −1.574∗∗∗ −1.757∗∗∗
−0.056
(0.159)
(0.155)
0.124∗∗
0.508∗∗∗
(0.043)
(0.029)
(0.097)
0.497∗∗∗
(0.028)
GWESP
−
2.354∗∗∗
(0.048)
−
−
In-Degree Receiver
1.699∗∗∗
(0.043)
0.007∗∗∗
(0.002)
0.009∗∗∗ Out-Degree Sender
(0.001)
0.199∗∗∗
(0.011)
Transitivity
Table 5: Email exchange network: Comparison of parameters obtained from the TERGM
(first column) and STERGM (Formation in the second column, Dissolution in the third
column). Standard errors in brackets and stars according to p-values smaller than 0.001
(∗∗∗), 0.05 (∗∗) and 0.1 (∗). Decay parameter of the geometrically weighted statistics is set
to log(2).
XV
|
1502.06682 | 2 | 1502 | 2015-02-26T15:31:34 | Maximizing Friend-Making Likelihood for Social Activity Organization | [
"cs.SI"
] | The social presence theory in social psychology suggests that computer-mediated online interactions are inferior to face-to-face, in-person interactions. In this paper, we consider the scenarios of organizing in person friend-making social activities via online social networks (OSNs) and formulate a new research problem, namely, Hop-bounded Maximum Group Friending (HMGF), by modeling both existing friendships and the likelihood of new friend making. To find a set of attendees for socialization activities, HMGF is unique and challenging due to the interplay of the group size, the constraint on existing friendships and the objective function on the likelihood of friend making. We prove that HMGF is NP-Hard, and no approximation algorithm exists unless P = NP. We then propose an error-bounded approximation algorithm to efficiently obtain the solutions very close to the optimal solutions. We conduct a user study to validate our problem formulation and per- form extensive experiments on real datasets to demonstrate the efficiency and effectiveness of our proposed algorithm. | cs.SI | cs |
Maximizing Friend-Making Likelihood for Social
Activity Organization
Chih-Ya Shen1, De-Nian Yang2, Wang-Chien Lee3, and Ming-Syan Chen1,2
1 Research Center for Information Technology Innovation, Academia Sinica
2 Institute of Information Science, Academia Sinica
3 Department of Computer Science and Engineering, The Pennsylvania State
University
{chihya,mschen}@citi.sinica.edu.tw,
[email protected],[email protected]
Abstract. The social presence theory in social psychology suggests that
computer-mediated online interactions are inferior to face-to-face, in-
person interactions. In this paper, we consider the scenarios of organiz-
ing in person friend-making social activities via online social networks
(OSNs) and formulate a new research problem, namely, Hop-bounded
Maximum Group Friending (HMGF), by modeling both existing friend-
ships and the likelihood of new friend making. To find a set of atten-
dees for socialization activities, HMGF is unique and challenging due
to the interplay of the group size, the constraint on existing friendships
and the objective function on the likelihood of friend making. We prove
that HMGF is NP-Hard, and no approximation algorithm exists unless
P = N P . We then propose an error-bounded approximation algorithm
to efficiently obtain the solutions very close to the optimal solutions. We
conduct a user study to validate our problem formulation and perform
extensive experiments on real datasets to demonstrate the efficiency and
effectiveness of our proposed algorithm.
1
Introduction
With the popularity and accessibility of online social networks (OSNs), e.g.,
Facebook, Meetup, and Skout4, more and more people initiate friend gatherings
or group activities via these OSNs. For example, more than 16 millions of events
are created on Facebook each month to organize various kinds of activities5,
and more than 500 thousands of face-to-face activities are initiated in Meetup6.
The activities organized via OSNs cover a wide variety of purposes, e.g., friend
gatherings, cocktail parties, concerts, and marathon events. The wide spectrum
of these activities shows that OSNs have been widely used as a convenient means
for initiating real-life activities among friends.
On the other hand, to help users expand their circles of friends in the cy-
berspace, friend recommendation services have been provided in OSNs to suggest
candidates to users who may likely become mutual friends in the future. Many
friend recommendation services employ link prediction algorithms, e.g., [10,11],
4 http://www.skout.com/
5 http://newsroom.fb.com/products/
6 http://www.meetup.com/about/
2
C.-Y. Shen, D.-N. Yang, W.-C. Lee, and M.-S. Chen
to analyze the features, similarity or interaction patterns of users in order to
derive potential future friendship between some users. By leveraging the abun-
dant information in OSNs, link prediction algorithms show high accuracy for
recommending online friends in OSNs.
As social presence theory [16] in social psychology suggests, computer-mediated
online interactions are inferior to face-to-face, in-person interactions, off-line
friend-making activities may be favorable to their on-line counterparts in cy-
berspace. Therefore, in this paper, we consider the scenarios of organizing face-
to-face friend-making activities via OSN services. Notice that finding socially
cohesive groups of participants is essential for maintaining good atmosphere for
the activity. Moreover, the function of making new friends is also an impor-
tant factor for the success of social activities, e.g., assigning excursion groups
in conferences, inviting attendees to housewarming parties, etc. Thus, for or-
ganizing friend-making social activities, both activity organization and friend
recommendation services are fundamental. However, there is a gap between ex-
isting activity organization and friend recommendation services in OSNs for the
scenarios under consideration. Existing activity organization approaches focus
on extracting socially cohesive groups from OSNs based on certain cohesive mea-
sures, density, diameter, of social networks or other constraints, e.g., time, spa-
tial distance, and interests, of participants [5,6,7,8]. On the other hand, friend
recommendation services consider only the existing friendships to recommend
potential new friends for an individual (rather than finding a group of people
for engaging friend-making). We argue that in addition to themes of common
interests, it is desirable to organize friend-making activities by mixing the "po-
tential friends", who may be interested in knowing each other (as indicated by
a link prediction algorithm), with existing friends (as lubricators). To the best
knowledge of the authors, the following two important factors, 1) the existing
friendship among attendees, and 2) the potential friendship among attendees,
have not been considered simultaneously in existing activity organization ser-
vices. To bridge the gap, it is desirable to propose a new activity organization
service that carefully addresses these two factors at the same time.
In this paper, we aim to investigate the problem of selecting a set of can-
didate attendees from the OSN by considering both the existing and potential
friendships among the attendees. To capture the two factors for activity organi-
zation, we propose to include the likelihood of making new friends in the social
network. As such, we formulate a new research problem to find groups with
tight social relationships among existing friends and potential friends (i.e., who
are not friends yet). Specifically, we model the social network in the OSN as
a heterogeneous social graph G = (V, E, R) with edge weight w : R → (0, 1],
where V is the set of individuals, E is the set of friend edges, and R is the set
of potential friend edges (or potential edges for short). Here a friend edge (u, v)
denotes that individuals u and v are mutual friends, while a potential edge [u′, v′]
indicates that individuals u′ and v′ are likely to become friends (the edge weight
w[u′, v′] quantifies the likelihood). The potential edges and the corresponding
Maximizing Friend-Making Likelihood for Social Activity Organization
3
0.4
g
0.8
0.5
e
0.7
f
0.9
a
0.6
b
0.3
d
a
0.4
g
0.8
0.5
e
0.2
c
0.7
f
0.9
g
0.7
f
c
d
g
0.8
0.7
f
0.9
e
d
(a) Input Graph G.
(b) H1.
(c) H2.
(d) H3.
Fig. 1. Illustrative Example.
edge weights can be obtained by employing a link prediction algorithm in friend
recommendation.
Given a heterogeneous social graph G = (V, E, R) as described above, we
formulate a new problem, namely, Hop-bounded Maximum Group Friending
(HMGF), to find a group that 1) maximizes the likelihood of making new friends
among the group, i.e., the group has the highest ratio of total potential edge
weight to group size, 2) ensures that the social tightness, i.e., hop count on
friend edges in G between each pair of individuals is small, and 3) is a suffi-
ciently large group, i.e., too small a group may not work well for socialization
activities.
Figure 1 illustrates the social graph and the interplay of the above factors.
Figure 1(a) shows a social graph, where a dash line, e.g., [a, b] with weight 0.6, is
a potential edge and a solid line, e.g., (c, d), is a friend edge. Figure 1(b) shows
a group H1:{a, e, f, g} which has many potential edges and thus a high total
weight. However, not all the members of this group have common friends as
social lubricators. Figure 1(c) shows a group H2:{c, d, f, g} tightly connected by
friend edges. While H2 may be a good choice for gathering of close friends, the
goal of friend-making in socialization activities is missed. Finally, Figure 1(d)
shows H3:{d, e, f, g} which is a better choice than H1 and H2 for socialization
activities because each member of H3 is within 2 hops of another member via
friend edges in G. Moreover, the average potential edge weight among them is
high, indicating members are likely to make some new friends.
Processing HMGF to find the best solution is very challenging because there
are many important factors to consider, including hop constraint, group size and
the total weight of potential edges in a group. Indeed, we prove that HMGF is an
NP-Hard problem with no approximation algorithm. Nevertheless, we prove that
if the hop constraint can be slightly relaxed to allow a small error, there exists a
3-approximation algorithm for HMGF. Theoretical analysis and empirical results
show that our algorithm can obtain good solutions efficiently.
The contributions made in this study are summarized as follows.
-- For socialization activity organization, we propose to model the existing
friendship and the potential friendship in a heterogeneous social graph and
formulate a new problem, namely, Hop-bounded Maximum Group Friending
(HMGF), for finding suitable attendees. To our best knowledge, HMGF is
the first problem that considers these two important relationships between
attendees for activity organization.
4
C.-Y. Shen, D.-N. Yang, W.-C. Lee, and M.-S. Chen
-- We prove that HMGF is NP-Hard and there exists no approximation al-
gorithm for HMGF unless P = N P . We then propose an approximation
algorithm, called MaxGF, with a guaranteed error bound for solving HMGF
efficiently.
-- We conduct a user study on 50 users to validate our argument for consider-
ing both existing and potential friendships in activity organization. We also
perform extensive experiments on real datasets to evaluate the proposed
algorithm. Experimental results manifest that HMGF can obtain solutions
very close to the optimal ones, very efficiently.
2 Problem Formulation
Based on the description of heterogeneous social graph described earlier, here we
formulate the Hop-bounded Maximum Group Friending (HMGF) tackled in this
paper. Given two individuals u and v, let dE
G(u, v) be the shortest path between
u and v via friend edges in G. Moreover, given H ⊆ G, let w(H) denote the total
weight of potential edges in H and let average weight, σ(H) = w(H)
H denote the
average weight of potential edges connected to each individual in H 7. HMGF is
formulated as follows.
Problem: Hop-bounded Maximum Group Friending (HMGF).
Given: Social network G = (V, E, R), hop constraint h, and size constraint p.
Objective: Find an induced subgraph H ⊆ G with the maximum σ(H), where
H ≥ p and dE
G(u, v) ≤ h, ∀u, v ∈ H.
Efficient processing of HMGF is very challenging due to the following reasons:
1) The interplay of the total weight w(H) and the size of H. To maximize σ(H),
finding a small H may not be a good choice because the number of edges in
a small graph tends to be small as well. On the other hand, finding a large H
(which usually has a high w(H)) may not lead to an acceptable σ(H), either.
Therefore, the key is to strike a good balance between the graph size H and the
total weight w(H). 2) HMGF includes a hop constraint (say h = 2) on friend
edges to ensure that every pair of individuals is not too distant socially from
each other. However, selecting a potential edge [u, v] with a large weight w[u, v]
may not necessarily satisfy the hop constraint, i.e., dE
G(u, v) > h which is defined
based on existing friend edges. In this case, it may not always be a good strategy
to prioritize on large-weight edges in order to maximize σ(H), especially when
u and v do not share a common friend nearby via the friend edges.
In the following, we prove that HMGF is NP-Hard and not approximable
within any factor. In other words, there exists no approximation algorithm for
HMGF.
Theorem 1. HMGF is NP-Hard and there is no approximation algorithm for
HMGF unless P = N P .
Proof. Due to the space constraints, we prove this theorem in the full version of
this paper (available online [1]).
7 Note that σ(H) = 0 if H = ∅.
Maximizing Friend-Making Likelihood for Social Activity Organization
5
3 Related Work
Extracting dense subgraphs or social cohesive groups among social networks is
a natural way for selecting a set of close friends for a gathering. Various social
cohesive measures have been proposed for finding dense social subgraphs, e.g.,
diameter [2], density [3], clique and its variations [4]. Although these social cohe-
sive measures cover a wide range of application scenarios, they focus on deriving
groups based only on existing friendship in the social network. In contrast, the
HMGF studied in this paper aims to extract groups by considering both the
existing and potential friendships for socialization activities. Therefore, the ex-
isting works mentioned above cannot be directly applied to HMGF tackled in
this paper.
Research on finding a set of attendees for activities based on the social
tightness among existing friends [5,6,7,8,9] have been reported in the literature.
Social-Temporal Group Query [5] checks the available times of attendees to find
the social cohesive group with the most suitable activity time. Geo-Social Group
Query [6,7] extracts socially tight groups while considering certain spatial prop-
erties. The willingness optimization for social group problem in [8] selects a set
of attendees for an activity while maximizing their willingness to participate. Fi-
nally, [9] finds a set of compatible members with tight social relationships in the
collaboration network. Although these works find suitable attendees for activi-
ties based on existing friendship among the attendees, they ignore the likelihood
of making new friends among the attendees. Therefore, these works may not be
suitable for socialization activities discussed in this paper.
Link prediction analyzes the features, similarity or interaction patterns among
individuals in order to recommend possible friends to the users [10,11,12,13,14].
Link prediction algorithms employ different approaches including graph-topological
features, classification models, hierarchical probabilistic model, and linear alge-
braic methods. These works show good prediction accuracy for friend recom-
mendation in social networks. In this paper, to estimate the likelihood of how
individuals may potentially become friends in the future, we employ link predic-
tion algorithms for deriving the potential edges among the individuals.
To the best knowledge of the authors, there exists no algorithm for activity
organization that considers both the existing friendship and the likelihood of
making new friends when selecting activity attendees. The HMGF studied in this
paper examines the social tightness among existing friends and the likelihood of
becoming friends for non-friend attendees. We envisage that our research result
can be employed in various social network applications for activity organization.
4 Experimental Results
We implement HMGF in Facebook and invite 50 users to participate in our user
study. Each user, given 12 test cases of HMGF using her friends in Facebook as
the input graph, is asked to solve the HMGF cases, and compare her results with
the solutions obtained by MaxGF. In addition to the user study, we evaluate the
performance of MaxGF on two real social network datasets, i.e., FB [15] and the
6
C.-Y. Shen, D.-N. Yang, W.-C. Lee, and M.-S. Chen
)
s
(
e
m
T
i
800
600
400
200
0
User(h=2)
MaxGF(h=2)
User(h=3)
MaxGF(h=3)
User_FeaRatio
MaxGF_FeaRatio MaxGF_ObjRatio
100%
User_ObjRatio
o
i
t
a
R
80%
60%
40%
10
14
18
22
V (p=2)
10
14
18
22
V (h=2, p=2)
User Satisfaction
o
i
t
a
R
100%
80%
60%
40%
20%
0%
User
DkS
MaxGF
(a) Required Time.
(b) FeaRatio and ObjRatio.
(c) User Satisfaction.
Fig. 2. User Study Results.
MS dataset from KDD Cup 20138. The FB dataset is extracted from Facebook
with 90K vertices, and MS is a co-author network with 1.7M vertices. We ex-
tract the friend edges from these datasets and identify the potential edges with
a link prediction algorithm [11]. The weight of a potential edge is ranged within
(0,1]. Moreover, we compare MaxGF with two algorithms, namely, Baseline and
DkS [3]. Baseline finds the optimal solution of HMGF by enumerating all the
subgraphs satisfying the constraints, while DkS is an O(V 1/3)-approximation
algorithm for finding a p-vertex subgraph H ⊆ G with the maximum density
on E ∪ R without considering the potential edges and the hop constraint. The
algorithms are implemented in an IBM 3650 server with Quadcore Intel X5450
3.0 GHz CPUs. We measure 30 samples in each scenario. In the following, Fea-
Ratio and ObjRatio respectively denote the ratio of feasibility (i.e., the portion
of solutions satisfying the hop constraint) and the ratio of σ(H) in the solutions
obtained by MaxGF or DkS to that of the optimal solution.
4.1 User Study
Figure 2 presents the results of the user study. Figure 2(a) compares the required
time for users and MaxGF to solve the HMGF instances. Users need much more
time than MaxGF due to challenges brought by the hop constraint and trade-
offs in potential edge weights and the group size, as explained in Section 2. As
V or h grows, users need more time because the HMGF cases become more
complicated. Figure 2(b) compares the solution feasibility and quality among
users and MaxGF. We employ Baseline to obtain the optimal solutions and de-
rive FeaRatio and ObjRatio accordingly. The FeaRatio and ObjRatio of users
are low because simultaneously considering both the hop constraint on friend
edges and total weights on potential edges is difficult for users. As shown, users'
FeaRatio and ObjRatio drop when V increases. By contrast, MaxGF obtains
the solutions with high FeaRatio and ObjRatio. In Figure 2(c), we ask each user
to compare her solutions with the solutions obtained by MaxGF and DkS, to
validate the effectiveness of HMGF. 74% of the users agree that the solution of
MaxGF is the best because HMGF maximizes the likelihood of friend-making
while considering the hop constraint on friend edges at the same time. By con-
trast, DkS finds the solutions with a large number of edges, but it does not
8 https://www.kaggle.com/c/kdd-cup-2013-author-paper-identification-
challenge/data
Maximizing Friend-Making Likelihood for Social Activity Organization
7
Baseline DkS MaxGF
DkS_FeaRatio MaxGF_FeaRatio
DkS_ObjRatio MaxGF_ObjRatio
1.E+05
1.E+03
1.E+01
1.E-01
o
i
t
a
R
100%
80%
60%
40%
20%
0%
o
i
t
a
R
100%
80%
60%
40%
20%
0%
50
100
150
200
250
V (h=4, p=15)
50
150
100
200
V (h=4, p=15)
250
50
150
100
200
V (h=4, p=15)
250
i
)
s
(
e
m
T
n
o
i
t
u
c
e
x
E
(a) Time of Diff. V .
(b) FeaRatio of Diff. V .
(c) ObjRatio of Diff V .
Baseline DkS MaxGF
DkS_FeaRatio MaxGF_FeaRatio
DkS_ObjRatio MaxGF_ObjRatio
1.E+05
1.E+03
1.E+01
1.E-01
o
i
t
a
R
100%
80%
60%
40%
20%
0%
o
i
t
a
R
100%
80%
60%
40%
20%
0%
2
4
3
5
h (V=150, p=15)
6
2
3
4
5
h (V=150, p=15)
2
3
4
5
h (V=150, p=15)
i
)
s
(
e
m
T
n
o
i
t
u
c
e
x
E
(d) Time of Diff. h.
(e) FeaRatio of Diff. h.
(f) ObjRatio of Diff. h.
Fig. 3. Comparisons with Optimal Solutions.
differentiate the friend edges and potential edges. Therefore, users believe that
the selected individuals may not be able to socialize with each other effectively.
4.2 Performance Evaluation
Baseline can only find the optimal solutions of small HMGF cases since it enu-
merates all possible solutions. Therefore, we first compare MaxGF against Base-
line and DkS on small graphs randomly extracted from FB. Figure 3(a) com-
pares the execution time of the algorithms by varying the size of input graph.
Since Baseline enumerates all the subgraphs H with H ≥ p, the execution
time grows exponentially. The execution time of MaxGF is very small because
the hop-bounded subgraphs and the pruning strategy effectively trim the search
space. Figures 3(b) and 3(c) present the FeaRatio and ObjRatio of the algo-
rithms, respectively. MaxGF has high ObjRatio because MaxGF iteratively re-
moves vertices with low incident weights from each hop-bounded subgraph Hv,
and extracts the solution SAP X with maximized σ(SAP X ) among different sub-
graphs in different Hv to strike a good balance on total edge weights and group
sizes as describe in Section 2. Moreover, the high FeaRatio and ObjRatio also in-
dicate that the post-processing procedure effectively restores the hop constraint
and maximizes the average weight accordingly. By contrast, DkS does not con-
sider the hop constraint and different edge types in finding solutions and thus
generates the solutions with smaller FeaRatio and ObjRatio.
Figures 3(d)-(f) compare execution time, FeaRatio and ObjRatio again but
by varying h. When h increases, the execution time of MaxGF grows slowly
because the pruning strategy avoids examining the hop-bounded subgraphs that
do not lead to a better solution. The FeaRatio and ObjRatio of MaxGF with
different h are high because MaxGF employs hop-bounded subgraphs to avoid
generating solutions with large hop distances on friend edges, and the post-
processing procedure effectively restores the hop constraint and maximizes the
objective function.
8
C.-Y. Shen, D.-N. Yang, W.-C. Lee, and M.-S. Chen
FB_FeaRatio
MS_FeaRatio
FB_GroupSize MS_GroupSize
FB
MS
o
i
t
a
R
100%
90%
80%
70%
70%
60%
e
z
i
S
p
u
o
r
G
G
250
200
150
100
50
50
0
2
3
4
5
6
2
3
4
5
6
h (p=60)
h (p=60)
i
)
s
(
e
m
T
n
o
i
t
u
c
e
x
E
1200
900
600
300
0
20
40
60
80
100
p (h=4)
(a) FeaRatio of Diff. h.
(b) SAP X of Diff. h.
(c) Time of Diff. p.
Fig. 4. Experimental Results on Different Datasets.
Figure 4 compares MaxGF in different datasets, i.e., FB and MS. Figures
4(a) and 4(b) present the FeaRatio and the solution group sizes with different h.
As h increases, MaxGF on both datasets achieves a higher FeaRatio due to the
post-processing procedure adjusts SAP X and further minimizes dE
G(u, v), ∀u, v ∈
SAP X . Moreover, it is worth noting that the returned group sizes grow when h in-
creases in MS. This is because MS contains large densely connected components
with large edge weights. When h is larger, MaxGF is inclined to extract larger
groups from these components to maximize the objective function. By contrast,
FB does not have large components and MaxGF thereby tends to find small
groups to reduce the group size for maximizing the objective function. In fact,
the solutions in FB are almost the same with different h. Finally, MaxGF needs
to carefully examine possible solutions with the sizes at least p, and thus Fig-
ure 4(c) shows that when p increases, the execution time drops because MaxGF
effectively avoids examining the candidate solutions with small group sizes.
5 Conclusion
To bridge the gap between the state-of-the-art activity organization and friend
recommendation in OSNs, in this paper, we propose to model the individu-
als with existing and potential friendships in OSNs for friend-making activity
organization. We formulate a new research problem, namely, Hop-bonded Max-
imum Group Friending (HMGF), to find suitable activity attendees. We prove
that HMGF is NP-Hard and there exists no approximation algorithms unless
P = N P . We then propose an approximation algorithm with guaranteed er-
ror bound, i.e., MaxGF, to find good solutions efficiently. We conduct a user
study and extensive experiments to evaluate the performance of MaxGF, where
MaxGF outperforms other relevant approaches in both solution quality and ef-
ficiency.
References
1. C.-Y. Shen, D.-N. Yang, W.-C. Lee, and M.-S. Chen. Maximizing Friend-Making
Likelihood for Social Activity Organization. arXiv:1502.06682, 2015.
2. S. Wasserman and K. Faust. Social Network Anlysis: Methods and Applications.
Cambridge University Press, 1994.
3. U. Feige, G. Kortsarz, and D. Peleg. The Dense k-Subgraph Problem. Algorithmica,
2001.
4. R. Mokken. Cliques, Clubs and Clans. Quality and Quantity: International Journal
of Methodology, 1979.
Maximizing Friend-Making Likelihood for Social Activity Organization
9
5. D.-N. Yang, Y.-L. Chen, W.-C. Lee, and M.-S. Chen. On Social-Temporal Group
Query with Acquaintance Constraint. VLDB, 2011.
6. D.-N. Yang, C.-Y. Shen, W.-C. Lee, and M.-S. Chen. On Socio-Spatial Group
Query for Location-based Social Networks. KDD, 2012.
7. Q. Zhu, H. Hu, J. Xu, and W.-C. Lee. Geo-Social Group Queries with Minimum
Acquaintance Constraint. arXiv:1406.7367v1, 2014.
8. H.-H. Shuai, D.-N. Yang, P. S. Yu, and M.-S. Chen. Willingness Optimization for
Social Group Activity. VLDB, 2014.
9. D. Surian, N. Liu, D. Lo, H. Tong, E.-P. Lim, and C. Faloutsos. Recommending
People in Developers' Collaboration Network. WCRE, 2011.
10. H. Kashima and N. Abe. A Parameterized Probabilistic Model of Network Evolu-
tion for Supervised Link Prediction. ICDM, 2006.
11. D. Liben-Nowell and J. Kleinberg. The Link Prediction Problem for Social Net-
works. Journal of the American Soceity for Information Science and Technology,
2007.
12. A. Clause, C. Moore, and M. Newman. Hierarchical Structure and the Prediction
of Missing Links in Network. Nature, 2008.
13. J. Kunegis and A. Lommatzsch. Learning Spectral Graph Transformations for Link
Prediction. ICML, 2009.
14. C. Leung, E.-P. Lim, D. Lo, and J. Weng. Mining Interesting Link Formation Rules
in Social Networks. CIKM, 2010.
15. B. Viswanath, A. Mislove, M. Cha, and K. Gummadi. On the Evolution of User
Interaction in Facebook. WOSN'09.
16. J. Short, E. Williams, and B. Christie. The Social Psychology of Telecommunica-
tions. London: Wiley, 1976.
|
1911.00067 | 1 | 1911 | 2019-10-31T19:26:21 | DNA: Dynamic Social Network Alignment | [
"cs.SI"
] | Social network alignment, aligning different social networks on their common users, is receiving dramatic attention from both academic and industry. All existing studies consider the social network to be static and neglect its inherent dynamics. In fact, the dynamics of social networks contain the discriminative pattern of an individual, which can be leveraged to facilitate social network alignment. Hence, we for the first time propose to study the problem of aligning dynamic social networks. Towards this end, we propose a novel Dynamic social Network Alignment (DNA) framework, a unified optimization approach over deep neural architectures, to unfold the fruitful dynamics to perform alignment. However, it faces tremendous challenges in both modeling and optimization: (1) To model the intra-network dynamics, we explore the local dynamics of the latent pattern in friending evolvement and the global consistency of the representation similarity with neighbors. We design a novel deep neural architecture to obtain the dual embedding capturing local dynamics and global consistency for each user. (2) To model the inter-network alignment, we exploit the underlying identity of an individual from the dual embedding in each dynamic social network. We design a unified optimization approach interplaying proposed deep neural architectures to construct a common subspace of identity embeddings. (3) To address this optimization problem, we design an effective alternating algorithm with solid theoretical guarantees.We conduct extensive experiments on real-world datasets and show that the proposed DNA framework substantially outperforms the state-of-the-art methods. | cs.SI | cs | DNA: Dynamic Social Network Alignment
Li Sun, Zhongbao Zhang, Pengxin Ji, Jian Wen, Sen Su
Beijing University of Posts and Telecommunications, Beijing, China
{l.sun, zhongbaozb, jpx, j.wen, susen}@bupt.edu.cn
Philip S. Yu
University of Illinois at Chicago, IL, USA
[email protected]
9
1
0
2
t
c
O
1
3
]
I
S
.
s
c
[
1
v
7
6
0
0
0
.
1
1
9
1
:
v
i
X
r
a
Abstract -- Social network alignment, aligning different social
networks on their common users, is receiving dramatic attention
from both academic and industry. All existing studies consider the
social network to be static and neglect its inherent dynamics. In
fact, the dynamics of social networks contain the discriminative
pattern of an individual, which can be leveraged to facilitate
social network alignment. Hence, we for the first time propose
to study the problem of aligning dynamic social networks.
Towards this end, we propose a novel Dynamic social Network
Alignment (DNA) framework, a unified optimization approach
over deep neural architectures, to unfold the fruitful dynamics
to perform alignment. However, it faces tremendous challenges
in both modeling and optimization: (1) To model the intra-
network dynamics, we explore the local dynamics of the latent
pattern in friending evolvement and the global consistency of the
representation similarity with neighbors. We design a novel deep
neural architecture to obtain the dual embedding capturing local
dynamics and global consistency for each user. (2) To model the
inter-network alignment, we exploit the underlying identity of
an individual from the dual embedding in each dynamic social
network. We design a unified optimization approach interplaying
proposed deep neural architectures to construct a common
subspace of identity embeddings. (3) To address this optimization
problem, we design an effective alternating algorithm with solid
theoretical guarantees. We conduct extensive experiments on real-
world datasets and show that the proposed DNA framework
substantially outperforms the state-of-the-art methods.
Index Terms -- social network alignment, dynamics, deep learn-
ing, optimization
I. INTRODUCTION
Social network alignment is to align different social net-
works on their common users, which is also known as anchor
link prediction [1]. It routinely finds itself in a wide spectrum
of applications, such as network fusion, link prediction, infor-
mation diffusion and cross-domain recommendation. On the
whole, social network alignment paves the way for the broad
learning across social networks. Thus, it is receiving increasing
attention from both academic and industry.
All existing studies [2], [3], [4], [5], [6], [7], [8], [9], [10],
[11], [12], [13], [14], [15], [16], [17], to the best of our knowl-
edge, consider social networks to be static. As shown in Fig. 1
(a), given static social networks with some anchor (common)
users, it is impossible to distinguish the counterpart of c in
social network A from c(cid:48) and d(cid:48) in network B. Indeed, static
social network alignment is still far away from the real-world
scenario, where there is nothing permanent except change.
Social networks are inherently dynamic as an individual more
often than not adds or deletes his or her friends. Hence,
can we leverage the dynamics of social networks to facilitate
Fig. 1: Difference between static and dynamic social network
alignment. Nodes denote users. Dotted lines denote intra-
network links and full lines denote anchor links. Fig. (a) shows
static social network alignment, where some anchor (common)
users, i.e., {a, a(cid:48)} and {b, b(cid:48)}, are known in advance, while
Fig. (b) shows dynamic social network alignment with social
networks evolving in the real-world scenario.
alignment? The answer is yes! We take advantage of the
order of friending in real-world dynamic social networks. The
order matters! In fact, friendship evolvement of an individual
across social networks is dominantly deterministic rather than
stochastic, as evidenced by the study of social psychology
[18]. Recall the toy example in Fig. 1 (b). We observe that
c first friends anchor user {a, a(cid:48)} and then friends {b, b(cid:48)}
in network A, and c(cid:48) also friends {a, a(cid:48)} prior to {b, b(cid:48)} in
network B. It shows that c and c(cid:48) share the same discriminative
pattern of dynamics. Thus, c is more likely to be aligned with
c(cid:48) than d(cid:48). Motivated by this observation, we for the first time
pose the problem of aligning dynamic social networks.
Towards this end, we propose a novel Dynamic social
Network Alignment framework, referred to as DNA, to unfold
the fruitful dynamics of social networks to perform alignment.
However, it faces tremendous challenges in both modeling and
optimization:
• The first and foremost challenge lies in how to model the
intra-network dynamics. An individual, more often than
not, expands or shrinks his or her friend list along with
the evolvement of dynamic social networks. However, no
existing studies focus on the representing users with such
complex dynamics for alignment. Thus, it is challenging
to encode the intra-network dynamics.
• The second challenge is how to model the inter-network
alignment. An intuitive idea is to discover the underly-
ing identity representation of users from different social
• Last but not
networks to facilitate alignment. Revealing such identity
representation is already nontrivial even if these social
networks are simplified to be static [4], [5], [19], [11],
[6]. Thus, it is tougher to reveal the underlying identity
from dynamic social networks.
it
the least, after modeling,
is always
a challenge to address the optimization problem. The
optimization problem tends to be nonconvex due to the
inherent complexity of dynamic network alignment. For
a nonconvex optimization, it lacks an effective method in
general to approach the optimal or suboptimal solution.
Thus, it renders this problem much more challenging.
To address the first challenge, we design a novel deep neural
architecture to model users in each dynamic social network. To
model users with dynamics, we exploit not only the temporal
evolvement in friending behaviors of a user, referred to as local
dynamics, but also representation similarity with the neighbors
of the user, referred to as global consistency. Specifically,
to capture local dynamics, we design an LSTM Autoencoder
zooming in the temporal friendship evolvement of each user
for encoding the evolvement pattern. To capture global con-
sistency, we design a consistency regularization and impose
it onto the heart of the LSTM Autoencoder for compensating
local dynamics. We utilize the hidden state in the heart of the
neural architecture, referred to as dual embedding, to jointly
embed local dynamics and global consistency for the users of
dynamic social network.
To address the second challenge, we design a unified
optimization approach to construct a common subspace from
different dynamic social networks. All behaviors of an individ-
ual across different social networks tend to share something
in common, termed as underlying identity. It can be leveraged
to identify this individual. Motivated by this fact, we first
introduce a projection matrix to transform the dual embedding
above into the identity embedding for the underlying identity
representation. Then, we formulate a unified optimization
approach to construct the common subspace of identity em-
beddings, where users are naturally aligned according to the
underlying identity. This optimization is constrained on the
collection of anchor users known in advance so that identity
embeddings of the same individual remain consistent. Fi-
nally, we obtain the unified optimization approach interplaying
proposed deep neural architectures, referred to as the DNA
framework, for dynamic social network alignment.
The optimization problem of DNA is not jointly convex
over the matrix variables of (1) neural architecture parameters
for dual embedding, (2) projection and (3) identity embed-
ding. To address this nonconvex optimization, we design an
effective alternating algorithm to approach a local optimum
of the objective, i.e., we update one of the matrix variables
while fixing others. Specifically, for the matrices of neural
architecture parameters, we employ the gradient method. For
the matrices of projections, we derive the closed-form solution
and prove that the derived solution is the global optimum.
For the matrices of identity embeddings, we first construct an
auxiliary function and then derive the multiplicative updating
rule. We further prove the correctness and convergence of the
multiplicative updating rule and draw corresponding theorems
with solid theoretical guarantees.
In DNA, the proposed deep neural architecture enjoys the
powerful nonlinear learning ability to capture the complex
dynamics, and the unified optimization takes advantage of its
inherent merit to naturally model the underlying identities. The
proposed DNA framework interplays two neural architectures
with the unified optimization,
thereby effectively aligning
dynamic social networks.
To evaluate DNA, we first restore the real-world dynamic
social network by framing a sequence of network snapshots via
the collected timestamps. Then, we conduct extensive experi-
ments on real-world datasets and demonstrate the superiority
of the proposed DNA framework.
Finally, we summarize noteworthy contributions as follows:
• To the best of our knowledge, we for the first time pose
the problem of aligning dynamic social networks.
• To address this problem, we design a novel DNA frame-
work. In DNA, we propose a novel deep neural archi-
tecture to capture the complex dynamics and a unified
optimization approach to construct the common subspace.
• To address the optimization of DNA, we design an
effective alternating algorithm to approach an optimum
with solid theoretical guarantees.
• We conduct extensive experiments on real-world datasets
and experimental results demonstrate evident superiority
of the proposed DNA framework.
The rest of this paper is organized as follows: in Sec. 2,
we formalize this problem. Sec. 3 and Sec. 4 introduce the
modeling and optimization of the proposed DNA framework,
respectively. Sec. 5 shows the experimental results and Sec. 6
summarizes the related work. Finally, we conclude our work
in Sec. 7.
II. PROBLEM DEFINITION
There are a source dynamic social network and a target
dynamic social network. While the terminology of social
network is commonly used, we use social graph in the rest of
this paper to avoid ambiguity with the terminology of neural
network. A dynamic social graph is naturally represented as
a graph tensor. We use the superscript of s (or t) to indicate
notations of the source graph tensor Gs (or target graph tensor
Gt). Take the social graph tensor Gs ∈ RN s×N s×M for
(·) in total and M snapshots in
instance. There are N s users us
time order. A slide of the tensor Gs,m ∈ RN s×N s represents
the mth snapshot of the dynamic social graph. The matrix
Gs,m can be asymmetric or symmetric and its element [Gs,m]ij
indicates directed or undirected connection strength between
the ith user and jth user at the time of mth snapshotting in Gs.
The value of [Gs,m]ij can be either binary {0, 1} or arbitrary
nonnegative value. The proposed framework, DNA, can align
any (un)directed (un)weighted dynamic social graphs indeed.
Without loss of generality, the source and target dynamic social
graphs are partially overlapped via the anchor users, who join
TABLE I: Main Notations and Definitions
Notations
G
A
U
V
P
Q
Du
Dc
α, β, γ
Definition
the dynamic social graph tensor
the anchor set with size of A
the matrix of dual embedding
the matrix of identity embedding
the indication matrix
the projection matrix
the dimension of dual embedding
the dimension of identity embedding
the model parameters
in both social graphs. These users are recorded in an anchor
set A = {(k, l)a}. Each element in A represents an anchor
user, i.e., the ath element (k, l)a describes the kth user in the
source graph Gs corresponds to the lth user in the target graph
Gt. We use A = {(k, l)a} to denote the size of A.
We summarize the main notations in Table 1 and formally
define the problem of aligning dynamic social networks:
Problem Definition (Aligning Dynamic Social Networks).
Given the source and target dynamic social graph tensors Gs
and Gt of M snapshots with an anchor set A, the problem
of aligning dynamic social networks is to find a mapping Φ
which maps a user to its owner of natural individual, i.e.,
(·)) holds if and only if the individual joins in
Φ(us
both dynamic social graphs.
(·)) = Φ(ut
III. DNA: FRAMEWORK
To address
the problem of aligning dynamic social
networks, we propose a novel Dynamic social Network
Alignment framework, referred to as DNA, with a unified op-
timization approach over deep neural architectures illustrated
in Fig. 2. In the DNA framework, to encode the intra-network
dynamics, we propose a novel deep neural architecture to
embed users of each dynamic social network. To address
the inter-network alignment, we design a unified optimiza-
tion framework interplaying proposed neural architectures to
construct a common subspace, where user embeddings of the
same individual are naturally aligned.
A. Modeling Users with Complex Dynamics
To model the complex dynamics, we explore the temporal
local dynamics.
evolvement of the user's friendship,
Moreover, user should own similar embedding with his or her
friends to facilitate alignment as pointed out in studies [8], [5].
That is, the global consistency, representation similarity with
the user's close neighbors, is of significant as well.
i.e.,
How can we capture both of them? Recently, LSTMs have
proved to be effective for encoding evolvement in the se-
quential data [20]. However, it largely remains open to model
users with complex dynamics for social network alignment. To
bridge this gap, we design a novel deep neural architecture,
LSTM Autoencoder with consistency regularization, where we
can obtain the dual embedding u(·)
i of dimension Du for each
user jointly embedding local dynamics and global consistency.
Since Gs is the same as Gt on encoding the dynamics, we use
Gs for elaboration in this subsection.
1) Ego Networks: What should be the input for capturing
the dynamics? We introduce the concept of an ego network
for each user and explore the temporal proximity evolvement
in his or her ego network via an LSTM Autoencoder.
i
i
i
,
and
i = ξr(ω−1)
r(ω)
probability that the random walker will not restart.
(1)
j[G]ij and ξ is the
ω r(ω)
D−1G + (1 − ξ)r(0)
Random walk with restart (RWR) [21] has been shown
as an effective strategy to identify the proximities. Here, we
leverage ω-step RWR to calculate the proximity between users.
Given an adjacency matrix G ∈ RN s×N s and a start node
i ∈ RN s, the RWR score vector
i, indicated by a binary r(0)
ri ∈ RN s encodes the proximities between user i and the
others. Formally, ri is defined as ri =(cid:80)
where D = diag({di}) with di = (cid:80)
aggregated adjacency matrix Gs,A = (cid:80)
Specifically, for the ith user, we first choose W friends with
high proximity to form the ego network. That is, we locate
the indexes {q1, q2, . . . , qW} of W largest RWR scores on the
m em−MGs,m with
exponential time decay for matrix G in Eq. (1). Now, the ego
network is formed by connecting each node to its W largest
RWR score neighbors over the aggregated adjacency matrix.
Next, we examine how the RWR scores to these W friends
evolve over the M snapshots. Specifically, we calculate the
∈ RW of user i for each snapshot Gs,m. The
ego vector xs,m
ego vector xs,m
contains the proximity between user i and his
(or her) W friends in the ego network, i.e., [xs,m
]w = [ri]qw
and ri is the RWR score vector on snapshot Gs,m, where
Gs,m serves as matrix G in Eq. (1). Finally, we obtain the
ego vector sequence {xs,m
}M
m=1 depicting the evolvement
regarding friendship distribution in the ego network for the
}M
given user i. The ego vector sequence {xs,m
m=1 is to feed
into the encoder LSTM.
i
i
i
i
i
i
i.e.,
2) Capturing Evolvement Pattern: How can we capture
local dynamics? For the ith
the evolvement pattern,
user, when ego vector sequence {xs,m
} is processed by the
encoder LSTM Cell recursively from 1 to M, the pattern of
temporal evolvement of his or her friendship is captured in
∈ RDu, where we have the function
the hidden state us,M
i = LSTMCell(us,m−1
us,m
) performed recursively and
us,0
is initiated to zero vector. In LSTM, gating mechanisms
(forget gate f, input gate i and output gate o) with cell state c
are leveraged for long- and short-term dependence. Formally,
we describe transformations in the LSTM cell as follows:
, xs,m
i
i
i
i
i
i
i
i
i = σ(Wf [us,m−1
f s,m
, xs,m
i = σ(Wi[us,m−1
is,m
, xs,m
i = σ(Wo[us,m−1
os,m
, xs,m
i = ReLU(Wc[us,m−1
cs,m
cs,m
i = f s,m
us,m
i = os,m
i (cid:12) cs,m−1
i (cid:12) ReLU(cs,m
i
i
i
i
i
] + bf ),
] + bi),
] + bo),
] + bc),
, xs,m
i
i (cid:12) cs,m
+ is,m
),
i
,
(2)
where [·,·] denotes concatenation. W(·) and b(·) are parameter
matrices and biases, respectively. σ is the sigmoid function and
(cid:12) is the Hadamard product. We utilize the activation function
ReLU to facilitate further optimization. The decoder LSTM
Fig. 2: An overview of DNA framework. The input are source and target dynamic social graphs with a collection of common
users known in advance, as shown in Fig. (a). For each dynamic graph, ego vector sequences are fed into the proposed deep
neural architecture to encode users with complex dynamics, as shown in Fig. (b), where superscript s or t is omitted for
clarity. Finally, interplaying the proposed deep neural architectures, DNA framework outputs a common subspace of identity
embeddings for alignment, as shown in Fig. (c).
i
decodes the hidden state xs,M
to recover ego vector reversely
from M to 1, and we append a full connection layer to the
decoder for dimension transformation, as shown in Fig. 2 (b).
For all users in the dynamic social graph Gs, we give
the reconstructed error of LSTM Autoencoder in the form of
tensor approximation: (cid:13)(cid:13)(cid:13)X s − X s(cid:13)(cid:13)(cid:13)2
,
F
i
i
(3)
where · F is the Frobenius norm. X s ∈ RN s×W×M is the
ego tensor. Its mth slide X s,m ∈ RN s×W is the matrix of ego
vector for the mth snapshot Gs,m, whose ith row is the ego
. X s is the reconstructed ego tensor from decoder
vector xs,m
LSTM. The evolvement pattern is embedded in the output
hidden state of the last encoder LSTM cell us,M
. According
to the study [22], this embedding space will preserve the local
dynamics in the graph tensor via minimizing the reconstructed
error even without an explicit preservation restriction.
ij = rij
(cid:13)(cid:13)us
i − us
(cid:104)
3) Consistency Regularization: Focusing on the temporal
evolvement of the user friendship, aforementioned embedding
us highlights local dynamics but lacks global consistency.
As a compensation, we impose a consistency regularization
Rs onto the heart of LSTM Autoencoder as shown in Fig.
2 (b), to incorporate the global consistency. Specifically, we
draw neighbors closer in the embedding space via a pairwise
penalty Rs
, where rij = [ri]j indicates
the proximity between the ith user and jth user in Gs. The
consistency regularization Rs is summation of Rs
ij over all
user pairs {(ij)} in Gs,
(cid:13)(cid:13)2
UsTGs,LUs(cid:105)
(4)
where Gs,L = D − Gs,A is the graph Laplacian and Us ∈
RN s×Du
is the dual embedding matrix of local dynamics and
global consistency, whose ith row corresponds to the ith user.
For dynamic social graph Gs, combining Eq. (3) and Eq.
(4), we obtain the loss function of the proposed deep neural
architecture as follows:
Ds =
(5)
where α is the weight parameter. The norm term is for local
dynamics and the trace term is for global consistency.
(cid:13)(cid:13)(cid:13)X s − X s(cid:13)(cid:13)(cid:13)2
Rs = T r
+ αRs,
+
F
j
2
,
B. Constructing Common Subspace
i ∈ RDc
All behaviors of an individual in both social graphs tend
to share a common nature, his or her underlying identity. It
can be leveraged to identify this individual. Motivated by this
fact, we design a unified optimization approach to construct
the common subspace of identity embedding vs
+ with
dimension Dc depicting such underlying identity.
Specifically, we first introduce a projection matrix Q(·) ∈
RDc×Du, for each dynamic social graph, to transform the
+ into the identity embedding vs
i .
dual embedding us
Use source dynamic graph Gs for elaboration. Formally, we
i Qs for the ith users. Thus, for Gs, we obtain
have us
the matrix of identity embedding Vs ∈ RN s×Dc
(Dc (cid:28)
min{N s, N t}), whose ith row is the identity embedding vs
i
for user i. We formulate this transformation as a loss term in
the form of matrix approximation:
i ∈ RDu
i = vs
+
(cid:107)Us − VsQs(cid:107)2
F .
(6)
Note that, Dc, without loss of generality, is not necessarily
equal to Du for dual embedding.
Then, we construct the common subspace of the identity
embeddings. Naturally, the identity embeddings of the same
individual in different networks remain equal, i.e., the matrices
of identity embedding Vs and Vt are aligned on the anchor
users. To formulate the constraint, we derive a pair of indi-
cation matrices {Ps, Pt} from the anchor set A = {(k, l)a},
where Ps ∈ RA×N s, Pt ∈ RA×N t and A is the size of
the anchor set A. The indication matrix Ps is defined as
a ∈ RN s is a binary
[ ps
vector with only one element valued 1, i.e., [ps
a]k = 1. The
a is to indicate the corresponding row, the identity
vector ps
k of kth user in Gs, for aligning. Similarly, Pt
embedding vs
a]l = 1 is the only
is defined as [ pt
a ∈ RN t. We enforce the
element valued 1 in binary vector pt
following equality
A ]T whose row vector ps
A ]T and [pt
2, . . . , ps
2, . . . , pt
1, ps
1, pt
(7)
for all anchor users in A to depict the alignment between
matrices of identity embeddings Vs and Vt.
PsVs = PtVt,
C. Optimization Objective
Now, we formulate a unified optimization interplaying the
proposed deep neural architectures for dynamic social network
alignment. Incorporating Eq. (5) and Eq. (6) for each dynamic
social graph G(·), we have following objectives:
J s = Ds + βUs − VsQs2
F ,
J t = Dt + βUt − VtQt2
F ,
(8)
where β weighs the significance of the projection.
Next, we rewrite the equality constraint in the form of
matrix approximation and reformulate the equality constraint
as a penalty term with a coefficient γ sufficiently large.
Finally, we obtain the unified optimization problem of DNA
framework, which interplays two deep neural architectures for
dynamic social network alignment, as follows:
J = J s + J t + γ(cid:13)(cid:13)PsVs − PtVt(cid:13)(cid:13)2
(9)
F ,
min
V(·),Q(·),Θ(·)
where Θ(·) denotes parameter matrices of the proposed neural
architecture. The users whose identity embeddings are posi-
tioned closely in the common subspace are regarded as good
candidates for the social network alignment.
IV. DNA: OPTIMIZATION
The optimization problem of DNA in Eq. (9) is not jointly
convex over matrix variables Q, V and Θ for both dynamic
social graphs. Thus, it is infeasible to obtain the closed-form
solution of the global optimum. To bridge this gap, we design
an effective alternating algorithm to approach an optimum. The
core idea is to approach a (local) optimal solution w.r.t. one
matrix variable while fixing others. Θ is readily to be solved
via gradient methods. Obviously, the key challenge lies in the
updating rules for Q and V. The derivation of updating rule of
the variables associated with Gt is the same as that of Gs and
thus omitted due to the limit of space. The overall procedure
of the DNA framework is summarized in Algorithm 1.
Algorithm 1: DNA Framework
Input: Source dynamic social graph tensor: Gs
Target dynamic social graph tensor: Gt
Anchor set: A
Output: Candidate lists for alignment between Gs and Gt
1 Generate ego tensor X s and X t for dynamic social graph
tensor Gs and Gt, respectively;
ego tensor via J s
2 Pretrain the proposed deep neural architecture for each
d and J t
d ;
3 Perform forward propagation to obtain Us(0) and Ut(0);
4 Initialize Vs(0) = Us(0) and Vt(0) = Ut(0);
5 while not converging do
6
7
Back-propagation to update Θs (or Θt);
for each dynamic social graph do
8
9
10
11
for each dynamic social graph do
(cid:16)
VsT Vs(cid:17)−1
Forward propagation to obtain Us (or Ut);
Update Qs =
rule of Qt is of the same structure);
Update Vs (or Vt) via Eqs. (14) and (15);
VsT Us (the updating
12 Generate candidate lists for alignment via computing the
distance between the rows in Vs and Vt.
where ⊗ denotes the Kronecker product. It is block diagonal,
[Hs
Q][k][k] = VsT Vs,
(13)
where Ik is a k-dimensional identity matrix and [Hs
Q][k][k]
is the (k, k)-th block in matrix Hs
Q. Obviously, the block
Q is always positive semidefinite as VsT Vs is
diagonal Hs
positive semidefinite. Thus, J is convex w.r.t. Qs. Therefore,
the global optimum is achieved by setting ∇s
Q = 0. Finally,
we obtain the closed-form updating rule for Qs in Eq. (10) as
the inverse matrix always exists for any positive semidefinite
matrix VsT Vs.
minJ (Qs), we give the closed-form
A. Updating Q
To solve Qs = arg
Qs
solution as follows:
Qs =
(cid:16)
VsT Vs(cid:17)−1
VsT Us
(10)
Theorem (Optimal Solution). Fixing other matrix variables,
the updating rule in the proposed DNA for Qs and Qt gives
the global optimal solution to minimize overall objective J
w.r.t. Qs and Qt, receptively.
Proof. First, we derive the matrix of partial differentials of J
w.r.t. Qs as follows:
∇s
Q = −2VsT Us + 2VsT VsQs
Then, we study the Hessian matrix Hs
Q = Ik ⊗ VsT Vs,
Hs
Q of J w.r.t. Qs:
(11)
(12)
B. Updating V
To solve Vs = arg
Vs
updating rule as follows:
(cid:115)
minJ (Vs), we give the multiplicative
Vs = Vs (cid:12)
where
β(Ψs + VsΓs) + γΛs
β(Υs + VsΦs) + γΠs ,
(14)
Ψs = (UsQsT )+ Υs = (UsQsT )−
Γs = (QsQsT )−
Φs = (QsQsT )+
Λs = PsT PtVt
Πs = PsT PsVs
√· are
Note that, (cid:12) is the Hadamard product. Both (·)
pairwise operators. Nonnegative operators (·)+ and (·)− are
defined as follows:
(·) and
(15)
[X+]ij =
[X]ij+[X]ij
2
[X−]ij =
[X]ij−[X]ij
2
,
(16)
TABLE II: Statistics of the datasets
Dataset
Twitter
Foursquare
#(Nodes)
5, 167
5, 240
#(Links)
164, 660
76, 972
#(Anchor Users)
2, 858
where · is the absolute value of a scalar. Hence, X+ and X−
are absolute value matrices of positive elements and negative
elements of an arbitrary matrix X, respectively.
Next, we prove the correctness of the multiplicative updat-
ing rule in Eqs. (14) and (15), and give the theorem below:
Theorem (Correctness). The limit point of multiplicative
updating rules of DNA for Vs (as well as Vt) in Eqs. (14)
and (15) satisfies the Karush-Kuhn-Tucker (KKT) condition.
Proof. We study the objective function J in Eq. (9) w.r.t. Vs,
referred to as J (Vs). Utilize the fact X2
F = tr(XT X) for
any real matrix X ∈ Rn×m. After some algebraic operations
and omitting constants, we obtain:
J (Vs) = − 2βtr((VsQs)T Us) + βtr((VsQs)T VsQs)
− 2γtr((PtVt)T PsVs) + γtr((PsVs)T PsVs).
(17)
We introduce the matrix Ω of the Lagrangian multipliers
to enforce the nonnegative constraint on Vs, leading to the
Lagrangian function of J (Vs),
L(Vs) = J (Vs) + tr(ΩVs).
(18)
Following the KKT complementary slackness condition, we
obtain the fixed point equation as follows,
where ∇s
[∇s
V]ij[Vs]ij = [Ω]ij[Vs]ij = 0,
V is the matrix of partial differentials:
∇s
V = − 2βUsQsT + 2βVsQsQsT
− 2γPtT
PtVt + 2γPsT PsVs.
(19)
(20)
The solution must satisfy Eq. (19) in the limit. Given the
multiplicative updating rule in Eqs. (14) and (15), we have,
in the limit, Vs(∞) = Vs(τ +1) = Vs(τ ) = Vs, where the
superscript (τ ) indicates the τ-th iteration. Utilizing the fact
X = X+ + X−, we obtain that:
V]ij[Vs]2
ij = 0.
[∇s
(21)
The right hand side in both Eq. (21) and Eq. (19) is 0. The left
hand side in both Eq. (21) and Eq. (19) is a multiplication of
two factors and the first factor in both equations is identical.
As [Vs]2
ij = 0 if and only if [Vs]ij = 0, we can claim that
Eq. (21) holds if and only if Eq. (19) holds. Thus, we reach
the Theorem (Correctness).
V. EXPERIMENT
A. Experimental Setup
Datasets: We enrich a benchmark dataset [4], [1], [11], [6],
[9], [7], Twitter-Foursquare (TF), by collecting time informa-
tion of friending. We restore the real-world dynamic social
(a) Precision@K under different K (b) Precision@3 under different λ
(c) MAP@K under different K
Fig. 3: Experimental results on Twitter-Foursquare dataset
(d) MAP@3 under different λ
network by framing a sequence of network snapshots via the
collected timestamps. The statistics are listed in Table II.
Comparison Methods: To evaluate the performance of the
proposed DNA framework, we chose several state-of-the-art
methods for comparison, which are introduced as follows:
• PALE [5]: It
leverages the structure information and
performs an embedding-matching framework to align
users across social networks.
• IONE [4]: It utilizes a unified optimization framework for
aligning social networks subject to hard or soft constraint
on the anchor users via an embedding approach.
• MASTER [11]: It constructs a common subspace of
multiple social networks via a semi-supervised optimiza-
tion framework. Note that, it only considers structure
information in this paper for fair comparison.
• DeepLink [6]: It introduces a neural framework, which
leverages dual learning to facilitate common subspace
construction to perform alignment.
Parameter Settings: In the proposed DNA framework, we set
ω to 3 for ω−step RWR. We utilize the l2 norm to regularize
the neural architecture parameters for alleviating overfitting
and employ the dropout trick in the training, whose empirical
value is set to 0.8.
Evaluation Metric: We evaluated all the comparison methods
in terms of the following metrics:
• Precision@K is defined as
(cid:80)NA
1i{success@K},
where 1i{success@K} = 1 if and only if
the
groundtruth exists in the candidate list of length K,
otherwise 1i{success@K} = 0. NA is the number of
groundtruth anchor users.
1
NA
, where Ranki
is the rank of the candidate hitting the ground truth in the
candidate list of length K. Specially, we set
= 0
for not hitting. NA is defined in Precision@K. Note that,
MAP highlights the rank of the hitting candidate in a
nonlinear way.
(cid:80)NA
• MAP@K is defined as
1
NA
Ranki
i=1
Ranki
i=1
1
1
Higher value of the metrics signifies better performance.
K13579111315Precision@K (%)203040506070PALEIONEMASTERDeepLinkDNAλ1020304050Precision@3 (%)203040506070PALEIONEMASTERDeepLinkDNAK13579111315MAP@K (%)1020304050PALEIONEMASTERDeepLinkDNAλ1020304050MAP@3 (%)1020304050PALEIONEMASTERDeepLinkDNA (a) Precision@3
(b) Precision@3
(a) Precision@3
(b) MAP@3
(c) MAP@3
(d) MAP@3
Fig. 4: Parameter sensitivity on Twitter-Foursquare dataset
B. Experimental Results
We first focus on verifying the importance of dynamics
and its reason behind, and then discuss the supervision for
DNA, where all the embedding dimensions are set to 128 as
default. Finally, we conduct sensitivity analysis on embedding
dimension. We repeated each experiment 10 times and report
the mean with 95% confidence interval.
The dynamics is important! We compare the proposed
DNA framework with static alignment methods to demonstrate
the importance of dynamics. To achieve this goal, we evaluate
DNA and its competitors under various settings. Specifically,
on both datasets, we first vary the length K of the candidate
list from 1 to 15 on 50%-overlap datasets, reported in Fig.
3 (a) and (c). We then vary the overlap rate λ from 10 to
50 while fixing K = 3, reported in Fig. 3 (d) and (d). The
overlap rate λ is measured by 2A
S+T , where A, S and T are
the number of anchor users, source network users and target
network users, respectively. A λ-overlap dataset is generated
by randomly deleting users. The performance in terms of both
precision and MAP are evaluated.
Obviously, DNA consistently outperforms all comparison
methods. For instance, DNA achieves 1.96× precision and
1.94× MAP of PALE on the 50%-overlap benchmark (TF)
dataset. The reasons are two-fold: (1) The DNA framework
captures the complex dynamics in the real-world social net-
work evolvement, which is discriminative for social network
alignment. However,
the other methods cannot effectively
identify the alignment due to limited user depiction, which
inevitably loses the fruitful information residing in network
dynamics. (2) DNA constructs a common subspace modeling
the underlying identities, and the users are thereby naturally
aligned in the common subspace. However, other methods,
such as IONE [4], PALE [5] and DeepLink [6], are some-
how powerless to reveal underlying identities of individuals
for alignment. To sum up, experimental results verify that
dynamics significantly facilitates social network alignment.
On the Snapshotting: The performance of DNA leverages
5 snapshots with the interval of 3 months in Fig. 3. Now,
is reasonable indeed. In general,
Fig. 5: Performance under different embedding dimensions
we zoom in the dynamics of social networks by varying
the frequency and number of snapshots. A natural question
arises that: the more the snapshots, the better the performance?
We evaluate the performance of DNA under various settings
of snapshotting, reported in Fig. 4 (a) and (c), where the
confidence interval is omitted for clarity. Both Precision@3
and MAP@3 of DNA increases as frequency and number of
snapshots increase in general. However, the performance tends
to be saturated when the information collected is adequate.
It
the higher frequency
and higher number of snapshots better restore the network
dynamics, i.e., friending behaviors in this context. It is of
high discriminative ability for alignment as it is dominantly
correlated to the patterns of individuals acting across social
networks as evidenced in the study [18]. For instance, if an
individual more often than not expands or shrinks his or her
friend list once a few weeks, it will reveal this pattern in a
few months. Thus, we need a time window long enough and a
snapshotting frequency high enough to capture the pattern of
friending or collaboration behavior. Moreover, we observed
that outdated information may not further help analysis as
stated in the study [18] as well. These results in fact verify the
motivation of our work and in turn shed light on the dynamics
of behavior pattern for the further study of social psychology.
On the Supervision Information: We analyze the impact
of training rate η by varying η from 1% to 5%. As shown in
Fig. 4 (b) and (d), the performance of DNA rises dramatically
as η increases from 1% to 3%, and then saturates when η
exceeds 3%. On both TF and AN datasets with 50% overlap
rate, DNA achieves better performance in terms of precision
and MAP than its competitors, leveraging only a few anchor
users. It is expected and the reason lies in two folds: (1)
In the DNA framework, the user dynamics are encoded in
a self-supervised way,
i.e., we can obtain the embedding
space without supervision. (2) The common subspace will be
constructed by aligning the embedding spaces on only a few
anchor users, which is one of the inherent merits of DNA. To
summarize, the experimental results on different training rates
verify the learning ability of the proposed DNA framework.
On the Embedding Dimension: We further discuss the
impact of the embedding dimension. The higher the em-
bedding dimension,
is not
always the case. We vary the dimension from 10 to 300 and
report the precision and MAP on both TF and AN datasets
in Fig. 5. As shown in these subfigures, both Precision@3
and MAP@3 of all these methods are undesirable when the
the better the performance? It
13011090Interval (day)70503010123Number456762595653Precision@3 (%)5354555657585960616263η12345Precision@3 (%)010203040506070PALEIONEMASTERDeepLinkDNA13011090Interval (day)70503010123Number456748454239MAP@3 (%)39404142434445464748η12345MAP@3 (%)01020304050PALEIONEMASTERDeepLinkDNADimension050100150200250300Precision@3 (%)010203040506070PALEIONEMASTERDeepLinkDNADimension050100150200250300MAP@3 (%)01020304050PALEIONEMASTERDeepLinkDNA dimension is low, and tends to be better as the dimension
increases. The reason lies in that when user embeddings with
low dimensionality tend to entangle densely, thus resulting in
the mislead of alignment. However, we will not obtain further
performance gain when embedding dimension is large enough
to layout user embeddings in the common embedding space.
Moreover, the DNA framework consistently outperforms all
comparison methods as the generated embeddings encode the
discriminative pattern residing in the dynamic social networks,
which significantly facilitates social network alignment.
VI. RELATED WORK
We summarize the related works in the following two fields:
A. Social Network Alignment
Social network alignment, a.k.a. anchor link prediction [23],
is to align different social networks on their common users. To
our knowledge, this problem is first proposed and addressed
by the study [24]. Some studies [2], [1], [25], [19], [26]
incorporate heterogeneous information to discover the latent
consistency of user identity, while others [3], [9], [4], [5], [6],
[7], [27] exploit the structure information to perform social
network alignment. Moreover, some studies [8], [10], [28],
[11], [12], [13], [14], [15] investigate to exploit both structure
and heterogeneous attribute information. However, all existing
studies consider the social network to be static for simplicity
and neglect its inherent dynamics. The essential distinction
between our work and all these studies above lies in that we
for the first time pose the problem of aligning dynamic social
networks and design the DNA framework to unfold the fruitful
dynamics for addressing this problem. In contrast to social
network alignment, biochemical network alignment [29], [30],
[31] aims to find similar conserved regions between networks
instead of exploring underlaying identities of nodes and thus
distinguishes itself with our work.
B. Network Embedding
Network embedding is to obtain a low-dimensional compact
vector representation for each node in a network. Inspired by
the success of word2vec, DeepWalk [32] conducts random
walks on the network and feeds generated node sequences
into Skip-gram. Following DeepWalk, node2vec [33] proposes
a novel random walk strategy to facilitate network embedding.
LINE [34] gives an explicit objective function of the first- and
second-order proximity for network embedding. Neural net-
works on graph, such as GCN [35], also obtain the embeddings
for each node. Some studies explore the structural regularity
regarding the regular equivalence [36], the embedding distribu-
tion [37] and the computational challenges [38], while others
attempt to incorporate the heterogeneous information [39],
[40]. Recently, a few works investigate the problem of dynamic
network embedding. However, they focus on positioning the
new arrivals [41], [42], facilitating the embedding in one
network [43], [44] or tailoring the embedding for other tasks,
e.g., anomaly detection [45]. The surveys [46], [47] give the
comprehensive summary. Distinguishing with these studies,
we focus on user embedding with complex dynamics for
dynamic social network alignment.
VII. CONCLUSION
To the best of our knowledge, this is the first study on the
problem of aligning dynamic social networks. We propose to
reveal the evolving pattern of the individuals in the dynamic
social network to facilitate aligning, inspired by the social
psychology. The dynamics in the social network contains
the discriminative pattern to facilitate alignment. However,
revealing such pattern faces significant challenges in both
modeling and optimization. Towards this end, we propose
a novel DNA framework, a unified optimization approach
over deep neural architectures, unfolding the fruitful dynamics
for alignment. In the DNA framework, to capture the intra-
network dynamics, we propose a novel neural architecture
in account of user dynamics. To address the inter-network
alignment, we design a unified optimization interplaying the
embedding spaces to construct
the common subspace. To
address this optimization problem, we design an effective alter-
nating algorithm with solid theoretical guarantees. We conduct
extensive experiments on real-world datasets and experimental
results demonstrate that DNA substantially outperforms the
state-of-the-art methods.
VIII. ACKNOWLEDGEMENT
This work was supported in part by: National Key Re-
search and Development Program of China under Grant
2018YFB1003804, National Natural Science Foundation un-
der Grant 61602050 and 61921003, Fundamental Research
Funds for the Central Universities 2019XD11 and NSF under
grants III-1526499, III-1763325, III-1909323, CNS-1930941,
and CNS-1626432.
REFERENCES
[1] X. Kong, J. Zhang, and P. S. Yu, "Inferring anchor links across multiple
heterogeneous social networks," in CIKM, 2013, pp. 179 -- 188.
[2] R. Zafarani and H. Liu, "Connecting users across social media sites: a
behavioral-modeling approach," in KDD, 2013, pp. 41 -- 49.
[3] S. Tan, Z. Guan, D. Cai, X. Qin, J. Bu, and C. Chen, "Mapping users
across networks by manifold alignment on hypergraph," in AAAI, 2014,
pp. 159 -- 165.
[4] L. Liu, W. K. Cheung, X. Li, and L. Liao, "Aligning users across social
networks using network embedding," in IJCAI, 2016, pp. 1774 -- 1780.
[5] T. Man, H. Shen, S. Liu, X. Jin, and X. Cheng, "Predict anchor links
across social networks via an embedding approach," in IJCAI, 2016, pp.
1823 -- 1829.
[6] F. Zhou, L. Liu, K. Zhang, G. Trajcevski, J. Wu, and T. Zhong,
"Deeplink: A deep learning approach for user identity linkage," in
INFOCOM, 2018, pp. 1313 -- 1321.
[7] C. Li, S. Wang, P. S. Yu, L. Zheng, X. Zhang, Z. Li, and Y. Liang, "Dis-
tribution distance minimization for unsupervised user identity linkage,"
in CIKM, 2018, pp. 447 -- 456.
[8] Y. Zhang, J. Tang, Z. Yang, J. Pei, and P. S. Yu, "Cosnet: Connecting
heterogeneous social networks with local and global consistency," in
KDD, 2015, pp. 1485 -- 1494.
[9] J. Zhang and P. S. Yu, "Multiple anonymized social networks alignment,"
in ICDM, 2015, pp. 855 -- 864.
[10] X. Cao and Y. Yu, "Bass: A bootstrapping approach for aligning
heterogenous social networks," in ECML-PKDD, 2016, pp. 459 -- 475.
[11] S. Su, L. Sun, Z. Zhang, G. Li, and J. Qu, "Master: across multiple social
networks, integrate attribute and structure embedding for reconciliation,"
in IJCAI, 2018, pp. 3863 -- 3869.
[43] L. Zhou, Y. Yang, X. Ren, F. Wu, and Y. Zhuang, "Dynamic network
embedding by modeling triadic closure process," in AAAI, 2018, pp.
571 -- 578.
[44] Y. Zuo, G. Liu, H. Lin, J. Guo, X. Hu, and J. Wu, "Embedding temporal
network via neighborhood formation," in KDD, 2018, pp. 2857 -- 2866.
[45] W. Yu, W. Cheng, C. C. Aggarwal, K. Zhang, H. Chen, and W. Wang,
"Netwalk: A flexible deep embedding approach for anomaly detection
in dynamic networks," in KDD, 2018, pp. 2672 -- 2681.
[46] P. Cui, X. Wang, J. Pei, and W. Zhu, "A survey on network embedding,"
IEEE Trans. on Knowl. and Data Eng., vol. 31, pp. 833 -- 852, 2018.
[47] D. Zhang, J. Yin, X. Zhu, and C. Zhang, "Network representation
learning: A survey," IEEE Trans. on Big Data, pp. 1 -- 25, 2018.
[12] Z. Zhong, Y. Cao, M. Guo, and Z. Nie, "Colink: An unsupervised
framework for user identity linkage," in AAAI, 2018, pp. 5714 -- 5721.
[13] J. Zhang, B. Chen, X. Wang, H. Chen, C. Li, F. Jin, G. Song, and
Y. Zhang, "Mego2vec: Embedding matched ego networks for user
alignment across social networks," in CIKM, 2018, pp. 327 -- 336.
[14] Y. Wang, C. Feng, L. Chen, H. Yin, C. Guo, and Y. Chu, "User
identity linkage across social networks via linked heterogeneous network
embedding," WWW Journal, pp. 1 -- 22, 2018.
[15] W. Xie, X. Mu, R. K.-W. Lee, F. Zhu, and E.-P. Lim, "Unsupervised user
identity linkage via factoid embedding," in ICDM, 2018, pp. 1338 -- 1343.
[16] J. Zhou and J. Fan, "Tanslink: User identity linkage across heterogeneous
social networks via translating embeddings," in INFOCOM, 2019.
[17] Y. Wang, H. Shen, J. Gao, and X. Cheng, "Learning binary hash codes
for fast anchor link retrieval across networks," in The Web Conf. (WWW).
ACM, 2019, pp. 3335 -- 3341.
[18] S. Utz, M. Tanis, and I. Vermeulen, "It is all about being popular: the
effects of need for popularity on social network site use," Cyberpsychol-
ogy Behavior & Social Networking, vol. 15, no. 1, pp. 37 -- 42, 2012.
[19] X. Mu, F. Zhu, E.-P. Lim, J. Xiao, J. Wang, and Z.-H. Zhou, "User
identity linkage by latent user space modelling," in KDD, 2016, pp.
1775 -- 1784.
[20] A. Graves and N. Jaitly, "Towards end-to-end speech recognition with
recurrent neural networks," in ICML, 2014, pp. 1764 -- 1772.
[21] H. Tong, C. Faloutsos, and J. Pan, "Fast random walk with restart and
its applications," in ICDM, 2006, pp. 613 -- 622.
[22] D. Wang, P. Cui, and W. Zhu, "Structural deep network embedding," in
KDD, 2016, pp. 1225 -- 1234.
[23] K. Shu, S. Wang, J. Tang, R. Zafarani, and H. Liu, "User identity linkage
across online social networks: A review," KDD Exp, pp. 5 -- 17, 2017.
[24] R. Zafarani and H. Liu, "Connecting corresponding identities across
communities," in ICWSM, vol. 9, 2009, pp. 354 -- 357.
[25] S. Liu, S. Wang, F. Zhu, J. Zhang, and R. Krishnan, "Hydra: Large-
scale social identity linkage via heterogeneous behavior modeling," in
SIGMOD, 2014, pp. 51 -- 62.
[26] Y. Ren, C. C. Aggarwal, and J. Zhang, "Meta diagram based active
social networks alignment," in ICDE, 2019.
[27] C. Li, Y. Wang, S. Wang, Y. Liu, P. S. Yu, Z. Li, and Y. Liang,
"Adversarial learning for weakly-supervised social network alignment,"
in AAAI, 2019.
[28] S. Zhang and H. Tong, "Final: Fast attributed network alignment," in
KDD, 2016, pp. 1345 -- 1354.
[29] J. Flannick, A. Novak, C. B. Do, B. S. Srinivasan, and S. Batzoglou,
"Automatic parameter learning for multiple local network alignment,"
Journal of computational biology, vol. 16, no. 8, pp. 1001 -- 1022, 2009.
[30] V. Vijayan, D. Critchlow, and T. Milenkovic, "Alignment of dynamic
networks," Bioinformatics, vol. 33, no. 14, pp. 180 -- 189, 2017.
[31] D. Aparicio, P. Ribeiro, T. Milenkovi, and F. Silva, "Temporal network
alignment via got-wave," Bioinformatics, pp. 1367 -- 4803, 2019.
[32] B. Perozzi, R. Al-Rfou, and S. Skiena, "Deepwalk: Online learning of
social representations," in KDD, 2014, pp. 701 -- 710.
[33] A. Grover and J. Leskovec, "node2vec: Scalable feature learning for
networks," in KDD, 2016, pp. 855 -- 864.
[34] J. Tang, M. Qu, M. Wang, M. Zhang, J. Yan, and Q. Mei, "Line: Large-
scale information network embedding," in WWW, 2015, pp. 1067 -- 1077.
[35] T. N. Kipf and M. Welling, "Semi-supervised classification with graph
convolutional networks," in ICLR, 2017.
[36] K. Tu, P. Cui, X. Wang, P. S. Yu, and W. Zhu, "Deep recursive network
embedding with regular equivalence," in KDD, 2018, pp. 2357 -- 2366.
[37] S. Pan, R. Hu, G. Long, J. Jiang, L. Yao, and C. Zhang, "Adversarially
regularized graph autoencoder for graph embedding," in IJCAI, 2018,
pp. 2609 -- 2615.
[38] D. Lian, K. Zheng, V. W. Zheng, Y. Ge, L. Cao, I. W. Tsang, and
X. Xie, "High-order proximity preserving information network hashing,"
in KDD, 2018, pp. 1744 -- 1753.
[39] C. Yang, Z. Liu, D. Zhao, M. Sun, and E. Chang, "Network representa-
tion learning with rich text information," in IJCAI, 2015, pp. 2111 -- 2117.
informed attributed network
[40] X. Huang, J. Li, and X. Hu, "Label
embedding," in WSDM, 2017, pp. 731 -- 739.
[41] L. Du, Y. Wang, G. Song, Z. Lu, and J. Wang, "Dynamic network
embedding: An extended approach for skip-gram based network em-
bedding," in IJCAI, 2018, pp. 2086 -- 2092.
[42] Z. Zhang, P. Cui, J. Pei, X. Wang, and W. Zhu, "Timers: Error-bounded
svd restart on dynamic networks," in AAAI, 2018, pp. 224 -- 231.
|
1704.05146 | 1 | 1704 | 2017-04-17T23:18:17 | On Predicting Geolocation of Tweets using Convolutional Neural Networks | [
"cs.SI"
] | In many Twitter studies, it is important to know where a tweet came from in order to use the tweet content to study regional user behavior. However, researchers using Twitter to understand user behavior often lack sufficient geo-tagged data. Given the huge volume of Twitter data there is a need for accurate automated geolocating solutions. Herein, we present a new method to predict a Twitter user's location based on the information in a single tweet. We integrate text and user profile meta-data into a single model using a convolutional neural network. Our experiments demonstrate that our neural model substantially outperforms baseline methods, achieving 52.8% accuracy and 92.1% accuracy on city-level and country-level prediction respectively. | cs.SI | cs |
On Predicting Geolocation of Tweets using
Convolutional Neural Network
Binxuan Huang and Kathleen M. Carley
School of Computer Science, Carnegie Mellon University,
5000 Forbe Ave., Pittsburgh, United States
{binxuanh,kathleen.carley}@cs.cmu.edu
Abstract. In many Twitter studies, it is important to know where a
tweet came from in order to use the tweet content to study regional user
behavior. However, researchers using Twitter to understand user behav-
ior often lack sufficient geo-tagged data. Given the huge volume of Twit-
ter data there is a need for accurate automated geolocating solutions.
Herein, we present a new method to predict a Twitter user's location
based on the information in a single tweet. We integrate text and user
profile meta-data into a single model using a convolutional neural net-
work. Our experiments demonstrate that our neural model substantially
outperforms baseline methods, achieving 52.8% accuracy and 92.1% ac-
curacy on city-level and country-level prediction respectively.
1
Introduction
Recently, there is growing interest in using social media to understand social phe-
nomena. For example, researchers have shown that analyzing social media reveals
important geospatial patterns for keywords related to presidential elections[26].
People can use Twitter as a sensor to detect earthquakes in real-time[24]. Recent
research also has demonstrated that Twitter data provides real-time assessments
of flu activity[1].
Using Twitter's API1, a keyword search can be done and we can easily
get tweet streams from across the world containing keywords of interest. How-
ever, we cannot conduct a fine-grained analysis in a specific region using such
a keyword-based search method. Alternatively, using the same API tweets with
geo-information can be collected via a bounding box. Since less than 1% of tweets
are tagged with geo-coordinates[8], using this location-based search means we
will lose the majority of the data. If we can correctly locate those ungeotagged
tweets returned from a keyword search stream, that would enable us to study
users in a specific region with far more information.
With this motivation, we are aiming to study the problem of inferring a
tweet's location. Specifically, we are trying to predict on a tweet by tweet ba-
sis, which country and which city it comes from. Most of the previous studies
rely on rich user information(tweeting history and/or social ties), which is time-
consuming to collect because of the Twitter API's speed limit. Thus those meth-
ods could not be directly applied to Twitter streams. In this paper, we study a
1 https://dev.twitter.com/docs
global location prediction system working on each single tweet. One data sam-
ple is one tweet JSON object returned by Twitter's streaming API. Our system
utilizes location-related features in a tweet, such as text and user profile meta-
data. We summarize useful features that can provide information for location
prediction in Table 1.
Table 1. Feature table
Type
Feature
Tweet content
Free text
User personal description Free text
Free text
User name
Free text
User profile location
Categorical
Tweet language(TL)
Categorical
User language(UL)
Timezone(TZ)
Categorical
UTC timestamp
Posting time(PT)
Recent research has shown that using bag-of-words and classical machine
learning algorithms such as Naive Bayes can provide us a text-based location
classifier with good accuracy[10]. Different from previous research, we intend to
use the convolutional neural network(CNN) to boost prediction power. Inspired
by the success of convolutional neural network in text classification[13], we are
going to use CNN to extract location related features from texts and train a
classifier that combines high-level text feature representations with these cat-
egorical features. To benchmark our method, we compared our approach with
a stacking-based method. Experimental results demonstrate that our approach
achieves 92.1% accuracy on country-level prediction and 52.8% accuracy on city-
level prediction, which greatly outperforms our baseline methods on both tasks.
2 Related work
Identifying demographic details of Twitter users[17] has been widely studied in
previous literature including inferring users' attributes like age, gender[6], po-
litical affiliation[18] and personalities like openness and conscientiousness[20].
Among these research, there is increasing interest in inferring Twitter user's lo-
cation, which is largely driven by the lack of sufficient geo-tagged data[8]. In
many situations, it is important to know where a tweet came from in order to
use the information in the tweet to effect a good social outcome. Key exam-
ples include: disaster relief[15], disaster management[4], earthquake detection[7],
predicting elections[25], and predicting flu trends[1].
A majority of previous works either focus on a local region e.g. United
States[5], Sweden[2], or using rich user information like a certain number of
tweets for each user[5], user's social relationship[12,19,27]. Different from these
works, this paper works on worldwide tweet location prediction. We only utilized
features in one single tweet without any external information. Thus this method
could be easily applied to real-time Twitter stream.
For fine-grained location prediction, there are several types of location rep-
resentation methods existing in literature. One typical method is to divide earth
into small grids and try to predict which cell one tweet comes from. Wing
and Baldridge introduced a grid-based representation with fixed latitude and
longitude[28]. Based on the similarity measured by Kullback-Leibler divergence,
they assign each ungeotagged tweet to the cell with most similar tweets. Because
cells in urban area tend to contain far more tweets than the ones in rural areas,
the target classes are rather imbalanced[9]. To overcome this, Roller et al. further
proposed an adaptive grid representation using K-D tree partition[22]. Another
type of representation is topic region. Hong et al. proposed a topic model to
discover the latent topic words for different regions[11]. Such parametric genera-
tive model requires a fixed number of regions. However, the granularity of topic
regions is hard to control and will potentially vary over time[10].
The representation we choose is city-based representation considering most
tweets come from urban area. One early work proposed by Cheng et al. using a
probabilistic framework to estimate Twitter user's city-level location based on
the content of tweets[5]. Their framework tries to identify local words with prob-
ability distribution smoothing. However, such method needs a certain number
of tweets(100) for each user to get a good estimation. Han et al. proposed a
stacking-based approach to predict user's city[9]. They combine tweet text and
meta-data in user profile with stacking[29]. Specifically, they train a multino-
mial naive Bayes base classifier on tweet text, profile location, timezone. Then
they train a meta-classifier over the base classifiers. More recently, Han et al.
further did extensive experiments to show that using feature selection method,
such as information gain ratio[21] and χ2 statistic could greatly improve the
classification performance.
3 Location prediction
In this section, we will introduce our location prediction approach. We first
briefly describe the useful features in a tweet JSON object. After that, we will
further explain how we utilize these features in our prediction model.
3.1 Feature Set
We have listed all useful information we want to utilize in Table 1. Tweet content,
user personal description, user name and profile location are four text fields
that we will use. Twitter users often reveal their home location in their profile
location and personal description. However, location indicating words are often
mixed with informal tweet text(e.g. chitown for Chicago). It is unrealistic to use
a gazetteer to find these words. In this work, we choose to apply CNN on these
four text fields to extract high-level representations.
In addition to these four text fields. there are another three categorical fea-
tures: tweet language, user language,and timezone. Tweet language is automati-
cally determined by Twitter's language detection tool. User language and time-
zone are selected by the user in his/her profile. These three categorical features
are particularly useful for distinguishing users at the country-level.
The last feature is UTC posting time. Using posting timestamp as a discrim-
inative feature is motivated by the fact that people in a region are more active
on Twitter at certain times during the day. For example, while people in United
Kingdom start to be active at 9:00 am in UTC time, most of the people in United
States are still asleep. We transform the posting time in UTC timestamp into
discrete time slots. Specifically, we divide 24 hours into 144 time slots each with
a length of 10 minutes. Thus each tweet will have a discrete time slot number in
the range of 144, which can be viewed as a categorical feature. In Figure 1, we
plotted the probability distribution of an user posting tweets in each time slot
in three different countries. As expected, there is a big variance between these
three countries.
Fig. 1. The probability of an user posting a tweet in different time slot in three different
countries: United States, United Kingdom, Japan.
3.2 Our Approach
Our approach is based on the convolutional neural network for sentence classi-
fication proposed by Kim[13]. Different from traditional bag-of-words method,
such convolutional neural networks take the word order into consideration. Our
model architecture is shown in Figure 2. We use this CNN architecture to ex-
i ∈ Rk be the
tract high-level features from four text fields in a tweet. Let xt
k-dimensional word vector corresponding to the i-th word in the text t, where
t ∈ {tweet content, user description, profile location, user name}. As a result,
one text of length n can be represented as a matrix
2 ⊕ .... ⊕ xt
(1)
where ⊕ is concatenation operator. In the convolutional layer, we apply each
filter w ∈ Rhk to all the word vector matrices, where h is the window size and
k is the length of a word vector. For example, applying filter w to a window of
i:i+h−1 + b). Here b ∈ R is a bias
word vectors xt
term and we choose f (x) as a non-linear ReLU function max(x, 0). Sliding the
filter window from the beginning of a word matrix till the end, we generated a
feature vector ct = [ct
n−h+1] for each text t. If we have m filters in the
convolutional layer, then we can produce m feature vectors for each text field
and 4m vectors in total.
i:i+h−1, we generated ct
i = f (w· xt
1, ct
2, ..., ct
X t
1:n = xt
1 ⊕ xt
n
Fig. 2. A diagram of the architecture of our neural model.
In the max-pooling layer, we apply a pooling operation over each feature
vector generated in the convolutional layer. Each pooling operation takes a fea-
ture vector as input and outputs the maximum value ct = max(ct). ct can be
viewed as the most representative feature generated by a filter on text t. Hence
we finally got a long vector θ ∈ R4m after the max-pooling layer. To avoid the
co-adaptation of hidden units, we apply dropout on the max-pooling layer that
randomly set elements in θ to zero in the training phase. After that, we append
four categorical features tweet language(TL), user language(UL), timezone(TZ)
and posting time(PT) with one-hot encoding at the end of θ and get θ. In the
last fully connected layer, we use a softmax function over this long vector θ to
generate the probability distribution over locations. Specifically, the probability
of one tweet coming from location li is
P (liθ) =
θ)
(2)
(cid:80)L
θ)
exp(βT
i
j=1 exp(βT
j
where L is the number of locations and βi are parameters in softmax layer. The
output predicted location is just the location with highest probability.
The minimization objective in the training phase is the categorical cross-
entropy loss. The parameters to be estimated include word vectors, weight vec-
tors w for each filters, the weight vectors β in softmax layer, and all the bias
terms. The optimization is performed using mini-batch stochastic gradient de-
scent and back-propagation[23].
4 Data
We used geo-tagged tweets collected from Twitter streaming API2 for training
and evaluation. In this study, we set the geographic bounding box as [-180, -90,
180, 90] so that we could get these geo-tagged tweets from the whole world.
Our collection started from January 7, 2017 to February 1, 2017. Because it is
very common for one user to post tweets from the same city, we randomly chose
one tweet for each city that one user has visited. This could ensure that there
is no strong overlap among our data samples. We only used tweets either with
specific geo-coordinates or a geo-bounding box smaller than [0.1,0.1]. For the
latter case, we used the center of one tweet's bounding box as its coordinates.
No other filtering was done. There are 3,321,194 users and 4,645,692 tweets in
total. For test data we used all tweets from 10% of the users who were randomly
selected. For the remaining 90% users, we picked tweets from 50,000 of them as
a development set and used the remaining tweets as training data.
There are two location prediction tasks we consider in this paper. The first
task is country-level location prediction. We adopted the country code in the
geo-tagged tweet as the label we want to predict. In our dataset, there are 243
countries and regions in total. The second task is city-level location classification.
We adopt the same city-based representation as Han et al.[3]. The city-based rep-
resentation consists of 3,709 cities throughout the world and was obtained by
aggregating smaller cities with the largest nearby city. We assigned the clos-
est city for each tweet based on orthodromic distance. Table 2 contains basic
statistics about our dataset. It is worth mentioning that this dataset is rather
imbalanced, where a majority of tweets are sent from a few countries/cities.
Table 2. Summaries about the dataset. Numbers in brackets are standard deviation.
# of
tweets
4645692 3321194
# of
users
# of
timezones
417
# of
lang.
103
# of countries
(or regions)
Tweets per
country
# of
cities
Tweets per
city
243
19118.0 (99697.1) 3709 1252.5(4184.5)
5 Experiments
5.1 Evaluation Measures
Following previous works of tweet geolocation prediction[9], we used four evalu-
ation measures listed below. One thing to note is that when we calculated the
error distance we used distance between predicted city and the true coordinates
in the tweet rather than the center of assigned closest city.
2 https://dev.twitter.com/streaming/reference/post/statuses/filter
– Acc: The percentage of correct location predictions.
– Acc@Top5: The percentage of true location in our top 5 predictions.
– Acc@161: The percentage of predicted city which are within a 161km(100
mile) radius of the true coordinates in the original tweet to capture near-
misses. This measure is only tested on city-level prediction.
– Median: The median distance from the predicted city to the true coordinates
in the original tweet. This measure is only tested on city-level prediction.
5.2 Baseline Method
We compared our approach with one commonly used ensemble method in pre-
vious research works [9,10]. We implemented an ensemble classifier based on
stacking[29] with 5-fold cross validation. The training of stacking consists of two
steps. First, five multinomial naive Bayes base classifiers are trained on different
types of data(tweet content, user description, profile location, user name and the
remaining categorical features). The outputs from the base classifiers are used
to train a multinomial naive Bayes classifier in the second layer. We call such
method STACKING in this paper. Same as [10], we also use information gain
ratio to do feature selection on text tokens. We call STACKING with feature
selection STACKING+.
5.3 Hyperparameters and Training
We used a tweet-specific tokenizer provided by NLTK3 to tokenize text fields. We
built our dictionary based on the words that appeared in text, user description,
and profile location. To reduce low-utility words and noise, we removed all words
that had a word frequency less than 10. For our proposed approach, we used filter
windows(h) of 3,4,5 with 128 feature vectors each, a dropout rate of 0.5 and batch
size of 1024. We initialize word vectors using word2vec4 vectors trained on 100
billion words from Google News. The vectors have dimensionality of 300 and were
trained using the continuous bag-of-words architecture[16]. For those words that
are not included in word2vec, we initialized them randomly. We also performed
early stopping based on the accuracy over the development set. Training was
done through stochastic gradient descent using Adam update rule with learning
rate 10−3[14]. For our baseline models, we applied additive smoothing with α =
10−2, which is selected on the development set. For STACKING+ method, we
first ranked these words by their information gain ratio value, then selected the
top n% words as our vocabulary. The tuning of n is based on accuracy over
the development set. We selected n as 40%, 55% for city-level prediction and
country-level prediction respectively.
5.4 Results
The comparison results between our approach and the baseline methods are
listed in Table 3 and Table 4. Our approach achieves 92.1% accuracy and 52.8%
3 http://www.nltk.org/api/nltk.tokenize.html
4 https://code.google.com/archive/p/word2vec/
accuracy on country-level and city-level location prediction respectively. Our
approach is consistently better than the previous model on the country-level
location prediction task as shown in Table 3. It greatly outperforms our baseline
methods over all the measures, especially on the city-level prediction task. It
could assign more than half of the test tweets to the correct city and gain more
than 20% relative improvement over the accuracy of the STACKING+ method.
Table 3. Country prediction results.
Table 4. City prediction results.
Acc Acc@Top5
STACKING 0.868 0.947
STACKING+ 0.871 0.950
Our approach 0.921 0.972
STACKING 0.389 0.573
STACKING+ 0.439 0.616
Our approach 0.528 0.692
Acc Acc@161 Acc@Top5 Median
77.5 km
47.2 km
28.0 km
0.595
0.629
0.711
Our approach performs better for countries with a large number of tweets. In
Figure 3, we plotted the precision and recall value for each country as a scatter
chart. The dot size is proportional to the number of tweets that come from that
country. Turkey appears to be the country with highest precision and recall.
These results suggest that our approach works better with more data samples.
The same graph is also plotted for city prediction in Figure 3. Because of
the skewness of our data and the difficulty of city-level prediction, our classifier
tends to generate labels towards big cities, which leads to high recall and low
precision for cities like Los Angeles.
Fig. 3. Two scatter graphs that show the performances for each country and cities.
The x-axis is precision, y-axis is recall. Each dot represents a country/city. The dot
size is proportional to the number of tweets that comes from the correponding location.
Some tiny invisible country outside of the scope are not shown in the figure.
For real world applications, people may ask how we could set a threshold to
get prediction results with high confidence. To answer this question, we further
examined the relation between prediction accuracy and the output probability.
Here the output probability is just the probability of our predicted location cal-
culated by equation 2. Figure 4 shows the distribution of tweets in terms of
output probability for two tasks. The grey bar represents the percentage of test-
ing tweets within certain range of output probability. The green bar represents
the percentage of correctly classified tweets. The number on the grey bar is the
accuracy for each output probability range.
As expected, the prediction accuracy increases as the output probability in-
creases. We get 97.2% accuracy for country-level prediction with output proba-
bility larger than 0.9. Surprisingly, the accuracy of city-level is as high as 92.7%
for the 29.6% of the tweets with output probability greater than 0.9. However,
the city-level accuracy for the remaining tweets with output probability less than
0.9 is only 48.4%. Unlike country-level prediction, the number of tweets decreases
as output probability increases, unless the output probability is larger than 0.9.
Fig. 4. Two bar charts that show the distribution of tweets in terms of the output
probability. The x-axis is the output probability associated with each prediction, the
y-axis is the percentage of tweets. The height of grey bar represents the percentage
of test data that has certain prediction output probability. The height of green bar
represents the percentage of correctly predicted tweets in each output probability range.
We listed the accuracy for each probability range above the bar. Take the rightmost
bar in country-level prediction for example, there are 81.8% test tweets' country are
predicted with output probability larger than 0.9. Among these 81.8% tweets, 97.2%
are predicted correctly.
6 Discussion and Conclusion
These experiments demonstrate that our approach is consistently better than
the prior method thus supporting more tweets to be accurately located by coun-
try, and city, of origin. At the country level, the more tweets that come from
the country, the better the prediction. Regardless of the number of tweets per
country, we can predict the country location for most tweets wih extremely high
confidence and accuracy. At the city level the results are more mixed. For a
small fraction of tweets we can get greater than 90% accuracy, but for the rest
of tweets the accuracy is less than 50%. For about half the tweets it is difficult
to infer the city location. This result is partially due to the fact that we base
the prediction on only a single tweet. Future work may consider using collection
of tweets per user. This result is also partially due to the fact that the data is
highly skewed toward a few cities. Future work should develop a training set that
is more evenly distributed across cities. Despite these limitations, this approach
shows promise.
This paper presents a method for geo-locating a single tweet based on the
information in a tweet JSON object. The proposed approach integrates tweet
text and user profile meta-data into a single model. Compared to the previous
stacking method with feature selection, our approach substantially outperforms
the baseline method. We developed the approach for both city and country level
and demonstrated the ability to classify tweets at both levels of granularity.
The results demonstrate that using a convolutional neural network utilizes the
textual location information better than previous approaches and boosts the
location prediction performance substantially.
Acknowledgments. This work was supported in part by the Office of Naval
Research (ONR) N000140811186, and the National Science Foundation(NSF)
00361150115291. The views and conclusions contained in this document are those
of the authors and should not be interpreted as representing the official policies,
either expressed or implied, of the Office of Naval Research or the National
Science Foundation. We want to thank tutors in the Global Communication
Center at Carnegie Mellon for their valuable advice.
References
1. Achrekar, H., Gandhe, A., Lazarus, R., Yu, S.H., Liu, B.: Predicting flu trends using
twitter data. In: Computer Communications Workshops (INFOCOM WKSHPS),
2011 IEEE Conference on. pp. 702–707. IEEE (2011)
2. Berggren, M., Karlgren, J., Ostling, R., Parkvall, M.: Inferring the location of
authors from words in their texts. arXiv preprint arXiv:1612.06671 (2016)
3. Bo, H., Cook, P., Baldwin, T.: Geolocation prediction in social media data by find-
ing location indicative words. In: Proceedings of COLING. pp. 1045–1062 (2012)
4. Carley, K.M., Malik, M., Landwehr, P.M., Pfeffer, J., Kowalchuck, M.: Crowd
sourcing disaster management: the complex nature of twitter usage in padang
indonesia. Safety science 90, 48–61 (2016)
5. Cheng, Z., Caverlee, J., Lee, K.: You are where you tweet: a content-based approach
to geo-locating twitter users. In: Proceedings of the 19th ACM international con-
ference on Information and knowledge management. pp. 759–768. ACM (2010)
6. Culotta, A., Kumar, N.R., Cutler, J.: Predicting the demographics of twitter users
from website traffic data. In: AAAI. pp. 72–78 (2015)
7. Earle, P.S., Bowden, D.C., Guy, M.: Twitter earthquake detection: earthquake
monitoring in a social world. Annals of Geophysics 54(6) (2012)
8. Hale, S., Gaffney, D., Graham, M.: Where in the world are you? geolocation and
language identification in twitter. Proceedings of ICWSM 12, 518–521 (2012)
9. Han, B., Cook, P., Baldwin, T.: A stacking-based approach to twitter user geolo-
cation prediction. In: ACL (Conference System Demonstrations). pp. 7–12 (2013)
10. Han, B., Cook, P., Baldwin, T.: Text-based twitter user geolocation prediction.
Journal of Artificial Intelligence Research 49, 451–500 (2014)
11. Hong, L., Ahmed, A., Gurumurthy, S., Smola, A.J., Tsioutsiouliklis, K.: Discover-
ing geographical topics in the twitter stream. In: Proceedings of the 21st interna-
tional conference on World Wide Web. pp. 769–778. ACM (2012)
12. Jurgens, D.: That's what friends are for: Inferring location in online social media
platforms based on social relationships. ICWSM 13, 273–282 (2013)
13. Kim, Y.: Convolutional neural networks for sentence classification. arXiv preprint
arXiv:1408.5882 (2014)
14. Kingma, D., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980 (2014)
15. Landwehr, P.M., Carley, K.M.: Social media in disaster relief. In: Data mining and
knowledge discovery for big data, pp. 225–257. Springer (2014)
16. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-
sentations of words and phrases and their compositionality. In: Advances in neural
information processing systems. pp. 3111–3119 (2013)
17. Mislove, A., Lehmann, S., Ahn, Y.Y., Onnela, J.P., Rosenquist, J.N.: Understand-
ing the demographics of twitter users. ICWSM 11, 5th (2011)
18. Pennacchiotti, M., Popescu, A.M.: Democrats, republicans and starbucks afficiona-
dos: user classification in twitter. In: Proceedings of the 17th ACM SIGKDD inter-
national conference on Knowledge discovery and data mining. pp. 430–438. ACM
(2011)
19. Qian, Y., Tang, J., Yang, Z., Huang, B., Wei, W., Carley, K.M.: A prob-
abilistic framework for location inference from social media. arXiv preprint
arXiv:1702.07281 (2017)
20. Quercia, D., Kosinski, M., Stillwell, D., Crowcroft, J.: Our twitter profiles, our
selves: Predicting personality with twitter. In: Privacy, Security, Risk and Trust
(PASSAT) and 2011 IEEE Third Inernational Conference on Social Computing
(SocialCom), 2011 IEEE Third International Conference on. pp. 180–185. IEEE
(2011)
21. Quinlan, J.R.: C4. 5: programs for machine learning. Elsevier (2014)
22. Roller, S., Speriosu, M., Rallapalli, S., Wing, B., Baldridge, J.: Supervised 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. pp. 1500–1510. Association for
Computational Linguistics (2012)
23. Rumelhart, D.E., Hinton, G.E., Williams, R.J.: Learning representations by back-
propagating errors. Cognitive modeling 5(3), 1 (1988)
24. Sakaki, T., Okazaki, M., Matsuo, Y.: Earthquake shakes twitter users: real-time
event detection by social sensors. In: Proceedings of the 19th international confer-
ence on World wide web. pp. 851–860. ACM (2010)
25. Shi, L., Agarwal, N., Agrawal, A., Garg, R., Spoelstra, J.: Predicting us primary
elections with twitter. URL: http://snap. stanford. edu/social2012/papers/shi. pdf
(2012)
26. Tsou, M.H., Yang, J.A., Lusher, D., Han, S., Spitzberg, B., Gawron, J.M., Gupta,
D., An, L.: Mapping social activities and concepts with social media (twitter) and
web search engines (yahoo and bing): a case study in 2012 us presidential election.
Cartography and Geographic Information Science 40(4), 337–348 (2013)
27. Wang, X., Xu, M., Ren, Y., Xu, J., Zhang, H., Zheng, N.: A location inferring
model based on tweets and bilateral follow friends. JCP 9(2), 315–321 (2014)
28. Wing, B.P., Baldridge, J.: Simple supervised document geolocation with geodesic
grids. In: Proceedings of the 49th Annual Meeting of the Association for Compu-
tational Linguistics: Human Language Technologies-Volume 1. pp. 955–964. Asso-
ciation for Computational Linguistics (2011)
29. Wolpert, D.H.: Stacked generalization. Neural networks 5(2), 241–259 (1992)
|
1909.01137 | 1 | 1909 | 2019-08-27T17:38:33 | Valuating User Data in a Human-Centric Data Economy | [
"cs.SI",
"cs.GT"
] | The idea of paying people for their data is increasingly seen as a promising direction for resolving privacy debates, improving the quality of online data, and even offering an alternative to labor-based compensation in a future dominated by automation and self-operating machines. In this paper we demonstrate how a Human-Centric Data Economy would compensate the users of an online streaming service. We borrow the notion of the Shapley value from cooperative game theory to define what a fair compensation for each user should be for movie scores offered to the recommender system of the service. Since determining the Shapley value exactly is computationally inefficient in the general case, we derive faster alternatives using clustering, dimensionality reduction, and partial information. We apply our algorithms to a movie recommendation data set and demonstrate that different users may have a vastly different value for the service. We also analyze the reasons that some movie ratings may be more valuable than others and discuss the consequences for compensating users fairly. | cs.SI | cs | Valuating User Data in a Human-Centric Data Economy
Marius Paraschiv
IMDEA Networks
Legan´es -- Madrid
Nikolaos Laoutaris
IMDEA Networks
Legan´es -- Madrid
[email protected]
[email protected]
9
1
0
2
g
u
A
7
2
]
I
S
.
s
c
[
1
v
7
3
1
1
0
.
9
0
9
1
:
v
i
X
r
a
Abstract -- The idea of paying people for their data is in-
creasingly seen as a promising direction for resolving privacy
debates, improving the quality of online data, and even offering
an alternative to labour-based compensation in a future domi-
nated by automation and self-operating machines. In this paper
we demonstrate how a Human-Centric Data Economy would
compensate the users of an online streaming service. We borrow
the notion of the Shapley value from cooperative game theory
to define what a fair compensation for each user should be for
movie scores offered to the recommender system of the service.
Since determining the Shapley value exactly is computationally
inefficient in the general case, we derive faster alternatives using
clustering, dimensionality reduction, and partial information.
We apply our algorithms to a movie recommendation data set
and demonstrate that different users may have a vastly different
value for the service. We also analyse the reasons that some
movie ratings may be more valuable than others and discuss
the consequences for compensating users fairly.
I. INTRODUCTION
Data, and the economy around it, are said to be driving
the fourth industrial revolution. Interestingly -- the people --
whose data is what moves the new economy, have a rather
passive role in it, as they are left outside the direct value
flow that transforms raw data into huge monetary benefits.
This is a consequence of the de facto understanding (or one
may say misunderstanding) between people and companies,
that
the former get unpaid access to online services in
exchange for unpaid access to their personal data. This is
increasingly being challenged by various voices who call
for the establishment of a new, renegotiated, relationship
between users and services. Indeed, a variety of pathologies
can be traced back to the way the data economy has been
working so far. Some are direct and obvious, such as privacy
risks for individuals, and market failures and dangers for the
economy from the rise of data monopolies and oligopolies.
Others are less obvious, and further reaching into the future,
such as mass unemployment due to data-driven automation.
It was estimated recently [1] that, if automation due to
artificial intelligence reaches maturity and fair remuneration
algorithms are set in place, a family of four could earn up to
$20,000 per year from their data. The idea of micropayments,
or providing small contributions to users in exchange for
their presence on a platform or for accessing a service,
is of course much older. In the pre-World Wide Web era,
France developed a videotex online service called Minitel,
that included micropayments as part of its design, but Jaron
Lannier brought it to public attention in 2013, in his book
"Who owns the future?"[2]. In it, he argues that we have
only underwent half of the Data Revolution, the part that
compensates users with implicit benefits, but not the part
that also compensates them with explicit monetary benefits.
There have been a series of proposed approaches for how
this compensation might materialise. The simplest, at least
in theory, would be to assign a context-free value to data, a
kind of dollar-per-bit measure. This has been proven to be
very hard [3], [4], [5], [12]. Indeed, since the value of data
is strongly connected to its intended use, it becomes very
difficult to argue about how to assign an a priori average
value. For traditional currencies, we are able to have a
context-free appreciation of their value for the simple reason
that we have been using these currencies long enough to
be able to do so. Although we clearly understand nowadays
that one's browsing and mobility patterns, social network, or
past purchases all have value, we are far from being able
to appreciate how much this value is in terms of dollars
or euros. The latter is further complicated by our inability
to tell in advance, by how many parties, and how many
times, a piece of data may be utilized. As an analogy, selling
an individual's data, or rather renting it temporarily, is as
difficult and risky as renting an infinitely fast vehicle, with no
gas and maintenance costs, and without any prior restrictions
with regard to mileage or the person driving it.
In a Human-Centric Data Economy, when a transaction,
or set of transactions,
is converted, a proportion of the
obtained revenue will be returned to the users. Defining the
right amount to be returned to the users is difficult, as it
depends on many market characteristics of a multilateral
value chain, such as competition and user loyalty [6]. In
this paper, we assume that the total amount of revenue to
be redistributed to users is given, e.g., 5% or 10%, or any
other number produced by the competition between services
A second proposed method has been to compensate users
for their privacy damage [13], [14]. Processing massive
amounts of data can lead to privacy infringements, such as
the leakage of habitual user behavior, their location or other
personal identifiable information (PII). Users are thus seen
as victims who must be compensated for their damage.
Our approach is different, we consider users as active
partners in the data value chain. Such a chain requires a
business model, smart predictive algorithms for extracting
useful information from raw data and online marketing for
attracting and retaining users, among many others. The
fundamental component of the value chain, however, is the
user, and it is ultimately a matter of common sense that they
should be rewarded in a fair manner, which may or may not
exceed the perceived privacy-related damages.
within a given sector. We thus turn our attention to the next
important question, namely: given a fixed total amount to be
distributed, how should this distribution be performed? How
should different users be compensated based on the value of
the data that they contribute to a specific service?
One obvious answer would be to split the sum by the
number of users on the platform. This may not be fair
however, because it
is unreasonable to assume that all
users contribute equally to the service. In the case of a
recommendation system, for example, some users may rate
and view hundreds of items, while others may have a
much sparser activity. Another example would be a traffic
application. Users who travel regularly through a given area
and constantly provide feed-back, may be more relevant to
the platform than occasional passers-by. Even at an equal
level of intensity, some movie scores, or location data may be
more important for a recommender than others. For example
a score for a blockbuster movie that
just aired may be
more important than a score for a well known (and voted)
blockbuster movie. Similarly, speed information for a known-
to-be-congested route is less useful to a navigation app than
speed data for an unknown alternative route with less traffic.
Another important point to understand is that the Data
Economy is not a zero-sum game. Paying users for data need
not be seen as a measure that will reduce the revenue or
profitability of online services. More and better data can lead
to better services and thus more revenue and profit. It is thus
not surprising that the vision of paying directly or through
taxes for data has received positive comments from industry
leaders such as Bill Gates [7], Elon Musk [8] and Mark
Zuckerberg [9]. Also, despite not happening today, there is a
realistic path that can lead to wide adoption. It only requires
that a small number of visionary companies start offering
micro-payments to get a competitive advantage in terms of
user retention, for others to follow, and the practice to get
traction in the online services market.
Our paper brings a threefold technical contribution towards
the realisation of a Human-Centric Data Economy:
• We define data payoff fairness in terms of accepted
economic notions. Specifically, we use the Shapley
value from collaborative game theory to define how
to split a total payback among all the users that have
provided data to a revenue-generating online service.
We also sketch a Contribution-Reward framework for
implementing such paybacks in practice.
• We develop two algorithms that can provide efficient
estimates of the Shapley value, which in its raw form
does not scale for large data sets. The first one applies
minibatch k-Means with N d-dimensional points, for
k clusters and t iterations directly on the definition
of Shapley to reduce the complexity from O(N !) to
O(N kdt + k!), where k, d, t (cid:28) N. The second is a
O(N 2) heuristic of local information that does not use
the Shapley definition.
• We apply the above algorithms to a movie scoring data
set and study how different users may be in terms of
their value for their service. We observe that some users
Fig. 1. The Contribution-Reward framework. In establishing the proportion
of each user's contribution, one first determines important features or
relevant measures that result in revenue. One such measure could be the
number of useful recommendations obtained from a recommender system.
After a measure or relevant feature has been identified, user contributions to
it are computed and users are ordered in a hierarchy. Based on their place
in the value hierarchy, they are remunerated.
may be as much as ×4 more valuable than others.
• Finally, we study how scoring behaviour impacts on
user value. We find that what differentiates a "good"
user from a "bad" one is that
the good user tends
to vote mainly for the most popular movies and is
consistent with the movie popularity hierarchy, namely
popular movies get high ratings and unpopular ones,
low ratings. By contrast, users at the other end of the
value distribution tend to vote and provide high scores
to unpopular movies as well.
While human centricity and fairness need to reside at the
core of the proposed framework, transparency must also be
considered. Users need a method of verifying for themselves
the amount of payment received and be able to connect their
behaviour on an online platform to the obtained revenue, in
a clearly interpretable manner. We discuss how to do this via
an accounting meta-data layer at the end of the article.
II. BACKGROUND
In the Background section, we introduce the Contribution-
Reward framework and describe the setting for our two algo-
rithms -- a recommender system generating recommendations
for the users of an online movie streaming service. We also
introduce the Shapley value, as the backbone of the first
proposed algorithm and also an accepted point of reference
in terms of fair credit distribution. The section ends with
a toy example that shows how the exact Shapley value is
computed in a simple two-user setting.
A. The Contribution-Reward Framework
The proposed Contribution-Reward framework, depicted
in Fig. 1, aims to provide users with a remuneration scheme
that is both fair and sustainable. With usage, the service
2
UsersDatasetFeature Importance andUser ContributionShapleyEstimatorScore to monetary valueconversion.Proportional user remunerationShapley scoresIndividual user contributionsto significant featuresData SampleUser-generated data gathers and stores relevant data. Based on the data type, a
number of important features are chosen and the contribution
of each user to the said features is estimated. Finally, the
user receives a repayment from the platform, proportional to
their estimated contribution. The difficulty of finding the said
estimation resides in the need to first observe which are the
relevant features that make a data set valuable. In the case
of recommendation data, for example, it could be that users
voting for the most popular movies, thus contributing to an
already existing hierarchy, are considered more valuable, or
on the contrary, users reviewing items which are initially
unpopular could bring an element of novelty, and thus a
higher overall value. On the other hand, if the accuracy and
update rate of the data set are essential, for example in traffic
applications, users with a higher contribution frequency may
be considered of higher value.
It thus becomes clear that user value depends on the struc-
ture of the data and its intended use. Due to this subjective
nature, deriving a generally-applicable valuation framework
is not a straightforward task. The question we raise here
is two-fold: first, how can one determine a hierarchy of
value for a given set of users and a particular use case
(recommender systems) and second, how can one quantify
the position of each user in the value hierarchy? An answer
to the second question is of particular importance, since if
one can assign scores to users, based on their contributions,
one can then define a mapping from these scores to an actual
financial amount.
For the rest of this paper, the setting is thus fixed: a
recommender system is trained on a training set, consisting
of movie reviews, and then makes predictions on a separate
test set. If the recommendations made to a user, on the test
set, have an error below a fixed threshold, we consider that
there is a high likelihood that the user will want to watch
the recommended item. We refer to these recommendations
as "clicks". As such, when we say: "user A has generated
5 clicks", this is to be interpreted as "the recommender has
made 5 recommendations to user A, all with an error below
our set threshold". Clicks here play the role of the important
feature in the Contribution-Reward framework.
The most naıve approach would be to train the recom-
mender system by removing one user at a time, and counting
the difference in the overall number of clicks obtained on the
test set. There are a series of disadvantages to this: first, a
data set may contain reviews from hundreds of thousands of
users, quickly making the leave-one-out training computa-
tionally unfeasible. The second drawback resides in the fact
that removing only a single user may not have the expected
result on the recommender. Indeed, if the user provides a
significant amount of novelty, it may happen that the system
produces more clicks in the absence of the user than in their
presence, leading to a negative assigned value. This would
further be difficult to map to a monetary contribution. In
the following sections, we present two alternative methods
which avoid these shortcomings and can be implemented in
a computationally-scalable way.
After a theoretical discussion of the proposed algorithms,
they are applied to a case study based on a subset of the
MovieLens data set [15], depicting how groups of users
can contribute differently to the overall performance of the
recommender system, and thus hold different values with
respect to the service.
B. Introducing the Shapley Value
Credit assignment in cooperative games has long been a
central problem of cooperative game theory. To this end,
Shapley [16], [17] proposed that players should be rewarded
in a manner proportional to their average marginal contribu-
tion to the payoff of any coalition they could join.
Let N be a set of N players and S ⊂ N be a coalition
with cost v(S). The Shapley value is a uniquely determined
vector of the form (φ1(v), ..., φn(v)), where the element
representing player i is given by
φi(v) =
1
N !
[v(S(π, i)) − v(S(π, i) \ i)],
(1)
(cid:88)
π∈SN
where π is a permutation representing the arrival order of
set N , while S(π, i) represents the set of players that have
arrived into the system before player i.
The Shapley value satisfies a series of important proper-
ties:
the total gain is completely distributed
• Efficiency:
among the players(cid:88)
i∈N
φi(v) = v(N )
(2)
• Symmetry: if i and j are two players who bring equal
contributions, in the sense that, for every subset S that
contains neither i nor j, v(S ∪ {i}) = v(S ∪ {j}), then
their respective Shapley values are also equal, φi(v) =
φj(v).
• Linearity: if two coalition games, denoted v and w are
combined, the resulting gain is the sum of the gains
derived from each game separately
φi(v + w) = φi(v) + φi(w)
(3)
and also φi(α · v) = α · φi(v), for any real α.
• Null player: the Shapley value of a null player is
zero. A player is considered null if they do not bring a
contribution to any possible coalition.
Unfortunately, the Shapley value has also been proven to
be NP-hard for many domains [18], [19], [20]. Since it takes
into account all possible coalitions, for each user, the number
of terms scales with N !, where N represents the number of
users, such that it quickly becomes computationally unfeasi-
ble.
In Ref. [21] the authors use Monte Carlo to approximate
the Shapley value for computing the cost contribution of
individual households to the peak hour traffic and costs
of an Internet Service Provider (ISP). In that case,
the
relatively simpler structure of the problem made Monte Carlo
an appropriate technique for approximating Shapley. Other
recent works have presented approximation algorithms for
3
Shapley for specific problems of lower complexity than
recommendation [10], [11]. In the context of the current
proposal, the inherent higher complexity of the considered
value functions v() that may represent the workings of com-
plex ML algorithms for things like recommendation, makes
using Monte Carlo inaccurate according to our preliminary
tests. Instead, we intend to use clustering to reduce the input
size of the problem. Our approach will be to first group
users in a number of clusters according to the similarity of
their movie ratings (or trajectories in the case of mobility
related applications), and then compute the Shapley value
of each cluster instead of each user. By controlling the
number of clusters we can make the computation as precise
and complex as our computing resources allow. In Section
III we will apply this method in the case of a movie
recommendation data set, containing an amount of users for
which the exact computation of Eq. 1 is not possible.
C. A Toy Example
The purpose of this section is to provide a very simple
example of computing the Shapley value for the case of
two users. Consider an artificial data set, containing movie
recommendations from two different users. The predictions
made by the recommender are the net contribution in our
case. We start from the marginal contributions V ({1}) = 12
and V ({2}) = 15, which would mean that the presence
of user 1 alone in the data set, causes the recommender to
produce 12 useful movie recommendations, and the presence
of user 2 alone results in 15 useful recommendations. Let
us further assume that the presence of both users, simultane-
ously, increases the number of recommendations to 28, hence
V ({1, 2}) = 28. Since it is clear that the two contributions
are not equal, the following question arises: how can we find
a factor, proportional to the user contributions, that would be
useful in determining a fair repayment ?
For this, we compute the Shapley value, as defined in
Eq. 1. We first note that
there are two possible orders
of arrival of the users, with equally likely probabilities of
occurrence: [1, 2] or [2, 1]. In the first situation, user 1 comes
first, bringing a contribution of V ({1}) = 12, followed
by user 2, who increases the overall useful contribution to
V ({1, 2}) = 28, thus, the net contributions of the two users,
for this particular order of arrival are
φ[1,2](1) = V ({1}) = 12,
φ[1,2](2) = V ({1, 2}) − V ({1}) = 28 − 12 = 16.
In the second case, the order of arrival is reversed, namely
[2, 1]. User 2 arrives first, bringing a contribution V ({2}) =
15, followed by user 1 who increases the net contribution to
V ({1, 2}) = 28, thus
φ[2,1](1) = V ({1, 2}) − V ({2}) = 28 − 15 = 13,
φ[2,1](2) = V ({2}) = 15.
Since the two cases are equally probable, the Shapley
value is the sum of the two marginal contributions multiplied
by the probability of the order of arrival (or of the factorial
of the number of users)
φ(1) =
φ(2) =
1
2!
1
2!
· (φ[1,2](1) + φ[2,1](1)) = 12.5,
· (φ[1,2](2) + φ[2,1](2)) = 15.5.
We thus see that the average marginal contribution of user
2 is higher than the one of user 1, and based on this factor,
we can devise a scheme of fair remuneration, to be presented
as part of the Contribution-Reward framework.
It is also worth noting that, in order to compute the Shap-
ley terms, we need to compute N ! marginal contributions,
which leads to scaling problems if an exact computation
of the Shapley value is required. The methods presented in
this paper provide ways of overcoming this obstacle through
various approximate approaches.
III. USER VALUE ESTIMATION
In this section we first present and then apply the two
proposed algorithms. We then discuss the results obtained on
a subset of the MovieLens data set. At the end of this section
and in the next, we will interpret user behaviour and try
to understand what is the relationship between votes given,
number of votes, types of movies reviewed and the assigned
user value.
A. Approximate Shapley Value Estimation (ASVE)
The first user contribution estimation method is based
on the Shapley value, described in the previous sections.
The algorithm is constructed around a recommender system
framework and we choose our value of interest to be the
number of "clicks" or useful recommendations that the model
provides, for a given data set. The pseudocode for this
algorithm is provided as Algorithm 1.
The input data consists of user identifiers, product iden-
tifiers (movie categorical IDs) and votes. From Eq. 1 we
observe that computing the Shapley value of each user in
the data set directly is unfeasible, as the computational
complexity is of order N ! where N (the number of users
in the data set) is typically extremely large (in the order
of hundreds of thousands or even millions for services like
YouTube and NetFlix). Clearly such an approach does not
scale.
There are two plausible directions one can pursue in order
to avoid the complexity barrier. The first is to attempt to
estimate the Shapley value using Monte Carlo sampling, as
done in Ref. [21]. The second, which we shall employ here,
takes advantage of the similarities between user behavioural
patterns. In general, when treating data related to user
preference, consumers tend to cluster into a limited number
of similar groups.
For example there may be individuals with a stronger
preference for action movies as opposed to romantic ones.
Exploiting such relationships allows us to greatly simplify
the estimate of a user's contribution to the overall data set.
4
# Cluster the users based on behaviour.
clusteredData, clusterLabels ← Cluster(inputData)
clusterCoalitions ← Compute(clusterLabels)
# Compute the total number of clicks.
K ← TrainPredict(inputDataset)
# Compute marginal contribution of every possible coalition.
# trainData and predData contain only clusters from coalition.
for S in clusterCoalitions do
Algorithm 1 Approximate Shapley Value Estimation
1: procedure ASVE
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
Vi ← V (π) − V (π − {i})
margContrib[i] += Vi
len(clusterLabels)! ∗ sum(margContrib[i])
# Create all possible permutations of complete cluster coalitions.
# and compute the marginal contribution of each cluster.
perm ← Permute(clusterLabels)
for π in perm do
VS ← TrainPredict(trainData, predData)
for i in clusterLabels do
for i in clusterLabels do
φi =
1
1+EuclideanDist(user, centroid(k)) )
k
φk
20:
21:
22:
23:
24:
25:
26:
# Compute the value of every user.
for userId in clusteredData do
φ1(userId) = ((cid:80)
userValSum += φ1(userId)
# Normalize user values to respect efficiency condition.
for userId in clusteredData do
φ1(userId) = φ1(userId) ·
K
userValSum
neighbors ← FindNeighbors(inputDataset, userId)
for id in neighbors do
Rec = 0
# Train model once to obtain the total number of clicks.
K ← TrainPredict(inputDataset)
# For all neighbors in a neighbourhood, compute the user's
# predictive potential
for userId in inputDataset do
Algorithm 2 Neighbourhood Similarity Value Estimation
1: procedure NSVE
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
# Scale the predictive potential by the total
# number of clicks divided by the sum of the predictive
# potentials, to respect the efficiency condition.
for userRec in RecList do
φ3(userId) ← K*userRec
j userRecj
newItems ← IntersectLeft(userId, id)
Rec+ = 1
RecList.append(Rec)
Rec ← 0
(cid:80)
The algorithm starts by clustering the input data, prior to
which, a dimensionality reduction method (such as PCA)
is applied. The next step is the calculation of all possible
coalitions of clusters. For example, for three clusters, these
are
{1},{2},{3},{1, 2},{1, 3},{2, 3},{1, 2, 3},
where {1} represents a data set where only the users corre-
sponding to the first clusters are present. The recommender
system is then trained on the filtered data sets, corresponding
only to the clusters in the above coalitions. A train-test
split is performed prior to this, but omitted in line 8 of
the pseudocode for clarity. During the prediction phase, the
model makes recommendations to users and also provides
an error estimate for each recommendation. We assert that,
for errors under a certain threshold, the recommendations
can be valuable, and we refer to these recommendations as
"clicks" for the rest of this paper, in analogy to advertising
recommendations where, if the user is interested in an advert,
they will click the banner.
After obtaining the number of clicks for every one of
the possible subsets of clusters, one must compute the
marginal contribution of the cluster. This takes into account
every possible complete coalition, where the order of arrival
matters. For example, in our three user case, all possible
complete coalitions are: {1, 2, 3},{2, 1, 3} and {3, 2, 1}. The
relevant code is between lines 11 - 19.
Finally, when the Shapley value (computed in line 19)
for each cluster is known, we can proceed to determine the
value of each individual user. For this, the centroid of the
cluster is labeled with the corresponding Shapley value of the
cluster. Thus, the value of each user is equal to the sum of all
individual cluster values divided by the Euclidean distance
between the point, representing the user, and the respective
cluster centroid, with one added for stability. The assigned
values, from both proposed algorithms, are then scaled with
the total number of clicks produced by the recommender, on
the complete data set, to ensure that the efficiency condition
is met. In this manner, to every point on a projected two-
dimensional surface, representing the space of all users, we
can assign an approximate Shapley value.
We have applied the ASVE algorithm to a subset of the
MovieLens data set, containing 92,394 total ratings on 4,180
movies from 610 users, with the intent on understanding if
the value hierarchy provided by the method can be intuitively
understood. First, the users were separated into three classes,
the best users (representing the ones with the highest scores
given by the ASVE method), average users and bad users.
We considered whether or not there is a relationship between
movie popularity (based on the overall number of ratings a
movie has received) and the vote distribution of the three
classes of users.
In Fig. 2, the user vote distributions are presented, for the
three classes of users. The movies are grouped into equal-
sized ordered buckets, with their popularity decreasing in
the positive direction of the horizontal axis. We observe that
5
Fig. 2. User vote distributions based on movie popularity for the ASVE method. Votes from the most valuable users are shown in red, those from average
users in blue and votes from the least valuable users are shown in green. For clarity, the movies have been grouped into ordered buckets, with popularity
decreasing in the positive direction of the horizontal axis.
Fig. 3. User vote distributions based on movie popularity for the NSVE method. Votes from the most valuable users are shown in red, those from average
users in blue and votes from the least valuable users are shown in green. For clarity, the movies have been grouped into ordered buckets, with popularity
decreasing in the positive direction of the horizontal axis.
users which the ASVE method considers as being the most
valuable, tend to vote mostly on popular movies, whereas
users considered to have low value, have a more widespread
distribution of votes. This reinforces our initial assumption
that the ASVE method gives high scores to users who agree
with a predetermined hierarchy, not necessarily to users with
a novel contribution.
One remark that must be made is that, by design, the
application of ASVE requires one prior check. When re-
moving a cluster of users, it is entirely possible that one or
more clusters provide a negative contribution. It is therefore
essential that one checks the values assigned to the clusters
for non-negativity, prior to applying the method to individual
users.
B. Neighbourhood Similarity Value Estimation (NSVE)
The second approach relies on first reducing the dimen-
sionality of the data set (similar to ASVE) and creating a
neighbourhood around each user. Based on the predictive
capability of the user on its neighbours, one can approximate
the user's importance.
The algorithm works as follows: first, the recommender
system is trained on the complete data set and a total number
of clicks (similar to the previous method) is obtained (line
4). For each user in the data set, we then construct a neigh-
bourhood and observe the number of items that the central
user has rated, and the neighbours have not. These elements
act as an estimator of the user's "predictive potential". For
example, if user A has seen and rated movies M1, M2 and
M5 and user B, which is in user A's neighbourhood, has only
seen movie M1 (out of the list that user A has rated overall),
then A could potentially recommend movies M2 and M5 to
B, so we say that "A has a predictive potential of 2 (movies)
over B). This predictive potential is then used, in line 18, to
proportionately assign clicks (from the total number K) to
each user, based on their respective predictive potentials.
One must keep in mind that the size of the neighbourhood
is a hyperparameter that must be determined through trial
and error. Furthermore, there may be other distance metrics,
other than the Euclidean distance, chosen here, that could
potentially offer a better estimate, however, for simplicity,
we restrict ourselves to Euclidean distances in the present
paper.
In order to compare the two methods, a proper scaling
must be found for the two corresponding distributions. The
unscaled distributions, shown in Fig. 4 are correlated, with
a correlation distance of 0.29 (0 representing a perfect
correlation) for the MovieLens data set. We were thus able
to learn an approximate mapping, that would project the
second distribution to the domain of the first, making a direct
comparison possible, as depicted in Fig. 5.
Due to the nature of this mapping being approximate,
one sees artefacts, such as the unusually tall peaks of
6
01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556Popularity-sorted Movie Buckets0200040006000800010000Votes01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556Popularity-sorted Movie Buckets0200040006000800010000Votes users (according to both methods) seem to offer high scores
to a wide array of movies, going against the established
popularity hierarchy. This further serves to confirm that users
reinforcing an existing order are rated higher.
One point to notice is that, in the case of the NSVE, by
design a user cannot have a negative assigned value, thus
no prior checks are required, as in the case of ASVE. A
potential problem here is the reliance of the method on
the neighbourhood radius hyperparameter, which strongly
depends on the data set. Furthermore, users outside a given
neighbourhood have zero contribution to the final rating of
the neighbourhood source. This is in stark contrast to the
smooth value assignment of ASVE, and does give rise to a
series of artefacts, one of which being that there are always
(or at least in all practical situations) users with an assigned
zero value, which raises the difficulty of mapping this to an
actual financial amount.
A second artefact can be observed once we ask the
question: how many clicks do movies generate (over all users
in a fixed test set) based on their popularity? As can be seen
in Fig. 7, the Neighbourhood-based method is very sensitive
to the relative distance between users. As movies become
less popular and users who vote for them get spread over
larger distances, the amount of clicks generated (which in
the case of NSVE represent movies that the central user
can recommend to its immediate neighbours) quickly falls to
zero. Thus, NSVE strongly favors users who vote for popular
movies, where users who favor them cluster together.
From a computational standpoint, the NSVE algorithm
only requires one training of the recommender, in order
to obtain the total number of clicks. The neighbourhood
comparison operations are linear. Another advantage of this
method is that, if applied to a production environment, the
model would continuously train as new user data becomes
available, and only be required to perform occasional infer-
ence, to update the total number of clicks.
Another potential impediment to both methods is their
sensitivity to badly-labelled data. If, for example, in our
movie data set, the tags describing the movie genre are
not consistent or worse, misleading,
the
distances between users and result in both methods providing
inaccurate value estimates. Therefore, data cleaning and a
careful analysis of the data set are important, before the
algorithms are employed.
this will affect
IV. DISCUSSION
In the previous section, we have presented two credit-
assignment methods that could become part of an online
service, and transform users into active participants in the
information value flow, from its raw form to the final gen-
erated revenue. As the main producers of raw data, fairness
requires that users share part of the resulting revenue, in a
manner proportional to the significance of their data.
We have proposed two different methods that are in
agreement on a few key points. The first method is based on
the commonly-used Shapley value, which is approximated
by being computed at cluster level. The users are then
Fig. 4. The ASVE distribution, shown in blue, and the unscaled NSVE
distribution, shown in orange. It is important to note that, even though the
variance of the two distributions is different, the user value hierarchy is
maintained to a large degree.
Fig. 5.
The ASVE distribution, in blue, and scaled NSVE distribution,
in orange, showing artefacts due to the approximate mapping between the
two.
the projected NSVE distribution, compared to its ASVE
counterpart. While this is an issue that needs to be considered
when comparing method predictions, the mapping process
would not be necessary in a real-world implementation, as
it's sole purpose is to aid the comparison, and would be
irrelevant for the purpose of credit assignment.
Having the two distributions on the same scale, we con-
tinued our investigation by asking whether or not the second
method produces the same distribution of user votes, based
on movie popularity. As seen in Fig 3, both methods agree
on the fact that users considered valuable tend to vote for
popular movies.
It is instructive, at this point, to also see how users voted,
based on movie popularity, that is, is it the case that more
popular movies got higher ratings than less popular ones?
Also, how do the three classes of users rate movies of various
levels of popularity?
In Fig. 6, we present the vote distribution by user cate-
gory. Rather surprisingly, there is somewhat little difference
between users considered very valuable and those considered
average, as both classes seem to give high scores to very pop-
ular movies and low scores to unpopular ones. Less valuable
7
020406080100User Value020406080100Number of Users0102030405060User Value020406080100120140Number of Users Fig. 6. Movie ratings given by the three user classes, sorted by movie popularity. The average ratings from the best users (over both methods) are depicted
in blue, those from users with an intermediate value, in orange and those from the least valuable users, in green.
Fig. 7. Movie votes and number of clicks generated by the corresponding bucket, according to the two credit-assignment methods. The overall number
of votes is shown in blue, the number of clicks that movies in the corresponding bucket generate are shown in orange, for ASVE and green for NSVE.
scored based on their relative distance from the cluster
centers. The second method exploits user similarity, creating
a neighbourhood around each user, where users with similar
preferences may be found. It is thus assumed that the items
the central user can recommend are close to the interests of
his neighbours, such that they become potential clicks. This
is a very crude, but also very effective method for performing
distance-based recommendations.
We have seen that the two algorithms provide consistent
predictions on key aspects such as the distribution of user
votes based on their attributed value, the score given by users
based on the same attributed value and the number of clicks
movies produce, based on their popularity. These come to
confirm our initial hypothesis, based on intuition, that users
receiving high scores vote in line with the item's popularity
level. There are, however, points where the two methods do
not agree. The NSVE is more sensitive to distances between
users in the projected plane. In areas where the users are
sparsely distributed, NSVE tends to assign zero values to
them. This is not
the case for the ASVE algorithm, as
belonging to a cluster ensures a non-zero contribution to the
final outcome.
Also as a result of the sharp cut-off at the neighborhood's
boundary, the scores assigned to users by the Shapley-based
method change in a much smoother and more continuous
manner from one region to another. This reflects itself in a
much narrower distribution (as seen in Fig. 4). In contrast, the
user score distribution of the NSVE method has a very high
variance, assigning very small scores to users who are far
apart and very high scores to the ones in densely-populated
regions.
One issue not yet discussed is how should new users
be treated, with respect
to their overall value. As their
contribution is minimal, a valid argument would be to start by
assigning them relatively small scores (hence small propor-
tional payments). This is indeed the case for both methods.
Considering ASVE, a new user would likely be far from
the center of any cluster, and, as the proportional payment
decreases linearly with the distance from the cluster center,
would get a small reward. For NSVE the situation is rather
similar, a new user would either be in a sparsely-populated
area of the projected space or in a densely-populated area
but having very few possible recommendations for their
neighbours. In either case, they would receive a small score.
Finally, what should be the impact of low votes? This
question is of particular importance if either of the two
proposed algorithms is to be implemented in a realistic
setting. Observing Fig. 6, we see that valuable users give
high votes to popular films and low votes for unpopular
ones. If this rule were to hold for a vast array of data sets,
users could attempt to game the system, by voting based on
popularity and hence obtaining a high reward. This is also a
problem that stems from the objective of the two methods,
namely maximizing the number of clicks for each item and
not necessarily considering the element of novelty brought
into the system by each individual.
8
0123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657Popularity-sorted Movie Buckets0.00.51.01.52.02.53.03.54.0Average Rating0123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657Popularity-sorted Movie Buckets0200040006000800010000Votes / Clicks V. CONCLUSION
The current business model for online services is reaching
the limits of sustainability. The exchange of free services for
free data raises concerns both in terms of safety, with users
allowing highly sensitive private information to be used as a
currency, and also in terms of fairness, as an ever increasing
number of voices call for an evaluation of the real worth of
an individual's online impact.
In the present paper, we have proposed a framework to
determine the value of every user for an online service,
based on their contribution to a quantity or metric of interest.
The proposed framework must be fair, such that users are
rewarded proportionally to their contribution and transparent,
the payments need to be interpretable and linked to user
behaviour on the platform.
By restricting ourselves to the specific case of a rec-
ommender system, we aim to gain an understanding of
what type of behaviour increases the overall worth of an
individual to the service. Once the profile of valuable users
has been studied on a series of particular cases, our goal is to
generalize and establish value in a broader sense. A second
direction of future research is the creation of a transparent
accounting meta-data layer that users can access to make
sure that they have been fairly compensated, and do that in
a manner that does not harm their privacy by leaking data.
As our goal is to ultimately expand our analysis to a broad
spectrum of domains,
the proposed Contribution-Reward
framework aims to be as wide-reaching as possible. There are
no mentions regarding the types of data or its use cases. The
framework simply states that, once a set of important features
or relevant metrics can be identified, based on data type and
use case, a value hierarchy for users can be established. The
selection of the metrics or features is done in terms of their
impact on the overall revenue produced by the service, and
this selection may not be a trivial procedure. Once this has
been achieved, however, the second step of the framework is
the ranking of users, and this can be done with algorithms
similar to the ones proposed in our case study. The obstacles
reside not only in determining the value of an individual, but
also in ensuring the safety of their data and fairness of their
remuneration. Once every user has been assigned a score, the
final step of the framework consists in mapping the scores
to actual financial quantities. This is certainly not a static
map, as it may depend on a series of continuously-varying
parameters, such as revenue produced, amount of revenue
to be distributed among users and possibly even exchange
rates between different currencies. In this paper we cover
the second problem, which is concerned with ranking.
Following this line of reasoning, we have presented two
algorithms for assigning proportional value to users, for a
specific task. We then saw that these algorithms do agree
on a number of points, despite the fact that both rely on
approximations. We also introduced the exact method for
determining proportional contributions, the Shapley value,
which in its raw form does not scale to large coalitions,
but can serve as a reference or the basis for approximate
9
methods.
The predictions of the two algorithms are in line with
common-sense assessments of the value of a particular user,
for example with respect to their votes for the most popular,
and hence revenue-producing movies. While this does offer
further validation of the two methods, their limitations have
also been highlighted, and these stem from the fact that
both are, to different degrees, estimates of an exact but
uncomputable estimate.
in its infancy,
As this field, which may aptly be named Data Economics
is still
the number of research directions
is rather considerable. One such path would be to treat
other particular cases, such as traffic data and the case of
a traffic application, and understand the reasons behind the
difference in value for various types of users. One can then
hope that, having understood a number of such cases, they
will provide some intuition into possible generalizations. A
different research direction would be to address the other
two components of the reward framework, namely feature
or metric identification and score-to-currency mapping.
An important aspect of our future work is designing and
implementing a layer of transparency in the Contribution-
Reward framework, with the explicit aim of allowing users to
verify the manner in which their repayments are distributed
and connected to their behaviour. Transparency is a funda-
mental characteristic, as it will both increase user confidence
in regard to participating on a platform which implements the
proposed framework; and growing confidence will, in turn,
boost user presence and, hence, generated revenue.
The purpose of the present paper has been to first pose a
series of essential questions in Data Economics, especially
related to fair credit assignment, and to show that, for a
particular use case, a transparent algorithmic solution can
be found, based on a commonly-accepted economic method,
the Shapley value. It is our hope that, as our data economy
reaches maturity, the remaining open problems in the field
will soon be addressed.
REFERENCES
[1] Posner E. A. and Weyl E. G., Radical Markets Uprooting Capitalism
and Democracy for a Just Society. Princeton University Press, 2018,
ISBN 9780691177502.
[2] Lanier J., Who owns the future?. Simon & Schuster, 2013, ISBN
9781451654967.
[3] Moody D. L. and Walsh P., Measuring the value of information-an
asset valuation approach. ECIS, (496 -- 512), 1999.
[4] King K., A case study in the valuation of a database. Journal of
Database Marketing & Customer Strategy, 14, 2, (110 -- 119), 2007.
https://doi.org/10.1057/palgrave.dbm.3250041
[5] Alstyne, M. W., A Proposal for Valuing Information and Instrumental
Goods. Proceedings of the 20th International Conference on Informa-
tion Systems, 20, (328 -- 345), 1999.
[6] Gyarmati L., Laoutaris N., Sdrolias K., Rodriguez P. and Courcou-
betis C., From advertising profits to bandwidth prices-A quantitative
methodology for negotiating premium peering. NetEcon'14, 2014.
"arXiv:1404.4208v4
[7] Delaney K. J., The robot that takes your job should pay taxes, says
Bill Gates. 2017. Retrieved from: https://qz.com/911968/bill-gates-the-
robot-that-takes-your-job-should-pay-taxes/
[8] Thomas L., Universal basic income debate sharpens as observers
from:
grasp
https://www.cnbc.com/2017/03/25/universal-basic-income-debate-
sharpens.html
2017. Retrieved
inequality.
solutions
for
to
[9] Gillespie
P.,
Mark
income. What
basic
https://money.cnn.com/2017/05/26/news/economy/mark-zuckerberg-
universal-basic-income/index.html
Retrieved
2017.
it?.
is
Zuckerberg
supports
universal
from:
and Context. ACM Transactions on Interactive Intelligent Systems, 5,
4, Article No. 19, 2016.
[16] Shapley L. S., A Value for n-Person Games. Annals of Mathematics
Study, 28, (307 -- 317), 1953.
[10] Cabello S. and Chan T. M., Computing Shapley values in the plane.
[17] Winter E., The Shapley Value. Published in The Handbook of Ga-
meTheory, North-Holland, 2002, ISBN 9780444894281.
[18] Papadimitriou C. H., Computatational Complexity, Addison-Wesley,
1994, ISBN 9780201530827.
[19] Deng X. and Papadimitriou C. H., On the Complexity of Cooperative
Solution Concepts. Mathematics of Operations Research, 19, 2, (257
-- 266), 1994.
[20] Bachrach Y., Elkind E., Meir R., Pasechnik D., Zuckerman M. Rothe
J. and Rosenschein J. S., The Cost of Stability in Coalitional Games.
Proceedings of SAGT, (112 -- 134), 2009.
[21] Stanojevic R., Laoutaris N. and Rodriguez P., On Economic Heavy
Hitters: Shapley value analysis of 95th-percentile pricing. IMC'10
Melbourne, Australia, November 1-3, 2010.
2018. arXiv:1804.03894
[11] Zhao K., Mahboobi S. H. and Bagheri S. R., Shapley Value
Methods for Attribution Modeling in Online Advertising. 2018.
arXiv:1804.05327
[12] Blackwell D., Equivalent Comparisons of Experiments. Annals of
Mathematical Statistics, 24, (265 -- 272), 1953.
[13] Carrascal J. P., Riederer C., Erramilli V., Cherubini M. and Oliveira
R., Your browsing behavior for a big mac: economics of personal
information online. Proceedings of the 22nd international conference
on World Wide Web, (189 -- 200), 2013.
[14] Acquisti A., Taylor C. R. and Wagman L., The Economics of Privacy.
Journal of Economic Literature, 54, 2, 2016.
[15] Harper F. M. and Konstan J. A., The MovieLens Datasets: History
10
|
1605.07861 | 1 | 1605 | 2016-05-25T12:46:02 | Consensus in the Presence of Multiple Opinion Leaders: Effect of Bounded Confidence | [
"cs.SI",
"cs.MA",
"eess.SY"
] | The problem of analyzing the performance of networked agents exchanging evidence in a dynamic network has recently grown in importance. This problem has relevance in signal and data fusion network applications and in studying opinion and consensus dynamics in social networks. Due to its capability of handling a wider variety of uncertainties and ambiguities associated with evidence, we use the framework of Dempster-Shafer (DS) theory to capture the opinion of an agent. We then examine the consensus among agents in dynamic networks in which an agent can utilize either a cautious or receptive updating strategy. In particular, we examine the case of bounded confidence updating where an agent exchanges its opinion only with neighboring nodes possessing 'similar' evidence. In a fusion network, this captures the case in which nodes only update their state based on evidence consistent with the node's own evidence. In opinion dynamics, this captures the notions of Social Judgment Theory (SJT) in which agents update their opinions only with other agents possessing opinions closer to their own. Focusing on the two special DS theoretic cases where an agent state is modeled as a Dirichlet body of evidence and a probability mass function (p.m.f.), we utilize results from matrix theory, graph theory, and networks to prove the existence of consensus agent states in several time-varying network cases of interest. For example, we show the existence of a consensus in which a subset of network nodes achieves a consensus that is adopted by follower network nodes. Of particular interest is the case of multiple opinion leaders, where we show that the agents do not reach a consensus in general, but rather converge to 'opinion clusters'. Simulation results are provided to illustrate the main results. | cs.SI | cs | IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
1
Consensus in the Presence of Multiple Opinion
Leaders: Effect of Bounded Confidence
Ranga Dabarera, Student Member, IEEE, Kamal Premaratne, Senior Member, IEEE,
Manohar N. Murthi, Member, IEEE, and Dilip Sarkar, Senior Member, IEEE
c(cid:13)2016 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works,
for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
6
1
0
2
y
a
M
5
2
]
I
S
.
s
c
[
1
v
1
6
8
7
0
.
5
0
6
1
:
v
i
X
r
a
Abstract-The problem of analyzing the performance of net-
worked agents exchanging evidence in a dynamic network has
recently grown in importance. This problem has relevance in
signal and data fusion network applications and in studying
opinion and consensus dynamics in social networks. Due to
its capability of handling a wider variety of uncertainties and
ambiguities associated with evidence, we use the framework of
Dempster-Shafer (DS) theory to capture the opinion of an agent.
We then examine the consensus among agents in dynamic net-
works in which an agent can utilize either a cautious or receptive
updating strategy. In particular, we examine the case of bounded
confidence updating where an agent exchanges its opinion only
with neighboring nodes possessing 'similar' evidence. In a fusion
network, this captures the case in which nodes only update their
state based on evidence consistent with the node's own evidence.
In opinion dynamics, this captures the notions of Social Judgment
Theory (SJT) in which agents update their opinions only with
other agents possessing opinions closer to their own. Focusing
on the two special DS theoretic cases where an agent state
is modeled as a Dirichlet body of evidence and a probability
mass function (p.m.f.), we utilize results from matrix theory,
graph theory, and networks to prove the existence of consensus
agent states in several time-varying network cases of interest.
For example, we show the existence of a consensus in which a
subset of network nodes achieves a consensus that is adopted
by follower network nodes. Of particular interest is the case of
multiple opinion leaders, where we show that the agents do not
reach a consensus in general, but rather converge to 'opinion
clusters'. Simulation results are provided to illustrate the main
results.
Index Terms-Consensus, opinion leaders, bounded confidence,
social judgement theory.
I. INTRODUCTION
Motivation. In multi-agent systems like social networks or
a group of mobile robots completing a task, agent behavior is
determined by local interactions between neighboring agents.
In distributed control problems, reaching a consensus state is
often necessary for the agents to achieve a control objective
Manuscript submitted in September 2015. This work is based on research
supported by the U.S. Office of Naval Research (ONR) via grant #N00014-10-
1-0140, and the U.S. National Science Foundation (NSF) via grant #1343430.
RD, KP, and MNM are with the Department of Electrical and Com-
puter Engineering of the University of Miami, Coral Gables, Florida (e-
mail: {ranga, kamal, mmurthi}@miami.edu); DS is with the Department of
Computer Scienceof the University of Miami, Coral Gables, Florida (e-mail:
[email protected]).
[1], [2]. Within the context of distributed sensor fusion, the
agents exchange evidence to eventually have a good estimate
of unknown parameters [3]. Agents in a social network typi-
cally exchange opinions in some manner with their neighbors,
and update their own opinions, possibly resulting in consensus.
The work in [4] discusses models of opinion diffusion among
rational agents and social learning, explaining how opinions
emerge or fail to emerge in contexts of economics and social
sciences. Understanding the conditions under which a group
of agents will reach a possible consensus in a network with
time-varying links is a challenging problem that has recently
garnered much interest.
Previous Work. Previous work within signal processing,
control, and data fusion geared towards analyzing agent con-
sensus formation has mostly modeled an agent state as a real-
valued vector [1]–[3], [5], [6] (see also [7] and references
therein). While such an assumption is useful in the context of
agents achieving consensus on a control vector, or sensing
network agents achieving a consensus signal estimate, the
assumption of a real-valued vector may not necessarily be
always suitable within higher-level fusion, or the analysis of
opinion dynamics in social networks. In higher-level fusion,
agents could reach a consensus assessment regarding a situ-
ation, while in opinion dynamics, agents gravitate to reach a
possible consensus opinion. The inherent uncertainty within
such applications may necessitate a more structured agent
state vector such as a probability mass function (p.m.f.) or an
imprecise probabilisitic formalism, such as Dempster-Shafer
(DS) theory [8].
Agent states/opinions in complex fusion environments and
social networks involve more qualitative and nuanced informa-
tion and hence are associated with higher and a wider variety
of uncertainties. Uncertainty modeling formalisms based on
imprecise probability, such as DS theory, provide a conve-
nient framework for such scenarios. The DS theoretic (DST)
framework stands apart from alternate uncertainty handling
frameworks (such as, fuzzy sets, rough sets, etc.) in that it
bears a close relationship to the probabilistic framework [9].
DST models converge to and/or can be converted to p.m.f.
models in limiting cases, thus allowing for a swift general-
ization of legacy techniques. Moreover, DST models provide
a more intuitive and convenient approach for handling the
0000–0000/00$00.00 c(cid:13) 2016 IEEE
2
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
types of uncertainties and the nuances that are characteristic of
agent states and opinions in complex fusion environments and
social networks [7], [10]. For example. complete ignorance
can be captured via the 'vacuous' model and lack of evidence
can be captured by allowing evidential support for non-
singleton propositions (see Section II). For the variety of data
imperfection types that DST models can conveniently capture,
see Table 1 in [11]. However, much of the work on opinion
dynamics in social networks has focused on modeling the
agent states as either scalar real numbers, or a vector of real
numbers [12], e.g., see the Hegselmann-Krause (HK) model
[13], [14] and the Deffuant-Weisbuch (DW) model [15], [16].
In our recent work [7], [17], we utilized the DST framework
to address consensus formation in asynchronous dynamic/ad-
hoc networks with applications to high-level fusion networks,
and opinion dynamics in social networks. By utilizing DS
theory, the methods by which an agent models an opinion
and updates its opinion are now equipped with a powerful
mechanism for grappling with the uncertainty inherent
in
the problem, whether in the form of vague agent opinions,
or imprecise data in a fusion network (e.g., vague witness
statements). In [7], we presented the foundations of ana-
lyzing opinion dynamics and revision of agent state beliefs
in asynchronous dynamic/ad-hoc, demonstrating conditions
under which consensus is achieved.
Our more recent work in [17] utilized ideas from psy-
chology, namely Social Judgment Theory (SJT) [18], which
examines the basic psychological processes underlying the
expression of attitudes and their modifiability through com-
munication. In particular, in [17], we specifically used the
notion of bounded confidence which stems from the concept of
latitude of acceptance in SJT. In a social network, agents may
only communicate and exchange opinions with their neighbors
who have similar opinions on a particular topic. In other
words, an agent may be willing to update its opinion with
the neighboring agent's opinion only if the 'distance' to that
opinion is within a certain bound of confidence. The opinion
exchange models in the HK and DW models account for
these bounded confidence notions. The work in [19] addresses
statistical estimation of the bound of confidence. In [20] agents
are treated as Bayesian decision-makers and Bayes' risk error
has been used to estimate the bounds. The bounded confidence
assumption is also useful within high-level fusion networks to
capture situations where agents may only exchange evidence
with agents that have states similar to their own and perhaps
avoiding the use of faulty outlier sensors/agents.
Contributions. As in [7], [17], this present work of ours on
agent consensus formation in dynamic networks is formulated
within the DST framework. This allows us to embrace a wider
variety of uncertainties and nuances that are more representa-
tive of agent states/opinions in complex fusion networks and
social networks.
Taking this DST setting as a springboard, for the main
contributions of this work, we focus on two special cases of
DST agent opinion models, viz., opinions modeled as a p.m.f.
and a Dirichlet body of evidence (which allows one to retain
DS theory's ability to capture complete ignorance with only
a slight increase in computational complexity) [21]. Utilizing
results from matrix theory, graph theory, and networks, we
provide a thorough theoretical analysis of the conditions for
consensus formation among a group of agents residing within
a dynamic network (i.e., the network connectivity graph is
dynamic, so that links can appear, or disappear, meaning that
the agents may not have the same neighbors at each time
instant). Of particular value are the following:
(1) Each agent may employ either cautious or recep-
tive opinion update strategies, which essentially determines
whether the agent acts as an opinion 'leader' or an opinion
'follower.'
(2) The agents employ bounded confidence notions so that
an agent exchanges and updates its state/opinion with only
those neighbors whose opinions are similar (as measured by
a suitable norm) to its own opinion.
(3) The notion of opinion dynamic chains and corresponding
results being presented in this work capture network topolo-
gies, and conditions under which, where one or two groups
of agents interact with another group of agents to drive them
toward a consensus opinion or opinion clusters.
(4) Focusing on the two cases of p.m.f. and DST Dirichlet
agent opinion models, the behavior of the network in the pres-
ence of multiple opinion leaders is explored. The analytical
results require significantly less restrictive assumptions than
those in [17], which, to the authors' knowledge, is the only
other work that deals with multiple opinion leaders.
(5) These general analytical results are then used to study
how opinion leaders interact with opinion followers. In par-
ticular, it is shown that the arrival at a consensus opinion or
opinion clusters is dependent on the the number of opinion
leaders present and the agents' bounds of confidence.
(6) Several cases of interest in both opinion dynamics and
fusion domains are studied. This includes the the leader-
follower problem [22] where a subset of the networked agents
achieve an opinion cluster which is then followed and adopted
by the remaining agents.
(7) Finally, experimental evaluations are carried out
to
explore whether these results on p.m.f. and Dirichlet agent
opinions apply to and are valid for agent opinions represented
via general DST models.
As elaborated upon earlier, DST framework offers a more
intuitive and convenient avenue to capture the types of uncer-
tainties and the nuances associated with agent opinions. This
current work and the results regarding p.m.f. and Dirichlet
agent opinions constitute an important step toward understand-
ing how agents whose states/opinions are captured via more
general DST models behave in terms consensus and opinion
cluster formation.
Organization of Paper. The paper is organized as fol-
lows: Section II covers basic background material. Section III
presents our opinion update model with the conditional update
equation (CUE) as the governing equation. Our main theoret-
ical contribution on opinion dynamic chains driven by one
group (1-ODC) and opinion dynamic chains driven by two
groups (2-ODC) are in Section IV where we focus on two
special cases of the DST agent opinion model: Section IV-A
deals with agent opinions captured via p.m.f.s which would
appeal to readers with background knowledge in probabilistic
DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
3
opinion representation; Section IV-B deals with DST Dirichlet
agent opinions. Section V provides experimental validation of
our results, including their applicability in scenarios where
agent opinions are captured via more general DST models.
II. PRELIMINARIES
We use N and R to denote the integers and reals, re-
spectively. Subscript ((cid:5))≥0 attached to these are their non-
negative counterparts; R[0,1] denotes the reals taking val-
ues in [0, 1]. The superscripts ((cid:5))N and ((cid:5))M×N denote N-
sized vector and (M × N )-sized matrix counterparts. For
(cid:80)N
X = {Xij} ∈ RM×N , (cid:107)X(cid:107) denotes its ∞-norm,
i.e.,
j=1 Xij. We use X > 0 and X ≥ 0 to
(cid:107)X(cid:107) = maxi∈1,M
denote a matrix/vector with positive and non-negative entries,
respectively. A matrix whose entries are all 0 s (and is of
compatible size) is denoted by 0; an N-element vector whose
elements are all 1 s is denoted by 1N .
such that (cid:80)
We use Θ to denote the cardinality of set Θ.
Relevant Notions from DS Theory. In DS theory, the frame
of discernment (FoD) refers to the set Θ = {θ1,··· , θM}
consisting of the mutually exclusive and exhaustive 'singleton'
propositions θi [8]. A singleton proposition θi ∈ Θ represents
the lowest level of discernible information; Θ = M is the
number of independent singleton propositions in the FoD Θ.
The power set of the FoD, 2Θ = {A : A ⊆ Θ}, denotes all the
possible subsets of Θ. For A ⊆ Θ, A denotes all singletons
in Θ that are not in A.
Basic Belief Assignment (BBA). A basic belief assignment
(BBA) or mass assignment is a mapping m(·) : 2Θ (cid:55)→ [0, 1]
A⊆Θ m(A) = 1 and m(∅) = 0. The BBA
measures the "support" assigned to proposition A ⊆ Θ.
Propositions that receive non-zero mass are referred to as focal
elements. The set of focal elements is the core F. The triplet
E = {Θ,F, m} is referred to as the body of evidence (BoE).
The mass vector corresponding to the BoE E is
m = [m(∅), m(θ1), . . . , m(θM ), m(θ1θ2), . . . , m(θ1θM ),
m(θ1θ2θ3), . . . , m(θ1θ2θM ), . . . , m(Θ)] ∈ R2M
[0,1]. (1)
Note that, masses are assigned to all propositions in 2Θ. DS
theory captures the notion of ignorance by allowing composite
(i.e., non-singleton) propositions to be focal elements, e.g., the
mass assignment m(θiθj) > 0 represents ignorance or lack
of evidence to differentiate between the singletons θi and θj.
The vacuous BBA which Θ as its only focal element (so that
m(Θ) = 1) captures complete ignorance.
Belief and Plausibility. Given a BoE, E ≡ {Θ,F, m}, the
B⊆A m(B),
represents the total support committed to A without also being
committed to its complement A. The plausibility Pl : 2Θ (cid:55)→
[0, 1], defined as Pl(A) = 1 − Bl(A), corresponds to the total
belief that does not contradict A. The uncertainty of A is
Θ : Bl(A) > 0}.
Un(A) = [Bl(A), Pl(A)]. We also use the notation (cid:98)F = {A ⊆
belief Bl : 2Θ (cid:55)→ [0, 1], defined as Bl(A) = (cid:80)
A BoE is called Bayesian if each focal element is a single-
ton. For a Bayesian BoE, the BBA, belief, and plausibility, all
reduce to the same probability (i.e., p.m.f.) assignment. The
expressive power that the DST framework wields a high cost:
For a given FoD Θ, where Θ = M, a DST model allocates
2M −2 mass assignments; only M −1 probability assignments
are required for a p.m.f. Much advances have been made for
the purpose of mitigating the associated computational burden
[23]–[25]. A special DST model which retains the ability to
capture complete ignorance with only a slight increase in
computational complexity is the Dirichlet BoE (so named
because of its close relationship with Dirichlet probability
distributions [21]). The singletons {θi} and Θ constitute the
only focal elements of a Dirichlet BoE, thus requiring only
M mass assignments.
DST Conditionals. Of the various notions of DST condi-
tionals that abound in the literature, the Fagin-Halpern (FH)
conditional offers a unique probabilistic interpretation and
constitutes a natural transition to the Bayesian conditional
notion [9], [25]. The extensive study in [26] identifies sev-
eral attractive properties of the FH conditionals including its
equivalence to other popular notions of DST conditionals.
Definition 1 (FH Conditionals). For the BoE E = {Θ,F, m}
[0, 1] and the conditional plausibility Pl(BA) : 2Θ (cid:55)→ [0, 1]
of B given A are, respectively,
and A ⊆ Θ s.t. A ∈ (cid:98)F, the conditional belief Bl(BA) : 2Θ (cid:55)→
Bl(BA) = Bl(A ∩ B)/[Bl(A ∩ B) + Pl(A ∩ B)];
Pl(BA) = Pl(A ∩ B)/[Pl(A ∩ B) + Bl(A ∩ B)].
The conditional core theorem [25] can be utilized to directly
identify the conditional focal elements to improve computa-
tional performance when applying FH conditionals.
DST Distance Measure. In our work, we need a distance
measure which captures the closeness between DST BoEs.
Among possible alternatives that have appeared in the litera-
ture, we use the DST distance measure in [27], [28].
Definition 2 (Distance Between BoEs). [27] The distance
between the two agent BoEs Ei = {Θ,Fi, mi} and Ej =
{Θ,Fj, mj} is
(cid:107)Ei − Ej(cid:107)J =
where mi, mj ∈ R2M
≥0 are the mass vectors associated with the
BoEs Ei and Ej, respectively; D = {dmn} ∈ R2M×2M
, with
dmn = Am∩ An/Am∪ An, Am, An ⊆ Θ, with ∅∩∅/∅∪
∅ ≡ 0.
(cid:105)1/2 ∈ R[0,1],
T D (mi − mj)
0.5 (mi − mj)
(cid:104)
≥0
Relevant Notions from Graph Theory. [29] We use Gk =
(V, Ek) to denote a directed graph at discrete-time (DT) instant
k ∈ N≥0. Here, eij ∈ Ek represents a unidirectional edge
from node Vj ∈ V to node Vi ∈ V . We use Ak to identify the
(N × N ) adjacency matrix associated with Ek.
Consider the directed graph Gk = (V, Ek). The out-
component of vertex Vi ∈ V is the set of vertices (including
vertex Vi itself) reachable via directed paths from vertex Vi.
The in-component of vertex Vi ∈ V is the set of vertices
(including vertex Vi itself) from which vertex Vi is reachable
via directed paths.
that X ∈ RN×N
and (cid:107)X(cid:107) = 1; we say that X ∈ RN×N
Relevant Notions from Stochastic Matrix Theory. We say
j=1 Xij = 1, ∀i ∈ 1, N
is sub-stochastic if
is stochastic if (cid:80)N
[0,1]
[0,1]
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
4
(cid:80)N
j=1 Xij ≤ 1, ∀i ∈ 1, N, and ∃ i ∈ 1, N s.t.(cid:80)N
j=1 Xij < 1.
Stochastic/sub-stochastic matrices and the limiting behavior of
their products play a critical role in our work.
III. DST MODELING OF OPINION DYNAMICS
In the work that follows, we consider N agents embedded
in the directed graph Gk = (V, Ek). Here, for i ∈ 1, N,
the node Vi ∈ V in Gk represents the i-th agent and the
directed edge eij ∈ Ek represents an unidirectional informa-
tion exchange link from the j-th agent to the i-th agent (i.e.,
the i-th agent can receive information from the j-th agent).
Unless otherwise mentioned, the opinion of the i-th agent at
time instant k ∈ N≥0 is taken to be captured via the BoE
Ei,k = {Θ,Fi,k, mi(·)k}, i ∈ 1, N. We assume that the agent
opinion BoEs are associated with the identical FoD Θ.
Definition 3 (Opinion Profile). Consider the agent BoEs
Ei,k = {Θ,Fi,k, mi(·)k}, i ∈ 1, N , k ∈ N≥0. The opinion
profile of B ⊆ Θ at k ∈ R≥0 is
πππ(B)k = [m1(B)k, . . . , mN (B)]T ∈ RN
[0,1],
with πππ(B)0 ∈ RN
[0,1] denoting its initial state.
A. Bounded Confidence
agents at time instant k:
For each agent, define the following sets of neighborhood
=(cid:8)Vj ∈ V : j ∈ 1, N , and eij ∈ Ek
Ni,k
Ni,k(εi) = {Vj ∈ Ni,k : (cid:107)Ei,k − Ej,k(cid:107)J ≤ εi} ,
(cid:9) ;
(2)
where (cid:107) · (cid:107)J refers to the distance measure in Definition 2
(while any valid norm applicable for DST BoEs could be
used); εi ≥ 0 is the latitude of acceptance or bound of
confidence associated with the i-th agent. So, Ni,k(εi) denotes
the neighbors of the i-th agent at time k left after 'pruning'
the links subjected to the bound of confidence requirement.
With εεε = [ε1, . . . , εN ]T , let
G†
k(εεε) = (V, E
†
k(εεε)),
(3)
where E
†
k(εεε) = {eij ∈ Ek : (cid:107)Ei,k − Ej,k(cid:107)J ≤ εi}.
The bounded confidence process of updating an agent's
opinion is as follows [5], [12]: the i-th agent updates its BoE
Ei in response to the opinion BoE Ej of its neighbor, the j-th
agent, only if j ∈ Ni(εi). In [15], the threshold εi is referred
to as an openness character. Another interpretation views εi
as an uncertainty, i.e., if the i-th agent possesses an opinion
with some degree of uncertainty εi, then it ignores the views
of those neighbor agents who fall outside its uncertainty range.
B. Opinion Updating and Consensus Formation
Opinion Updating. In what follows, we will use S to
identify the indices corresponding to a subset of the agents
in V , i.e., S ⊆ {1, 2, . . . , N} s.t. Vi, ∀i ∈ S, identifies a
subset of agents in V . To proceed, we adopt the following
Definition 4 (Opinion Clusters, Consensus). Let Ei,k, i ∈
1, N , k ∈ N≥0, denote the opinions of N agents embed-
ded within the network Gk = (V, Ek), where each agent
repeatedly updates its state by iterative opinion exchange. Let
S ⊆ {1, 2, . . . , N} identify a subset of the agents in V .
(i) Suppose limk→∞ (cid:107)Ei,k − Ej,k(cid:107)J = 0, ∀i, j ∈ S (or
limk→∞ Ei,k ≡ E∗, ∀i ∈ S) and suppose
equivalently,
limk→∞ Ei,k (cid:54)= E∗, ∀i ∈ 1, N \ S. Then, the agents in S
are said to form an opinion cluster.
(ii) The agents are said to reach a consensus if S = V , i.e.,
all the agents in V form a single opinion cluster.
Henceforth, our results will only be stated with the forma-
tion of a consensus in mind (e.g., see Lemma 1). Due to (ii)
above, these results can easily be reformulated so that they
pertain to the formation of opinion clusters.
In our work, we assume that each agent updates its opinion
in accordance with the conditional update equation (CUE):
Definition 5 (Conditional Update Equation (CUE)). [7], [30],
[31] Suppose the i-th agent updates its opinion Ei by taking
its neighboring agents j ∈ Ni,k(εi), where
into account
εi > 0 is the i-th agent's bound of confidence. The CUE-
based updated opinion of the i-th agent is
Bli(B)k+1 = αi,kBli(B)k +
βij(A)kBlj(BA)k.
Here, the index set Ii,k identifies the agents in Ni,k(εi); the
CUE parameters αi,k, βij(·)k ∈ R[0,1] satisfy
(cid:88)
j∈Ii,k\i
A∈ Fj,k
(cid:88)
(cid:88)
(cid:88)
αi,k +
βij(A)k = 1.
j∈Ii,k\i
A∈ Fj,k
The above CUE can also be expressed directly in terms
of DST mass values. Alternately, the updated mass values
can be computed from the updated belief values. Note that,
with agents utilizing CUE-based opinion updating, the edge
'weights' along both directions of an edge are not equal in
general because the CUE weights βij(·) and βji(·) are not
necessarily the same. It is for this reason that we treat the
underlying graph Gk as directed.
The work in [7], [30], [31] also provides various strategies
for selecting the CUE parameters. Of particular importance
are two particular strategies:
Definition 6 (Receptive and Cautious Update Strategies).
Consider the i-th agent updating its opinion Ei according to
the CUE-based update in Definition 5.
(i) The i-th agent is said to employ receptive updating if
βi,j(A)k ∝ mj(A)k, and receptively updating agents are
referred to as opinion leaders.
(ii) The i-th agent is said to employ cautious updating
if βij(A)k ∝ mi(A)k, and cautiously updating agents are
referred to as opinion followers.
Receptive updating 'weighs' the incoming evidence accord-
ing to the support each focal element receives from the incom-
ing BoEs. This has an interesting Bayesian interpretation: it
reduces to a weighted average of the p.m.f.s of the BoEs.
Cautious updating 'weighs' the incoming evidence according
to the support each focal elements receives from the BoE being
updated. We consider three cases:
DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
5
(1) No Opinion Leaders: This is the most common scenario
that appears in the literature [12], [14], [15]. Here, all agents
are receptively updating and no opinion leaders are present.
(2) Single Opinion Leader: Here, all agents employ recep-
tive updating except one opinion leader. This is the scenario
considered in typical
leader-follower models [6], and the
recent work in [7]).
(3) Multiple Opinion Leaders: Here,
there are multiple
cautiously updating opinion leaders, generally with different
initial opinions. To our knowledge, this case has not been
addressed prior to the work in [17] which was based on
the work in [13], [32]. It provides sufficient conditions for
consensus/cluster formation under certain strong assumptions
(e.g., the agent opinions can be ordered in what is referred to
as an εεε-chain). We will relax such assumptions by utilizing
properties of products of stochastic/sub-stochastic matrices.
IV. OPINION DYNAMIC CHAINS WITH OPINION LEADERS
While the DST framework provides a powerful tool for
grappling with the types of uncertainties and nuanced infor-
mation prevalent in complex fusion and social networks, the
expressive power inherent in DST models levy a high compu-
tational burden. A Dirichlet BoE [21] is a special type of DST
BoE which retains DS theory's ability to capture complete
ignorance with only a slight increase in computational effort.
With our DST opinion model in place, we now focus on
probabilistic (i.e., p.m.f.) and DST Dirichlet agent opinions.
αi,k,
A. Probabilisitic Agent Opinions
With the DST BoEs Ei,k, i ∈ 1, N , k ∈ N≥0, possessing
only singleton focal elements, we have probabilistic agent
opinions. In this case,
the CUE-based opinion update in
Definition 5 reduces to the DT dynamic system
πππ(θp)k+1 = Wk πππ(θp)k, p ∈ 1, M ,
(4)
where πππ((cid:5))k ∈ RN
is row-
stochastic [33]. When all agents employ receptive updating,
[0,1], Wk = {wij,k} ∈ RN×N
[0,1]
wij,k =
(1 − αi,k)/Ni,k(εi),
0,
for i = j;
for j ∈ Ni,k(εi);
otherwise.
(5)
As in [34], we refer to Wk as the confidence matrix because
wij,k represents the weight
the i-th agent attaches to the
opinion of the j-th agent at
time step k. Note that, Wk
constitutes the weighted adjacency matrix of G†
k(εεε) in (3).
We proceed with
Proposition 1. In an environment where agents possess prob-
abilistic agent opinions, the opinion of a cautiously updating
agent is invariant.
Proof. Suppose the i-th agent employs a cautiously updating
strategy. Its CUE parameters satisfy (see Definition 6)
otherwise [25]. Using this and the fact that for singletons
propositions belief and masses are equal, the CUE-based up-
date in Definition 5 for the i-th agent reduces to mi(B)k+1 =
mi(B)k, ∀k ∈ N≥0.
Definition 7 (Left (or Backward) Products). [35]
{Wk}, Wk ∈ RN×N , is
(i) Left Product: The left product of the sequence of matrices
I,
Wk:(cid:96) =
W(cid:96),
WkWk−1 ··· W(cid:96),
for k < (cid:96);
for k = (cid:96);
for k > (cid:96).
(ii) Left-Converging Product: The left product Wk:0 is said
to be left-converging if limk→∞ Wk:0 exists, in which case we
write W∞ = limk→∞ Wk:0.
Note that the dynamic system in (4) can be expressed as
πππ(θp)k+1 = Wk:0 πππ(θp)0. Thus, whenever W∞ exists,
k→∞ πππ(θp)k+1 = W∞ πππ(θp)0.
lim
(6)
Clearly, the convergence of the agent opinions depends on
the existence and the nature of W∞. When W∞ exists, let us
denote the converged opinion profile for θp as πππ∗(θp) ∈ RN
[0,1].
Consensus (in the sense of Definition 4) is a special case of
a converged opinion profile.
Lemma 1. The agents form a consensus iff ∃ ηηη = {ηp} ∈
[0,1] s.t., πππ∗(θp) = ηp1N ,∀p ∈ 1, M , ∀i ∈ 1, N.
RM
Proof. Suppose πππ∗(θp) = ηp1N ,∀p ∈ 1, M , ∀i ∈ 1, N,
for some ηηη = {ηp} ∈ RM
[0,1]. This clearly means that
limk→∞ Ei,k ≡ E∗, ∀i ∈ 1, N. Thus,
the agents form a
consensus. Conversely, if limk→∞ Ei,k ≡ E∗, ∀i ∈ 1, N, we
must clearly have πππ∗(θp) = ηp1N ,∀p ∈ 1, M , ∀i ∈ 1, N.
The limiting behavior of the stochastic matrix product {Wk}
plays a crucial role in consensus analysis when DST BoEs
possess only singleton focal elements.
Lemma 2. Consider the stochastic chain {Wk}, k ∈ N≥0,
s.t. W∞ = 1 vT for some stochastic vector v ∈ RN
[0,1]. Then,
the agents reach the consensus πππ∗(θp) = (vT πππ(θp)0) 1, where
πππ(θp)0 ∈ RN
[0,1] denotes the initial opinion profile.
Proof. Use (6): limk→∞ πππ(θp)k+1 = 1 vT πππ(θp)0 = ηp1,
where ηp = vT πππ(θp)0. So, from Lemma 1, we achieve a
consensus.
1) No Opinion Leaders: This is the most widely studied
scenario and many consensus-related results applicable to this
case are available [6], [12], [34], [36], [37]. The work in
[34] studied convergence when each element in the stochastic
chain {Wk} has positive diagonals. From a graph theoretic
viewpoint, this implies that there is a path from each vertex to
itself; from an opinion dynamics perspective, this is referred
to as having the self-communicating property [34].
(cid:88)
j(cid:54)=i
(cid:88)
A∈ Fj,k
βij(B)k = µij,kmi(A)k; αi,k +
µij,k = 1,
2) Single Opinion Leader: We first introduce
for singletons A, B ∈ Θ. But, for singleton propositions
A, B ∈ Θ, Bl(BA) = 1 only if B = A, and Bl(BA) = 0
Definition 8 (Opinion Dynamics Chain Driven By One Group
†
(1-ODC)). The directed dynamic network G†
k(εεε))
in (3) is said to be an opinion dynamics chain driven by one
k(εεε) = (V, E
6
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
group (1-ODC) if its corresponding confidence matrix Wk in
(4) can be expressed as the lower block triangular matrix
(cid:20)Ak
(cid:21)
0
Ck Dk
Wk =
and Dk ∈ RNout×Nout
,
[0,1]
where Ak ∈ RNC×NC
matrices have compatible sizes.
[0,1]
, and the other
A, n ∈ N≥0, we express this as
for n ∈ N≥0. Here, we have used (10). Employing the notation
∆Pn = Pn − 1NoutvT
∆Pn+1 = Cn+1[An:0 − 1NC vT
Then, we may bound (cid:107)∆Pn+1 − Dn+1∆Pn(cid:107) as
(cid:107)∆Pn+1(cid:107) − (cid:107)Dn+1∆Pn(cid:107)
A] + Dn+1∆Pn, n ∈ N≥0.
≤ (cid:107)∆Pn+1 − Dn+1∆Pn(cid:107) = (cid:107)Cn+1[An:0 − 1NC vT
A](cid:107)
≤ (cid:107)An:0 − 1NC vT
A(cid:107).
(11)
A, where
vA is a NC-sized stochastic vector, implies that the agents in
VC converges to a consensus (see Lemma 2). Hence, given an
arbitrary A > 0, ∃NA ∈ N≥0 s.t.
We proceed by noting that limn→∞ An:0 = 1NC vT
(cid:107)An:0 − 1NC vT
A(cid:107) < A, ∀n ≥ NA.
(12)
From (11) and (12), we can obtain the following: given an
arbitrary A > 0, ∃NA ∈ N≥0 s.t.
(cid:107)∆Pn+1(cid:107) − (cid:107)Dn+1∆Pn(cid:107) < A, ∀n ≥ NA.
So, for n ≥ NA, we have
(cid:107)∆Pn+1(cid:107) < (cid:107)Dn+1∆Pn(cid:107) + A < ρ(cid:107)∆Pn(cid:107) + A,
where we use the fact that (cid:107)Dn+1(cid:107) ≤ ρ < 1, ∀n ≥ NA. This
upper bound for (cid:107)∆Pn+1(cid:107) yields
L−1(cid:88)
(cid:107)∆PNA+L(cid:107) < A
L−1(cid:88)
(cid:32)
(cid:96)=0
We note that
(cid:33)
ρ(cid:96) + ρL(cid:107)∆PNA(cid:107)
A
lim
L→∞
=
i.e., given an arbitrary P > 0, ∃LP ∈ N≥0 s.t.
(cid:96)=0
A
1 − ρ
,
ρ(cid:96) + ρL(cid:107)∆PNA(cid:107), L ≥ 0.
(13)
ρ(cid:96) + ρL(cid:107)∆PNA(cid:107) − A
1 − ρ
< P , ∀L ≥ LP .
Use this in (13): (cid:107)∆PNA+L(cid:107) < P + A/(1 − ρ), ∀L ≥ LP .
In other words,
lim
n→∞(cid:107)∆Pn(cid:107) = lim
(cid:20)πππA(θp)n+1
n→∞(cid:107)Pn − 1NoutvT
(cid:20)An:0 0NC×Nout
A(cid:107) = 0.
(cid:21)
(cid:21)(cid:20)πππA(θp)0
(cid:21)
Using (8) in (7), we get
πππD(θp)n+1
Pn
Dk:0
πππD(θp)0
=
Let us take the consensus among agents in VC as
πππ∗
A(θp) = (1NC vT
A) πππA(θp)0.
(14)
(15)
However, from (14), we know that limn→∞ Pn = 1Nout vT
A.
Then, by (15), we can write
(cid:20)πππ∗
(cid:21)
A(θp)
πππ∗
D(θp)
=
0NC×Nout
A
A 0Nout×Nout
1Nout vT
(cid:20) 1NC vT
(cid:124)
(cid:123)(cid:122)
1NC +Nout v∗T
(cid:20)πππA(θp)0
(cid:21)
πππD(θp)0
(cid:21)
(cid:125)
,
(16)
where v∗ is a (NC + Nout)-sized stochastic vector created by
concatenating the vectors vA and 0Nout×1, and
πππ∗
D(θp) = 1NoutvT
AπππA(θp)0.
(17)
L−1(cid:88)
(cid:96)=0
Proof. The opinion update strategy yields the dynamic
A
system
Fig. 1: An opinion dynamics chain driven by one group (1-
ODC).
C,k(εεε) = (VC, E
out,k(εεε) = (Vout, E
One may view an 1-ODC as consisting of a 'central'
†
C,k(εεε)) (with NC agents) and
†
out,k(εεε)) (with Nout
out,k(εεε) belongs to the in-component
C,k(εεε), for any k ∈ N≥0. See Fig. 1. Note that, Ak and
C,k(εεε)
component G†
another component G†
agents) s.t. no agent in G†
of G†
Dk correspond to the confidence matrices of agents in G†
and G†
Theorem 1. Consider agents embedded in a 1-ODC employ-
ing a CUE-based update strategy. Furthermore, suppose that
is a
stochastic vector so that the agents in VC reach their own
consensus, and
A, where vA ∈ RNC
(a) limn→∞ An:0 = 1NC vT
out,k(εεε), respectively.
(b) (cid:107)Dk(cid:107) ≤ ρ < 1, ∀k ∈ N≥0.
Then, the agents in V (i.e., agents in VC and Vout) reach
[0,1]
a consensus at the consensus reached by the agents in VC.
where k ∈ N≥0, θp ∈ Θ, and
(cid:21)
(cid:20)πππA(θp)k+1
(cid:21)
(cid:20)Ak
πππD(θp)k+1
0
Ck Dk
= Wk
,
(7)
(cid:20)πππA(θp)k
(cid:21)
(cid:20)An:0
πππD(θp)k
(cid:21)
0
Wk =
=⇒ Wn:0 =
Pn Dn:0
.
(8)
Here the sub matrices have sizes compatible with Wk. Use
Wn+1:0 = Wn+1Wn:0 to get
Pn+1 = Cn+1An:0 + Dn+1Pn, P0 = C0,
(9)
for n ∈ N≥0. Due to the row-stochasticity of Wk, we have
1Nout = Ck1NC + Dk1Nout, ∀k ∈ N≥0.
(10)
Subtract 1Nout vT
Pn+1 − 1NoutvT
A
A from both sides of (9):
= Cn+1An:0 + Dn+1Pn − 1NoutvT
= Cn+1An:0 + Dn+1Pn − [Cn+11NC + Dn+11Nout ] vT
= Cn+1[An:0 − 1NC vT
A] + Dn+1[Pn − 1Nout vT
A],
A
A
GC,k(ε)†Gout,k(ε)†Ck DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
7
Since (16) satisfies the conditions of Lemma 2, all agents in
V achieve a consensus. Moreover, as (16) and (17) show, this
consensus is the same consensus achieved within VC.
An immediate consequence of Theorem 1 is
k(εεε) = (V, E
Corollary 1. Consider the network G†
†
k(εεε)) in
(3) populated with receptively updating opinion followers
and a single cautiously updating opinion leader. Suppose
Dk corresponds to the confidence matrix of the receptively
updating opinion followers with (cid:107)Dk(cid:107) ≤ ρ < 1, ∀k ∈ N≥0.
Then, with a CUE-based update strategy, all the agents reach
a consensus opinion at the opinion of the opinion leader.
Proof. Construct a 1-ODC as in Definition 8 with G†
C,k(εεε)
containing the opinion leader only and G†
out,k(εεε) populated
†
with the opinion followers. Proposition 1 implies that V
C
(which consists of the only cautiously updating agent) gen-
erates a consensus. So, from Theorem 1, all the agents in V
reach a consensus at the opinion of the opinion leader.
Fig. 2: An opinion dynamics chain driven by two groups (2-
ODC).
3) Two Opinion Leaders: The situation turns out to be sig-
nificantly more complicated when there are multiple opinion
leaders. To look at the two opinion leader case, let us introduce
Definition 9 (Opinion Dynamics Chain Driven By Two
Groups (2-ODC)). The directed dynamic graph G†
k(εεε) =
†
k(εεε)) in (3) is said to be an opinion dynamics chain
(V, E
driven by two groups (2-ODC) if its corresponding confidence
matrix Wk in (4) can be expressed as the lower block
triangular matrix
Wk =
A(1)
k
0
C (1)
k
.
0
0
Dk
0
A(2)
k
C (2)
k ∈ RNC2×NC2
[0,1]
k
, and Dk ∈
k ∈ RNC1×NC1
[0,1]
, A(2)
Here, A(1)
RNout×Nout
; the other matrices have compatible sizes.
C1,k(εεε) = (VC1, E
out,k(εεε) = (Vout, E
[0,1]
One may view a 2-ODC as consisting of two 'central'
†
components G†
C1,k(εεε)) (with NC1 agents)
†
and G†
C2,k(εεε)) (with NC2 agents), plus a
C2,k(εεε) = (VC2, E
†
third component G†
out,k(εεε)) (with Nout
agents) s.t. no agent in G†
out,k belongs to the in-components
of either G†
C1,k(εεε) or G†
C1,k(εεε). See Fig. 2. Note that, A(1)
k ,
A(2)
k , and Dk correspond to the confidence matrices of agents
in G†
C1,k(εεε), G†
We now demonstrate that the agents in even a 2-ODC cannot
reach a consensus in general, even though the agents in VC1
and VC2 may have achieved their own consensus opinions.
C2,k(εεε), and G†
out,k(εεε), respectively.
(a)
n:0 = 1NC1vT
Theorem 2. Consider agents embedded in a 2-ODC employ-
ing a CUE-based update strategy. Furthermore, suppose that
n:0 =
[0,1] are stochastic
1NC2 vT
vectors so that the agents in VC1 and VC2 achieve their own
consensus opinions, and
limn→∞ A(1)
2 , where v1 ∈ RNC1
[0,1] and v2 ∈ RNC2
1 and limn→∞ A(2)
(b) (cid:107)Dk(cid:107) ≤ ρ < 1, ∀k ∈ N≥0.
Then, the following are true:
(i) The agents in V (i.e., agents in VC1, VC2, and Vout)
reach a consensus iff the consensus opinions of the agents in
VC1 and VC2 are equal.
(ii) When the consensus opinions of the agents in VC1
and VC2 are not equal, a consensus among the agents in
Vout occurs if ∃ λ(1)
k = 1 and
λ(1)
k C (1)
k 1NC1 = λ(2)
Proof. Note that we may write
k ∈ (0, 1) s.t. λ(1)
k 1NC2, ∀k ∈ N≥0.
k , λ(2)
k C (2)
k + λ(2)
A(1)
n:0
0
P (1)
n
,
0
A(2)
n:0
P (2)
n
0
0
Dn:0
Wn:0 =
where, for n ∈ N≥0,
(18)
(19)
P (1)
n+1 = C (1)
P (2)
n+1 = C (2)
n+1A(1)
n+1A(2)
n:0 + Dn+1P (1)
n:0 + Dn+1P (2)
n , P (1)
n , P (2)
0 = C (1)
0 ;
0 = C (2)
0 .
0
A(1)
(cid:21)
(cid:20)
(i) Suppose the two consensus opinions of the agents in VC1
and VC2 are equal. One may then recast the 2-ODC as a 1-
DOC with the confidence matrices corresponding to the central
component and the other component taken as
and
Dk, respectively. Noting that the central component reaches
a common consensus opinion, apply Theorem 1 to show that
all the agents must reach a consensus which is identical to
the common consensus opinion formed within the central
component. Conversely, if the two consensus opinions of the
agents in VC1 and VC2 are not equal, no consensus is possible
among the agents in VC1 and VC2 because these two sets of
agents do not update from each other.
k
0 A(2)
(ii) Suppose the two consensus opinions of the agents in VC1
and VC2 are not equal. The row stochasticity of Wk implies
k
(20)
k C (2)
k 1NC1 = λ(2)
k 1NC2 and λ(1)
k 1NC2 + Dk1Nout ;
k Dk1Nout, k ∈ N≥0,
k 1NC1 + C (2)
k 1NC1 + λ(2)
k C (1)
1Nout
= C (1)
k 1Nout = C (1)
λ(2)
where we used λ(1)
k +
λ(2)
k = 1. Now, proceed as we did in the proof of Theorem 1.
Subtract λ(2)
1 from both sides of (19) and substitute
k 1Nout vT
for 1Nout from (20):
n+1 − λ(2)
P (1)
= C (1)
n − λ(2)
for n ∈ N≥0. As before, we use the notation ∆P (1)
λ(2)
k 1NoutvT
n:0 − 1NC1 vT
1 ] + Dn+1[P (1)
1 ],
n −
k 1NoutvT
k 1Nout vT
n = P (1)
n+1[A(1)
1
1 to express this as
n+1 = C (1)
n+1[A(1)
n:0 − 1NC1 vT
∆P (1)
1 ] + Dn+1∆P (1)
n ,
GC1,k(ε)†Gout,k(ε)†GC2,k(ε)†Ck(1)Ck(2) 8
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
for n ∈ N≥0. Now, bound (cid:107)∆P (1)
(cid:107)∆P (1)
n+1(cid:107) − (cid:107)Dn+1∆P (1)
n (cid:107)
n+1 − Dn+1∆P (1)
n:0 − 1NC1vT
1 (cid:107),
≤ (cid:107)∆P (1)
≤ (cid:107)A(1)
n+1 − Dn+1∆P (1)
n (cid:107) as
n (cid:107) = (cid:107)C (1)
n+1[A(1)
n:0 − 1NC1 vT
1 ](cid:107)
where we used the sub-stochasticity of C (1)
n .
Now, as in the proof of Theorem 1, use the fact that the
agents in VC1 and VC2 each achieve a consensus to show that
2 (cid:107)
n→∞(cid:107)P (1)
k 1NoutvT
n→∞(cid:107)P (2)
n − λ(1)
n − λ(2)
k 1Nout vT
lim
1 (cid:107) = lim
= 0.
In other words,
B. Dirichlet Agent Opinions
Next, let us suppose that agents opinions are modeled via
Dirichlet BoEs. Then, using the properties Bl(θiθi) = 1,
Bl(θiθj) = 0, i (cid:54)= j, and Bl(BΘ) = B, ∀B ⊆ Θ, one
can easily show that the CUE-based update mechanism in
Definition 5 retains the Dirichlet property of the updated BoEs
at each step. For this Dirichlet BoE case, the CUE-based
opinion update reduces to the following DT dynamic system:
πππ(θp)k+1 = Wk πππ(θp)k, p ∈ 1, M .
(21)
Here πππ(·)k is as in (4) and Wk = { wij,k} ∈ RN×N where
wij,k, i, j ∈ 1, N , k ∈ N≥0. When the i-th agent is receptively
updating,
n = λ(2)
n→∞ P (1)
lim
Denote the consensus among the agents in VC1 and VC2 as
k 1Nout vT
1 ;
k 1Nout vT
2 .
n→∞ P (2)
lim
n = λ(1)
wij,k =
πππ∗
A(1)(θp) = (1NC1 vT
πππ∗
A(2)(θp) = (1NC2 vT
1 ) πππA(1)(θp)0;
2 ) πππA(2)(θp)0,
=
πππ∗
respectively. Then we have
1
1NC1vT
(cid:104)
0
A(1)(θp)
πππ∗
A(2)(θp)
πππ∗
D(θp)
where πππ∗
πππ∗
D(θp) = 1Nout
D(θp) is given as
0
1NC2vT
2
k 1Nout vT
λ(2)
k 1NoutvT
1 λ(1)
λ(2)
k vT
1
λ(1)
k vT
2
,
0
0
2 0
πππA(1)(θp)0
(cid:105)(cid:20)πππA(1)(θp)0
(cid:21)
πππA(2)(θp)0
πππD(θp)0
.
πππA(2)(θp)0
k
k
Hence, from Lemma 2, we conclude that the agents in Vout
reach a consensus if λ(1)
k = 1, ∀k ∈ N≥0.
k + λ(2)
and C (2)
One may interpret this result in the following manner: the
matrices C (1)
signify the 'weights' or 'bias' that
agents in Vout give to the agents in VC1 and VC2, respectively.
To reach a consensus, each agent in Vout must give the same
proportion of weights to the agents in VC1 and VC2 for all
k ∈ N≥0: λ(2)
implies that a higher weight is given to
the opinions of the agents in VC1 than to the opinions of the
agents in VC2. This might be due to stronger interconnections
between the agents in VC1 and Vout or it could simply be due
to a bias towards the opinions of the agents in VC1.
k > λ(1)
k
As an immediate consequence of Theorem 2 we get
Corollary 2. Consider the network G†
†
k(εεε)) in (3)
populated with receptively updating opinion followers and two
cautiously updating opinion leaders. Suppose Dk corresponds
to the confidence matrix of the receptively updating opinion
followers with (cid:107)Dk(cid:107) ≤ ρ < 1, ∀k ∈ N≥0. Then, with a CUE-
based update strategy, the following are true:
k(εεε) = (V, E
(i) All the agents reach a consensus iff the opinions of the
two opinion leaders are equal.
(ii) When the opinions of the opinion leaders are not equal,
the opinion followers form an opinion cluster if the proportion
of weights each receptively updating agent gives to the two
opinion leaders is identical and no opinion leader is given
zero weight.
0,
αi,k,
(1 − αi,k)(1 + mj(Θ)k)
Ni,k(εi)
,
for i = j;
for j ∈ Ni,k(εi);
otherwise;
(22)
when the i-th agent is cautiously updating,
wij,k =
1,
(1 − αi,k)mi(Θ)k
Ni,k(εi)
,
0,
for i = j;
for j ∈ Ni,k(εi);
otherwise.
(23)
While we may still refer to Wk as the corresponding confi-
dence matrix, unlike Wk, Wk is not necessarily stochastic.
To proceed, we take inspiration from [31], where it is shown
that, under mild conditions, the masses for complete ambiguity
Θ vanish when two Dirichlet agents mutually update each
other. The same result turns out to hold true for multiple BoEs.
Lemma 3. Consider the CUE-based updating of the Dirichlet
BoEs as in (21). If
αi,k +
βij(Θ)k ≤ ρ < 1, ∀i, j ∈ 1, N , ∀k ∈ N≥0,
(cid:88)
j(cid:54)=i
then limn→∞ mi(Θ)n = 0, ∀i ∈ 1, N.
Proof. Observe that the update of πππ(Θ) can be written as
where
ΓΓΓk =
πππ(Θ)k+1 = ΓΓΓkπππ(Θ)k,
β21(Θ)k
α1,k
...
β12(Θ)k
α2,k
...
β13(Θ)k
β23(Θ)k
...
βN 1(Θ)k βN 2(Θ)k βN 3(Θ)k
.
···
···
...
···
β1N (Θ)k
β2N (Θ)k
...
αN,k
The condition in the statement implies that (cid:107)ΓΓΓk(cid:107)∞ ≤ ρ <
1, ∀k ∈ N≥0, which guarantees the claim.
Note that, each BoE being updated possessing at least one
singleton focal element ensures that the condition in Lemma 3
is satisfied, which in turn, ensures that the mass for each
completely ambiguous proposition vanishes in the limit. When
this occurs, one may show the following results which apply
to Dirichlet agents. These are the counterparts to Corollaries 1
and 2 which apply to probabilistic agents. Rigorous proofs of
DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
9
k(εεε) = (V, E
these results are not provided; they can be carried out in a
manner which parallel the proofs of Corollaries 1 and 2.
Corollary 3. Consider the network G†
†
k(εεε))
in (3) populated with receptively updating Dirichlet opinion
followers and a single cautiously updating Dirichlet opinion
leader. Suppose that the condition in Lemma 3 is satisfied. Let
Dk denote the confidence matrix of the receptively updating
opinion followers with (cid:107)Dk(cid:107) ≤ ρ < 1, ∀k ≥ ND, for some
ND ∈ N≥0. Then all the agents reach a consensus opinion.
†
Corollary 4. Consider the network G†
k(εεε))
in (3) populated with receptively updating Dirichlet opinion
followers and two cautiously updating Dirichlet opinion lead-
ers. Suppose that the condition in Lemma 3 is satisfied. Let
Dk denote the confidence matrix of the receptively updating
opinion followers with (cid:107)Dk(cid:107) ≤ ρ < 1, ∀, k ≥ ND, for some
ND ∈ N≥0. Then, with a CUE-based update strategy, the
following are true:
k(εεε) = (V, E
(i) All
the agents reach a consensus iff
opinions of the two opinion leaders are equal.
the converged
(ii) When the opinions of the opinion leaders are not equal,
the opinion followers form an opinion cluster if the proportion
of weights each receptively updating agent gives to the two
opinion leaders is identical for k ≥ ND and no opinion leader
is given zero weight.
V. EMPIRICAL EVALUATIONS AND DISCUSSION
We now present some typical results obtained through
extensive simulations of scenarios where agent opinions are
captured via p.m.f.s (see Section V-A), Dirichlet BoEs (see
Section V-B), and more general DST BoEs (see Section V-C).
The results confirm our theoretical analysis in Section IV and
demonstrate the applicability of these results for the more
general DST models presented in Section III.
The agents in all the simulations employ a CUE-based
opinion update strategy with αi = 0.50, ∀i ∈ 1, N. The
agents' bounds of confidence are taken as identical,
i.e.,
εi = ε, ∀i ∈ 1, N. Note that, for αi > 0, ∀i ∈ 1, N, and for
sufficiently large ε, all agents satisfy the self-communicating
or the strong-aperiodic property [34], [37]. Even though being
embedded in a static network, the agents must accommodate
the bounds of confidence as the opinions are updated. In effect,
this creates a dynamic network G†
For ease of visualization, consensus/cluster formation are
displayed using bifurcation diagrams that depict the state of
consensus/cluster formation in the limit density versus ε [12].
k(εεε).
A. Probabilistic Agent Opinions
For our simulations we embed seven agents on a graph
of seven nodes and 100 agents on Erdos-R´enyi (ER) random
graphs of 100 nodes.
1) Simulations with Seven Agents: The FoD of the opinion
BoE of each agent is Θ = {θ1, θ2, θ3}. For the results shown
in Figures 3 to 6, initial opinion profile of θ1 is selected
as πππ(θ1)0 = [0.80, 0.78, 0.76, 0.40, 0.80, 0.10, 0.20]T ; the re-
maining masses are equally distributed among the opinion
profiles πππ(θ2)0 and πππ(θ3)0.
No Opinion Leaders. Fig. 3(a) shows the network topology
of the seven receptively updating agents Ri, i ∈ 1, 7. As
the corresponding bifurcation diagram in Fig. 3(b) shows, for
smaller values of ε, seven 'opinion clusters' are generated
because agents are essentially isolated. As the value of ε is
increased, the number of opinion clusters decreases because
the agents are updating their opinions based on opinions
of their neighbors who are within their confidence bounds.
Eventually, for ε > 0.46 (approx.), as expected from our
analytical results in Section IV-A1, a consensus emerges.
(a) Network topology.
(b) Bifurcation diagram for πππ(θ1).
(c) For ε = 0.5, first
iteration:
(R3, R6) and (R5, R7) do not ex-
change opinions because the cor-
responding opinion distances are
above ε = 0.5.
(d) For ε = 0.5, second iteration:
All neighbor agents exchange opin-
ions because all corresponding opin-
ion distances are below ε = 0.5.
Fig. 3: Probabilistic agents: Simulation results for seven recep-
tively updating agents (Ri, i ∈ 1, 7) and no opinion leaders.
A consensus occurs for ε > 0.46 (approx.). For ε = 0.5,
Figs 3(c) and 3(d) show the 'directions' of opinion exchange
at the first two iterations. Edge labels indicate the distance
between the opinions of the corresponding agent pair.
Even though the underlying network topology shown in
Fig. 3(a) is static, opinion updating occurs in a dynamic
network G†
k(εεε). To illustrate this further, consider the case
when ε = 0.5. As Fig. 3(c) illustrates, agent pairs (R3, R6)
and (R5, R7) do not exchange opinions because their opinion
distances exceed the bound of confidence ε = 0.5. However,
after the first iteration of opinion exchanges, distances among
agents change. As Fig. 3(d) illustrates, the opinion distances
of all agents (including (R3, R6) and (R5, R7)) are now well
within ε = 0.5 and, at the second iteration, all agents exchange
opinions with their neighboring agents.
Single Opinion Leader. We create this scenario by replacing
agent R1 in Fig. 3(a) with a cautiously updating agent C1
0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)R1R2R3R4R5R6R7(0.035)(0.017)(0.329)(0.017)(0.312)(0.346)(0.52)>ε=0.5(0.572)>ε=0.5(0.087)R1R2R3R4R5R6R7(0.146)(0.088)(0.089)(0.065)(0.146)(0.154)(0.421)<ε(0.351)<ε(0.078) 10
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
thus obtaining the graph in Fig. 4(a). For this topology the
bifurcation diagram for πππ(θ1) is depicted in Fig.4(b). As
before, for smaller values of ε, each agent forms its own
'opinion cluster'. For larger values of ε, in accordance with
Corollary 1, a consensus emerges, and this consensus opinion
is the opinion of the opinion leader C1 (viz., m(θ1) = 0.80).
As Fig. 4(b) indicates, this consensus begins to emerge for
ε > 0.46 (approx.). The dynamic nature of opinion exchange
in the first three iterations is illustrated in Figs 4(c), 4(d), and
4(e).
(a) Network topology.
(b) Bifurcation diagram for πππ(θ1).
(c) For ε = 0.46, first
iteration: (R3, R6) and
(R5, R7) do not ex-
change opinions.
(d) For ε = 0.46, sec-
ond iteration. (R3, R6)
now exchange opinions;
(R5, R7) still do not
exchange opinions.
(e) For ε = 0.46, third
iteration. All agents ex-
change opinions with
their neighbors.
Fig. 4: Probabilistic agents: Simulation results for one opinion
leader (C1) and six receptively updating agents (Ri, i ∈ 2, 7).
A consensus occurs for ε > 0.46 (approx.), at C1's opinion
(i.e., m(θ1) = 0.80). For ε = 0.46, Figs 4(c), 4(d), and
4(e) show the 'directions' of opinion exchange at the first
three iterations. Edge labels indicate the distance between the
opinions of the corresponding agent pair.
Two Opinion Leaders. Here we replaced the two agents
{R1, R7} in Fig. 3(a) by the cautiously updating agents
{C1, C7}, respectively. Figs 5(a) and 5(b) show the corre-
sponding network topology and bifurcation diagram for πππ(θ1),
respectively. In accordance with Corollary 2, no consensus
is reached because the two opinion leaders {C1, C7} possess
different opinions.
With the opinions of two opinion leaders being different, the
number of opinion clusters created depends on the bound of
confidence ε. For 0.31 < ε < 0.42 (approx.), we observe two
opinion clusters, the minimum number of clusters possible.
We have achieved this by picking the agent BoEs carefully so
that the network separates into two components, each with its
own opinion leader, for the aforementioned values of ε. For
these values, the network gets separated into two components
(a) Network topology.
(b) Bifurcation diagram for πππ(θ1).
Fig. 5: Probabilistic agents: Simulation results for two opinion
leaders (C1 and C7) and five receptively updating agents
(Ri, i ∈ 2, 6). With the two opinion leaders possessing
different opinions, no consensus is achieved. A minimum of 2
opinion clusters are achieved for 0.31 < ε < 0.42 (approx.).
There is no consensus among the receptively updating agents.
{C1, R2, R3, R,R5} and {C7, R6} because (cid:107)E3 − E6(cid:107) > ε
and (cid:107)E5 − E7(cid:107) > ε, for 0.31 < ε < 0.42 (approx.). The
ensuing network generates two opinion clusters at the opinions
of the two opinion leaders C1 and C7. For larger values of ε,
some (or all) receptively updating agents get influenced by
both opinion leaders which creates different opinion clusters
that are influenced by both opinion leaders.
(a) Altered topology.
(b) Bifurcation diagram for πππ(θ1).
Fig. 6: Probabilistic agents: Simulation results for two opinion
leaders (C1 and C7) and five receptively updating agents
(Ri, i ∈ 2, 6) embedded in a topology that generates a con-
sensus among the receptively updating agents. This consensus
appears for ε > 0.43 (approx.).
and C (2)
k
k
As asserted in Corollary 2, when the two opinion leaders
possess different opinions, the receptively updating agents will
reach a consensus if the matrices C (1)
in Definition 9
k 1NC2, ∀k ∈ N≥0. Fig. 6(a)
satisfy λ(1)
shows a network topology which satisfies this condition, and
the corresponding bifurcation diagram in Fig. 6(b) shows the
emergence of a third opinion cluster around m(θ)1 = 0.50
(approx.) for ε > 0.43 (approx.).
k 1NC1 = λ(2)
k C (2)
k C (1)
2) Simulations with 100 Agents: Here we embed 100
agents in random graphs of 100 nodes generated using the
Erdos-R´enyi (ER) random graph model. As is well known,
for an ER random graph, the phase transition for network
0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)C1R2R3R4R5R6R7(0.017)(0.035)(0.329)(0.017)(0.312)(0.346)(0.52)>ε=0.46(0.572)>ε=0.46(0.087)C1R2R3R4R5R6R7(0.067)(0.104)(0.115)(0.024)(0.078)(0.091)(0.472)>ε=0.46(0.459)<ε(0.001)C1R2R3R4R5R6R7(0.079)(0.176)(0.055)(0.029)(0.042)(0.027)(0.456)<ε(0.272)<ε(0.115)0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1) DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
11
(a) No opinion leaders.
(b) One opinion leader.
(c) Two opinion leaders.
Fig. 7: Probabilistic agents: Simulation results for 100 agents embedded in an Erdos-R´enyi random graph with p = 0.10 and
agent BoEs sampled from Dir(1, 1, 1). Consensus can be seen in Fig. 7(a) and 7(b), for ε > 0.26 and ε > 0.21, respectively.
(a) No opinion leaders.
(b) One opinion leader.
(c) Two opinion leaders.
Fig. 8: Dirichlet agents: Simulation results for 7 agents with no opinion leaders, one opinion leader and two opinion leaders
embedded in the graphs in Figs 3(a), 4(a), and 5(a) respectively. Consensus can be seen in Fig. 8(a) and 8(b), for ε > 0.51
and ε > 0.5, respectively.
connectivity occurs when the edge formation probability p
exceeds ln n/n [29]. With n = 100, ln n/n = 0.046 and
we used p = 0.10 for generating all our random graphs.
Moreover, every random graph was tested for connectedness
at initialization. For sufficiently large values of ε, the graph
G†
k(ε)ε)ε) is therefore essentially the same as Gk, and thus it is
connected as well.
The BoE of each agent was sampled from the symmetric
Dirichlet distribution Dir(1, 1, 1), which is equivalent to a
uniform distribution over the open standard 2-simplex [38].
As Figs 7(a) and 7(b) show, for ε > 0.26 (approx.) and
ε > 0.21 (approx.), a consensus appears when there are no
opinions leaders and when only one opinion leader is present,
respectively. In accordance with Corollary 2, Fig. 7(c) shows
that there is no consensus among the 100 agents when the two
opinion leaders have different opinions.
Fig. 8 shows the corresponding bifurcation diagrams. As is
evident, and in consistent with Corollary 3, a consensus can
be seen in Figs 8(a) and 8(b) for ε > 0.51 and ε > 0.5,
respectively. However,
there is no consensus
among the agents. This is consistent with Corollary 4(i)
because the cautious agents do not possess the same converged
opinion. The minimum number of opinion clusters appear for
0.26 < ε < 0.52 (approx.).
in Fig. 8(c),
Interestingly, even for higher values of ε, no consensus
emerges even among the receptive agents. Indeed, one would
expect that the receptive agents who are now less restrained
to exchange opinions with their neighbors would form an
opinion clusters of their own. In contrast, when the agents
are embedded in the graph shown in Fig. 6(a), a consensus
emerges among the receptive agents for ε > 0.46 (approx.).
See Fig. 9. This is because the graph topology in Fig. 6(a)
satisfies the condition in Corollary 4(ii).
B. Dirichlet Agent Opinions
Here, we repeat the experiments conducted with the 7-agent
topologies in Section V-A1 but with Dirichlet agent opinions.
For all the agents, we kept the same mass vectors πππ(θ2)0 and
πππ(θ3)0 as those in Section V-A1 while we used πππ(Θ)0 = 0.1;
the remaining masses were assigned to πππ(θ1)0.
C. General DST Agent Opinions
1) Simulations with Seven Agents: In this study, we as-
signed random DST mass assignments for the seven agents
embedded within the topology in Fig. 6(a). For this purpose,
we utilized the Dirichlet distribution which has been widely
0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)0.00.20.40.60.81.0Confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)0.00.20.40.60.81.0confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)BifurcationDiagram0.00.20.40.60.81.0confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)BifurcationDiagram0.00.20.40.60.81.0confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)BifurcationDiagram 12
IEEE TRANSACTIONS ON SIGNAL AND INFORMATION PROCESSING OVER NETWORKS
(a) No opinion leaders.
(b) One opinion leader.
(c) Two opinion leaders.
Fig. 10: General DST agents: Simulation results for seven receptively updating agents embedded within the network topology
in 6(a) and DST mass values sampled from Dir(4, 4, 4, 2, 2, 2, 1). A consensus appears for the no opinion leader and one
opinion leaders cases in Fig. 10(a) and 10(c) for ε > 0.18 (approx.) and ε > 0.17 (approx.), respectively.
1(θ2), m∗
2(θ1), m∗
formed by {R1, . . . , R5} converge to the probabilistic opinion
1(θ3)} = {0.63, 0.19, 0.18}; the second
{m∗
1(θ1), m∗
cluster formed by R6 and R7 converge to the general DST
opinion {m∗
2(θ2, θ3)} = {0.15, 0.85} which allows
no further 'refinement' between the singletons θ2 and θ3. Such
emergent behavior is qualitatively different than what appears
in prior models [12]–[16].
TABLE I: Initial and Converged Opinions (with ε = 0.30)
Agent
DST Mass Values
(θ1, θ2)
θ3
θ2
θ1
(θ2, θ3)
Initial Opinions:
R1
R2
R3
R4
R5
R6
R7
0.60
0.62
0.51
0.57
0.60
0.10
0.20
0.10
0.04
0.05
0.03
0.10
–
–
Converged Opinions (with ε = 0.30):
{R1, . . . , R5}
0.18
–
0.10
0.11
0.12
0.15
0.10
–
–
{R6, R7}
0.63
0.15
0.19
–
0.10
0.11
0.12
0.15
0.10
–
–
–
–
0.10
0.12
0.20
0.10
0.10
0.90
0.80
-
0.85
VI. CONCLUSION
In this paper, we use the DST framework for representing
agent opinions and explore the formation of consensus and
opinion clusters when agents residing within a network ex-
change and update their opinions. In particular, we explore
the effect that opinion leaders have on these processes. Our
opinion model accounts for aspects from SJT and possesses
the ability to capture a wider variety of uncertainties and
nuances in agent opinions, an advantage inherited from its
DST basis. Theoretical analysis, which focuses on probabilis-
tic and Dirichlet agent opinions, provides conditions for the
emergence of consensus and opinion clusters in the presence
of opinion leaders. Our results show that a consensus can be
formed when the number of opinion leaders is no more than
one and with a sufficiently high bound of confidence of the
agents. With two or more opinion leaders possessing different
opinions, no consensus can be reached in general. We also
explore the conditions for opinion cluster formation among
the opinion followers.
Fig. 9: Dirichlet agents: Simulation results for two opinion
leaders (C1 and C7) and five receptively updating agents
(Ri, i ∈ 2, 6) embedded in the graph in Fig. 6(a) generates
a consensus among the receptively updating agents. This
consensus appears for ε > 0.46 (approx.).
employed in opinion modeling [39]–[41]. In particular, for
each agent in each trial, the DST masses for θ1, θ2, θ3, (θ1θ2),
(θ1θ3), (θ2θ3), and Θ = (θ1θ2θ3) were sampled from the
Dirichlet distribution Dir(4, 4, 4, 2, 2, 2, 1). Figs 10(a), 10(b),
and 10(c) show the bifurcation diagrams when the network
contains no opinion leaders, one opinion leader, and two
opinion leaders, respectively.
As Fig. 10(a) shows, with no opinion leaders, a consensus
appears for ε > 0.18 (approx.). Fig. 10(b) shows bifurcation
diagram when only one opinion leader is present, and we
can see that a consensus appears for ε > 0.17 (approx.). As
Fig. 10(c) shows, with two opinion leaders, no consensus is
reached among the agents. However, it is interesting to note
that three opinion clusters emerge for ε > 0.21 (approx.).
As mentioned earlier, the main reason for utilizing the DST
framework for capturing agent opinions is its ability to capture
the types of uncertainties and the nuances that are characteris-
tic of agent states and opinions. DST agent opinions can also
generate new emergent behavior which cannot be captured via
probabilistic agents. For example, consider 7 receptive agents
embedded within the topology in Fig. 3(a). The initial opinions
and the converged opinions for ε = 0.30 appear in Table I.
Notice that two opinion clusters have emerged: the first cluster
0.00.20.40.60.81.0confidencebound,ε0.00.10.20.30.40.5DSmass,m(θ1)BifurcationDiagram0.00.20.40.60.81.0confidencebound,ε0.00.10.20.30.40.5DSmass,m(θ1)BifurcationDiagram0.00.20.40.60.81.0confidencebound,ε0.00.10.20.30.40.5DSmass,m(θ1)BifurcationDiagram0.00.20.40.60.81.0confidencebound,ε0.00.20.40.60.81.0DSmass,m(θ1)BifurcationDiagram DABARERA, et al.: CONSENSUS IN THE PRESENCE OF MULTIPLE OPINION LEADERS ...
13
Our current work involves extending this theoretical analysis
to scenarios where agent opinions are captured via more
general DST BoEs, which may require recourse to tools
from paracontractions theory [7]. It is also noteworthy that
we have taken all agents to possess the identical bound of
confidence value. When this is not
the opinion
exchange mechanism itself would be directional (because an
agent with a lower bound of confidence may update itself
from its neighbor agent who may not update itself because
of a higher bound of confidence value). An interesting future
research problem is the study of networked agents whose
bounds of confidence values are different. Another interesting
issue to be addressed is the assessment of the convergence
speed of our algorithms [42].
the case,
REFERENCES
[1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and coopera-
tion in networked multi-agent systems," Proc. IEEE, vol. 95, no. 1, pp.
215–233, Jan. 2007.
[2] Z. Qu, J. Wang, and R. A. Hull, "Cooperative control of dynamical
systems with application to autonomous vehicles," IEEE Transactions
on Automatic Control, vol. 53, no. 4, pp. 894–911, 2008.
[3] L. Xiao, S. Boyd, and S. Lall, "A scheme for robust distributed sensor
fusion based on average consensus," Int. Symp. on Inf. Proc. in Sensor
Netw. (IPSN), pp. 63 – 70, Apr. 2005.
[4] C. Chamley, A. Scaglione, and L. Li, "Models for the diffusion of beliefs
in social networks: An overview," IEEE Signal Processing Magazine,
vol. 30, no. 3, pp. 16–29, May 2013.
[5] S. Fortunato, V. Latora, A. Pluchino, and A. Rapisarda, "Vector opinion
dynamics in a bounded confidence consensus model," Int. J. of Modern
Phys. C, vol. 16, no. 10, pp. 1535–1551, Oct. 2005.
[6] W. Ren, R. W. Beard, E. M. Atkins, and A. G. Theory, "A survey
of consensus problems in multi-agent coordination," Amer. Cont. Conf.
(ACC), pp. 1859–1864, 2005.
[7] T. Wickramarathne, K. Premaratne, M. Murthi, and N. Chawla, "Con-
vergence analysis of iterated belief revision in complex fusion environ-
ments," IEEE J. of Selected Topics in Sig. Proc., Special Issue on Sig.
Proc. for Social Netw., vol. 8, no. 4, pp. 598–612, Aug. 2014.
[8] G. Shafer, A Mathematical Theory of Evidence. Princeton Univ. Press,
1976.
[9] R. Fagin and J. Y. Halpern, "A new approach to updating beliefs," Proc.
Conf. on Uncertainty in Art. Intell. (UAI), pp. 347–374, 1991.
[10] S. Blackman and R. Popoli, Design and Analyis of Modern Tracking
Systems. Norwood, MA: Artech House, 1999.
[11] T. Wickramarathne, K. Premaratne, M. Kubat, and D. Jayaweera,
"CoFiDS: A belief-theoretic approach for automated collaborative filter-
ing," IEEE Transactions on Knowledge and Data Engineering, vol. 23,
no. 2, pp. 175–189, Feb. 2011.
[12] J. Lorenz, "Continuous opinion dynamics under bounded confidence: A
survey," Int. J. of Modern Phys. C, vol. 18, no. 12, pp. 1819–1838, Jul.
2007.
[13] U. Krause, "A discrete nonlinear and non-autonomous model of con-
sensus formation," Comm. Diff. Equations, no. 1, pp. 227–236, 2000.
[14] R. Hegselmann and U. Krause, "Opinion dynamics and bounded confi-
dence models, analysis and simulation," J. of Art. Societies and Social
Simulation, vol. 5, no. 3, Jun. 2002.
[15] G. Deffuant, D. Neau, F. Amblard, and G. Weisbuch, "Mixing beliefs
among interacting agents," Adv. in Complex Syst., vol. 03, no. 04, pp.
87–98, Jan. 2000.
[16] G. Weisbuch, "Bounded confidence and social networks," The European
Physical J. B - Condensed Matter, vol. 38, no. 2, pp. 339–343, Mar.
2004.
[17] R. Dabarera, R. Nunez, K. Premaratne, and M. Murthi, "Dynamics of
belief theoretic agent opinions under bounded confidence," Int. Conf. on
Inf. Fusion (FUSION), pp. 1–8, Jul. 2014.
[18] M. Sherif and C. Hovland, Social Judgment: Assimilation and Contrast
Effects in Communication and Attitude Change, ser. Yale Studies in
Attitude and Communication. Greenwood Pub. Group, Inc., 1980.
[19] A. A. Rad, H. Rahmandad, and M. Yahyanejad, "Estimating the dynam-
ics of individual opinions in online communities," in Proceedings of the
International System Dynamics Conference, Delft, 2014.
[20] K. Varshney, "Opinion dynamics with bounded confidence in the Bayes
risk error divergence sense," in Acoustics, Speech and Signal Processing
(ICASSP), 2013 IEEE International Conference on, May 2013, pp.
6600–6604.
[21] A. Josang, "Cumulative and averaging fusion of beliefs," Information
Fusion, vol. 11, no. 2, pp. 192–200, Apr. 2010.
[22] X. Lu, R. Lu, S. Chen, and J. Lu, "Finite-time distributed tracking
control for multi-agent systems with a virtual leader," IEEE Trans. Circ.
and Syst. I: Regular Papers, vol. 60, no. 2, pp. 352–362, Feb. 2013.
[23] M. Bauer, "Approximation algorithm and decision making in the
Dempster-Shafer theory of evidence-an empirical study," International
Journal of Approximate Reasoning, vol. 17, no. 2/3, pp. 217–237,
Aug./Oct. 1997.
[24] N. Wilson, "Algorithms for Dempster-Shafer theory," in Algorithms for
Uncertainty and Defeasible Reasoning, ser. Handbook of Defeasible
Reasoning and Uncertainty Management Systems, D. M. Gabbay and
P. Smets, Eds. Kluwer, 2000, vol. 5, pp. 421–475.
[25] T. Wickramarathne, K. Premaratne, and M. Murthi, "Toward efficient
computation of the Dempster-Shafer belief theoretic conditionals," IEEE
Transactions on Cybernetics, vol. 43, no. 2, pp. 712–724, Apr. 2013.
[26] D. Denneberg, "Conditioning (updating) non-additive measures," Ann.
of Operations Research, vol. 52, no. 1, pp. 21–42, Mar. 1994.
[27] A.-L. Jousselme, D. Grenier, and E. Boss´e, "A new distance between
two bodies of evidence," Inf. Fusion, vol. 2, no. 2, pp. 91–101, Jun.
2001.
[28] A. Martin, A.-L. Jousselme, and C. Osswald, "Conflict measure for the
discounting operation on belief functions," 11th Int. Conf. on Inf. Fusion
(FUSION), pp. 1–8, Jun. 2008.
[29] M. E. J. Newman, Networks: An Introduction. New York, NY: Oxford
University Press, 2012.
[30] K. Premaratne, D. A. Dewasurendra, and P. H. Bauer, "Evidence com-
bination in an environment with heterogeneous sources," IEEE Trans.
on Syst., Man, and Cyber. - A: Syst. and Humans, vol. 37, no. 3, pp.
298–309, 2007.
[31] K. Premaratne, M. Murthi, J. Zhang, M. Scheutz, and P. Bauer, "A
Dempster-Shafer theoretic conditional approach to evidence updating for
fusion of hard and soft data," Proc. Int. Conf. on Inf. Fusion (FUSION),
Seattle, WA, pp. 2122–2129, Jul. 2009.
[32] J. Dittmer, "Consensus formation under bounded confidence," Nonlinear
Anal.: Theory, Methods & Applicat., vol. 47, pp. 4615–4621, 2001.
[33] D. J. Hartfiel, Nonhomogeneous Matrix Products. World Scientific,
2002.
[34] J. Lorenz, "Convergence of products of stochastic matrices with positive
diagonals and the opinion dynamics background," in Positive Systems,
ser. Lecture Notes in Control and Inform. Sciences.
Springer, 2006,
vol. 341, pp. 209–216.
[35] I. Daubechies and J. C. Lagarias, "Sets of matrices all infinite products
of which converge ," Linear Algebra and its Applicat. , vol. 161, no. 0,
pp. 227 – 263, 1992.
[36] S. Chatterjee and E. Seneta, "Towards consensus: Some convergence
theorems on repeated averaging," J. of Appl. Probability, vol. 14, no. 1,
pp. 89–97, Mar. 1977.
[37] B. Touri and A. Nedic, "Product of Random Stochastic Matrices," IEEE
Transactions on Automatic Control, vol. 59, no. 2, pp. 437–448, Feb.
2014.
[38] C. M. Bishop, Pattern Recognition and Machine Learning. Secaucus,
NJ, USA: Springer-Verlag New York, Inc., 2006.
[39] K. Chaloner and G. T. Duncan, "Some properties of the Dirichlet-
multinomial distribution and its use in prior elicitation," Communications
in Statistics - Theory and Methods, vol. 16, no. 2, pp. 511–523, 1987.
[40] E. Regazzini and V. V. Sazonov, "Approximation of laws of multinomial
parameters by mixtures of Dirichlet distributions with applications to
Bayesian inference," Acta Applicandae Mathematica, vol. 58, no. 1-3,
pp. 247–264, 1999.
[41] R. E. Zapata-V´azquez, A. O'Hagan, and L. S. Bastos, "Eliciting expert
judgements about a set of proportions," Journal of Applied Statistics,
vol. 41, no. 9, pp. 1919–1933, 2014.
[42] A. Olshevsky and J. N. Tsitsiklis, "Convergence speed in distributed
consensus and averaging," SIAM J. Control Optim., vol. 48, no. 1, pp.
33–55, Feb. 2009.
|
1511.05618 | 1 | 1511 | 2015-11-17T23:36:46 | Topic Model Based Behaviour Modeling and Clustering Analysis for Wireless Network Users | [
"cs.SI"
] | User behaviour analysis based on traffic log in wireless networks can be beneficial to many fields in real life: not only for commercial purposes, but also for improving network service quality and social management. We cluster users into groups marked by the most frequently visited websites to find their preferences. In this paper, we propose a user behaviour model based on Topic Model from document classification problems. We use the logarithmic TF-IDF (term frequency - inverse document frequency) weighing to form a high-dimensional sparse feature matrix. Then we apply LSA (Latent semantic analysis) to deduce the latent topic distribution and generate a low-dimensional dense feature matrix. K-means++, which is a classic clustering algorithm, is then applied to the dense feature matrix and several interpretable user clusters are found. Moreover, by combining the clustering results with additional demographical information, including age, gender, and financial information, we are able to uncover more realistic implications from the clustering results. | cs.SI | cs | Topic Model Based Behaviour Modeling and
Clustering Analysis for Wireless Network Users
Bingjie Leng, Jingchu Liu, Huimin Pan, Sheng Zhou, and Zhisheng Niu
Tsinghua National Laboratory for Information Science and Technology
Department of Electronic Engineering
Email: {lengbj14, liu-jc12, phm13}@mails.tsinghua.edu.cn, {sheng.zhou, niuzhs}@tsinghua.edu.cn
Tsinghua University, Beijing 100084, China
5
1
0
2
v
o
N
7
1
]
I
S
.
s
c
[
1
v
8
1
6
5
0
.
1
1
5
1
:
v
i
X
r
a
Abstract—User behaviour analysis based on traffic log in wire-
less networks can be beneficial to many fields in real life: not only
for commercial purposes, but also for improving network service
quality and social management. We cluster users into groups
marked by the most frequently visited websites to find their
preferences. In this paper, we propose a user behaviour model
based on Topic Model from document classification problems. We
use the logarithmic TF-IDF (term frequency - inverse document
frequency) weighing to form a high-dimensional sparse feature
matrix. Then we apply LSA (Latent semantic analysis) to deduce
the latent topic distribution and generate a low-dimensional
dense feature matrix. K-means++, which is a classic clustering
algorithm, is then applied to the dense feature matrix and several
interpretable user clusters are found. Moreover, by combining
the clustering results with additional demographical information,
including age, gender, and financial information, we are able to
uncover more realistic implications from the clustering results.
Keywords—traffic log, user behaviour modeling, clustering
analysis, topic model.
I.
INTRODUCTION
Thanks to the wide adoption of smart devices such as
smart phones and tablets, nowadays people can perform an
unprecedented number of tasks online, ranging from news
and finance to social and gaming. As a consequence, Internet
browsing log in wireless networks has become an essential
source of information for analyzing users’ hidden preferences
and inferring their real life behaviour. With a deeper under-
standing on the usage pattern of mobile users, network service
providers are able to provide more personalized services and
improve the service quality as well. Users’ browsing interests
are also helpful
in fields such as urban planning, mobile
advertisement, transportation, education, etc [1–3].
The most naive way to extract user behaviour from the
Internet browsing dataset is to observe the long-term global
statistics of various websites. But in this situation, individual
differences will be covered up. On the contrary, if we focus on
the analysis on one single user, the similarity between users’
browsing habits will be ignored. Hence, clustering becomes an
efficient method to strike a balance between these two extremes
and extract the average behaviour of a group of users who have
similar browsing history. Therefore, we design and implement
a process to cluster similar users into groups, each of which
is labeled by the type of frequently visited websites.
To apply clustering algorithms, the first step is to represent
users with a profile vector through user behaviour modeling.
In this paper, we propose a user behaviour modeling method
based on the topic model, which is originally proposed for
paper:
document classification, to generate an original profile matrix.
To enhance the discriminative power of the original matrix, we
apply TF-IDF (term frequency - inverse document frequency)
weights to regenerate a feature matrix with large dimension-
ality. With methods in Latent semantic analysis (LSA) [4],
we are able to get a low-dimensional feature matrix reflecting
the distribution of different topics of all the users. Finally,
clustering algorithms such as K-means++ can be applied to
the final feature matrix and the clustering results are analyzed.
Concretely, we make the following contributions in this
• We analyze the similarity and differences between
network user modeling and document classification,
and propose to utilize text mining algorithms for
network user modeling problems.
Based on the analysis on our dataset, we utilize
logarithmic TF-IDF to generate sparse feature matrix
and use LSA for topic discovery and dimensionality
reduction. To our knowledge, this is the first study to
analyze user behaviour with a combination of these
tools.
• We extract users’ interests by clustering users with
similar browsing habits into groups. We also examine
our clustering results with additional demographical
information including age, gender, and financial in-
formation on the campus during five months. Obvious
preference differences are found between different
genders and age. It helps us explain our clustering
findings accordingly and proves that our algorithm can
work effectively. Moreover, our findings can help with
campus management in many aspects.
•
The rest of the paper is outlined as follows. Section
II introduces related work about user behaviour analysis in
WLAN. Section III presents the network user behavior mod-
eling problem and explain its analogy with topic modeling
in document classification. In Section IV, we introduce our
datasets and the details of our algorithm implementation. In
Section V, we present the clustering results and explain the
findings. Finally, in Section VI, we conclude and discuss future
work.
II. RELATED WORK
With the rapid development of wireless networks,
the
potential of user behaviour analysis has brought up tremen-
dous attention recently. The most common method for user
clustering is by applying K-means on raw profile matrix.
For example, the web browsing similarity among users of
TABLE I.
ANALOGY BETWEEN USER BEHAVIOUR MODELING AND TOPIC MODELING IN DOCUMENT CLASSIFICATION
Elements in Network User Behavior Modeling
Elements in Document Classification
Tuple of Domains/ Locations/ Time
Value for Tuple(Total Bytes/ Duration/ Request Numbers)
User
User Clustering
Document
Words
Frequency of Words
Document Classification
a university laboratory is examined using web log data in
in [5]. Besides, in [6], the authors evaluate the effectiveness
of K-Means and spectral clustering and discuss the effect of
the profile size to the clustering results. The authors in [7]
discuss this personalization problem from another perspective.
They apply K-Means to construct user profiles using device-
side traffic logs. The authors also point out that the extra
information on gender may help with the problem in future
work.
Besides K-means, other clustering algorithms are also
investigated in the literature. The authors in [8] present a
Greedy Clustering Using Belief Function (GCB) algorithm.
The representatives of the clusters are picked iteratively, so
that
the current representative is well separated from the
former ones. In [9], the authors analyze the online activity
and mobility for thousands of mobile users. They use Self-
Organized Map (SOM) for discovering mobile users’ trends.
In [10], the authors use a information theoretic co-clustering
method on user-domain matrix and consider the effect of
location.
Many similarity measures between users are studied. The
authors in [11] use dataset containing traces of 34 users and
mainly deal with temporal dimension. In their work, different
similarity matrices including Cosine Similarity, Hellinger dis-
tance and SVD are compared with each other. The authors in
[8] apply a belief function as the similarity measure.
A closely recent related work is [12]. Same as us, the
authors study user behaviour modeling in a perspective of
topic modeling. By way of analogy,
the authors find that
there are many similarities between user behaviour profiling
and document classification in natural language processing,
which is also used in our modeling. But in their work, the
authors generate their feature matrix by oversampling the
URLs, while we use TF-IDF weights. The data comes from
cellualr networks in their paper. But the datasets we use are
clollected from WLAN. In [13], the author proposes a similar
procedure in analyzing user preferences with our algorithm for
webpage prefetching. But the data set used in the paper is small
and there are only around 100 domains included. Therefore,
user behaviour cannot be fully presented.
III. USER BEHAVIOUR MODELING
There are many similarities between user behaviour anal-
ysis and document classification. We will introduce our user
behaviour modeling by comparing the nature of user behaviour
profiling with the properties of a document in this section.
The main purpose of document classification is to find the
hidden topics of the corpus of documents by unsupervised
learning algorithms. In Table I, the correspondences between
user behaviour modeling and topic modeling are listed. In
user behaviour modeling, we plan to cluster users with simi-
lar internet browsing records, which is similar as clustering
documents with similar topics. The content of the internet
browsing records can be considered as words existing in a
document to some degree, although the statistic properties of
internet records are more complicated than term frequency
in a document. These findings encourage us to use tools in
document classification to classify users’ internet browsing
preferences.
Due to this similarity, tools designed for document classi-
fication may also be fit to network user behaviour modeling.
In this work, we use TF-IDF statistics and LSA to cluster
users with similar network usage behavior and preferences. TF-
IDF is a nonlinear weighting method to reveal the correlation
between words and topics. LSA is an efficient technique to
mine topics and reduce complexity at the same time.
IV. DATASETS AND CLUSTERING IMPLEMENTATION
In this section, we introduce the properties of the campus
WiFi dataset used in this paper. We also outline the implemen-
tation of our algorithm to extract user behaviour, with detailed
explanation for each algorithm step. The steps include data
pre-processing, user profile matrix generalization, generating
the feature matrix with TF-IDF, LSA and clustering.
A. Overview of Datasets
In this paper, we use the Internet browsing data extracted
from a WiFi network in a Chinese university campus. This
dataset has been collected from September, 2014 to January,
2015, and contains the browsing records of over 28,000
anonymous users.
User’s browsing activity is aggregated into network ses-
sions, which are defined to be the period in which users
utilize the network with pauses that are less than 5 minutes.
If a user stops exchanging information with the network for
more than 5 minutes, the session ends and activities from that
moment on are recorded in the following session. Each session
contains the information of network browsing activity of a
user, including user ID, session starting time, session duration,
location, first-level or second-level domains, ISP names, HTTP
request number and predefined service classes.
Except for the internet browsing dataset, a complementary
dataset containing users’ personal profile is also given. Each
row in this dataset represents for one user’s profile, including
user ID, gender, the year of birth, the year of enrolment and
type of degree. Besides, we are also provided with a dataset
containing all trading information of users’ campus cards dur-
ing the same period of time. These additional information gives
us an opportunity to interpret the results of user behaviour
clustering and learn more about users’ behaviour.
B. Data Pre-processing and Overall Statistics
To facilitate algorithm processing, we extract the network
browsing information of each user from the dataset and form
a user profile matrix. Each row represents one user and each
column represents the values of one profile dimension. Based
on the types of information available in the original dataset,
there are several ways for us to formulate the user profile,
including the domain names, time of day or week, locations
or encapsulating a multiply of these informations. In this work,
corresponded to the ratio between the total bytes consumed on
each domain to the total bytes consumed on all the domains
that appear in the dataset. Still, the traditional form of TF
cannot be directly applied to the network traffic log because
of the large dynamic range across the total communication
bytes of different domains: the largest and the smallest can be
different to 8 orders of magnitude. To cope with this problem,
we apply logarithmic operation on TF terms to narrow down
the dynamic range. Concretely, assume that there are Nu users
and Nd domains involved. Let Bij denote the total bytes of
the jth domain for the ith user, where j ∈ {1, 2, ..., Nd} and
i ∈ {1, 2, ..., Nu}. The TF weight for the jth domain and user
i, i.e. TFij can be expressed as:
Bij(cid:80)Nd
j=1 Bij
Fig. 1. Histogram of total bytes for user browsing.
TFij = 1 + log(
).
(1)
we only use the domain names of the websites that the users
access over the WiFi.
As for the values in the feature matrix, four statistics of
user behaviour are available, including the total communication
bytes, the aggregated duration, the cumulative HTTP request
number, and the cumulative number of sessions of one user
on a certain domain during these five months. In this paper,
we choose the total communication bytes as the profile matrix
entry value.
After pre-processing, we find that
there are 893 first-
level domains covered in the dataset. We investigate the basic
statistics of the resulting profile matrix. We find that
the
medians of rows (or equivalently, users) are all zeros, which
indicates that all the users only visit a highly limited range
of domains. Besides, only 5% of the domains have non-zero
median browsing activity among users. This indicates that most
of these domains are only accessed by a small amount of users
and users’ domain preferences are quite centralized. We rank
the domains with respect to their median browsing intensity
and show the histogram of browsing intensity of three of the
top-10 domains in Fig. 1. As can be seen, most of the users
would visit these domains with high traffic. The top domains
which are visited most include baidu.com, qq.com, sinaimg.cn,
etc.
C. Generation of Feature Matrix With TF-IDF
The matrix generated in the last section needs to be
normalized before being used for clustering. But with simple
normalization of each row, the weights in the matrix (i.e. TF)
cannot represent the importance of various domains to one
user precisely. This is because there are some domains that are
almost visited by everyone, such as search sites like baidu.com
and portal sites like qq.com. It is obvious that their normalized
weights for most users would dominate over other domains
and user preferences would be biased toward these dominating
domains. For this reason, we use TF-IDF as a method to reduce
the weights of this kind of popular domains and generate a new
feature matrix which will be used in the following steps.
TF-IDF is a non-linear transformation which is widely used
in document classification to unveil the true discriminative
power of a certain term to some document. It is the product of
term frequency (TF) and inverse document frequency (IDF).
In document classification tasks, TF means the frequency for a
word to appear in a document. We extend this concept to user
clustering with network traffic log. Specially, TF-IDF is now
The inverse document frequency (IDF) is introduced to
show how much information a domain provides. IDF can be
expressed as:
IDFij = log(
Nu
nj
),
(2)
where nj denotes the number of users ever visited the jth
domain.
Using the above notations, we can get the new feature
matrix F , in which each element is given by:
Fij = TFij · IDFij.
(3)
D. Dimensionality Reduction Using Truncated SVD
To analyze user preferences by clustering, the distances
between feature vectors should be computed. But the large
(row) dimension of the user feature matrix may render com-
mon distance measures, e.g. cosine distance, useless due to
the so called ”curse of dimensionality”. Besides, the large
dimensionality also makes clustering algorithms slow. To solve
this problem, we assume network activity to follow a topic
model and apply LSA to reduce the dimensionality of the
sparse feature matrix and get a dense one with much smaller
number of dimensions.
Topic modeling assumes the following generative process
for text documents: each document can be expressed as a
distribution over a collection of T topics {pt, t = 1, 2...T},
and each topic is manifested through a probability distribution
over a collection of W words {pwt, w = 1, 2, 3..., W}. The
overall probability of a word w to appear in a document is
t=1 pwtpt. In user behaviour analysis, the traffic of
each user on domains can be assumed to follow a similar
generative process.
pw = (cid:80)T
A common way to recover the hidden parameters of such
latent distribution is LSA. Specially, we can factorize the
feature matrix F using SVD. The decomposition of F is given
by:
F = U ΣV T ,
(4)
where U and V are orthogonal matrices and Σ is a diagonal
matrix. If we further assume the domain corpus contians
only a small number of topics, we can perform the truncated
SVD to extract part of the singular vectors and values. The
approximation of F can be expressed as follows:
FM = UM ΣM V T
M ,
(5)
TABLE II.
RUNNING TIME OF THE WHOLE PROCESS
M
100
200
300
400
500
600
700
800
Running Time (s)
11.66
22.37
30.13
43.06
56.82
84.07
93.15
112.17
Fig. 3. TF-IDF distribution of domains in each cluster.
different clustering results, K is chosen to be 8 as an example
to explain the clustering results in the following.
The average TF-IDF distribution of domains in each user
cluster is shown in Fig. 3. The horizontal axis lists 61 domain
names which include all the top-10 domains with the largest
TF-IDF weights of each cluster. Each cluster is marked on the
Y-axis with the summarized topic or by the domain name with
the largest TF-IDF weight within this cluster. As we can see,
there are several hot domains in each cluster, which indicates
the preferences of users within this cluster. The clusters include
preferences for automobiles, games, social networks, online
shopping websites and so on. To be clearer, we list three typical
domain names in each cluster with the largest TF-IDF weights
and the topic of each cluster is summarized based on these
domains in Table III. For clusters with no obvious topic, but
only preferring some general websites such as WeChat and
Baidu, the topic of these clusters are the services correspoding
to the largest TF-IDF weight in each of them.
As a comparison with our implementation, we list the top
three domains in each cluster without TF-IDF weighting in
Table IV. The values in the input matrix of LSA here are
the simple row normalization results of the original feature
matrix generated in Section IV-B. From the clustering results,
we find that all the top domains are some general websites
within multiple service classes. Unlike Table I, it is hard to
define a particular topic or interest for each cluster, which
indicates the process with TF-IDF raises the effectiveness of
clustering a lot.
Fig. 4 shows the percentage of males in each cluster. Note
that the overall percentage of males is 67.9 %. We find that
there are more boys than girls visiting game websites, while
there are much more girls than boys shopping online. These
findings are quite intuitive in the sense that girls pay more
attention on their appearance and most girls love shopping,
while boys are more keen on different kinds of games, from
Fig. 2. The trend of inertia changing with the number of clusters K.
where the M largest singular values in Σ and their correspond-
ing singular vectors from U and V are selected. UM can be
treated as the final feature matrix used for clustering, which is
of much lower dimensionality than the original sparse feature
matrix. In the meantime, observations would be extracted
from VM , which contains information about the relationship
between different domains.
E. K-Means Clustering
As for the clustering algorithm, we finish the user clus-
tering based on the classic K-Means algorithm. The K-Means
algorithm divides all the users into K clusters and guarantees
that the within-cluster sum of squared distances, a.k.a. inertia,
is minimized. Due to the randomness of initial seeds selection
in standard K-Means algorithm, we use K-Means++ [14]
instead, which performs multiple K-Means runs with different
random initialization and selects the best clustering results.
V. EVALUATION OF IMPLEMENTATION AND
EXPERIMENTAL RESULTS
In this section, we first evaluate and choose the parameters
of our implementation based on numerical analysis. Then,
we describe the experimental results and give corresponding
advice on campus management.
A. Evaluation of Implementation
The selection of the cluster number K is an essential issue,
as this may influence the clustering results a lot. Fig. 2 shows
the values of inertia when different values of K are applied,
from 1 to 13. As we can see, the value of inertia goes down
with the increase of K.
Except for K, the number of the selected singular values
M in LSA must be decided as well. In Table II, the running
time of the whole implementation process with different values
of M ia given. As M goes up, the complexity of the algorithm
increases, so that the running time increases. But the smaller
M is, the more information of the original feature matrix will
be ignored. Therefore, a modest M should be selected because
of the tradeoff between complexity and information integrity.
B. Experimental Results
Based on the analysis above, we select the top 80 singular
components in LSA, i.e., M = 80, to reduce the computation
complexity and get enough information for clustering at the
same time. Since different numbers of clusters K may incur
TABLE III.
LIST OF SUMMARIZED TOPICS AND TOP THREE DOMAINS IN EACH CLUSTER WITH TF-IDF WEIGHTING.
Topics
Top Three Domains
Big Games
Social Networks
Casual Games
Shopping
Recruiting
WeChat
Automobiles
Microblogging
geilicdn.com (mobile shopping), koudai.com (mobile shopping), meitu.com
yingjiesheng.com, 51job.com, zhaopin.com (recruitment sites)
duowan.com (game news), yy.com (game webcast), battlenet.com.cn
xnimg.com, renren.com, xnpic.com (all related to renren.com)
66game.com.cn, 7.qq.com (webgame), zg.qq.com (webgame)
weixin.qq.com (WeChat), apple.com, qq.com
autohome.com.cn, auto.suhu.com, auto.qq.com
sinaimg.cn, weibo.cn, blog.sina.cn (mobile blog)
TABLE IV.
LIST OF TOP THREE DOMAINS IN EACH CLUSTER WITHOUT
TF-IDF WEIGHTING.
Cluster No.
Top Three Domains
1
2
3
4
5
6
7
8
sinaimg.cn, sina.cn, sinastorage.com
qq.com, gtimg.com, speed.qq.com
baidu.com, bdimg.com, baidu-img.cn
xnimg.com, renren.com, xnpic.com
apple.com, verisign.com, autonavi.com
weixin.qq.com, apple.com, qq.com
360.cn, microsoft.com, mi.com
163.com, umeng.com, youdao.com
Fig. 5. Contour map of user birth year distributions in each cluster.
Fig. 6. Distribution of average campus card consumption in five months in
each cluster.
suburb of the city with few business areas and restaurants.
On the contrast, the graduate students are living in the urban
campus site. The restaurants and supermarkets are attractive
for them to spend more money outside the campus, while the
undergraduate students barely have chances. Reasons may also
be that the graduate students are more likely to get married
and spend less time on campus, which will cause the decrease
in the consumption. Therefore, to increase the turnover of
the university, administrators may pay efforts to improve the
attractiveness of campus dining halls, especially for professors
and graduate students.
By changing the number of clusters K, other interesting
phenomenons come out. For example, when K = 6,
the
clustering results show that users holding the same brand of
mobile devices, Apple and MI, are clustered into the same
group. The reason may be that different brands of smart
phones may tend to visit the APP domains that are within
its own ecosystem. When combining with their gender, it is
Fig. 4. Percentage of males in each cluster based on domains.
casual web games such as Plants vs. Zombies to big games
such as DotA.
Fig. 5 shows the user birth year distributions of each cluster
on a contour map. The clusters are the same as Fig. 3. There are
two highlighted regions in this figure. The upper one shows
that in the clusters of games and social networks, there are
more undergraduate students. The lower one shows that in the
cluster preferring recruiting websites, there are more graduate
students who are likely to be searching for jobs for themselves.
The results are quite straightforward that young students prefer
spending more time on entertainment, while graduate students
would like to pay more attention on job hunting and their
future direction. Combining with Fig. 4, these findings warn
the university to supervise the male undergraduates and avoid
them from being addicted to the games and abandoning their
studies. For graduate students, more recruitment information
can be targeted to them.
Fig. 6 shows the distribution of the average amount of
money consumed on the campus in five months, by users
of each cluster. The unit of the amount is RMB. Comparing
the statistics with the age distribution in Fig. 5, we can
find that undergraduate students spend more money inside
the campus than graduate students. As far as we know,
students in this university are divided into two campus sites.
The undergraduate students are living on the campus in the
IEEE 18th International Symposium on.
pp. 1–6.
IEEE, 2007,
[7] W. Paireekreng and K. Wong, “Client-side mobile
user profile for content management using data min-
ing techniques,” in Natural Language Processing, 2009.
SNLP’09. Eighth International Symposium on.
IEEE,
2009, pp. 96–100.
[8] Y. Xie and V. V. Phoha, “Web user clustering from access
log using belief function,” in Proceedings of
the 1st
international conference on Knowledge capture. ACM,
2001, pp. 202–208.
[9] S. Moghaddam and A. Helmy, “Multidimensional model-
ing and analysis of wireless users online activity and mo-
bility: A neural-networks map approach,” in Proceedings
of the 14th ACM international conference on Modeling,
analysis and simulation of wireless and mobile systems.
ACM, 2011, pp. 401–408.
[10] S. Moghaddam, A. Helmy, S. Ranka, and M. Somaiya,
“Data-driven co-clustering model of internet usage in
large mobile societies,” in Proceedings of the 13th ACM
international conference on Modeling, analysis, and sim-
ulation of wireless and mobile systems. ACM, 2010, pp.
248–256.
[11] M. ElSherief, T. ElBatt, A. Zahran, and A. Helmy, “The
quest for user similarity in mobile societies,” in Pervasive
Computing and Communications Workshops (PERCOM
Workshops), 2014 IEEE International Conference on.
IEEE, 2014, pp. 569–574.
[12] R. Giri, H. Choi, K. S. Hoo, and B. D. Rao, “User
behavior modeling in a cellular network using latent
dirichlet allocation,” in Intelligent Data Engineering and
Automated Learning–IDEAL 2014. Springer, 2014, pp.
36–44.
[13] A. Sundin, “Word space models for web user clustering
and page prefetching,” 2012.
[14] D. Arthur and S. Vassilvitskii, “k-means++: The advan-
tages of careful seeding,” in Proceedings of the eighteenth
annual ACM-SIAM symposium on Discrete algorithms.
Society for Industrial and Applied Mathematics, 2007,
pp. 1027–1035.
obvious that males prefer MI, while females prefer Apple.
Therefore, when an update for either of the brands releases,
the administrators for the campus networks can deploy cache
for Apple updates at girls’ dorms and cache for MI updates at
boys’ dorms.
VI. CONCLUSION
In this paper, we propose a method to analyze users’ inter-
net browsing behaviour and preferences. The user behaviour
modeling is based on an analogy between user behaviour pro-
filing and topic modeling in document classification. Methods
including TF-IDF and LSA are part of our process. With K-
Means++ algorithm, we are able to extract hidden browsing
habits by partitioning similar users into groups.
Combined with additional demographic information such
as age and gender of users, the clustering results reveal more
insights on user interests, which indicates the effectiveness of
our proposed algorithm. The browsing habits of users vary
with age and the differences between males and females are
obvious. This helps us to give advice for campus management,
including student management, school development and net-
work optimization.
With larger scaled dataset, e.g., the data in a district or
even a city, we may get more general user behaviour and these
findings can be used in many aspects, such as commercial rec-
ommendation and network operation management. Especially,
user behaviour may help with resource allocation and energy
saving problems.
ACKNOWLEDGMENT
This work is sponsored in part by the National Basic
Research Program of China (No. 2012CB316001), and the
Nature Science Foundation of China (No. 61201191 and
61401250),
the Creative Research Groups of NSFC (No.
61321061), the Sino-Finnish Joint Research Program of NSFC
(No. 61461136004), and Hitachi R&D Headquarter. We also
want to express our sincere appreciation for Shanghai Jiaotong
University to provide all the datasets used in this work.
REFERENCES
[1] N. Lin, G. Chen, K. Zheng, and Y. Tang, “Clusoaf:
A cluster-based semantic oriented analyzing framework
for user behaviors in mobile learning environment,” in
Human Centered Computing. Springer, 2015, pp. 340–
351.
[2] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi, “Un-
derstanding individual human mobility patterns,” Nature,
vol. 453, no. 7196, pp. 779–782, 2008.
[3] N. Bicocchi and M. Mamei, “Investigating ride sharing
opportunities through mobility data analysis,” Pervasive
and Mobile Computing, vol. 14, pp. 83–94, 2014.
[4] S. T. Dumais, “Latent semantic analysis,” Annual review
of information science and technology, vol. 38, no. 1, pp.
188–230, 2004.
[5] J. Xu and H. Liu, “Web user clustering analysis based
on kmeans algorithm,” in Information Networking and
Automation (ICINA), 2010 International Conference on,
vol. 2.
IEEE, 2010, pp. V2–6.
[6] P. N. Karamolegkos, C. Z. Patrikakis, N. Doulamis,
and E. Z. Tragos, “User-profile based communities as-
sessment using clustering methods,” in Personal, Indoor
and Mobile Radio Communications, 2007. PIMRC 2007.
|
1712.06979 | 2 | 1712 | 2018-09-05T10:31:33 | An Automorphic Distance Metric and its Application to Node Embedding for Role Mining | [
"cs.SI",
"physics.soc-ph"
] | Role is a fundamental concept in the analysis of the behavior and function of interacting entities represented by network data. Role discovery is the task of uncovering hidden roles. Node roles are commonly defined in terms of equivalence classes, where two nodes have the same role if they fall within the same equivalence class. Automorphic equivalence, where two nodes are equivalent when they can swap their labels to form an isomorphic graph, captures this common notion of role. The binary concept of equivalence is too restrictive and nodes in real-world networks rarely belong to the same equivalence class. Instead, a relaxed definition in terms of similarity or distance is commonly used to compute the degree to which two nodes are equivalent. In this paper, we propose a novel distance metric called automorphic distance, which measures how far two nodes are of being automorphically equivalent. We also study its application to node embedding, showing how our metric can be used to generate vector representations of nodes preserving their roles for data visualization and machine learning. Our experiments confirm that the proposed metric outperforms the RoleSim automorphic equivalence-based metric in the generation of node embeddings for different networks. | cs.SI | cs |
An Automorphic Distance Metric and its Application to Node Embedding for
Role Mining
V´ıctor Mart´ınez∗
Research Center for Information and Communications Technologies, University of Granada, Spain
Fernando Berzal† and Juan-Carlos Cubero‡
Department of Computer Science and Artificial Intelligence, University of Granada, Spain
Role is a fundamental concept in the analysis of the behavior and function of interacting entities
represented by network data. Role discovery is the task of uncovering hidden roles. Node roles are
commonly defined in terms of equivalence classes, where two nodes have the same role if they fall
within the same equivalence class. Automorphic equivalence, where two nodes are equivalent when
they can swap their labels to form an isomorphic graph, captures this common notion of role. The
binary concept of equivalence is too restrictive and nodes in real-world networks rarely belong
to the same equivalence class.
Instead, a relaxed definition in terms of similarity or distance
is commonly used to compute the degree to which two nodes are equivalent.
In this paper,
we propose a novel distance metric called automorphic distance, which measures how far two
nodes are of being automorphically equivalent. We also study its application to node embedding,
showing how our metric can be used to generate vector representations of nodes preserving their
roles for data visualization and machine learning. Our experiments confirm that the proposed
metric outperforms the RoleSim automorphic equivalence-based metric in the generation of node
embeddings for different networks.
I. INTRODUCTION
Role discovery is defined as the process of finding sets of nodes following similar connectivity patterns or structural
behaviors (Rossi and Ahmed, 2015). The role of a node can be understood as the function that node plays in the
network. Different studies have shown the importance of roles in different domains, including predator-prey food webs
(Luczkovich et al., 2003), international relations (Hafner-Burton et al., 2009), or the function of proteins in proteomes
(Holme and Huss, 2005).
Unfortunately, this problem has received limited attention when compared to community detection (Fortunato,
2010; Lancichinetti and Fortunato, 2009; Papadopoulos et al., 2012), despite the fact that role discovery identifies
complementary information and has found application in several useful network data mining tasks. For example,
roles can be used to model and characterize the behaviors of entities in a network to predict structural changes and
detect anomalies (Rossi et al., 2013). Since the same roles can be observed across different networks, this information
has been successfully exploited for transfer learning (Henderson et al., 2012). Role information can also be used for
enhanced visualization of interesting patterns in graphs (Xing et al., 2010). Additional applications of role discovery
are covered in more detail in (Rossi and Ahmed, 2015).
Formally, two nodes have the same role if, given an equivalence relation, they belong to the same equivalence class
(Burt, 1990; White et al., 1976). Different equivalence classes haven been studied for nodes in networks.
Structural equivalence, where two nodes play the same role if they are connected to exactly the same neighbor nodes,
has been widely studied (Lorrain and White, 1971; Sailer, 1978). These nodes will have exactly the same topological
properties, such as degree, clustering coefficient, or centrality, since they are indistinguishable from a structural point
of view. However, different authors have pointed out the limitations of structural equivalence for modeling roles or
positions (the name that roles receive in sociology), since structural equivalence is more related to the concept of
locality than the actual concept of role (Borgatti and Everett, 1992). If the constraint of needing to be connected to
exactly the same neighbors is relaxed to being connected to neighbors with exactly the same topological function,
we obtain automorphic equivalence classes, where two nodes are equivalent if they can swap their labels to form
an isomorphic graph (Friedkin and Johnsen, 1997; Pattison, 1988). Automorphically equivalent nodes will also have
exactly the same topological properties but, without the requirement of locality imposed by structural equivalence,
∗Electronic address: [email protected]
†Electronic address: [email protected]
‡Electronic address: [email protected]
2
pairs of nodes at distances larger than two can still have the same role. Therefore, automorphic equivalence is more
closely related to the intuitive concept of role, understood as the function of a node within a network.
Other equivalence classes,
less relevant than the ones previously mentioned, are not covered in this work.
Regular equivalence deserves special mention due to its importance as a relaxation of automorphic equivalence
that only requires being connected to nodes with the same function, omitting the actual count of connections
(Everett and Borgatti, 1994). Regular equivalence does not preserve topological properties and is more suited to
hierarchically-organized networks (Luczkovich et al., 2003).
These binary equivalences are strict mathematical abstractions that rarely occur in real-world networks, leading to
all nodes being assigned a different role. In practice, these equivalences are relaxed to similarities, allowing two nodes
to play the same role by partially satisfying the constraints imposed by the mathematical definition of structural,
automorphic, or regular equivalence.
In this paper, we present a novel automorphic distance metric, capturing distances between nodes in terms of
automorphic equivalence. According to the network structure, two nodes will be at a distance that is proportional
to how far they are from being automorphically equivalent. This leads to a softer definition of automorphic roles,
instead of forcing all nodes to fit in strict classes of roles. However, when needed, these distances can be used to
discover and instantiate specific role classes. Our distance function satisfies metric axioms, as we prove below, does
not require external parameters nor feature engineering, and is computable for nodes across different networks. We
also present different applications of our proposal, with special emphasis on generating node embeddings that preserve
node roles. Node embeddings are vector representations of nodes capturing relevant information in terms of pair-wise
distances (Goyal and Ferrara, 2017). Much work has been done in embedding techniques that preserve neighborhoods
or communities (Grover and Leskovec, 2016; Zheng et al., 2016), but, to the best of our knowledge, our work is the
first one on role-preserving embeddings.
Our paper is structured as follows. In Section 2, we discuss the relevant related work. In Section 3, we describe our
proposed automorphic distance metric and study its admissibility as a distance metric, as well as its computational
complexity. In Section 4, we analyze its application to the generation of node embeddings that preserve roles and
show how it outperforms previously proposed approaches. Finally, conclusions and suggestions for future research are
presented in Section 5.
II. RELATED WORK
Different metrics have been proposed to measure node similarity. One of the most popular metrics is SimRank
(Jeh and Widom, 2002), which iteratively computes similarity scores based on the hypothesis that two nodes are
similar if they link to similar nodes. Different extensions of SimRank have been proposed (Hamedani and Kim, 2016).
SimRank recursively computes similarity of two nodes according to the average similarity of all their neighbor pairs,
which can also be interpreted, as suggested by its original authors, as how soon two random walkers will meet if they
start from these nodes. Thus, this definition is not suitable as a metric of similarity capturing automorphic equivalence
because it requires the two nodes to be close to play the same role. Other similarity measures not based on SimRank
have been proposed, such as PageSim (Lin et al., 2006) and Leicht's vertex similarity (Leicht et al., 2006). These
similarities are formally rejected as valid metrics for capturing automorphic equivalence in (Jin et al., 2014).
Since automorphic equivalence ensures the same topological properties, some authors have tried to capture auto-
morphic equivalence by defining a similarity function over a set of network topological properties (Li et al., 2015).
The problem of these feature-based methods is that they require combining different complex hand-crafted features
obtained by experts, which is far from trivial in practice. In addition, they cannot guarantee which set of features
will correctly approximate automorphic equivalence, leading to a very limited approach to automorphic equivalence
discovery.
As far as we know, RoleSim (Jin et al., 2011, 2014) is the only proposed metric that tries to formally capture the
concept of automorphic equivalence without using limited approximations based on hand-crafted topological features.
Omitting the decay factor they introduce, by setting it to 0 in order to capture the global network topology, this
similarity measure is iteratively computed until convergence as
s(x, y) = max
M(x,y)
P(u,v)∈M(x,y) s(u, v)
deg(x) + deg(v) − M (x, y)
where deg(n) is the degree of a node n and M (x, y) is the optimal assignment of nodes in the neighborhood of x to
nodes in the neighborhood of y maximizing the expression; that is, the pairs of neighbors of x and y with maximal
similarity. In their work, Jin et al. prove that this function satisfies the axioms required to be considered a valid
role similarity metric. RoleSim is a form of generalized Jaccard coefficient based on a recursive definition of the
similarity of neighbor roles. Despite the admissibility of RoleSim, their approach presents several limitations. The
RoleSim similarity can be considered an automorphic distance by taking its complementary or Jaccard distance:
d(x, y) = 1 − s(x, y). The problem is that the Jaccard coefficient is a normalized metric, which leads to a normalized
distance. As will be shown in our experimentation, this normalization has a negative impact on the results obtained
by RoleSim. In addition, this similarity function exhibits serious inconsistencies. For example, in the graph shown in
Figure 1, where node d has a one-to-many relationship to xi nodes, the node pair (a, c) has the same exact similarity
than any pair (a, xi), independently of the number of xi nodes. This simple example shows the limitations of RoleSim
when trying to capture automorphic similarity.
3
a
b
d
c
x1
x2
...
xn
FIG. 1: Example graph where RoleSim yields inconsistent values. Node d has a one to many relation to xi nodes.
As far as we know, no distance metric has been proposed that is able to capture the concept of automorphic
distance in a consistent way, without relying on approximations based on extracted topological properties nor forcing
the normalization of the distance function.
III. A NOVEL AUTOMORPHIC DISTANCE METRIC
An isomorphism is a bijection between the nodes of two graphs where two nodes are adjacent in one graph if and only
if the nodes that result from applying the bijective function are also adjacent in the other graph. An automorphism
is an isomorphism from one graph to itself. Therefore, two nodes are automorphically equivalent if there exists an
automorphism creating a correspondence between them.
One form of testing for automorphic equivalence is computing the canonical form of graphs. Graph canonization
is the task of computing a labeling for nodes in a graph such that every isomorphic graph yields the same canonical
labeling. Given a canonized graph, two automorphically-equivalent nodes must have been assigned the same label. As
previously stated, automorphic equivalence is too restrictive to appear in real-world networks, leading to most nodes
having different canonical labels.
The solution that we propose to this problem is the definition of distances between labels, which ultimately allows
the definition of distances between nodes based on the concept of automorphic equivalence. This distance will be
proportional to the number of changes that need to be done in the network to transform one label or equivalence class
into another. A zero distance implies that two nodes are automorphically equivalent and play exactly the same role.
According to this distance d, we can say that nodes x and y are more automorphically similar or have a more similar
role than u and v if d(x, y) < d(u, v). In order to propose a valid distance metric, we must also prove that our metric
satisfies the distance metric axioms.
Our work is based on the 1-dimensional Weisfeiler-Lehman test of isomorphism (Furer, 2001; Weisfeiler and Lehman,
1968), also known as color refinement, which is an algorithm to compute the canonical labeling of graphs. These canon-
ical labels can be used to solve related problems, such as the computation of efficient graph kernels (Shervashidze et al.,
2011). The Weisfeiler-Lehman algorithm works by initially assigning a label to each node according to its degree, so
nodes with the same degree have the same initial label. Then, the algorithm iteratively updates these labels by the
following procedure. First, it takes the labels from neighbor nodes, concatenates them according to certain arbitrary
order (the same ordering must be applied for all nodes), and finally appends the label of the node at the beginning
of the obtained list. Each different sequence is substituted by a newly generated unique label so nodes exhibiting
exactly the same sequence are assigned the same label. This refinement process is repeated until labels stabilize, that
is, when every pair of nodes with the same label in the previous iteration have the same label in the current iteration.
Therefore, after m iterations, which depend on the network diameter, the canonical form is achieved and an additional
iteration is required for testing the stabilization condition. These final labels are the canonical form of the graph and,
4
therefore, two nodes with the same final label are automorphically equivalent. An example of running the algorithm
in a simple graph is shown in Figure 2.
2
1
5
7
3
2
4
4
(b) Initial labeling.
6
8
5
9
(a) Original graph.
32,2,4
23,4
41,2,3,4
41,1,2,4
14
14
(c) First iteration.
(d) First relabeling.
65,5,8
56,9
11
12
79
79
10
14
85,6,7,9
95,7,7,8
15
16
(e) Second iteration.
(f) Second and final relabeling.
FIG. 2: The Weisfeiler-Lehman canonization algorithm applied to a simple graph.
1
1
7
7
13
13
23,4
14
56,8
78
It can be noted that some pairs of the labels appearing in the same iteration of the Weisfeiler-Lehman algorithm
are more similar than others. The automorphic distance between two nodes can be defined as the distance between
their canonical labels. We propose a scheme to compute distances between the labels that are obtained by the
Weisfeiler-Lehman algorithm. Since distances are only defined for labels appearing in the same iteration, a special
label associated to nodes of degree 0, which we call the empty label ℓ∅, is considered for convenience. Isolated nodes
are directly assigned this label and left out of the iterative process.
Since labels created in the initial assignment are based on node degree, we define the distance of labels of nodes x
and y as the number of links that must be added to or removed from node x to transform it into node y. This can be
easily computed as their absolute degree difference:
d(ℓ0(x), ℓ0(y)) = deg(x) − deg(y),
(1)
where ℓ0(n) is the initially-assigned label to node n. This definition of distance for initial labels is also valid for
isolated nodes, with degree 0, which have been assigned the empty label.
Given these distances for initial labels, the distance between labels for the subsequent iterations can be computed
as the distance of the optimally-matched pairs of labels of their neighbors from the previous iteration. The distance
of labels from the i-th iteration can be computed as
d(ℓi(x), ℓi(y)) = min
Mi−1(x,y)
X
d(ℓi−1(u), ℓi−1(v)),
(u,v)∈Mi−1(x,y)
5
(2)
where Mi−1(x, y) is the optimal assignment of neighbors of x to neighbors of y that minimizes the expression and,
therefore, it is just the sum of distances between neighbors of x and y. If the neighborhood of one node is larger than
the neighborhood of the other, leading to unmatched nodes, these nodes are directly matched with virtual nodes,
which are labeled with ℓ∅. The distances of unmatched nodes to the empty label can be seen as the cost, in terms
of distance, of inserting and transforming a virtual isolated node to obtain a node with the label of the unmatched
node. The optimal assignment, which would consider O(n!) alternatives using a naive brute force approach, can be
computed in polynomial time using the Hungarian algorithm (Kuhn, 1955).
Initially, Equation 1 and, in subsequent iterations, Equation 2 are used to compute a distance table. At any given
time, only distances from two iterations need to be maintained: the distances currently being computed and the
distances from the most recent previous iteration.
The described iterative process is carried out for each iteration of the 1-dimensional Weisfeiler-Lehman algorithm
until label stabilization. The automorphic distance between a pair of nodes is defined as the distance between their
canonical labels. It should be noted that labels can be represented using any set of symbols. However, for simplicity,
we represent labels as positive integers.
Algorithm 1 Automorphic Distance Algorithm
procedure Automorphic Distance
Input: Set of nodes N of an undirected graph.
Initialize i = 0, stabilized = f alse, r = HashT able().
Set ℓi(x) = deg(x) ∀x ∈ N .
Set d(ℓi(x), ℓi(y)) = deg(x) − deg(y) ∀x, y ∈ N × N .
while not stabilized do
Set m = i. Iterate i = i + 1.
Set hi(x) = concatenate(sort(neighbors(x))).
Set ci(x) = concatenate(ℓi−1(x), hi(x)).
Set ℓi(x) = unique(ci(x)).
Use Hungarian algorithm to compute Mi−1(x, y) ∀x, y ∈ N × N .
Set d(ℓi(x), ℓi(y)) = minMi−1(x,y) P(u,v)∈Mi−1(x,y) d(ℓi−1(u), ℓi−1(v)) ∀x, y ∈ N × N .
Set stabilized = true.
Store in r the label of x in iteration i − 1 ∀x ∈ N as r[ℓi(x)] = ℓi−1(x).
If the entry r[ℓi(x)] was already set with a different label, set stabilized = f alse.
end while
Set d(x, y) = d(ℓm(x), ℓm(y)) for each pair of nodes x, y ∈ N × N .
Output: Pairwise automorphic distances d(x, y) for each pair of nodes x, y ∈ N × N .
end procedure
The complete algorithm is shown in Algorithm 1. The function neighbors(x) returns the set of neighbor nodes of
node x. The function sort(s) sorts a set of elements. The ordering among elements is not relevant for the algorithm,
but the same ordering must always be applied. The function concatenate(x1, . . . , xn) returns the concatenation of
elements x1, . . . , xn. Finally, the function unique(s) generates and returns a unique symbol, such an integer, for each
observed unique string s, where unique(s) = unique(s′) if and only if s = s′.
A. Case Example
In this section, we show an illustrative example applying the proposed automorphic distance metric to the network
shown in Figure 2a.
The proposed algorithm for computing the automorphic distance initially assigns a label to each node according to
its degree, as shown in Figure 2b. Therefore, two nodes will have the same label if and only if they have the same
degree. The initial distance table, represented as an upper triangular matrix due to the symmetry of distances, as will
be proved in Section III.B.3, is shown in Table I. This distance table is computed using Equation 1 according to the
initial label assignations. For example, the distance between the labels 1 and 4 is 3, since this value is the absolute
degree difference of the corresponding nodes.
After the initialization, the algorithm enters into its main loop and performs its first iteration. For each node, the
labels of its neighbors are ordered and concatenated with its own label, as shown in Figure 2c. For example, the only
6
ℓ/ℓ ℓ∅ 1 2 3 4
ℓ∅ 0 1 2 3 4
0 1 2 3
1
0 1 2
0 1
0
2
3
4
TABLE I: Initialization of the distance table.
node with label 3 has the associated string 32, 2, 4, since its neighbors have labels 2, 4, and 2. These concatenated
strings are replaced by a new label, chosen so that two nodes are assigned the same new label if and only if they had
the same concatenated string. This process generates a new labeling as shown in Figure 2d. Given these new labels,
the algorithm computes their pairwise distances using Equation 2, which are shown in Table II.
ℓ/ℓ ℓ∅ 5 6 7 8 9
ℓ∅ 0 7 8 4 10 8
0 3 3 3 3
5
0 4 4 2
0 6 4
0 2
0
7
8
9
6
TABLE II: Distance table after the first relabeling.
For example, in order to compute the distance between labels 5 and 9, the optimal assignment between their
In the
neighbors minimizing the summation of distances, according to the previous iteration, must be obtained.
previous iteration, 3 and 4 were the labels of the neighbors of nodes with label 5. Likewise, 1, 1, 2, and 4 were the
labels of the neighbors of nodes with label 9. The Hungarian algorithm matches these neighbor labels to minimize
their sum of distances: (3, 2) and (4, 4) according to Table I. Since the two neighbor labels 1 of label 9 were left
unmatched, they are both matched with the empty label as (1, ℓ∅). Given this optimal assignment, we can compute
the distance between labels 5 and 9 as:
d1(5, 9) = d0(3, 2) + d0(4, 4) + d0(1, ℓ∅) + d0(1, ℓ∅) = 1 + 0 + 1 + 1 = 3
Following this iterative process, the algorithm performs the second iteration. Concatenated strings are computed
as shown in Figure 2e and labels are updated as shown in Figure 2f. It can be easily seen that labels have stabilized,
obtaining the canonical labeling of this graph. The stabilization condition can be tested by performing an additional
iteration and observing that nodes with label 13 are assigned the same label, while the other nodes are assigned an
unique new label. This new labeling would be equivalent to the labeling obtained in this iteration, the condition
required to achieve stabilization. The pairwise distances computed in this iteration are shown in Table III.
11
ℓ/ℓ ℓ∅ 10 11 12 13 14 15 16
ℓ∅ 0 18 24 16 8 10 27 25
0 10 2 12 8 13 11
10
0 12 16 14 13 7
8 10 11 13
2 19 17
0
0 21 15
6
0
12
14
13
15
16
0
0
TABLE III: Distance table after the second and final relabeling.
For instance, let us compute the distance between labels 11 and 16. We start by finding the optimal assignment
that minimizes the pairwise distances of their neighbor labels in the previous iteration, which are 5, 5, and 8 for label
7
11 and 5, 7, 7, and 8 for label 16. The Hungarian algorithm obtains the optimal matching (5, 5), (5, 7), and (8, 8),
with an additional (ℓ∅, 7), due to the difference of the node degrees associated to labels 11 and 16. Once this optimal
matching has been obtained, we can easily compute the distance between labels 11 and 16 using Equation 2 as:
d2(11, 16) = d1(5, 5) + d1(5, 7) + d1(8, 8) + d1(ℓ∅, 7) = 0 + 3 + 0 + 4 = 7
The automorphic distance between a pair of nodes is defined as the distance between their canonical labels, which
are the final labels assigned in the iteration where stabilization is achieved. Therefore, in our example, the distance
between nodes is given by Table III. For example, we can see how nodes with canonical labels 13 and 14 are close to
being automorphically equivalent, since their automorphic distance is only 2. In contrast, nodes with canonical labels
14 and 15 have a large automorphic distance, equal to 21, which indicates that they are far from being automorphically
equivalent.
B. Metric Admissibility
In this section, we prove that the distance function that we have defined is a valid metric or distance function. In
order to assert this statement, we must prove the following four conditions: non-negativity, identity of indiscernibles,
symmetry, and triangle inequality.
To prove these conditions, we note that Equation 2 can be recursively decomposed as
d(ℓm(x), ℓm(y)) = min
Mm−1(x,y)
X
d(ℓm−1(u), ℓm−1(v))
(u,v)∈Mm−1(x,y)
d(ℓm−2(u′), ℓm−2(v′))
= min
Mm−1(x,y)
Mm−2(u,v)
X
(u,v)∈
X
(u′,v′)∈
Mm−1(x,y)
Mm−2(u,v)
= . . .
= X
(x′,y ′)∈M ′(x,y)
d(ℓ0(x′), ℓ0(y′))
= X
deg(x′) − deg(y′)
(x′,y ′)∈M ′(x,y)
(3)
where M ′(x, y) is the set of pairs of nodes that appear in the recursive summation at the deepest level of recursion
as a result of choosing the optimal assignment in each iteration.
1. Proof of Non-Negativity
Non-negativity requires that the distance function satisfies d(x, y) ≥ 0 for any possible pair of nodes x and y. Given
the decomposition of our metric as shown in Equation 3, it is straightforward to see that the summation of absolute
values is guaranteed to be always equal or greater than 0.
2. Proof of the Identity of Indiscernibles
The identity of indiscernibles implies that the distance function satisfies d(x, y) = 0 if and only if x ≡ y. The
Weisfeiler-Lehman algorithm guarantees that automorphically equivalent pairs of nodes are assigned the same canon-
ical label, and non-automorphically equivalent pairs of nodes are assigned different canonical labels.
Equation 3 is only equal to zero when deg(x′) = deg(y′) for every pair of nodes in M ′(x, y). Two nodes can
only have the same canonical label if they are automorphically equivalent, as guaranteed by the Weisfeiler-Lehman
algorithm. If two nodes are assigned the same canonical label, their neighbors must have been assigned equivalent
labels in all the iterations. Since the distance of a label to itself is 0, we can see that the summation yields 0, leading
to a zero distance for nodes when x ≡ y.
On the other side, when two nodes have different canonical labels, their neighbors must have been assigned different
labels during the execution of the algorithm. This implies that, in the recursive decomposition shown in Equation 3,
at least one pair of nodes will not match nodes with the same initial labels, leading to a distance greater than 0 for
nodes x 6≡ y.
8
3. Proof of Symmetry
The condition of symmetry requires that the proposed distance function must satisfy the property d(x, y) = d(y, x).
We can prove that Equation 3 is symmetric, as
d(ℓm(x), ℓm(y)) = X
deg(x′) − deg(y′)
(x′,y ′)∈M ′(x,y)
= X
deg(y′) − deg(x′)
(y ′,x′)∈M ′(y,x)
= d(ℓm(y), ℓm(x))
since M ′(x, y) is equal to M ′(y, x) when we swap the nodes. The order of the nodes in each pair does not affect the
result of our distance function.
4. Proof of the Triangle Inequality
The triangle inequality requires that the inequality d(x, y) ≤ d(x, z)+d(z, y) is satisfied by the proposed automorphic
distance function.
By the definition of the proposed distance, we know that
d(x, y) ≤ d(x, z) + d(z, y) =⇒
d(ℓm(x), ℓm(y)) ≤ d(ℓm(x), ℓm(z)) + d(ℓm(z), ℓm(y)) =⇒
deg(a) − deg(c) + X
deg(a) − deg(b) ≤ X
X
(a,b)∈M ′(x,y)
(a,c)∈M ′(x,z)
(c,d)∈M ′(z,y)
deg(c) − deg(d).
We know that the absolute value satisfies the triangle inequality, thus the lower value that the right side can take is
X
deg(a) − deg(b) ≤ X
deg(a) − deg(d)
(a,b)∈M ′(x,y)
(a,d)∈M ′′(x,y,z)
where M ′′(x, y, z) is the set of pairs resulting from chaining or combining M ′(x, z) and M ′(z, y) so that (a, d) ∈
M ′′(x, y, z) if and only if, (a, c) ∈ M ′(x, z) and (c, d) ∈ M ′(z, y).
For this inequality to hold, it requires the non-existence of a pairing of nodes better than the matching done in the
left side. Since the Hungarian algorithm ensures that matchings are optimal, minimizing their sum of distances, the
matching at the right side cannot be better than optimal and, therefore, the value of the right side can only be equal
to or greater than the value on the left side, satisfying the triangle inequality condition.
C. Metric Computational Complexity
In this section, we analyze the temporal and spatial computational complexity of our proposed metric.
The initialization of labels based on the degree of nodes has O(n) temporal and spatial complexity, where n is
the number of nodes in the network. The computation of the table for the initial distances has O(n2) temporal and
spatial complexity, since the distance is computed for every pair of nodes.
Each iteration of the algorithm requires computing the sorted list of labels of the neighbors for each node. This
task can be accomplished for each node with computational and spatial complexity O(k), where k is the degree of the
node, when using radix, bucket, or counting sort. Thus, computing these strings for all nodes has O(nk) temporal
and spatial complexity. Renaming these labels can be done in O(n) time using hash-based data structures. To
compute the pairwise distances between labels in the current iteration, the Hungarian algorithm, with computational
complexity O(k3), must be computed for each pair of nodes, leading to O(n2k3) temporal complexity and O(n2)
spatial complexity. Finally, checking if the labels have stabilized can be done in O(n) using a hash-based index.
The number of iterations, m, required for 1-dimensional Weisfeiler-Lehman algorithm to converge is closely related
to the diameter of the network (Furer, 2001). Even though the number of iterations is theoretically bounded by n, it
has been widely observed that real-world networks tend to exhibit the small-world phenomenon, presenting a small
9
diameter (Travers and Milgram, 1967; Watts and Strogatz, 1998) and leading to a small number of iterations required
for convergence.
Therefore, by combining these partial results, the total temporal complexity is O(n + n2 + mn2k3) ≈ O(mn2k3),
where n is the number of nodes in the graph, k is the degree of nodes, and m is the number of iterations required for
convergence. The spatial complexity of the algorithm is O(n2), since only the distances and labels from the previous
and the current iteration must be maintained at any given time. In practice, the algorithm can handle large networks,
given that m and k tend to remain small in real-world networks due to their sparse and small-world nature.
In
addition, most of these steps can be easily parallelized, since most of them are independent for each node and are
only based on the results from the previous iteration. For example, the initial labels for each node can be assigned
independently. Once we have assigned these labels, the computation of their pairwise distances can be split among
the available processors, since they are independent. The iterative assignment of labels in each loop iteration can be
also parallelized using a concurrent data structure to ensure that the new labels are properly generated. Furthermore,
the pairwise distances between these new labels can be easily computed in a parallel way, since they are completely
independent as they only rely on the distance table computed in the previous iteration.
IV. EXPERIMENTAL EVALUATION
The experimental evaluation of role discovery techniques is a complicated task due to the lack of available evaluation
datasets. Most role mining research projects use private datasets, which are not released and made publicly available
to the scientific community. To overcome this limitation and perform an illustrative comparison of automorphic
distances, we propose a novel experimental evaluation, which is presented in this section. As a collateral result of this
experimentation, a new method for representing node roles using feature vectors is presented.
A central problem in machine learning is finding representations that ease the visualization or extraction of useful
information from data (Bengio et al., 2013). A common solution is the computation of embeddings that represent
complex objects in a vector space preserving certain properties (Goldberg and Levy, 2014). Node embedding, also
known as graph embedding, is the task of mapping each node in a graph to a dimensional space trying to preserve the
similarity or distance between pairs of nodes. Therefore, similar nodes will be located in similar regions of the space.
Node embeddings have lately gained attention since they have achieved good results in different machine learning
tasks (Goldberg and Levy, 2014). Several models have been proposed for node embedding. However, these techniques
try to preserve the connectivity of the network by obtaining embeddings that preserve structural equivalence, the
neighborhood, or the community of nodes (Grover and Leskovec, 2016; Perozzi et al., 2014; Zheng et al., 2016). This
information has proven to be useful due to the presence of homophily, also known as assortativity, in real-world
networks (McPherson et al., 2001), where entities tend to be connected to similar ones, a feature that allows us to
explain certain features of the nodes. Even though the connectivity information captured by these techniques is
relevant, these techniques fail to capture information related to the role or function of the nodes in the network, which
is a highly-valuable information that is complementary to the information obtained by locality-based embedding
techniques.
We propose a new kind of embedding by exploiting this complementary information. Our node embeddings capture
the roles of nodes by placing nodes that play a similar function in the network close in the resulting vector space.
To compute these embeddings, we apply the classical multidimensional scaling (MDS) (Borg and Groenen, 2005;
Wickelmaier, 2003) to the distance matrices computed by the different approaches. In the following section, we show
how 2-dimensional embeddings can capture relevant information in different real-world networks. We compare the
results obtained by our distance metric with the results obtained by RoleSim, which can be interpreted as a distance
function as previously described.
A. Zachary's Karate Club Network
Zachary's karate club network is a popular social network representing the 34 members (as nodes) of a university
karate club and their interactions (as edges) outside the club (Zachary, 1977). During the study carried out by Wayne
W. Zachary, a conflict arose between the two club administrators, leading to the split of the club into two groups
according to the leader each member decided to follow. For this reason, this network has served as a prototypical case
study for community detection algorithms and some network analysis techniques.
If each member is assigned to a binary class according to the leader it decided to follow, Zachary observed that
this property is highly homogeneous and assortative. Therefore, nodes tend to be connected to nodes that took the
same decision. In this context, previously proposed embedding techniques generate embeddings that clearly separate
nodes according to the leader they decided to follow (Perozzi et al., 2014). This information is crucial when the task
10
is related to community detection. However, these embeddings fail, by nature, to reveal the role of each node in the
network.
To illustrate how our technique captures the roles of nodes, we assigned a class to each node according to objective
role-related properties. The two leaders are colored in red. Nodes interacting with the two leaders are colored in
green. Nodes not interacting with any of the two leaders are colored in yellow. Finally, the remaining nodes, which
interact with only one of the leaders, are colored in blue. Figure 3 shows the network drawn using the Kamada-Kaway
layout algorithm (Kamada and Kawai, 1989) and the embeddings obtained by applying our automorphic distance and
RoleSim. It can be seen the four classes of roles are linearly separated in the embedding generated using our distance
metric. In addition, it can be seen how the two leaders are clearly mapped as outliers, placed significantly apart from
the other nodes representing normal members of the club. However, RoleSim fails to clearly separate the different
role classes. The two leaders are placed pretty close to normal club members, without capturing their unique function
in the network.
(a) Kamada-Kawai layout.
(b) Automorphic embedding.
(c) RoleSim embedding.
FIG. 3: Zachary's karate club network and its node embeddings, with node role coloring.
B. World Trade Network
In network data mining, homophily is commonly exploited in node classification tasks, since nodes tend to exhibit
the same class than their neighbors (Bhagat et al., 2011). Even though this situation occurs in a large number of
networks from very different domains, homophily-based classification techniques fail when the classes of the nodes are
defined by the role they play in the network, instead of by the community they belong to.
An illustrative example of this situation is a network containing data on trade of miscellaneous manufactures of
metal among 80 countries 1, according to data gathered in 1993 and 1994 from the Commodity Trade Statistics
published by the United Nations (Nooy et al., 2011). Each country is represented by a node in the network. Each
commercial relationship is represented by an arc, which we consider an undirected edge in practice. In this case, arcs
correspond to trading high technology products or heavy manufactures between countries.
In addition, the authors of this dataset annotated countries in the network with their structural world economic
position in 1994. World economic positions are a classification of countries in the context of the world-system theory
that explains some complex dynamics observed in the real world (Chirot and Hall, 1982; Smith and White, 1992).
This classification splits countries into three possible categories: core countries (colored in green), semi-periphery
countries (colored in blue), and periphery countries (colored in red). In short, core countries have a high economical,
military, and political power, which allows them to control the world economic system. The periphery is composed
of less developed countries, owning a disproportionately small share of global wealth. Finally, the semi-periphery is
conformed by countries that do not clearly fall in the previous two categories and exhibit a more intermediate status.
Figure 4 shows the trade network drawn using the Kamada-Kaway layout algorithm and the embeddings obtained
by applying our automorphic distance and the RoleSim-based distance. It can be seen that both metrics achieve a good
separation of core and semi-periphery countries. However, our distance metric is clearly superior in the separation
of semi-periphery and periphery countries.
In addition, the embedding generated using distances computed with
1 The dataset can be downloaded from http://vlado.fmf.uni-lj.si/pub/networks/data/esna/metalWT.htm
RoleSim exhibits an artificial curved-line shape, which indicates that only a dimension would be required to represent
the information captured by RoleSim. In contrast, the embedding generated using our proposed automorphic distance
exploits the two dimensions to represent role-related node properties and achieves a better separation of the different
classes.
11
(a) Kamada-Kawai layout.
(b) Automorphic embedding.
(c) RoleSim embedding.
FIG. 4: World trade network and its corresponding node embeddings.
V. CONCLUSIONS
In this paper, we have proposed a novel distance metric for nodes that relaxes the strict concept of automorphic
equivalence. To the best of our knowledge, this is the first work to propose a consistent non-normalized distance
metric that captures the concept of automorphic equivalence without approximating it using feature engineering.
In addition, we have shown that the proposed distance function is a valid distance metric by proving the required
conditions. Finally, we have shown how our metric can be exploited to generate node embeddings that capture role
information in contrast to previously proposed embedding techniques, which capture locality-based information. We
have also shown how our metric is superior to RoleSim in the generation of automorphic node embeddings, leading
to a better separation of nodes according to their roles.
Our proposal creates new opportunities in problems related to role discovery. Future work includes exploiting our
distance metric in problems related to anomaly detection in networks and transfer learning based on roles shared by
nodes across different networks.
Acknowledgments
This work is partially supported by the Spanish Ministry of Economy and the European Regional Development
Fund (FEDER), under grant TIN2012-36951 and the program "Ayudas para contratos predoctorales para la formaci´on
de doctores 2013" (grant BES-2013-064699).
References
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 35(8):1798 -- 1828, 2013.
Smriti Bhagat, Graham Cormode, and S. Muthukrishnan. Node classification in social networks, pages 115 -- 148. Springer US,
Boston, MA, 2011. ISBN 978-1-4419-8462-3.
Ingwer Borg and Patrick JF Groenen. Modern multidimensional scaling: Theory and applications. Springer Science & Business
Media, 2005. ISBN 978-0-387-28981-6.
Stephen P Borgatti and Martin G Everett. Notions of position in social network analysis. Sociological Methodology, pages 1 -- 35,
1992.
Ronald S Burt. Detecting role equivalence. Social Networks, 12(1):83 -- 97, 1990.
Daniel Chirot and Thomas D Hall. World-system theory. Annual Review of Sociology, 8(1):81 -- 106, 1982.
Martin G Everett and Stephen P Borgatti. Regular equivalence: General theory. Journal of Mathematical Sociology, 19(1):
29 -- 52, 1994.
Santo Fortunato. Community detection in graphs. Physics Reports, 486(3):75 -- 174, 2010.
Noah E Friedkin and Eugene C Johnsen. Social positions in influence networks. Social Networks, 19(3):209 -- 222, 1997.
12
Martin Furer. Weisfeiler-Lehman refinement requires at least a linear number of iterations. In International Colloquium on
Automata, Languages, and Programming, pages 322 -- 333. Springer, 2001.
Yoav Goldberg and Omer Levy. word2vec explained: deriving Mikolov et al.'s negative-sampling word-embedding method.
arXiv preprint arXiv:1402.3722, 2014.
Palash Goyal and Emilio Ferrara. Graph embedding techniques, applications, and performance: A survey. arXiv preprint
arXiv:1705.02801, 2017.
Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, pages 855 -- 864. ACM, 2016.
Emilie M Hafner-Burton, Miles Kahler, and Alexander H Montgomery. Network analysis for international relations. Interna-
tional Organization, 63(3):559 -- 592, 2009.
Masoud Reyhani Hamedani and Sang-Wook Kim. SimRank and its variants in academic literature data: Measures and
evaluation. In Proceedings of the 31st Annual ACM Symposium on Applied Computing, pages 1102 -- 1107. ACM, 2016.
Keith Henderson, Brian Gallagher, Tina Eliassi-Rad, Hanghang Tong, Sugato Basu, Leman Akoglu, Danai Koutra, Christos
Faloutsos, and Lei Li. Rolx: Structural role extraction & mining in large graphs. In The 18th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD '12, Beijing, China, August 12-16, 2012, pages 1231 -- 1239,
2012.
Petter Holme and Mikael Huss. Role-similarity based functional prediction in networked systems: Application to the yeast
proteome. Journal of the Royal Society Interface, 2(4):327 -- 333, 2005.
Glen Jeh and Jennifer Widom. SimRank: A measure of structural-context similarity. In Proceedings of the 8th ACM SIGKDD
international conference on Knowledge discovery and data mining, pages 538 -- 543. ACM, 2002.
Ruoming Jin, Victor E Lee, and Hui Hong. Axiomatic ranking of network role similarity. In Proceedings of the 17th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 922 -- 930. ACM, 2011.
Ruoming Jin, Victor E Lee, and Longjie Li. Scalable and axiomatic ranking of network role similarity. ACM Transactions on
Knowledge Discovery from Data (TKDD), 8(1):3, 2014.
Tomihisa Kamada and Satoru Kawai. An algorithm for drawing general undirected graphs. Information Processing Letters, 31
(1):7 -- 15, 1989.
Harold W Kuhn. The Hungarian method for the assignment problem. Naval Research Logistics (NRL), 2(1-2):83 -- 97, 1955.
Andrea Lancichinetti and Santo Fortunato. Community detection algorithms: A comparative analysis. Physical Review E, 80
(5):056117, 2009.
Elizabeth A Leicht, Petter Holme, and Mark EJ Newman. Vertex similarity in networks. Physical Review E, 73(2):026120,
2006.
Longjie Li, Lvjian Qian, Victor E. Lee, Mingwei Leng, Mei Chen, and Xiaoyun Chen. Fast and accurate computation of role
similarity via vertex centrality, pages 123 -- 134. Springer International Publishing, 2015. ISBN 978-3-319-21042-1.
Zhenjiang Lin, Michael R Lyu, and Irwin King. PageSim: A novel link-based measure of web page aimilarity. In Proceedings
of the 15th International Conference on the World Wide Web, pages 1019 -- 1020. ACM, 2006.
Francois Lorrain and Harrison C White. Structural equivalence of individuals in social networks. The Journal of Mathematical
Sociology, 1(1):49 -- 80, 1971.
Joseph J Luczkovich, Stephen P Borgatti, Jeffrey C Johnson, and Martin G Everett. Defining and measuring trophic role
similarity in food webs using regular equivalence. Journal of Theoretical Biology, 220(3):303 -- 321, 2003.
Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in social networks. Annual Review
of Sociology, 27(1):415 -- 444, 2001.
Wouter De Nooy, Andrej Mrvar, and Vladimir Batagelj. Exploratory social network analysis with Pajek. Cambridge University
Press, New York, NY, USA, 2011. ISBN 0521174805, 9780521174800.
Symeon Papadopoulos, Yiannis Kompatsiaris, Athena Vakali, and Ploutarchos Spyridonos. Community detection in social
media. Data Mining and Knowledge Discovery, 24(3):515 -- 554, 2012.
Philippa E Pattison. Network models: Some comments on papers in this special issue. Social Networks, 10(4):383 -- 411, 1988.
Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. DeepWalk: Online learning of social representations. In Proceedings of the
20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 701 -- 710. ACM, 2014.
Ryan A Rossi and Nesreen K Ahmed. Role discovery in networks. IEEE Transactions on Knowledge and Data Engineering,
27(4):1112 -- 1131, 2015.
Ryan A Rossi, Brian Gallagher, Jennifer Neville, and Keith Henderson. Modeling dynamic behavior in large evolving graphs.
In Proceedings of the 6th ACM International Conference on Web Search and Data Mining, pages 667 -- 676. ACM, 2013.
Lee Douglas Sailer. Structural equivalence: Meaning and definition, computation and application. Social Networks, 1(1):73 -- 90,
1978.
Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M. Borgwardt. Weisfeiler-Lehman
graph kernels. Journal of Machine Learning Research, 12:2539 -- 2561, 2011.
David A Smith and Douglas R White. Structure and dynamics of the global economy: Network analysis of international trade
1965 -- 1980. Social Forces, 70(4):857 -- 893, 1992.
Jeffrey Travers and Stanley Milgram. The small -- world problem. Phychology Today, 1:61 -- 67, 1967.
Duncan J Watts and Steven H Strogatz. Collective dynamics of small -- world networks. Nature, 393(6684):440 -- 442, 1998.
Boris Weisfeiler and AA Lehman. A reduction of a graph to a canonical form and an algebra arising during this reduction.
Nauchno-Technicheskaya Informatsia, 2(9):12 -- 16, 1968.
Harrison C White, Scott A Boorman, and Ronald L Breiger. Social structure from multiple networks. I. Blockmodels of roles
and positions. American Journal of Sociology, 81(4):730 -- 780, 1976.
Florian Wickelmaier. An introduction to MDS. Sound Quality Research Unit, Aalborg University, Denmark, 46, 2003.
Eric P Xing, Wenjie Fu, Le Song, et al. A state-space mixed membership blockmodel for dynamic network tomography. The
Annals of Applied Statistics, 4(2):535 -- 566, 2010.
Wayne W Zachary. An information flow model for conflict and fission in small groups. Journal of Anthropological Research, 33
(4):452 -- 473, 1977.
Vincent W Zheng, Sandro Cavallari, Hongyun Cai, Kevin Chen-Chuan Chang, and Erik Cambria. From node embedding to
community embedding. arXiv preprint arXiv:1610.09950, 2016.
13
|
1701.08799 | 1 | 1701 | 2017-01-30T19:52:25 | Scalable Bicriteria Algorithms for the Threshold Activation Problem in Online Social Networks | [
"cs.SI"
] | We consider the Threshold Activation Problem (TAP): given social network $G$ and positive threshold $T$, find a minimum-size seed set $A$ that can trigger expected activation of at least $T$. We introduce the first scalable, parallelizable algorithm with performance guarantee for TAP suitable for datasets with millions of nodes and edges; we exploit the bicriteria nature of solutions to TAP to allow the user to control the running time versus accuracy of our algorithm through a parameter $\alpha \in (0,1)$: given $\eta > 0$, with probability $1 - \eta$ our algorithm returns a solution $A$ with expected activation greater than $T - 2 \alpha T$, and the size of the solution $A$ is within factor $1 + 4 \alpha T + \log ( T )$ of the optimal size. The algorithm runs in time $O \left( \alpha^{-2}\log \left( n / \eta \right) (n + m) |A| \right)$, where $n$, $m$, refer to the number of nodes, edges in the network. The performance guarantee holds for the general triggering model of internal influence and also incorporates external influence, provided a certain condition is met on the cost-effectivity of seed selection. | cs.SI | cs | Scalable Bicriteria Algorithms for the Threshold
Activation Problem in Online Social Networks
1
Alan Kuhnle, Tianyi Pan, Md Abdul Alim, and My T. Thai
Department of Computer & Information Science & Engineering
University of Florida
Gainesville, Florida 32611
Email: {kuhnle, tianyi, alim, thai}@cise.ufl.edu
7
1
0
2
n
a
J
0
3
]
I
S
.
s
c
[
1
v
9
9
7
8
0
.
1
0
7
1
:
v
i
X
r
a
Abstract -- We consider the Threshold Activation Problem
(TAP): given social network G and positive threshold T , find a
minimum-size seed set A that can trigger expected activation of at
least T . We introduce the first scalable, parallelizable algorithm
with performance guarantee for TAP suitable for datasets with
millions of nodes and edges; we exploit the bicriteria nature
of solutions to TAP to allow the user to control the running
time versus accuracy of our algorithm through a parameter
α ∈ (0, 1): given η > 0, with probability 1 − η our algorithm
returns a solution A with expected activation greater than
T − 2αT , and the size of
the solution A is within factor
1 + 4αT + log(T ) of the optimal size. The algorithm runs in time
O(cid:0)α−2 log (n/η) (n + m)A(cid:1), where n, m, refer to the number
of nodes, edges in the network. The performance guarantee holds
for the general triggering model of internal influence and also
incorporates external influence, provided a certain condition is
met on the cost-effectivity of seed selection.
I. INTRODUCTION
With the growth of online social networks, viral marketing
where influence spreads through a social network has become
a central research topic. Users of a social network can activate
their friends by influencing them to adopt certain behaviors
or products. In this context, the influence maximization (IM)
problem has been studied extensively [1] -- [8]: given a budget,
the IM problem is to find a seed set, or set of initially
activated users, within the budget that maximizes the expected
influence. Much recent work [3], [4], [9], [10] has developed
scalable algorithms for IM that are capable of running on
social networks with millions of nodes while retaining the
provable guarantees on the quality of solution; namely, that
the algorithm for IM will produce a solution with expected
influence within 1 − 1/e of the optimal activation.
However, a company with a specific target in mind may
adopt a more flexible approach to its budget:
instead of
having a fixed budget k for the seed set, it is natural to
minimize the size of the seed set while activating a desired
threshold T of users within the network. For example, suppose
a company desires a certain level of exposure on social media;
such exposure could boost the sales of any of its products.
Alternatively, suppose a profit goal T for a product must
c(cid:13)2017 IEEE. Personal use of this material is permitted. Permission from
IEEE must be obtained for all other uses, in any current or future media,
including reprinting /republishing this material for advertising or promotional
purposes, creating new collective works, or resale or redistribution to servers
or lists, or reuse of any copyrighted component of this work in other works.
be met with the least expense possible. Thus, we consider
the following threshold activation problem (TAP): given a
threshold T , minimize the size of the set of seed users in order
to activate at least T users of the network in expectation.
Goyal et al. [11] provided bicriteria performance guarantees
for a greedy approach to TAP based upon Monte Carlo
sampling at each iteration to select the best seed node, an
algorithm reminiscent of the greedy algorithm for IM in
Kempe et al. [1]; this approach is inefficient and impractical
for large networks. Although TAP is related to IM, scalable
solutions that already exist for IM are unsuitable for TAP: the
TIM [10] and IMM [9] algorithms require knowledge of the
size k of the seed set ahead of time; the SKIM algorithm [3]
for average reachability has been shown to be effective for IM
in specialized settings, but it is unclear how to apply SKIM to
more general situations or to TAP while retaining performance
guarantees.
Moreover, empirical studies have shown that in the viral
marketing context, it is insufficient to consider internal prop-
agation alone; external activation, i.e. activations that cannot
be explained by the network structure, play a large role in
influence propagation events [12] -- [15], and recent works on
scalable algorithms for IM [3], [9], [10] have neglected the
consequences of external
influence. For internal diffusion,
two basic models have been widely adopted, the independent
cascade and linear threshold models; Kempe et al. [1] showed
these two models are special cases of the triggering model,
a powerful, general model that has desirable properties in a
viral marketing context.
Motivated by the above observations, the main contributions
of this work are:
• We establish a new connection between the triggering
model and a concept of generalized reachability that al-
lows a natural combination of external influence with the
triggering model. We show any instance of the triggering
model combined with any model of external influence is
monotone and submodular.
• We show how to use the generalized reachability frame-
work to efficiently estimate the expected influence of
the triggering model combined with external influence,
leveraging scalable estimators of average reachability by
Cohen et al. [3], [16], [17]. This efficient estimation
results in a parallelizable algorithm (STAB) for TAP
with performance guarantee in terms of user-adjustable
trade-offs between efficiency and accuracy. The desired
accuracy is input as parameter α ∈ (0, 1) which de-
termines running time as O(cid:0)α−2 log(n/η)(n + m)A(cid:1),
where n, m are number of nodes, edges in the network,
and A is the seed set returned by STAB. With probability
1 − η, the expected activation is guaranteed to be with
2αT of threshold T , and the size of the seed set A is
guaranteed to be within factor 1 + 4αT + log(T ) of
the optimal size. If the cost-effectivity of seed selection
falls below 1, this performance bound may not hold; we
provide a looser bound for this case.
• Through a comprehensive set of experiments, we demon-
strate that on large networks, STAB not only returns a
better solution to TAP, but it runs faster than existing
algorithms for TAP and algorithms for IM adapted to
solve TAP, often by factors of more than 103 even
for the state-of-the-art IMM algorithm [9]. In addition,
we investigate the effect of varying levels of external
influence on the solution of STAB.
The rest of this paper is organized as follows: in Section
II, we introduce models of influence, including the triggering
model and our concept of generalized reachability. We prove
these two concepts are equivalent. In Section III, we formally
define TAP and first prove bicriteria guarantees in a general
setting. Next, we employ the generalized reachability concept
to show how combination of triggering model and external
influence can be estimated efficiently, and we present and
analyze STAB, our scalable bicriteria algorithm. In Section
IV, we analyze STAB experimentally and compare with prior
work. We discuss related work in Section V.
II. MODELS OF INFLUENCE
A social network can be modeled as a directed graph
G = (V, E), where V is the set of users and directed edges
(u, v) ∈ E denote social connections, such as friendships,
between the users u, v. In this work, we study the propagation
of influence through a social network; for example, say a user
on the Twitter network posts a message to her account; this
message may be reposted by the friends of this user, and
as such it propagates across the social network. In order to
study such events from a theoretical standpoint, we require
the concept of a model of influence propagation.
Intuitively, the idea of a model of influence propagation
in a network is a way by which nodes can be activated
given a set of seed nodes. In this work, we use σ to denote
a model of influence propagation. Such a model is usually
probabilistic, and the notation σ(S) will denote the expected
number of activations under the model σ given seed set
S ⊆ V . Kempe et al. [1] studied a variety of models in
their seminal work on influence propagation on a graph,
including the Independent Cascade (IC) and Linear Threshold
(LT) models. For completeness, we briefly describe these two
models. An instance of influence propagation on a graph G
follows the IC model if a weight can be assigned to each
edge such that the propagation probabilities can be computed
as follows: once a node u first becomes active, it is given
a single chance to activate each currently inactive neighbor
v with probability proportional to the weight of the edge
2
(u, v). In the LT model each network user u has an associated
threshold θ(u) chosen uniformly from [0, 1] which determines
how much influence (the sum of the weights of incoming
edges) is required to activate u. u becomes active if the total
influence from its activated neighbors exceeds the threshold
θ(u).
These well-studied models are both examples of the Trig-
gering Model, also introduced in [1]: Each node v ∈ G
independently chooses a random "triggering set" Tv according
to some probability distribution over subsets of its neighbors.
A seed set S is activated at time t = 0; a node v becomes
active at time t if any node in Tv is active at time t − 1.
Two important properties that it is desirable for a model σ
of influence propagation to satisfy are firstly the submodularity
property of the expected activation function: for any sets
S, T ⊆ V , σ(S ∪ T ) + σ(S ∩ T ) ≤ σ(S) + σ(T ), and secondly
the monotonicity property: if S ⊆ T ⊆ V , σ(S) ≤ σ(T ).
These properties together allow a greedy approach to have
a performance ratio for the influence maximization problem
(IM): given k, find a seed set A of size k such that the
expected activation of A is maximized. Kempe et al. [1]
showed that the triggering model is both submodular and
monotonic. Both properties are also important
in proving
performance guarantees for TAP, the problem studied in this
work and defined in Section III.
It is #P-hard to compute the exact influence of a single seed
node under even the restricted version of IC where each edge
is assigned the probability 0.5 [18]. Therefore, it is necessary
to estimate the value of σ(S) by sampling the probability
distribution determined by σ. Sampling efficiently such that
the estimated value σ(S) satisfies σ(S) − σ(S) < ε for
all seed sets is a difficult problem; we discuss this problem
when σ is a combination of the triggering model and external
influence in Section III-B2. Because of the errors associated
with estimating the value σ(S), we introduce slightly general-
ized versions of the above two properties. First, let us define
∆xσ(A) = σ (A ∪ {x}) − σ(A), for any model σ, and subset
A ⊂ V . The following property is equivalent to σ satisfying
submodularity and monotonicity together.
Property 1 (Submodularity and monotonicity). For all A ⊆
B, and x ∈ V , ∆xσ(A) ≥ ∆xσ(B).
Let us define a σ to be ε-approximately monotonic and
submodular if the following property is satisfied instead:
Property 2 (ε-submodularity and monotonicity). Let ε > 0.
For all A ⊆ B, and x ∈ V , ε + ∆xσ(A) ≥ ∆xσ(B).
A. Triggering model
reachability
from the perspective of generalized
Next we define a class of influence propagation models that
naturally generalize the notion of reachability in a directed
graph: that is, these models generalize the simple model that
a node is activated by a seed set S if it is reachable from the
seed set by edges in the graph. Somewhat surprisingly, the
triggering model is equivalent to this notion of generalized
reachability.
of directed graphs {(Gi = (V, Ei), pi)}l
set V , and associated probabilities such that (cid:80)l
Suppose instead of a single directed graph G, we have a set
i=1 on the same vertex
i=1 pi = 1.
Then define an influence propagation model σ in the following
way: when a seed set S is activated, graph Gi is chosen with
probability pi. Then, influence is allowed to propagate from
seed set S in Gi according to the directed edges of Gi. Let
σi(S) be defined as the number of vertices reachable from S
in Gi. Then the expected activation of a seed set S is given
i=1 piσi(S). We will term a model σ of this
form a model of generalized reachability, since it generalizes
the notion of simple reachability on a directed graph. We have
the following important proposition:
Proposition 1. Generalized reachability is equivalent to the
Triggering Model.
by σ(S) = (cid:80)l
(cid:81)
Proof: Suppose we have an instance σ of the triggering
model. For each node v, triggering set Tv is chosen indepen-
dently with probability p(Tv). If all nodes choose a triggering
set, then define graph HT for this choice by adding directed
edges (u, v) for each u ∈ Tv. Assign graph HT probability
v p(Tv) and we have instance of generalized reachability
with the same expected activation.
Conversely, suppose σ is an instance of generalized reach-
ability. For v ∈ V , let Tv be any subset of nodes excluding v
itself. Assign p(Tv) to be the sum of the probabilities of the
graphs in which the in-neighbors of v are exactly the set Tv.
Then we have an instance of the triggering model.
B. External influence
In this section, we outline our model of external influence
σext in a social network. We wish to capture the idea that
users in the network may be activated by a source external
to the network; that is, these activations do not occur through
friendships or connections within the social network. The most
general model of external influence is simply an arbitrary
probability distribution on the set of subsets of nodes. That
is, for any S ⊂ V , there is a probability pS that S is activated
from an external source. In this work, we adopt this model and
denote such a model of external influence by σext. In order
to consider both external and internal influence in our social
networks, we next define the concept of combining models of
influence together.
Definition 1 (Combination of σ1, σ2). Let σ1 and σ2 be two
models of influence propagation. We define the combination
σ of these two models in the following way: At any timestep
t, if set At ⊂ V is activated new nodes may be activated from
At according to either σ1 or σ2; that is, if in the next timestep
t+1∪
σ1 activates A1
t+1. We denote the combination model as σ1 ⊕ σ2 and write
A2
σ1 ⊕ σ2(A) for the expected number of activations resulting
from seeding A under this model.
t+1, then At+1 ≡ A1
t+1 and σ2 activates A2
In this work, we are most interested in combining external
activation with the triggering model. That is, if σext is a model
of external influence as defined above, and σtrig is an instance
of the triggering model, then we consider σtrig ⊕ σext.
3
1) Submodularity of σtrig ⊕ σext: In this subsection, we
establish the submodularity and monotonicity of Φ ≡ σtrig ⊕
σext. First, we require the following proposition.
Proposition 2. Let σ be a submodular, monotone-increasing
model of influence propagation. Define σA(T ) to be the ex-
pected influence of seeding set T when A is already activated;
that is σA(T ) = σ(A∪T ). Then, for any A, σA is submodular
and monotone increasing.
Proof: Let T1, T2 ⊆ V . We prove submodularity only,
monotonicity is similar.
σA(T1 ∪ T2) + σA(T1 ∩ T2)
= σ(T1 ∪ T2 ∪ A) + σ(T1 ∩ T2 ∪ A)
= σ((T1 ∪ A) ∪ (T2 ∪ A)) + σ((T1 ∪ A) ∩ (T2 ∪ A))
≤ σ(A ∪ T1) + σ(A ∪ T2)
= σA(T1) + σA(T2).
Theorem 1. Let σtrig be any instance of the Triggering Model,
and σext be any instance of external influence. Then, Φ ≡
σtrig ⊕ σext is submodular and monotonic.
Proof: For any A ⊆ V , let pA be the probability that
(cid:80)
A is activated via σext, the external influence. Then, Φ(S) =
A⊂V pAσtrig(A∪S). By Prop. 2, σtrig,A(S) = σtrig(A∪S)
for fixed A is monotone and submodular. Since a non-negative
combination of submodular and monotonic set functions is
also submodular and monotonic, the result follows.
III. THRESHOLD ACTIVATION PROBLEM
The framework has been established to define the problem
we consider in this work. We suppose a company wants to
minimize the number of seed users while expecting a certain
level of activation in the social network. Formally, we have
Problem 1 (Threshold activation problem (TAP)). Let G =
(V, E) be a social network, n = V . Given influence propa-
gation model σ and threshold T such that 0 ≤ T ≤ n, find a
subset A ⊆ V such that A is minimized with σ(A) ≥ T .
First, we consider performance guarantees for a greedy
approach to TAP with only the assumption that the influence
propagation σ is ε-submodular. We do not discuss how to
sample σ for these results. Subsequently, we specialize to the
case when σ is the the estimated value of the combination of
the triggering model and external influence, which is approxi-
mately submodular up to the error of estimation, and we show
how to efficiently estimate to a desired accuracy in Sections
III-B2 and III-B3b. We detail the algorithm STAB in Section
III-B1, a scalable algorithm with performance guarantees for
TAP utilizing this estimation and analysis.
A. Results when σ is ε-submodular
We analyze the greedy algorithm to solve TAP, which adds
a node that maximizes the marginal gain at each iteration to its
chosen seed set A until σ(A) ≥ T , at which point it terminates.
One might imagine in analogy to the set cover problem that
there would be a log n-approximation ratio for the greedy
algorithm for TAP -- however, this result only holds for integral
submodular functions [11]. We next give a bicriteria result for
the greedy algorithm that incorporates the error ε inherent in
ε-approximate submodularity into its bounds. In this context,
a bicriteria guarantee means that the algorithm is allowed to
violate the constraints of the problem by a specified amount,
and also to approximate the solution to the problem. In the
viral marketing context, this means that we may not activate
the intended threshold T of users, but we will guarantee to
activate a number close to T . Furthermore, we will not achieve
a solution of minimum size, but there is a guarantee on how
large the solution returned could be.
Theorem 2. Consider the TAP problem for σ when σ is ε-
approximately submodular. Then the greedy algorithm that
terminates when the marginal gain is less than 1 returns
a solution A of size within factor 1 + ε + log(cid:0) T
(cid:1) of
the optimal solution size, OP T , and the solution A satisfies
σ(A) ≥ T − (1 + ε) · OP T.
OP T
Proof: Let Ai = {a1, . . . , ai} be the greedy solution after
i iterations, and let Ag be the final solution returned by the
greedy algorithm. Let o = OP T be the size of an optimal
solution C = {c1, . . . , co} satisfying σ(C) ≥ T . Then
T − σ(Ai) ≤ σ(Ai ∪ C) − σ(Ai)
=
o(cid:88)
≤ o(cid:88)
j=1
j=1
∆cj σ (Ai ∪ {c1, . . . , cj−1})
∆cj σ(Ai) + oε
(by Property 2)
≤ o · [σ(Ai+1) − σ(Ai) + ε] .
Therefore, T − σ(Ai+1) − ε ≤(cid:0)1 − 1
(cid:19)i
(cid:19)i
T − σ(Ai) ≤ T
1 − 1
o
(cid:18)
(cid:18)
o
+ ε
(cid:1) (T − σ(Ai)). Then
(cid:18)
i−1(cid:88)
(cid:19)j
1 − 1
o
(1)
≤ T
1 − 1
o
j=0
+ εo.
From here, there exists an i such that the following differences
satisfy
T − σ(Ai) ≥ o(1 + ε), and
T − σ(Ai+1) < o(1 + ε),
Thus, by inequalities (2) and (3), o ≤ T exp(cid:0)−i
i ≤ o log(cid:0) T
o(cid:0)1 + ε + log(cid:0) T
(cid:1) , and
(cid:1) . By inequality (4) and the assumption on
(cid:1)(cid:1) . Finally, if the algorithm terminates be-
the termination of the algorithm, the greedy algorithm adds
at most o(1 + ε) more elements, so g ≤ i + o(1 + ε) ≤
fore σ(Ag) ≥ T , then the marginal gain is less than 1. Hence,
by (1), σ(A) ≥ T − (1 + ε)o.
o
o
o
Notice that the above argument requires only that σ is a
ε-submodular set function; in particular, it did not use the fact
that σ represents expected influence propagation on a social
network.
(2)
(3)
(4)
4
1) Approximation ratios: Next, we consider ways in which
the bicriteria guarantees of Theorem 2 can be improved. In
viral marketing, we may suppose a company seeks to choose
a threshold T such that the marginal gain to reach T is always
at least 1; seeding nodes with a marginal gain of less than 1
would be cost-ineffective. In other words, it would cost more
to seed a node than the benefit obtained from seeding it. There
is little point in activating T users if the marginal gain drops
too low; intuitively, the company has already activated as many
as it cost-effectively can.
We term this assumption the cost-effectivity assumption
(CEA): In an instance of TAP, if B ⊆ V such that σ(B) < T ,
there always exists a node u such that ∆uσ(B) ≥ 1. Under
CEA, the greedy algorithm in Theorem 2 would be an approx-
imation algorithm; that is, it would ensure σ(A) ≥ T , with the
same bound on solution size as stated in the theorem. To see
this fact, notice that once inequality (4) above is satisfied, the
algorithm must add at most o(1 + ε) more elements before
σ(A) ≥ T , by CEA.
More generally, each node v ∈ V has an associated
reticence rv ∈ [0, 1]; rv is the probability that v will remain
inactive even if all of the neighbors of v are activated. Then
we have the following theorem, whose proof is analogous to
Theorem 2.
Theorem 3. Let σ be ε-approximately submodular and let
r∗ = minv∈V rv. Suppose r∗ > 0. Then the greedy algorithm
for TAP is an approximation algorithm which returns solution
A within factor 1+ε
(cid:1) of optimal size.
r∗ + log(cid:0) T
OP T
B. Scalable bicriteria algorithm for σint ⊕ σext
In this subsection, we detail the scalable bicriteria Algorithm
1 for TAP when the propagation is given by an instance of
the triggering model in the presence of external influence;
that is, when σ = σtrig ⊕ σext. We describe our scalable
algorithm STAB first and then discuss the necessary sampling
and estimation techniques in the subsequent sections.
1) Description of algorithm: As input, the algorithm takes
a graph G = (V, E) representing a social network, an exter-
nal influence model σext, internal influence model σtrig, an
instance of the triggering model, and the desired threshold of
activated users T . In addition, the user specifies the fractional
error α, on which the running time of the algorithm and
the accuracy of the solution depend. Using α, in line 1 the
algorithm first determines (cid:96), the number of graph samples it
requires according to Section III-B2.
Next in the for loop on line 2, the algorithm constructs
a collection of oracles which will be used to estimate the
average reachability in the sampled graphs Hi, which is used
to approximate the expected influence. Each graph Hi
is
needed only while updating the oracle collection in iteration
i; once this step is completed, Hi may be safely discarded.
Since the samples are independent, this process is completely
parallelizable.
Once the set of oracles has been constructed, a greedy
algorithm is performed in attempt to satisfy the threshold T
of expected activation with a minimum-sized seed set. The
estimation in line 11 may be done in one of two ways: using
estimator C1 or C2; both are described in detail in Section
III-B3b. The estimator chosen has a strong effect on both the
running time and performance of the algorithm: given the same
oracles, C1 can be computed in time O(k), and in practice
is much faster to compute than C2. However, the quality of
C1 degrades with the size of seed set. On the other hand,
C2 takes time O(kB log B) time to compute, where B is
the seed set for which the average reachability is estimated;
our experimental results show that the quality of C2 is vastly
superior to C1 for larger seed set sizes B; however its
running time increases.
This algorithm achieves the following guarantees on perfor-
Then,
mance:
Theorem 4. Suppose we have an instance of TAP with σ =
σtrig ⊕ σext and that T has been chosen such that CEA holds.
the solution
returned by Alg. 1 satisfies the following two conditions, with
probability at least 1 − η:
if η > 0, by choosing δ = η/n3,
1) σ(A) ≥ T − 2αT
2) If A∗ is an optimal solution satisfying σ(A∗) ≥ T ,
A
A∗ ≤ 1 + 4αT + log(T ).
If Assumption CEA is violated, the algorithm can detect this
violation by terminating when the marginal gain drops below
1. In this case, bound 1 above becomes σ(A) ≥ T − (1 +
ε)A∗.
yields running time O(cid:0)log(n/η)(n + m)A/α2(cid:1). If estimator
Using estimator C1 for average reachability in line 11
C2 is used, a factor of A log A is multiplied by this bound.
Proof: Let A∗ be a seed set of minimum size satisfying
σ(A∗) ≥ T . Then, as discussed in Section III-B2, if δ = η/n3,
then with probability at least 1 − η, A∗ satisfies σ(A∗) ≥
T − αT by the choice of (cid:96) in Alg. 1, and the analysis in
Section III-B2. Hence A∗ ≥ B∗, where B∗ is a set of
minimum size satisfying σ(B∗) ≥ T − αT .
Notice σ is 4αT -approximately submodular on the sets
considered by the greedy algorithm with probability at least
1− η. By Section III-A1, in this case, the solution A returned
by Alg. 1 satisfies A ≤ (1+4αT +log T )B∗ ≤ (1+4αT +
log T )A∗. Furthermore, σ(A) ≥ σ(A) − α ≥ T − 2αT,
if CEA holds, otherwise the alternate bound follows from
Theorem 2.
Next, we consider the running time of Alg. 1. Let m be
the number of edges that have a nonzero probability to exist
in one of the reachability instances (m is at most the number
of edges in input graph G). Lines 2 and 3 clearly take time
(cid:96)(n + m). By Cohen et al. [3], line 4 takes time O(k(cid:96)m). The
while loop on line 9 executes exactly A times, and the for
loop on line 10 requires time O(kn) if estimator C1 is used,
and time O(kA log An) if estimator C2 by Section III-B3b.
By the choices of k and (cid:96) on line 1, we have the total running
2) Estimation of σtrig ⊕ σext(A): Let σtrig be a model of
times bounded as stated.
internal influence propagation, which in this section will be
an instance of the triggering model. Let σext be the model of
external influence activation, and let σ = σtrig ⊕ σext be the
combination of the two as defined above. We use the following
version of Hoeffding's inequality (T will be the threshold that
5
Algorithm 1: Scalable TAP Algorithm with Bicriteria
guarantees (STAB)
Data: G, σtrig, σext, T, α, δ
Result: Seed set A
1 Choose (cid:96) ≥ log(2/δ)/α2, k ≥ (2 + c) log n(αT )−2 as
discussed in Section III-B2;
2 for i = 1 to (cid:96) do
3
4
5
Sample graph Gi from σtrig;
Sample external seed set Aext
Construct Hi = Gi − σi(Aext
i
σi(Aext
Update the oracle collection {Xu : u ∈ V } for τ
according to Hi as in Section III-B3a;
), as described in Section III-B3a;
for σext;
), and store the value of
i
i
6
8 A ≡ ∅, O ≡(cid:80)(cid:96)
7 end
i=1 σi(Aext
9 while σ(A) < T − αT do
10
11
for u ∈ V do
i
) /(cid:96);
Estimate ∆u = {τ (A ∪ {u}) − τ (A)} using one
of the estimators C1 or C2 as described in
Sections III-B3b, III-B3d;
end
A = A ∪ {u∗}, where u∗ maximizes ∆u;
Compute σ(A) = τ (A) + O by Lemma 2;
12
13
14
15 end
.
T 2
(cid:17)
is input to Alg. 1).
Theorem 5 (Hoeffding's inequality). Let X1, . . . , X(cid:96) be in-
dependent random variables in [0, T ]. Let X be the empirical
mean of these variables, X = 1
i=1 Xi. Then for any t
(cid:96)
(cid:80)(cid:96)
(cid:16)− 2(cid:96)t2
P r(cid:0)(cid:12)(cid:12)X − E[X](cid:12)(cid:12) ≥ t(cid:1) ≤ 2 exp
for any seed set A, σtrig(A) = (cid:80)
Since σtrig is an instance of the triggering model, by
Theorem 1, there exists a set {(Gj, pj) : j ∈ J} such that
j∈J pjσj(A), where σj
is simply the size of the set of nodes reachable from A in
Gj. Thus, if A is fixed, then by taking independent samples
of graphs from the probability distribution on {Gj}, we get
independent samples of σi(A).
In the general model of external influence presented above,
every seed set B ⊂ V has a probability pB of being activated
by the external influence. By sampling from this distribution
(cid:80)
on subsets of nodes, and independently sampling as above
from σtrig, σ(A) could be computed exactly in the following
j∈J pjpBσj(A ∪ B). In most cases,
this sum cannot be computed in polynomial time, and certainly
it has Ω(2n) summands. Accordingly, we estimate its value
by independently sampling (cid:96) externally activated sets {Aj
ext}
from σext; we also independently sample (cid:96) reachability graphs
Gj according to σtrig and estimate by averaging the size of
the reachability from a seed set A in this context: σ(A) ≈
σ(A) ≡ 1
within error αT with probability at least 1−δ from Hoeffding's
inequality we require (cid:96) ≥ log(2/δ)
(cid:1) . To estimate σ(A) in this way
way: σ(A) = (cid:80)
(cid:0)A ∪ Aext
such samples.
(cid:80)(cid:96)
j=1 σj
B⊂V
(cid:96)
j
Now, in our analysis of the greedy algorithm in Theorem 2
only at most n3 sets were considered. All that is required for
2α2
the analysis to be correct as it that those sets were estimated
within the error αT ; if δ < η/n3, where η < 1, then by the
union bound, with probability at least 1 − η, the analysis for
Theorem 2 holds. In practice, we were able to get good results
with much higher values of δ, see Section IV.
3) Estimation of σ: In the previous section, we describe an
approach to estimate σ(A), based upon independent samples
{Gi} of graphs from the triggering model, and independent
samples of externally activated nodes {Aext
i }. Next, we need
to compute the value of the estimator σ(A). One method
would be to compute it directly using breadth-first search
from the sets A ∪ Aext
in each graph Gi. This method would
unfortunately add a factor of Ω(n + m) to the running time,
which would result in a running time of Alg. 1 of Ω(n2), too
large for our purposes. Thus, we would like to take advantage
of estimators formulated by Cohen et al. [3] for the average
reachability of a seed set over a set of graphs. However,
because the external seed sets Aext
for each graph Gi vary
with i, we must first convert the problem into an average
reachability format.
i
i
a) Conversion to an average reachability problem:
(cid:80)(cid:96)
Suppose we have sampled as above (cid:96) pairs of sample graphs
)} To compute
and external seeds: {(G1, Aext
i=1 σi(A ∪ Aext
) efficiently, we first convert
σ(A) = 1
(cid:96)
this sum to a generalized reachability problem: we construct
graph Hi from Gi by removing all nodes (and incident edges)
reachable from Aext
). The average
reachability of a set A in the graphs {Hi}, which we term
τ (A), is
1 ), . . . , (G(cid:96), Aext
i
: Hi = Gi − σi(Aext
(cid:96)
i
i
(cid:96)(cid:88)
i=1
τ (A) =
1
(cid:96)
τi(A),
(5)
where τi(A) is the size of the set reachable from A in Hi. The
estimators formulated by Cohen et al. are suitable to estimate
the value of τ, and the following two lemmas show how we
can compute σ(A) from τ (A).
Lemma 1. The size of the reachable set from A in Hi can
be computed from reachability σi in Gi as follows: τi(A) =
σi(A ∪ Aext
) − σi(Aext
).
i
i
Proof: Suppose, in Gi, x is reachable from A, but not
. This is true iff there exists a path from A to x in
), which is equivalent to the path existing
i
from Aext
Gi avoiding σi(Aext
in in Hi.
i
The next lemma shows explicitly how to get σ(A) from
τ (A):
Lemma 2. τ (A) = σ(A) − 1
(cid:80)(cid:96)
6
(cid:16){r(i)
v (v, i) ∈ τi(A)}(cid:17)
(5). For convenience, we refer to this method of estimation in
the rest of the paper as method C1.
Each pair (v, i) ∈ V × {1, . . . , (cid:96)}, consisting of a node v
in graph Hi, is assigned an independent, random rank value
r(i)
v uniformly chosen on the interval [0, 1]. The reachability
sketch for a set A is defined as follows: let k be an integer,
and consider XA = bottom-k
, where
bottom-k(S) means to take the k smallest values of the set
S, and R(i)
A is the set of nodes reachable from A in graph
Hi. The threshold rank of a set A ⊂ V is then defined to be
γA = max XA, if XA = k, and γA = 1 if XA < k. The
estimator for τ (A) is then τ (A) = (k − 1)/((cid:96)γA)
If k = (2 + c)()−2 log n, the probability that this estimator
has error greater than is at most 1/nc [3]. The bounds on k
needed for the proof of Theorem 4 are determined by taking
= αT .
c) Computation of XA: First, we compute Xu for all
u ∈ V using Algorithm 2 of [3] in time O(k(cid:96)m), where m
is the maximum number of edges in any Hi. This collection
{Xu : u ∈ V } is referred to as oracles.
Next, we discuss how to compute, for an arbitrary node
u, XA∪{u} given that XA has already been computed. This
computation will take time O(k) and is necessary for the
bicriteria algorithm: given that Xu and XA are both sorted,
we compute XA∪{u} by merging these two sets together until
the size of the new set reaches k values.
(cid:80)
z∈(cid:83)
d) Estimation of τ (A), method C2: Alternatively to
estimator C1, we can estimate τ (A) from the oracles
{Xu : u ∈ V } in the following way. Let γu be the
threshold rank as defined above for Xu. Then τ (A) =
. For convenience,
1
(cid:96)
we refer to the estimator in the rest of the paper as estimator
C2; it was originally introduced in [17].
maxu∈Az∈Xu\{γu} γu
v∈A Xv\{γv}
1
IV. EXPERIMENTAL EVALUATION
In this section, we demonstrate the scalability of STAB as
compared with the current state-of-the-art IMM algorithm [9]
and with the greedy algorithm for TAP in Goyal et al. [11].
The methodology is described in Section IV-A, comparison to
existing IM algorithms is in Section IV-B, and investigation of
the effect of external influence on the performance of STAB
is in Section IV-C. All experiments were run on an Intel(R)
Core(TM) i7-3770K CPU @ 4.0GHz CPU with 32 GB RAM.
i=1 σi(Aext
i
(cid:96)
).
A. Datasets and framework
Proof: This statement follows directly from Lemma 1 and
the definitions of τ, σ.
In Alg. 1, for each i, σi(Aext
) is computed in the con-
struction of Hi; its size can be stored as instructed on line 5,
and used in the computation of O for line 14 in the greedy
algorithm's stopping criterion.
i
b) Estimation of τ (A), method C1: In this subsection,
we utilize methods developed by Cohen et al. [3] to estimate
efficiently the average reachability problem τ (A) defined in
We evaluated the following algorithms in our experiments.
1) CELF:
the greedy algorithm by naive sampling of
Kempe et al. [1] can be modified to find a solution to TAP, as
shown in Goyal et al. [11]. The modified algorithm performs
Monte Carlo sampling at each step to select the node with the
highest marginal gain into the seed set until the threshold T is
satisfied. The Cost-Effective Lazy Forward (CELF) approach
by Leskovec et al. [19] improves the running time of this
algorithm by reducing the number of evaluations required.
2) IMM: The IMM algorithm [9] is the current state-of-
the-art algorithm to solve the IM problem, where the number
of seeds k is input. Since TAP asks to minimize the number
of seeds, this algorithm cannot be applied directly. For the
purpose of comparison to our methods, we adapt the algorithm
by performing a binary search on k in the interval [1, T ], where
T is the threshold given in TAP. At each stage of the search,
IMM utilizes the value of k in question until the minimum
k as estimated by IMM is found. Since binary search can
identify the minimum in at most log T iterations, we chose
this approach over starting at k = 1 and incrementing by one
until the minimum is found, which in the worst case would
require T iterations.
3) STAB: the STAB algorithm (Alg. 1) using estimators C1
and C2, referred to as STAB-C1, STAB-C2 respectively. Since
these are greedy algorithms with an approximately submodular
function, we also use the CELF approach to reduce the number
of evaluations performed by STAB-C2; for STAB-C1, we
found this optimization unnecessary.
Network topologies: We generated networks according to
the Erdos-Renyi (ER) and Barabasi-Albert (BA) models. For
ER random graphs, we used varying number of nodes n; the
independent probability that an edge exists is set as p = 2/n.
The BA model was used to generate scale-free synthesized
graphs; the exponent in the power law degree distribution was
set at −3 for all BA graphs.
The following topologies of real networks collected by
the Stanford Network Analysis Project [20] were utilized: 1)
Facebook, a section of the Facebook network, with n = 4039,
m = 88234; 2) Nethept, high energy physics collaboration
network, n = 15, 229, m = 62, 752; 3) Slashdot, social
network with n = 77, 360, m = 828, 161; 4) Youtube,
from the Youtube online social network, n = 1, 134, 890,
m = 2, 987, 624; and 5) Wikitalk, the Wikipedia talk (com-
munication) network, n = 2, 394, 385, m = 5, 021, 419.
ER 1000
BA 15000
Nethept
Slashdot
Youtube
Wikitalk
α = 0.1
1.0
6.4
954
25
385
1704
α = 0.2
0.4
1.6
230
6.4
91
274
α = 0.3
0.2
0.9
102
3.3
41
122
TABLE I
ORACLE COMPUTATION TIME (SEC)
ipmax
1
0.1
1
0.01
0.01
0.01
Models of internal and external influence: In all experi-
ments, we used the independent cascade model for internal
influence propagation: each edge e in the graph is assigned
a uniform probability ipe ∈ [0, ipmax]. For synthesized net-
works, we usually set ipmax = 1; for real networks, we
observed that if ipmax = 1, then in most cases a single
node can activate a large fraction of the network (often more
than 33%). However, a large number of empirical studies have
confirmed that most activation events occur within a few hops
of a seed node [12] -- [14]; these works indicate that ipmax = 1
is not a realistic parameter value for the IC model. Therefore,
we also ran experiments using lower values for ipmax.
For external influence, we adopted in all experiments the
7
model that each node u is activated externally independently
with uniformly chosen probability epu ∈ [0, epmax]. The set-
ting of epmax is discussed in the context of each experiment.
Unless otherwise stated, we set δ = 0.01, which we found
sufficient to return a solution within the guarantee provided in
Theorem 4 in most cases with estimator C1, and in all cases
with estimator C2.
B. Performance comparison and demonstration of scalability
In this section, we compare the performance of STAB-C1
and STAB-C2 to CELF and IMM as described above; we
experimented on the datasets listed in Table I, where the total
CPU time required to compute the oracles is shown. The
oracle computation is parallelizable, and in our experiments
we used 7 threads of computation. The oracles for each value
of α were computed once and stored, and thereafter when
running STAB the oracles were simply read from a file. The
running times we report for the various versions of STAB
do not include the oracle computation time unless otherwise
specified. Also in Table I are the values of ipmax used for each
dataset in this set of experiments. Since IMM and CELF do
not consider external influence, the experiments in this section
had no external influence; that is, epmax = 0 for all datasets.
To evaluate the seed set returned by the algorithms we used
the average activation from 10000 independent Monte Carlo
samples.
We show typical results in Fig. 1 on the following four
datasets: ER 1000 (n = 1000), Nethept, Youtube, and Wik-
italk. The first row of Fig. 1 shows the expected activation,
normalized by the threshold value T , of the seed set returned
by each algorithm plotted against T . Thus, a value of 1
indicates the algorithm successfully achieved the threshold of
activation. The second row of the figure shows the running
time in minutes of each algorithm against T , and the third row
shows the size of the seed set returned by each algorithm.
For the ER 1000 network results are shown in the first
column of Fig. 1; this dataset was the only one on which CELF
finished under the time limit of 60 minutes. We see that IMM
is consistently returning a larger seed set than STAB and CELF
and overshooting the threshold value in activation by as much
as a factor of 2.5. Thus, it has poor performance of minimizing
the size k of the seed set for TAP. This behavior appears to
be a result of IMM underestimating the influence of its seed
set internally. As expected, CELF performs very well in terms
of the size of the seed set and meeting the threshold T , but
is running on a timescale larger than the other algorithms by
factor of at least 100 and as large as 104. Notice that STAB-
C2 with α = 0.1 has virtually identical activation and size of
seed set to CELF, while running at a much faster speed; as
expected, the quality of solution of STAB deteriorates as α is
raised, but the running time decreases drastically. In addition,
notice that none of the versions of STAB seed too many nodes
and overshoot the threshold as IMM does; instead, STAB
errs on the side of seeding too few nodes and only partially
achieving the threshold T of activation. Finally, it is evident
that STAB-C1 runs faster than STAB-C2 and has a similar
amount of activation when the seed set required is relatively
8
(a) ER, activation
(b) Nethept, activation
(c) Youtube, activation
(d) Wikitalk, activation
(e) ER, running time
(f) Nethept, running time
(g) Youtube, running time
(h) Wikitalk, running time
(i) ER, number of seeds
(j) Nethept, number of seeds
(k) Youtube, number of seeds
(l) Legend
Fig. 1.
versus T .
First row: expected activation normalized by threshold T versus T . Second row: running time in minutes versus T . Third row: size of seed set
small. However, the larger the seed set required, the farther
is STAB-C1 from achieving the threshold T while STAB-C2
does not suffer from this drawback.
The results from Nethept are shown in the second column
of Fig. 1; with n = 15, 229, CELF was unable to run within
the time limit, and IMM was able to complete its binary
search only for T ≤ 7000; for the higher threshold values,
IMM exceeded the 32 GB memory usage limit imposed in
our setup. On this network, STAB-C2 again exhibits the best
performance. Despite an initial decrease in activation relative
to T shown in Fig. 1(b), the trend reverses around T = 8000
and seed set size 100 and the algorithm gets closer to achieving
the threshold. This behavior is explained by lower coefficient
of variation of estimator C2 as analyzed in [17]; the CV
can be lower than estimator C1 by up to a factor of (cid:112)A,
where A is the seed set. In stark contrast, STAB-C1 was
unable to proceed past T = 8000 even for α = 0.1 because
the estimator C1 appears to lose accuracy as the number of
seeds increases. By this mechanism, STAB-C1 had achieved
its maximum estimation of influence of any set and thereby
could not increase it before reaching an estimated activation
of T , for T > 8000.
Next, the Youtube network is shown in the third column
of Fig. 1. As in the ER network, IMM is underestimating the
influence of its seed set and thereby picking too many seeds
and overshooting the threshold, as shown in Fig. 1(c); IMM
picks nearly twice as many seed nodes as STAB-C1, α = 0.1,
as shown in Fig. 1(k). In Fig. 1(g), the scalability of STAB is
demonstrated as the most precise version, STAB-C2 with α =
0.1, runs faster than IMM by as much as a factor of 50. On
our largest dataset, the Wikitalk network, IMM again exceeded
32 GB memory after T = 6000; notice that the running time
of STAB in all cases is under 2 minutes and STAB-C2 with
α = 0.3 maintained activation greater than 0.6T while running
in less than 5 seconds, as shown in Fig. 1(h). With inclusion
of the parallelizable and reusable oracle computation time of
122 seconds from Table I, the total time taken by STAB-C2
is less than 3 minutes. The total running time for α = 0.1 of
STAB-C2 including the oracles at T = 10000 is less than 30
minutes.
Choice of α: The above discussion demonstrates that α =
0.1 provides the close activation to the threshold T while
maintaining high scalability. If faster running time is desired,
α may be increased, which results in a loss of accuracy shown
clearly in Fig. 1(d), for α ∈ {0.1, 0.2, 0.3}. On the other hand,
if activation closer to T is required, smaller values of α may
be used at higher computational cost.
C. External influence
In this section, we analyze the performance of STAB when
external influence is present in the network; that is, when
epmax > 0. For this section, we considered a BA network
with 100,000 nodes, with a threshold of T = 1000, and the
Facebook network with T = 1500. Results on other topologies
were qualitatively similar. For all experiments in this section,
we set α = 0.1.
9
guarantees; Goyal et al. [11] studied the TAP problem with
monotonic and submodular models of influence propagation;
their bicriteria guarantees differ from ours, and provide no
method of efficient sampling required for scalability. Chen et
al. [23] considered external influence in a viral marketing con-
text. However, their model of external influence is much less
general than ours. Furthermore, they restrict external influence
to only pass through seed consumers and have no discussion
of sampling, scalability, or the TAP problem. Nguyen et al.
[24], [25] studied methods to restrain propagation in social
networks.
VI. CONCLUSION
We establish equivalency between the triggering model
and generalized reachability, which allows incorporation of
external influence into our efficient sampling techniques. We
gave precise trade-off between accuracy and running time with
a bound on the number of samples required to solve TAP.
Our algorithm is highly scalable and outperforms adaptations
to TAP of the current state-of-the-art algorithm for the IM
problem.
This work is supported in part by the NSF grant #CCF-
ACKNOWLEDGMENT
1422116.
REFERENCES
[1] David Kempe, Jon Kleinberg, and ´Eva Tardos. Maximizing the spread
of influence through a social network. In Proceedings of the 9th ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 137 -- 146, 2003.
[2] Thang N Dinh, Huiyuan Zhang, Dzung T Nguyen, and My T Thai. Cost-
effective viral marketing for time-critical campaigns in large-scale social
networks. IEEE/ACM Transactions on Networking (TON), 22(6):2001 --
2011, 2014.
[3] Edith Cohen, Daniel Delling, Thomas Pajor, and Renato F. Werneck.
Sketch-based Influence Maximization and Computation: Scaling up with
Guarantees. In Proceedings of the 23rd ACM International Conference
on Conference on Information and Knowledge Management, pages 629 --
638, 2014.
[4] Christian Borgs, Michael Brautbar, Jennifer Chayes, and Brendan Lucier.
Maximizing Social Influence in Nearly Optimal Time. In Proceedings of
the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms,
pages 946 -- 957, 2014.
[5] Huiyuan Zhang, Thang N Dinh, and My T Thai. Maximizing the spread
of positive influence in online social networks. In Distributed Computing
Systems (ICDCS), 2013 IEEE 33rd International Conference on, pages
317 -- 326. IEEE, 2013.
[6] Huiyuan Zhang, Dung T Nguyen, Huiling Zhang, and My T Thai. Least
cost influence maximization across multiple social networks. IEEE/ACM
Transactions on Networking, 24(2):929 -- 939, 2016.
[7] Thang N Dinh, Dung T Nguyen, and My T Thai. Cheap, easy, and
massively effective viral marketing in social networks: truth or fiction?
In Proceedings of the 23rd ACM conference on Hypertext and social
media, pages 165 -- 174. ACM, 2012.
[8] Huiyuan Zhang, Huiling Zhang, Alan Kuhnle, and My T Thai. Profit
Maximization for Multiple Products in Online Social Networks. In IEEE
International Conference on Computer Communications, 2016.
[9] Youze Tang. Influence Maximization in Near-Linear Time : A Martin-
gale Approach. In Proceedings of the 2015 ACM SIGMOD International
Conference on Management of Data, pages 1539 -- 1554, 2015.
[10] Youze Tang, Xiaokui Xiao, and Yanchen Shi. Influence maximization:
Near-optimal time complexity meets practical efficiency. In Proceedings
of the 2014 ACM SIGMOD International Conference on Mangement of
Data, pages 75 -- 86, 2014.
[11] Amit Goyal, Francesco Bonchi, Laks V S Lakshmanan, and Suresh
Venkatasubramanian. On minimizing budget and time in influence
propagation over social networks. Social Network Analysis and Mining,
3(2):179 -- 192, 2013.
[12] Sharad Goel, Ashton Anderson, Jake Hofman, and Duncan J Watts. The
Structural Virality of Online Diffusion. Management Science, 62(1):180 --
196, 2016.
[13] Sandra Gonz´alez-Bail´on, Javier Borge-Holthoefer, and Yamir Moreno.
Broadcasters and Hidden Influentials in Online Protest Diffusion. Amer-
ican Behavioral Scientist, 57:943 -- 965, 2013.
(a) BA 100,000
(b) Facebook
Fig. 2. The effect of external influence on activation and size of seed set of
STAB-C2.
In 2(a), we plot the expected activation (Act) of the seed
set returned by STAB-C2 normalized by the threshold T , as
epmax varies from 0 to 0.006. As in the previous section, the
expected activation of the returned seed set A is estimated by
independent Monte Carlo sampling. We also plot the expected
fraction of T activated by the external influence, along with
the size of the seed set returned by STAB-C2, normalized by
its maximum value. As the effect of external influence in the
network increases, the algorithm requires fewer seed nodes
to ensure the expected activation is within the specified error
tolerance to threshold T .
In 2(b), we show an analogous plot for the Facebook
network; interestingly, the size of the seed set chosen by
STAB-C2 nearly doubles as epmax increase from 0 to 0.01,
before beginning to decrease to 0. This increase differs from
what we expected; it is counterintuitive that the algorithm
would require more seeds to reach the threshold as external
influence increases. One possible explanation for this effect
is that the external influence both decreases and distributes
the marginal gain more evenly among seed nodes, so that
the greedy algorithm has a more difficult time identifying the
best seed nodes, especially in the presence of the error of
estimation.
V. RELATED WORK
Kempe et al. introduced the triggering model in a seminal
work on the IM problem [1], where they exhibited a Monte
Carlo greedy sampling algorithm that achieves 1− 1/e perfor-
mance ratio for IM; this algorithm, although it runs in polyno-
mial time, is very inefficient and cannot scale well. Since the
maximum coverage problem prohibits performance guarantee
better than 1 − 1/e + o(1) under standard assumptions, the
ratio for IM likely cannot be improved, but much work has
improved the scalability of the algorithm while retaining the
guarantee. Leskovec et al. [19] introduced the CELF method to
exploit submodularity and improve the running time. Reverse
Influence Sampling (RIS) was introduced in [4] to further
improve the greedy performance; algorithms using RIS include
[4], [9], [10], [21], [22]; the current state-of-the-art are the
SSA [21] and the IMM algorithm [9] to which we compare
in this paper. Cohen et al. [3] introduced a methodology for a
highly scalable IM algorithm SKIM; in this work, we extend
this methodology to solve the TAP problem with the triggering
model and external influence.
As compared with IM, much less effort has been devoted
to scalable solutions to TAP while maintaining performance
10
[14] Seth A. Myers and Jure Leskovec.
The bursty dynamics of the
Twitter information network. In Proceedings of the 23rd International
Conference on World Wide Web, pages 913 -- 924, 2014.
[15] Seth A. Myers, Chenguang Zhu, and Jure Leskovec.
Information
In Proceedings of the
diffusion and external influence in networks.
18th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pages 33 -- 41, 2012.
[16] Edith Cohen. Size-Estimation Framework with Applications to Transi-
tive Closure and Reachability. Journal of Computer and System Sciences,
55(3):441 -- 453, 1997.
[17] Edith Cohen, H Kaplan, and Tel Aviv. Leveraging discarded samples
for tighter estimation of multiple-set aggregates. ACM SIGMETRICS
Performance Evaluation Review, 37(1):251 -- 262, 2009.
[18] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization
for prevalent viral marketing in large-scale social networks. In Proceed-
ings of the 16th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pages 1029 -- 1038, 2010.
[19] Jure Leskovec, Andreas Krause, Carlos Guestrin, Christos Faloutsos,
Jeanne VanBriesen, and Natalie Glance. Cost-effective Outbreak De-
the 13th ACM SIGKDD
tection in Networks.
international conference on Knowledge discovery and data mining,
pages 420 -- 429, 2007.
In Proceedings of
[20] Jure Leskovec and Andrej Krevl. SNAP Datasets: Stanford large network
dataset collection. http://snap.stanford.edu/data, June 2014.
[21] H.T. Nguyen, M. T. Thai, and T. N. Dinh. Stop-and-Stare: Optimal
Sampling Algorithms for Viral Marketing in Billion-Scale Networks. In
ACM SIGMOD/POSD Conference, 2016.
[22] H. T. Nguyen, T. N. Dinh, and M. T. Thai. Cost-aware Targeted Viral
Marketing in Billion-scale Networks. In IEEE International Conference
on Computer Communications, 2016.
[23] Wei Chen, Fu Li, Tian Lin, and Aviad Rubinstein. Combining Tradi-
tional Marketing and Viral Marketing with Amphibious Influence Max-
imization. In Proceedings of the 16th ACM Conference on Economics
and Computation, pages 779 -- 796, 2015.
[24] N. P. Nguyen, Y. Xuan, and M. T. Thai. A novel method for worm
containment on dynamic social networks. In Military Communications
Conference, pages 2180 -- 2185, 2010.
[25] Nam P. Nguyen, T. N. Dinh, D. T. Nguyen, and M. T. Thai. Overlapping
community structures and their detection on social networks. In Privacy,
Security, Risk, and Trust (PASSAT) and 2011 Third International Con-
ference on Social Computing, 2011.
|
1811.01638 | 1 | 1811 | 2018-11-05T12:22:16 | Identifying influential patents in citation networks using enhanced VoteRank centrality | [
"cs.SI",
"cs.DL",
"physics.soc-ph"
] | This study proposes the usage of a method called VoteRank, created by Zhang et al. (2016), to identify influential nodes on patent citation networks. In addition, it proposes enhanced VoteRank algorithms, extending the Zhang et al. work. These novel algorithms comprise a reduction on the voting ability of the nodes affected by a chosen spreader if the nodes are distant from the spreader. One method uses a reduction factor that is linear regarding the distance from the spreader, which we called VoteRank-LRed. The other method uses a reduction factor that is exponential concerning the distance from the spreader, which we called VoteRank-XRed. By applying the methods to a citation network, we were able to demonstrate that VoteRank-LRed improved performance in choosing influence spreaders more efficiently than the original VoteRank on the tested citation network. | cs.SI | cs | Identifying influential patents in citation networks using enhanced
VoteRank centrality
Freitas, João C.S. *, Barbastefano, Rafael **, Carvalho, Diego **
* FGV -- Getulio Vargas Foundation. Praia de Botafogo, 190 -- Botafogo -- Rio de Janeiro/RJ - Brazil
[email protected]
** Cefet -- Federal Centre for Technological Education Celso Suckow da Fonseca. Av. Maracanã, 229 -
Maracanã -- Rio de Janeiro/RJ -- Brazil
[email protected]
[email protected]
Abstract
This study proposes the usage of a method called VoteRank, created by Zhang et al. (2016), to identify
influential nodes on patent citation networks. In addition, it proposes enhanced VoteRank algorithms,
extending the Zhang et al. work. These novel algorithms comprise a reduction on the voting ability of the
nodes affected by a chosen spreader if the nodes are distant from the spreader. One method uses a
reduction factor that is linear regarding the distance from the spreader, which we called VoteRank-LRed.
The other method uses a reduction factor that is exponential concerning the distance from the spreader,
which we called VoteRank-XRed. By applying the methods on a citation network, we were able to
demonstrate that VoteRank-LRed improved performance in choosing influence spreaders more efficiently
than the original VoteRank on the tested citation network.
Keywords
Social networks, centrality
Classification Codes
MSC: 91D30 , 68W01
JEL: O33, C63, C90,
1 - Introduction
Patent citations are often used as a direct means for distinguishing technologies among themselves, but
the amount of times a patent is cited is also an indicative of its value (Jang, Lo, & Chang, 2009).
Moreover, a patent citation usually represents knowledge contained in one patent and extended by a
newer one, meaning a knowledge flow (Jaffe, Trajtenberg, & Henderson, 1993). Patent citations have also
been used to detect technology diffusion and forecasting (Chang, Lai, & Chang, 2009) or knowledge
spillover (Xiang, Cai, Lam, & Pei, 2013).
Since citations establish relationships between patents (actors), social networks theory can be applied to
build patent citations networks and acquire better comprehension about citations (Chang, Lai, & Chang,
2009). Therefore, many studies have been made using patent citation networks, such as (Hung & Wang,
2010) or (Kim, Cho, & Kim, 2014). Xiang et al. (2013) applied a spillover technique to increase the
amount of relationships among patents in citation networks. The spillover technique creates an arc in the
network from an older to a newer patent if they share at least one common author, even if there was no
citation between them on the original network. The goal is to explicit the relationship between these
patents, because tacit knowledge acquired during the first work was implicitly used by the common
author on the newer patent.
Centrality is a key concept in social networks. The most important or prominent actors of a network are
said to be central (Wasserman & Faust, 2006). A node centrality can be calculated in many different
ways, the most usual being degree, closeness, and betweenness, which can be found on most social
network analysis books, such as (Wasserman & Faust, 2006) or (De Nooy, Mrvar, & Batagelj, 2005).
However, increasingly more ways of expressing the centrality of nodes, like LocalRank, ClusterRank,
PageRank, H-Index and others, have been appearing lately, as gathered by Lü et al. (2016).
Among these newer methods, lies the VoteRank algorithm (Zhang, Chen, Dong, & Zhao, 2016), known
for its swiftness and efficiency and for providing several spreaders simultaneously. In a patent citation
network, a spreader can be understood as a document that is cited throughout an innovation chain. Since
the efficient detection of important members from a network is a core need for the understanding of
knowledge flow through citations or patents values, this study discloses this new method to the field of
patent citation network analysis. It proposes some modifications on VoteRank in order to find better
spreaders.
This study is structures as follows: section 2 presents the Centrality, VoteRank, and SIR criterion; section
3 discloses two variations of VoteRank; and section 4 displays the achieved results.
2 -- Centrality
Centrality is a key concept in social network analysis. One of most important applications of graph theory
in SNA is the identification of prominent vertices (actors) in a network (Wasserman & Faust, 2006).
Networks are abstract concepts; thus, the definition of centers becomes difficult, since there is no
similarity with the physical concept of centroid.
The first centrality concept is related to the number of ties connected to a vertex. A central actor with
several connections is involved in more relations than an actor with fewer connections. This definition is
both topological and local. Other traditional centrality measures are related to the average distance from
one vertex to the others (closeness), or to network information flows (betweenness).
The relative importance of each actor within the network can be recognized through centrality metrics. In
the scope of this study, the following centrality measures were employed: degree, closeness, and
betweenness. The so-called giant component is yet another interesting variable to observe in a citation
network. The following are the centrality metrics and the giant component definitions as seen on
(Wasserman & Faust, 2006) and (De Nooy, Mrvar, & Batagelj, 2005).
Let G={V, E} be a undirected graph representing a network, where V is the set of nodes and E is the set of
edges in the network. The degree centrality Dc of a node is measured by the number of nodes directly
connected to the node v: . On directed networks, we can also have the in-degree
1
which is defined by the number of incoming arcs on the node, and also the out-degree, defined by the
number of outgoing arcs from the node.
The closeness centrality Cc of a node v is defined by the inverse of the average of all shortest paths
(geodesic paths) from v to all other n nodes on the network v:
. The shorter the sum of
distances from v to other nodes is, the higher the closeness centrality gets. Unlike the degree centrality,
which is local, the closeness centrality is a global measure, since it is defined by the entire network
topology.
The betweenness centrality Bc of a node v indicates the percentage of information flows on the network
that goes through v. It is defined as a proportion of the amount T(x,v,y) of shortest paths between every
pair of nodes in the network where v lies in the path and the amount T(x,y) of all shortest paths in the
network:
. The betweenness centrality is measurable only on directed networks
since the direction of the flow is relevant. It is also a global measure, like the closeness centrality.
On citation or co-authoring networks, it is common to verify that a node is not connected to every other
node in the network. On these cases, most analysis happens on a so-called giant component that
comprises many nodes of the network. The giant component can be defined as the largest connected
component of a network and normally it contains a significant fraction of the nodes.
2.1 - VoteRank
The VoteRank algorithm is a method to calculate the centrality of nodes on directed or undirected
networks. In this context, the centrality itself reflects the node's capacity for spreading news or influence
over a network. The VoteRank algorithm provides a better choice of influential nodes than other
comparable methods, such as PageRank or Degree, as demonstrated by Zhang, Chen, Dong & Zhao
(2016).
The VoteRank algorithm creates a concept of votes between connected nodes. Each node can be voted by
its immediate neighbors, each of these neighbors having its own voting ability. The voting score of a node
is the sum of the voting ability of its neighbors. A node with high voting score is more relevant than a
node with a low voting score.
The VoteRank algorithm begins with every node having a voting ability of 1. On each turn, the VoteRank
algorithm computes the voting score for every node in the network, and picks the node with the highest
voting score as an influence spreader. On undirected networks, the voting score is the sum of all
neighbors' voting ability within a distance to the spreader equal to one. On directed networks, the voting
score is the sum of the voting abilities only from the neighbors that receive information from the spreader,
which Zhang et al. refer to as out-neighbors.
Once a spreader is chosen, its voting ability and voting score are zeroed, and the voting ability of each of
its neighbors is reduced by a factor of
, where <k> is the average degree of the network, or the average
out-degree on directed networks. If there are more spreaders to choose from the network, the VoteRank
algorithm once again computes the voting scores for the network's nodes, excluding the spreaders that
had already been chosen, and picks the current node with the highest voting score as a spreader. Once
more, the spreader's score and voting ability are zeroed, and the voting ability of its immediate neighbors
is decreased. The method repeats this cycle until a satisfying number of spreaders is chosen.
The VoteRank method is computationally efficient. On each iteration, the method reduces the voting
ability values for a small set of nodes, and then computes the new voting scores for their neighbors alone.
The method is quite swift, since it only operates on a small section of the network at a time. Moreover,
the VoteRank algorithm provides an interactive form for choosing important nodes on the network, and it
is more efficient than other methods explored in the study, according to the SIR spreading model.
2.2 -- SIR
The methods comparison on the study by Zhang et al. was performed using the SIR spreading model. The
SIR spreading model is a de facto standard used for comparing algorithms that search for the most
2
important or central nodes on networks. It has been extensively utilized on several studies such as (Wang,
Zhang, Zhao, & Yi, 2016), (Lü, Zhou, Zhang, & Stanley, 2016), (Liu, Tang, Zhou, & Do, 2015) or
(Malliaros, Rossi, & Vazirgiannis, 2016), among others. The following information in this section is
based on the study of Pastor-Satorras et al. (2015).
The standard susceptible-infected-recovered (SIR) epidemic model is applied to simulate how fast and
extensively an infection can be spread over a network.
On this model, all nodes on the network are marked as susceptible, infected or recovered. A susceptible
node is a node that can be infected. An infected node is a node that can spread the infection to susceptible
nodes. A recovered node is a node that once was but no longer is infected, and became immune to
infection. There are variations, e.g. the recovered can be infected again (SIS), but the SIR spreading
model is more commonly used.
The method starts with a set of one or more spreaders being defined as infected. All other nodes on the
network are marked as susceptible. On each turn the infected nodes try to spread their infection to
neighbors with a success probability µ. After the spreading, and still on the same turn, the newly infected
nodes can become no longer infected with probability β and, in case of success, get marked as recovered.
This ends the current turn, and a new turn begins, with new infections occurring and infected nodes being
recovered. The simulation ends when all nodes are either susceptible or recovered.
In our scenario, the infection is the spreading of information. Clearly, the spreaders are the nodes chosen
with VoteRank or any other method. Simulations using several ratios between µ and β can be made,
determining how fast the spreading occurs and/or how extensive the spreading is. This ratio is usually
expressed as λ (λ
). If λ>0, the number of infected nodes grow exponentially, so simulations usually
have λ<0.
The methods are comparable in terms of how long the simulation took to end, or how fast the spreading
grew. But the most important factor is how extensive was the spreading, in terms of the percentage of the
network size that was affected, e.g. how many people received information broadcasted by a set of
spreaders. It can be expressed by the function R(t)
, where r is the number of recovered nodes and n
is the number of nodes on the network. When t = tend, the simulation is over and there are no infected
nodes left.
Generally, a method is considered to be more efficient than others if it determines a set of spreaders that
affects a more significant portion of the network than the set computed by the other methods, provided
the sets have a same size and the simulation parameters β and µ are equivalent. In a nutshell, the methods
compete among themselves for the best choice of spreaders, and the SIR method judges among them.
3 - Alternative VoteRank algorithms
The VoteRank algorithm, as it was originally proposed, assumes that only the nearest neighbors have
their voting ability reduced after a spreader has been chosen. On directed acyclic graphs, such as regular
citation networks, even fewer nodes are affected, because only nodes connected with arcs leaving the
spreader are computed (out-neighbors). This feature implies that a node on the network only spreads
influence on its immediate neighborhood, via direct communication, arcs or edges.
However, we consider that a node spreads influence to a larger area of the network, both indirectly and
through its closest neighbors. Assume that any given node A broadcasts information on the network. If
information broadcasted by A reaches another node B being at a distance greater than one from node A,
that is because of the spreading influence of node A anyway. Therefore, it is hypothesized that the voting
ability for node B could also be affected by A, as B was influenced by A. Thus, the VoteRank algorithm
can be somewhat extended.
We hereby propose an alternative VoteRank algorithm, where more neighbors can be affected by a
chosen spreader node A. In theory, if there is a path between nodes A and B on the network, B could also
have its voting ability reduced, regardless of how far apart they are.
3
The original VoteRank algorithm applies a reduction on the voting ability of its direct neighbors at a
factor of
, where <k> is the average degree of the network, or the average out-degree on directed
networks. We hypothesize that this reduction will also affect other, more distant neighbors, and that this
reduction will depend on the distance from the spreader. The farther a node is from a spreader, the less its
voting ability will be affected.
The first proposed alternative, called VoteRank-LRed, is applying a linear reduction on the voting ability
at a factor of
, where <k> is the average degree of the network, or the average out-degree on directed
voting ability
networks, and d is the distance from the node to the spreader. It is straightforward to verify that a node
that is very far from the spreader nearly isn't affected by it.
The second proposed alternative, called VoteRank-XRed, is applying an exponential reduction on the
where, again, <k> is the average degree of the network, or the average out-degree on
directed networks, and d is the distance from the node to the spreader. This variation applies the reduction
at a much faster pace, and the spreader's impact on distant neighbors is smaller than on VoteRank-LRed,
though higher than on the original VoteRank.
Since the reduction factor decays with distance, the algorithm applies the reduction on the voting ability
only up to a maximum distance from the spreader, in order to avoid excessive calculation. We decided to
apply the reduction factor on nodes with a distance of at most <k> from the spreader, where <k> is the
average degree node of the network, or the average out-degree on directed networks. On the next section,
we will apply VoteRank, VoteRank-LRed, and VoteRank-XRed on some patent citation networks and
compare their performances using the SIR model.
4 - Results
In order to apply and compare the VoteRank algorithms, we selected one real patent citation network
created by the authors on previous study.
3DPRN describes citations amongst patents regarding 3D printing technology; it was obtained
from Derwent World Patent Index on Aug-2014, and contains patents published until 2013.
Table 1 summarizes the main topological features of the network.
Features
3DPRN
nodes
arcs
density
average degree
653
1416
0,003321
4,336907
average out-degree
2,168453
nodes on giant component
442
% nodes on giant component
67,7%
Table 1 -- topological features of the test network
We tested the network with the Degree, Closeness, Betwenness, VoteRank, VoteRank-LRed and
VoteRank-XRed methods. For every test, we selected an arbitrary number of spreader nodes expressed in
terms of a percentage p of the number of nodes on the network. The tests were performed with p values
ranging from 0.0001 to 0.04. Many SIR simulations, with varying µ and β parameters, were performed
for each network, method and p amount of spreaders. For each choice of p, µ and β, 1.000 independent
simulations were performed. Therefore, the following graphs express average values from 1.000
simulations when p, µ and β were defined, or even more simulations when these parameters were not
4
fully detailed. On this study, we performed over 3 million SIR simulations overall. Figure 1 shows the
function R(t) for the percentage of nodes that recovered from the infection during the simulation.
Figure 1 -- Extent of infection on networks 3DPRN where p=0.02, µ=0.3 and β =0.2
Figures 1 shows that VoteRank-LRed had the best performance under these conditions.
Figure 2 shows how R(tend) varies according to different values of p, given µ and β are fixed.
Figure 2 -- Variation of infection extent due to amount of spreaders chosen, given µ =0.4 and β =0.2
Results were consistent, across several values of p, that Betweenness was the weakest method, and the
strongest were VoteRank-LRed, VoteRank-XRed, and VoteRank, in this order. On most simulations,
results were inconclusive when the number of spreaders was smaller than 15, varying a lot and pending to
Closeness as the method of choice. Above this threshold, VoteRank-LRed had the best performance, and
VoteRank-XRed came second between the leading method and VoteRank.
So far, we have only observed the behavior of the methods in the variation of time and number of
spreaders. Now we show in Figure 3 the comparison between the methods when the factor λ is changed.
We fixed p=0.01 and µ=0.5. Changing values of β, we had the following results. We show only the
VoteRank methods on this figure for better visibility due to the scale of the graphs. As expected, as λ
grows, more nodes are infected. The three methods behave equally.
5
Figure 3 -- Variation of infection extent due to changes in SIR parameter β, with p=0.01 and µ =0.5.
Table 2 shows the distribution of victories of each method along with all the tests for each value of p and
for each network. We created 66 possible combinations of values µ and β for each value of p. Again, it is
clear that networks with smaller amounts of nodes were more affected by smaller values of p. As p
increased, the usefulness of VoteRank-LRed increased as well. The EOLIC network is the largest one
tested, and with p equal to or larger than 0.001, i.e. 15 spreaders, results consistently point to VoteRank-
LRed as the best choice.
Table 2 -- Distribution of victories of each method through all tests.
VoteRank-LRed and VoteRank-XRed are somewhat more computationally intensive than original
VoteRank. In the step where voting abilities are reduced, more nodes are searched and updated. The
procedures needed to reach these other nodes up to distance <k> must be performed, and are not present
on the original method. However, using good tree implementations and caching techniques, we can
overcome the difference with no major impact. Moreover, it is a small price to pay if we are interested in
achieving better results.
5 - Conclusion
The obtained results revealed that VoteRank-LRed had better results than original VoteRank in most
scenarios. The variation hereby proposed is useful, and the hypothesis that the influence spread by the
nodes on larger areas of the network should be taken in account has proven to be valid.
6
Further investigations can be conducted regarding an extension of VoteRank-LRed. This version of the
method reduces the voting ability of nodes up to a <k> distance from the spreader, where <k> is the
average degree of the network. The <k> value was rather arbitrary, so simulations could be run to try to
determine if there are better approximations for a distance up to which the neighbors should be affected.
Other variations of the method could be tried, with other types of reductions instead of a linear one.
Lastly, the voting process for choosing a spreader by calculating its score considers only its closest
neighbors, i.e., those with a distance equal to one. Other versions of the method could be applied having
the score take into account other neighbors, as we did with the reduction of the voting ability, perhaps up
to a <k> distance as well.
We are of the opinion that the application of these new methods on patent citation networks can bring
new insights regarding the discovery of the most relevant patents in a network, and also help future
analysis of the values and influences of both patents and scientists.
References
Chang, S.-B., Lai, K.-K., & Chang, S.-M. (2009). Exploring technology diffusion and classification of
business methods: Using the patent citation network. Technological Forecasting & Social
Change, 76, 107-117; doi:10.1016/j.techfore.2008.03.014.
de Nooy, W., Mrvar, A., & Batagelj, V. (2005). Exploratory network analysis with Pajek. New York:
Cambridge University Press.
Hung, S.-W., & Wang, A.-P. (2010). Examining the small world phenomenon in the patent citation
network. Scientometrics, 82, 121-134.
Jaffe, A. B., Trajtenberg, M., & Henderson, R. (1993). Geographic Localization of Knowledge Spillovers
as Evidenced by Patent Citations. The Quarterly Journal of Economics, 108(3), 577-598.
Jang, S.-L., Lo, S., & Chang, W. H. (2009). How do latecomers catch up with forerunners?
Scientometrics, 79(3), 563-591; doi: 10.1007/s11192-007-2032-1.
Kim, E., Cho, Y., & Kim, W. (2014). Dynamic patterns of technological convergence in printed
electronics technologies: patent citation network. Scientometrics, 98, 975-998.
Liu, Y., Tang, M., Zhou, T., & Do, Y. (2015). Improving the accuracy of the k-shell method by removing
redundant links: From a perspective of spreading dynamics. Sci. Rep, 5, 13172; doi:
10.1038/srep13172.
Lü, L., Chen, D., Ren, X.-L., Zhang, Q.-M., Zhang, Y.-C., & Zhou, T. (2016). Vital nodes identification
in complex networks. Physics Reports, 650, 1-63.
Lü, L., Zhou, T., Zhang, Q.-M., & Stanley, H. E. (2016). The H-index of a network node and its relation
to degree and coreness. Nat. Commun., 7, 10168; doi: 10.1038/ncomms10168.
Malliaros, F. D., Rossi, M.-E. G., & Vazirgiannis, M. (2016). Locating influential nodes in complex
networks. Sci. Rep., 6, 19307; doi: 10.1038/srep19307.
Pastor-Satorras, R., Castellano, C., Van Mieghem, P., & Vespignani, A. (2015). Epidemic processes in
complex networks. Rev. Mod. Phys., 87(3), 925-979.
Wang, X., Zhang, X., Zhao, C., & Yi, D. (2016). Maximizing the Spread of Influence via Generalized
Degree Discount. PLoS ONE, 11(10), e0164393; doi:10.1371/journal.pone.0164393.
Wasserman, S., & Faust, K. (2006). Social network analysis: Methods and applications. New York:
Cambridge University Press.
7
Xiang, X.-Y., Cai, H., Lam, S., & Pei, Y.-L. (2013). International knowledge spillover through co-
inventors: An empirical study using Chinese assignees' patent data. Technological Forecasting
and Social Change, 80(1), 161-174; doi:10.1016/j.techfore.2012.07.003.
Zhang, J.-X., Chen, D.-B., Dong, Q., & Zhao, Z.-D. (2016). Identifying a set of influential spreaders in
complex networks. Sci. Rep., 6, 27823; doi: 10.1038/srep27823.
8
|
1603.04010 | 1 | 1603 | 2016-03-13T08:39:19 | Using Twitter to Understand Public Interest in Climate Change: The case of Qatar | [
"cs.SI",
"cs.CY"
] | Climate change has received an extensive attention from public opinion in the last couple of years, after being considered for decades as an exclusive scientific debate. Governments and world-wide organizations such as the United Nations are working more than ever on raising and maintaining public awareness toward this global issue. In the present study, we examine and analyze Climate Change conversations in Qatar's Twittersphere, and sense public awareness towards this global and shared problem in general, and its various related topics in particular. Such topics include but are not limited to politics, economy, disasters, energy and sandstorms. To address this concern, we collect and analyze a large dataset of 109 million tweets posted by 98K distinct users living in Qatar -- one of the largest emitters of CO2 worldwide. We use a taxonomy of climate change topics created as part of the United Nations Pulse project to capture the climate change discourse in more than 36K tweets. We also examine which topics people refer to when they discuss climate change, and perform different analysis to understand the temporal dynamics of public interest toward these topics. | cs.SI | cs | Using Twitter to Understand Public Interest in Climate Change:
The case of Qatar
Sofiane Abbar, Tahar Zanouda, Laure Berti-Equille, Javier Borge-Holthoefer
Qatar Computing Research Institute
Hamad Bin Khalifa University
PO 5825, Doha, Qatar
{sabbar, tzanouda, lberti, jborge}@qf.org.qa
6
1
0
2
r
a
M
3
1
]
I
S
.
s
c
[
1
v
0
1
0
4
0
.
3
0
6
1
:
v
i
X
r
a
Abstract
Climate change has received an extensive attention from pub-
lic opinion in the last couple of years, after being considered
for decades as an exclusive scientific debate. Governments
and world-wide organizations such as the United Nations are
working more than ever on raising and maintaining public
awareness toward this global issue.
In the present study,
we examine and analyze Climate Change conversations in
Qatar's Twittersphere, and sense public awareness towards
this global and shared problem in general, and its various re-
lated topics in particular. Such topics include but are not lim-
ited to politics, economy, disasters, energy and sandstorms.
To address this concern, we collect and analyze a large dataset
of 109 million tweets posted by 98K distinct users living in
Qatar – one of the largest emitters of co2 worldwide. We use
a taxonomy of climate change topics created as part of the
United Nations Pulse project to capture the climate change
discourse in more than 36K tweets. We also examine which
topics people refer to when they discuss climate change, and
perform different analysis to understand the temporal dynam-
ics of public interest toward these topics.
Introduction
In 2011, the state of Qatar was designated by the World Bank
as the highest per-capita emitters of CO2 in the world with
more than 44.02 metric tons per person [19]. This surpris-
ingly high rank is actually due to Qatar's massive energy
sector and relatively small population. With a population of
only 2.4 million in 2015 (1.8 million in 2011), Qatar was in
2013 the world's fourth-largest dry natural gas producer, and
world's first liquefied natural gas (LNG) exporter in 2014
[10]. Given the serious implications that such high emis-
sions of greenhouse gases may have on population's health
and wellbeing and on the national eco-system and biodiver-
sity equilibrium, the government of Qatar has taken many
steps toward limiting the country's ecological footprint [1].
For instance, Qatar hosted in 2012 the Conference of the
Parties (COP18) of the United Nations Framework Conven-
tion on Climate Change where an agreement was signed to
extend the life of Kyoto protocol till 2020 [15].
Qatar's mission to limit the country's carbon footprint
is part of a global movement to act on the urgent warn-
Copyright c(cid:13) 2015, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
ings emanating from the scientific communities about Cli-
mate Change. Governments and world-wide organizations
such as the United Nations are working more than ever on
raising and maintaining public awareness toward this issue.
While sensing the engagement of governments to mitigate
the causes of climate changes such as dioxide carbon emis-
sions is relatively straightforward by looking at the interna-
tional treaties and protocols ratified, it is not clear how to
achieve the same objective at citizens' scale, who are part of
the causes and consequences of this shared problem. Tra-
ditional surveys such as field interviews are costly and do
not scale for obvious reasons. To address this concern, we
analyze Twitter's content to sense the awareness of people
toward climate change and investigate their interest toward
its multiple topics. The recent multiplication of hazardous
natural disasters around the world and the news coverage
they received have helped disseminate awareness among in-
dividuals about the consequences of climate change prob-
lems, which happens to be amplified by social network plat-
forms.
Main contributions.
In this study, we are interested in
sensing the public interest in climate change of people liv-
ing in Qatar. This is achieved by analyzing a large collection
of tweets posted between 2011 and 2016. Unlike previous
studies that focused either on a limited set of climate change
related keywords or on a set of geo-coded tweets (represent-
ing only 2% of twitter traffic), our approach aims at captur-
ing most of the discourse related to climate change under
its different aspects ranging from politics and economy to
energy and disasters. Our main contributions include: (i)
We identified of 117K users (5% of the entire population)
claiming they live in Qatar and collected all their climate
change related tweets using a rich taxonomy of topics de-
veloped by the United Nations; (ii) We propose to use net-
work backbones analysis to better understand relationships
between hashtags in co-occurrence graphs; (iii) We analyze
different time series of different topics and infer the most
driving topics discussed in the context of climate change.
Challenges. The main challenge we encountered in this
study was identifying a relevant and accurate taxonomy of
terms to use in order to retrieve climate change related posts.
Many studies have focused on the usage of a single word,
i.e., climate in [4], climate change in [7], and some of them
added a handful of other terms to it such as global warm-
ing in [6, 17]. We finally decided to use a taxonomy de-
veloped by a United Nations work group including climate
change experts and social scientists [16]. The taxonomy is
composed by a set of topics, each one is defined by a set
of keywords or hashtags that should co-occur in a particular
way. The experts have tried as much as possible to keep the
set of terms that define topics disjoint. The other challenge
was to identify correctly users living in Qatar and crawl all
their tweets. We have provided a heuristic based on follow-
ing relationship and a gazetteer of locations to find relevant
users. Parsing a collection of 109 million tweets to extract
climate change related ones, and working to keep the level of
noise to its minimum possible level was a challenging task
as many manual checking was required.
Context of the study.
In this study, we focus on Twitter
conversations related to environmental issues posted by peo-
ple living in Qatar. Qatar is a small peninsula in the Arabian
Gulf. It is the home of a native population of approximately
300,000 Qataris, who live among a total population of 2.4
million [13]. Internet penetration is high in Qatar; 85% of
the population has Internet access. In addition, social me-
dia are popular and widely used among the native popula-
tion –65% of Qataris have an Instagram account, 44% are
on Facebook, and 46% use Twitter [11].
Roadmap. The rest of the paper is organized as follows.
We start by providing a literature review on related work.
Then, we define the methods used for the analysis. Next, we
describe the datasets we collected. Then, follows a section
that presents and discusses our findings. Finally, we con-
clude the paper with some insights.
Literature Review
Climate change is a major challenge facing the world, with
myriad causes and consequences. One of the leading con-
tributors to the rising levels of CO2 emissions is the activity
of humans in urban places, as a consequence, the unmanage-
able rising of CO2 emissions altered global climatic condi-
tions and indirectly harms the environmental ecosystem 1.
In order to strategically raise public awareness on climate
change issues, and protect the world's shared natural en-
vironment for sustainable and spiritual benefit, researchers
have analyzed public opinions to identify public awareness,
examine this trend on the social media sphere and investi-
gate lasting public awareness plans.
As most people have limited resources to get accurate data
on air pollution and climate change, Wang and Paul [18]
have investigated the value of social media as a sensor of
air quality and public response. Authors analyzed more than
93 million messages from Sina Weibo, China's largest mi-
croblogging service and found a high correlation with par-
ticle pollution levels, which makes the use of social media
1http://www.un.org/climatechange/
data valuable to augment existing air pollution surveillance
data, especially perception.
Alternatively, other studies have demonstrated the utility
of social media data sources to analyze public awareness
about climate change, and the influence of mass media to
interpret global climate problems from the different angles
(causes and remedies) [5]. While Veltri and Atanasova [17]
highlighted the potential opportunity of using social media
to analyze and understand public opinion dynamics. Oltre-
anu et al. [9] compared the coverage of climate change in
both, social media and mainstream news. This comparison
uncovered significant differences between triggers, actions,
and news values of events covered in both types of media.
For instance, mainstream news sources frequently feature
extreme weather events framed as being a consequence of
climate change, as well as high-profile government publica-
tions and meetings. In contrast, actions by individuals, legal
actions involving governments, and investigative journalism
feature frequently as viral events in social media.
As most of people have no deep scientific understanding
on climate change, their perceptions are most likely influ-
enced by framing strategies [20] , which leads the research
community to question public opinions over time and lo-
cation, along with the role that mass media played. Cody
et al. [4] studied user sentiment toward climate change by
analyzing environment-related tweets posted between 2008
and 2014. The study shed a light on how climate news,
events, and disasters affects people's sentiment. An et al.
[2] also conduct an opinion analysis of climate change re-
lated tweets. Authors aimed at understanding whether Twit-
ter data mining can complement and supplement insights
about climate change perceptions, especially how such per-
ceptions may change over time upon exposure to climate
related hazards. A combination of techniques drawn from
text mining, hierarchical sentiment analysis and time series
methods is employed for this purpose. At a large scale, Jang
and Hart [6] investigated how people incorporate climate
change into everyday conversation by the stream of Twitter
over two years. Their study shows geographical differences
in the use of frames and terms concerning climate change,
and how certain frames that promote skepticism about cli-
mate change are widely circulated by users within specific
regional and political contexts.
In this paper, we analyze climate change conversations
and air quality trends in Qatar through the lens of social
media data, and investigate the ability of measuring public
awareness, along with the impact of environmental factors
on public conversations.
Methods
We describe in this section the methods we used in this
study.
Building the Network of Co-occurrences
The relationship between different concepts in the dataset
can be suitably represented as a network. Nodes in the net-
work represent concepts, and a link from node i to node j
implies that the corresponding concepts of the two nodes
co-occur in the same tweet. More importantly, each node
bears an associated quantity a that stands for the proportion
of tweets in which it appears. On the other hand, each link
from i to j has an associated weight (i, j, w), which states
the fraction of tweets in which these two nodes co-occur.
Network backbone extraction
Networks can be described from different levels of analy-
sis. At the micro level, the focus lies on single nodes and
their specific positions within the overall structure; this level
can be described in terms of node degree, strength or clus-
tering coefficient, among other metrics. At the macro level,
the focus shifts to the aggregation of those metrics and the
properties of their distribution. Between these two extremes,
we have a third level of analysis, the mesoscale, which aims
at accounting for the complexity of networks between the
position of individual nodes and the relational properties of
the collectives they form. It is at this level where reduction
techniques like backbone extraction operate.
Network backbone extraction refers to the filtering tech-
niques aimed at uncovering relevant information; in general,
such techniques aim at pruning the links of a network, keep-
ing only those which are statistically relevant. Ideally, the
reduced structure is computationally more tractable while it
retains most of the interesting features of the original one.
We apply the state of the art backbone extraction algorithm
proposed in [12] to the network of hashtag co-occurrences.
In their work, Serrano et al. exploit the empirical trend by
which link weights are heavily fluctuating, i.e., only few
links carry the largest proportion of the node's strength.
Linear Correlation
We use Pearson product-moment coefficient to measure the
linear correlation between different pairs of variables (en-
vironment topics and weather conditions) of interest in this
study. Pearson correlation score ranges in [−1, 1] where 1
(resp. -1) means the existence of total positive (resp. nega-
tive) correlation and 0 means no correlation at all. Pearson
formula between two time series X and Y is given below:
ρX,Y =
E[(X − µX )(Y − µY )]
σX σY
(1)
where: µX , µY are the means of X, Y respectively, and
σX , σY are the standard deviations of the two corresponding
time series.
Datasets
To conduct this study, we need to combine data from two
different sources: Twitter and Weather Underground.
We use Twitter to collect user generated content about
environment related topics. As there is not solicitation for
users to post on Twitter, we claim that this data reflects the
actual and natural responses to users toward different envi-
ronmental topics as they get exposed to them through news
media or witness them happening during natural disasters,
heavy rains, drought, etc. Underground Weather is used
to collect daily weather conditions in Doha, the capital of
Qatar. It is worth noticing that Qatar is a relatively small
(a) Tweets
(b) Census 2010
Figure 1: Qatar Maps
country (area: 11,586 km2) with more than 80% of its pop-
ulation based in the city of Doha. Thus, the weather does
not vary that much from a location to another in the coun-
try. We explain in the following the processes by which we
collected and curated the different datasets.
Identifying users of Qatar
First, we created a gazetteer of locations in Qatar including
the names of main cities and districts in both English and
Arabic, and under different spellings. We matched this list to
a 45 day sample of the Twitter Decahose (a sample of 10% of
all Twitter traffic.) For each user whose tweet was captured,
we requested the list of their followers based on the intuition
that for an average user, followers are more likely to be from
the same country.
Then, we requested the profiles of the followers and filter
out those users who did not mention a location correspond-
ing to Qatar's gazetteer. The same process was iteratively
executed on the list of newly added users; i.e., find their fol-
lower IDs, then follower profiles, then filter out followers
with irrelevant locations, until no new users are identified.
To expand the list of users even further, a social analytics
tool (FollowerWonk.com) was used to obtain lists of users
who mention a location associated with our gazetteer. This
list was then merged with the list obtained earlier from Dec-
ahose and Twitter API. The resulting list consists of over
117K users who claim to be living in Qatar in their Twitter
profiles.
Collecting user tweets
We use Twitter Historical API2 to collect up to 3,200 tweets
posted (the maximum allowed by Twitter API) for each of
the users of Qatar identified in the previous step. This col-
lection has been conducted twice. The first time was in July
2https://dev.twitter.com/, last accessed on Feb,2016
2014 and the second time was in February 2016. Hence,
we could obtain for each user up to 6,400 tweets he posted
over a period of nine years (from 2007 to 2016). This col-
lection generated a total of 109.6 million tweets posted by
more than 98,066 different users. The remaining missing
users have either a disabled or protected account. As Twitter
was launched in 2006 only, the activity of the users of Qatar
was very marginal before 2011. Thus, we focus in this study
on the time period spanning from January 1rst, 2011 to Jan-
uary 1rst, 2016.
Weather data
We use Weather Underground (WU) API [14] to collect
data about daily weather conditions in Qatar for the same
time period. Out of the sixty different measurements WU
API returns, we were particularly interested in the ones that
are more likely to sprinkle climate change related tweets,
temperature (◦C), humidity
these measurements include:
(%), wind speed (km/h), visibility (km), and precipitation
(mm) (including: rain, snow, etc.)
Extracting environment and climate change tweets
In order to identify the climate change related content, we
use a taxonomy of keywords grouped into topics on inter-
est. Previous studies have used simple keywords to filter
climate change tweets. Cody and Reagan for instance used
only one word climate [4]. Veltri and Atanasova used two
keywords climate change and global warming [17] while
Andrei and Svetlana added translations of the aforemen-
tioned words in German, Portuguese, Russian and Spanish
[7]. Jang and Har also used the keywords climate change
and global warming paired with some specific terms related
to different frames [6].
We use in this paper a taxonomy of climate change
topics[16] developed by the UN Global Pulse team. The
taxonomy consists of set of topics, each of which is defined
with a list of terms and hashtags that should co-occur in a
particular way. According to Ren´e Clausen Nielsen – data
innovation specialist in UN Global Pulse – the taxonomy is
created by climate change experts as follows: (1) find a list
of relevant words. (2) find synonyms, plurals, and abbre-
viations by searching Twitter and dictionaries.
(3) Group
words into distinct topics. (4) Create negative filters to get
rid of occurring noise [8].
topics are carefully defined not
An important feature of the used taxonomy is the fact
that
to allow overlap-
ping. These topics are: General, Risk&Disaster ,Poli-
tics&Opinion, Economy, Energy, Oceans&Water ,Weather
,Agriculture&Forestry, Arctic , Campaigns , Negotia-
tions&Summits. For a tweet to be labeled as belonging to
a particular topic, it should satisfy a list of conditions. For
instance, a tweet is about Politics&Opinions if and only if
it CONTAINS ANY {climate, enviro, environment, carbon}
AND NOT CONTAINS ANY {monoxide, fork, copy, fiber}
AND CONTAINS ANY {politics, group of eight, . . .}.
The dusty construction sites in the middle of the city, and
the flat dusty landscape stretching out of the capital's bor-
ders, both contribute to complicated air quality conditions.
We added Air Quality topic separately to captures conver-
sations around air quality and sense how people talk about
the air they breath by tracking words that refer to air and air
pollution (e.g., air quality, air pollution ). Like Countries in
the region, Qatar has built cities in the middle of the desert.
Heavy sandstorms hit the country regularly, and keep fine
dust hanging in the air for days. Thus we added Sandstorm
topic composed of terms such as duststorm, dusty, and sand
storm. A full list of topics with their respective definitions,
terms, and matching conditions can be downloaded from [3].
Data Curation
Next, we started doing a qualitative analysis on our dataset
to determine content that might not be relevant in our analy-
sis. We have noticed that some irrelevant keywords such as
Syria, California, Colorado, etc. were appearing to the top
of co-occurrence lists. The manual inspection allowed us
to identify some bots that were re-tweeting all tweets about
Syria with by appending US state names as hashtags.
At the end of this process, we obtained a clean set of
36,612 environment relevant tweets posted by 8,470 differ-
ent users (also known as tweeps).
Figure 1 shows the geo-coded tweets related to environ-
ment that we have gathered, and a population census from
2010 for comparison. The most populated area is definitely
around the capital city of Doha, with beaches and main roads
being other popular posting sites. A shallow visual inspec-
tion reveals that posting proficiency is correlated with popu-
lation density in general.
Table 1: Description of the final set of topics related to envi-
ronment and climate changes with some basic statistics.
Negotiations/Summits
Energy
Agriculture/Forestry
Category
General
Risk/Disaster
Politics/Opinion
Economy
Ocean/Water
Campaigns
Arctic
Air quality
Sandstorm
Weather
#Tweets
3828
320
11124
2615
7576
1716
1119
458
470
132
69
13124
1139
#Users
1293
129
3192
1038
1080
722
659
289
159
112
41
5605
793
T/U
2.96
2.48
3.48
2.51
7.01
2.37
1.69
1.58
2.95
1.17
1.68
2.34
1.43
%T
10%
8%
30%
07%
20%
4%
3%
1%
1%
0.3%
0.1%
30%
0.03%
Table 1 illustrates the final set of topics related to envi-
ronment and climate changes with some basic statistics in
terms of number of tweets, number of users, ratio of tweets
per user and percentage of tweets related to each category.
Results
In this section, we present the results of different analy-
sis performed on the set of environment-related tweets de-
scribed in the previous section. First, we show the temporal
trends of people reactions to different environmental topics.
We show also the most prominent hashtags associated with
different topics. Then, we look at the co-occurrence graph of
these hashtags to unveil different hashtag relationships and
graph topologies. Finally, we analyze user profiles to check
whether users posting about environment demonstrate any
specific characteristics.
Temporal dynamics
Figure 2 shows the trends of the monthly total number of
tweets (a) and unique users (b) observed in the Qatar "twee-
tosphere". The blue curves represent the trends of all the
tweets posted while the green curves represent environment
related tweets only. To ease the reading, the y-axis is plotted
in log scale. The first striking observation is that even in log
scale, users' activity around environmental topics fluctuates
a lot compared to the smooth general discussions on Twitter.
Most of the spikes observed in Fig 2(a) can be explained by
looking at the tweets posted around the spike period. For in-
stance, one of the most noticeable burst, the one happening
in November-December 2012 is a response to the Confer-
ence of the Parties (COP18) that took place at Qatar late in
2012. That conference to which more than 50 presidents
from around the world took part received a large news cov-
erage locally and internationally. Yet, people living in the
small peninsula of Qatar have extensively experienced the
conference as it has directly impacted their everyday life in
terms of commute deviations, security enforcement, and ac-
cess limitations to some hotels. All this has generated an
unprecedented excitement on Twitter for the event. This ob-
servation is corroborated by Fig 2(b) in which it is easy to
see the important burst of the number of unique users who
took part of discussions and debates around COP18 con-
ference. The second sharper spike that takes place around
March 2015 is due to a sudden strong sandstorm that hap-
pened in Qatar in that same period, causing schools and in-
stitutions to close doors on that day as it was almost impos-
sible to drive. As an example, the below-mentioned tweet
is posted by Qatar's Ministry of Interior on 1st April 2015
3: "Visibility is almost zero in all roads inside and on high-
ways. Motorists are requested to follow safety instructions.
#MoI Qatar."
Continuing further with the temporal analysis of our set
of tweets, we plot in Figure 3 the breakdown of this trend
into different topics by providing the daily time series of
the number of tweets associated with each topic. The first
figure to the left (Global) plots the daily time series of all
environment-related tweets. Some observations can be done
here. First, the different topics vary a lot in terms of volume
and trends. As highlighted in Table 1, the sandstorm topic
seems to overtake all other topics. It is clear that this topic
constitute the most discussed environment related topic on
Twitter by the people of Qatar as this has a direct and imme-
diate impact on their lives. The figure shows also that when
the people of Qatar discuss about environment on Twitter, it
is mostly about Politics&Opiniom, Negotiations&Summits,
and climate change in general. The less covered topics in-
clude Oceans&Water, Arctic and surprisingly Air Quality.
3https://twitter.com/MOI_QatarEn/status/
583359974278967296
(a) Tweets
(b) Users
Figure 2: Monthly volume of the total number of tweets
and unique users in log scale in Qatar from January 2011
to January 2016. We see that climate change curves present
a noticeable variance over time unlike the general smoother
Twitter pattern.
Second, one can easily see that topic time series follow dif-
ferent trends. Indeed, while some topics such as Economic
and Energy maintain a decent level of activity throughout
the years, other topics such as Negotiations&Summits and
Arctic are sparse and only acquire users' attention when a
related event happens.
Content Analysis
In order to visually inspect the content of tweets in differ-
ent topics, we plot Figure 4 in for each topic, we gener-
ate its word cloud of the most frequent hashtags mentioned
in the corresponding tweets. The font size of hashtags is
proportional to their frequencies in different topics. These
word clouds reveal for instance that the hashtag #COP18
is dominating in our dataset of tweets (global word cloud)
and is obviously dominating in tweets belonging to Nego-
tiations&Summits. We also found the hashtags #environ-
ment, #climate, and #carbon to be dominating in Economy
and Energy as #carbon is relevant to both topics (the carbon
economy). #environment dominates also Ocean&Water and
Agriculture&Forestry, #climate dominates in Weather, #ac-
tonclimate dominates in Campaigns, #climatechange domi-
nates in Arctic, and #airpollution dominates in Air Quality.
To further analyze the content, we focused on the relation-
ships between hashtags within topics and we built the co-
occurrence networks of hashtags for the tweets of each topic.
Because the generated networks were dense, it was not pos-
sible to perform any kind of visual observations on them.
Thus, we apply the backbone extraction algorithm explained
in the method section on all generated graphs. The algorithm
201120122013201420152016Time0246810121416Number of tweets (log scale)All tweetsEnvironment related tweets201120122013201420152016Time024681012Number of users (log scale)General Twitter usersEnvironment-aware users Figure 3: Daily distributions of the total number of tweets for different climate change related topics ordered by volume from
top left to bottom right. Notice that y-axises are not normalized across the figures.
removes all noisy edges from the input graph and retains
only relevant once. Figure 6 plots the backbone networks
of each topic. We use Gephi4 tool to draw the graphs. Col-
ors in the graphs correspond to the different communities of
hashtags identified by Gephi's internal modularity detection
algorithms. The first graph (Global) is generated using all
the climate change related hashtags. We clearly see the ex-
istence of four big communities of hashtags around: #envi-
ronment, #climatechange, #qatar. Interestingly, we see that
the #climatechange community includes two heavy nodes:
#cop18 and #cop21. While #cop18 is heavily connected to
the community around #qatar just as the other hashtag #doha
(the capital), #cop21 (which happened this year in Paris)
does not seem to be connected to Qatar. As the other COP
editions didn't attract media's attentions comparing to the
above-mentioned ones, our study dataset reflects the same
pattern.
The backbone network corresponding to general topic is
nicely clustered too.
In fact, it has a giant component of
hashtags around #climatechange and a small yet accurate
one around #drought containing hashtags such as: Califor-
nia, Haiti, Somelia, rain, and water. All these hashtags
are related in a way or another to drought. The absence
of rain and water are the main causes of drought, and the
mentioned countries and cities are the ones that are strug-
4http://gephi.org, last accessed on Feb, 2016
gling with ongoing drought. Negotiations&Summits net-
work is also very informative. The modularity detection
algorithm used was able to partition the network into four
main communities: #cop17, #cop18, #cop19, and #cop21,
with a high polarization around #cop18 (Doha) and #cop21
(Paris) which are both associated with some distinctive hash-
tags.
Interestingly one can easily see that COP17 (Dur-
ban, South Africa), COP19 (Warsaw, Poland) and COP20
(Lima, Peru) did not attract as much attention as COP18
(Doha, Qatar) and COP21 (Paris, France) if not at all. While
COP18 is heavily covered for obvious reasons, we believe
that the Paris' COP21 local attention is mainly due to the
fact that Paris was already catching the worldwide attention
for the sad Paris attacks event that happened two weeks be-
fore the start of COP21 5, thus benefiting from a large in-
ternational media coverage that was already in place. It is
also interesting to see that both conferences are associated
with distinctive sets of hashtags. For instance, COP18 is as-
sociated with hashtags such as #climatelegacy, #kyoto, #no-
justicenodeal, #ethiopia, and #philippines whereas COP21
is associated with hashtags such as #climatejustice, #car-
bon, # renewables, #climatefinance, #energy, #morocco, and
#china. Interestingly, the hashtag #Aljazeera which refers to
the name of the famous Qatar-based media network appears
5November 2015 Paris attacks: https://en.wikipedia.
org/wiki/November_2015_Paris_attacks
2011201220132014201520160100200300400500600700800900global2011201220132014201520160100200300400500600700800900sand_storm2011201220132014201520160100200300400500600negotiations_summits201120122013201420152016020406080100120140160180politics_opinion201120122013201420152016010203040506070general201120122013201420152016010203040506070weather20112012201320142015201605101520economy201120122013201420152016024681012141618energy201120122013201420152016024681012risk_disaster2011201220132014201520160246810121416campaigns201120122013201420152016024681012agriculture_forestry201120122013201420152016012345ocean_water2011201220132014201520160.00.51.01.52.02.53.0arctic2011201220132014201520160.00.51.01.52.02.53.0air_quality on the COP21 side. Politics&Opinion backbone network
presents an interesting topology with a strong emphasize on
#environment which is linked to three smaller communities:
#qatar, #climate, and #waste (represented in blue). The re-
maining graphs are self-explanatory. However, it is worth
notice that the relatively small networks of Oceans&Water
and Risk&Disaster have only one community. Arctic and
Air Quality networks are not represented here because the
former returned an empty backbone graph whereas the latter
had only two nodes, namely: #airpollution and #health.
Figure 4: Word clouds of hashtags corresponding to differ-
ent environment-related topics. The first word cloud to the
left is generated from all environment-related tweets.
Environment-aware users
After our content-centered analysis, we work on analyzing
users. We use Users endpoint of Twitter API to request the
profiles of 8K users who posted at least one tweet about en-
vironment. Each user profile comes with basic information
such as id, screen name, full name, bio and some statistics
about the number of posts, followers, and friends. We an-
alyze the profiles of users, particularly their self-identified
biographies. Notice that while not all users provide a biog-
raphy in their profiles, most of them do. Our objective is to
get insights into the interests of users who tweet about envi-
ronment and climate change. Thus, we merged all users' bi-
ographies into one file and perform some cleaning tasks such
as the removal of Arabic and English stop words. We show
in Figure 5 the most frequent 50 terms found in the biogra-
phies of top 1% (a), top 5% (b) most tweeting users, and that
of all users (c). Unsurprisingly, the more users who tweets
about climate change, the more likely they mention a climate
change related term in their biographies. In other words, this
means that prolific users in our dataset (top 1%) are defi-
nitely environmental activists who identify themselves with
words such as: climate, change, development, environment
and sustainable. The second tier of users (top 5%), present
a media and news oriented profiles. These users mention
in their biographies words such as: news, views, journalist,
aljazeera (which is one of the biggest news groups in the
world). When all the users are taken together, all signals re-
(a) Top 1% most prolific users.
(b) Top 5% most prolific users.
(c) All users.
(d) Complementary cumulative
distribution function.
Figure 5: Results for users analysis.
(a) through (c) rep-
resent word clouds of terms found in user biographies re-
trieved from Twitter. We see that the more a user tweets
about climate change, the more likely she is to have climate
change related terms in her profile. (d) shows the skewed
distribution of number of climate change related posts per
users, with more than half users posting only once.
ferring to climate change fade out. In fact, at scale, we found
that more than half of user users (51.08%) have tweeted only
once about climate change in a course of 5 years (2011 -
2015). These users tend to mention broader topics in their
biographies such as: love, life, world, etc. Notice that the
word Qatar has been removed from all three word clouds
as it dominated all of them. Figure 5(d) plots the comple-
mentary cumulative distribution function for the probabil-
ity (y-axis, log scale) that a random user would post more
than a given number of tweets (x-axis). The figure shows a
rapidly decreasing trends with most users posting only one
tweet, and very few (p=0.005) who posted a hundred tweets
or more.
Correlations
The correlation analysis performed on the 18 variables (11
UN climate change topics, 2 Qatar specific topics, and 5
weather measurements) revealed the existence of a signif-
icant dependency between some of them. Figure 7 shows
the Pearson correlation score computed over the daily time
series of all the pairs.
Looking solely at correlations between climate change
topics, we find that Politics&Opinion and Negotia-
tions&Summits achieve the highest correlation score of ρ =
0.73. This definitely makes sense as many political activities
and "pourparlers" take place during the negotiations. The
second and third highest correlation scores are achieved be-
tween Politics&Opinion and Economy (ρ = 0.62) and Poli-
tics&Opinion and Energy (ρ = 0.61). This is also expected
100101102103Number of posted tweets10-410-310-210-1100CCDF (a) Global
(b) General
(c) Risk&Disaster
(d) Summits&Negotiations
(e) campaigns
(f) Politics&Opinion
(g) Sandstorm
(h) Energy
(i) Economy
(j) Agriculture&Forestry
(k) Oceans&Water
(l) Weather
Figure 6: Graphs representing hashtag co-occurrence backbones of different climate change topics (b) through (l). (a) is the
global backbone graph computed from the whole set of climate change related tweets.
globalwindyunepfollowdohanewsparisopportunitiesenvironmentsotuqnfsprecyclesafetykahramaawomenclimatemarchcmp8oceansdmkrworldbankelectronicsearthdayqfradioinstadohahawaiigreenbuildingsandsaudisciencedemocratsmyqatarinstagramriyadhbiomassstreetwastemanagementbusinessinvestingroadtoparismiddleeastquranearthtoparissomaliabritish_columbiaspecialistiranwaterchinasustainabledevelopmentchangebiodiversityforumdustyomanalaskasocialactionglfcop19loveegyptafricanowreadingreddlimasanantonioworkinggotvlevantfranceliveiphoneperuindianaactonclimateclimatelegacyindiaoregonestedamaqatarconnecticutqeeriyourpowercharitymeanimalscardohaexhibitioncenterworksandstormnebraskabeautifulbeijingcrazyclimatsdgsworldenvironmentdaycleantechuaeinstaqatargccunitedwinterzchizzerzhotasmfanimaldust_stormidahoeconomyfitnbeautysustainabilityfbcop18volunteerssb42climatefinanceearthhourpost2015coalgreennew_mexicowindhelpdrivesafelyyoungohellbirdsislamclimatechangeclimatechangeconferencesocialmediarhode_islandwalesfinancefoodcop18photoicymiinvestorqgbcunfccccountmeinajclimatesaopaulocsrphilippinesenergyestedamaeuagriculturesandycanadakuwaitoklahomalossanddamageresearchhealthsmogarizonadurbanecowcsj2011fridayabudhabinew_jerseyzerovisibilityundpcaliforniapapersmartmarineorganicdoha2012dohadisasterfloridaplaceaspiresyriakccwgtarsheedfeaturedwestbaydohasandstormantarcticcityplasticsouth_carolinawarmingstormpalestineadp2015wastesenvironmentaladvancedglobaltradingjordaninyteftobamaipad50by30treeyouthfuturehotnewsglobalwarmingnojusticenodealwarsaw1o5cairpollutionbiogassomalilandnova_scotig7summithealthy_tipscomedyhourchangeclimatechangemontanadroughtayiccclimateactiondrivekpministry_of_environmentpartofthesolutionindcfacthungerdesertlivingaycmearthindcsexpo2015gulfngo14febksacop20menaawarenessbahraindotreeshaitiqumorningbadqfsustainableohioviewinstapicfamineseasportbestbrazilco2nhpoliticsarkansaseuropejobsar5naturejobjowtfqfrecruitmentcarboncitiesifnotusthenwhoaljazeeraenergyefficiencyclimatesandstormsgogreenpovertypakistanduststormrecyclingcommunitytennesseeadaptationnewsipadgamesconferenceislamicmitigationcop15cop17cop18cop19wasterenewableslifeprotectlpaahmcgeorgiagameinsightairhealthyeidseemydohaalmatydohayearinspaceadp2alsaddbagsthepearlqatard12smokefreeauspoleventsdaydevelopmentoillatviacop18fossilfuelgocop21arcticrainiondustrenewableqatarfoundationaction2015pennsylvaniapolandindonesiaccsclimate2014alabamasummerparisagreementcop21photoforeststhinklandscapehumandohagreenupmoi_qatarwefadpskycornichealkhorsavepublichealthaustraliasif15virginiaarabtrafficlcaecosystemsdubaiiphone5instagramqataralamointeqbigblockofcheesedaygivesyouwingsecologypeopleroad2dohaunfaoqnfspqatartransportqatarqnccdelawareglfcop21foginnovationpowerkyotocfcc15qatarsandstormrenewableenergycop21pariswildlifelifestyleusaworldentrepreneursqsciencerecognizedafternoonfoodsecuritypollutionfreefoodcoloradogarbagewashingtoninterstellarweatherresiliencebangladeshcdnpoliconservationlowboliviaclimatejusticeparis2015stealthrangerwbgyouthsummitheatsolaragchatozusplanetmoroccounemissionsukageducationwed2015solarenergypeacefrackingillinoisecotipcoy8dohaoasisqatarlivingcleanweekenddesertbuildingipccsamsungdtff12summitroadtodohaagthungercop21cop20globalthinklandscapedemocratsearthhourfollowclimatechangewefparisngoenvironmentmenanhpoliticsbahraindoaustraliafoodarabicymiecosystemshaitiunfcccphilippinesohioeconomyenergyestedamafamineearthagriculturebrazilplanetqataroklahomaroadtoparismiddleeastglfcop21sotuhealthsomaliairanwcsj2011waterchinarenewableenergycarboncitiesifnotusthenwhochangecop21parisclimateworldislamicfloridasocialcaliforniafoodsecuritypollutionpakistancoloradoafricanowreadingweatherreddantarcticresilienceaglevantfrancecop17cop18cop19renewablesactonclimateparis2015wbgyouthsummitjordaninyteftobamageorgiaagchatozusyouthunemissionsglobalwarmingdohasdgsindiaeducationpovertycsrsomalilandoilfossilfuelarcticraindeforestationsustainabilitydroughtcdnpolirenewableipccaction2015climateactionindonesiagreenpartofthesolutionfactqatarclimatechangedohacop21climateestedamaenvironmentgreencop18citiesglobalwarmingparisagreementcop21cop20youngoaycmdohagreenupearthhelpindcsclimatechangeadpparisngoenvironmentqnfspcdmwomendofinanceclimatemarchcmp8sif15dmkrarabcop18photolcaworldunfcccthinklandscapephilippinescountmeinajclimateqfinteqenergyeuqnfspqatarclimatechangeconferenceqatarlimakuwaitlossanddamageglfcop21nhpoliticsar5durbankyotowaterchinapapersmartcarboncitiesaljazeeracop21pariscopclimatedoha2012dohadisasterqscienceactionsb42glfcop19freefoodkccwgreddqnccconferenceclimatfrancelivechangeperucop17cop18cop1950by30renewablesadp2015cop15actonclimateclimatejusticeclimatelegacyparis2015lpaasolarayiccinyteftobamacomedyhourdohaexhibitioncenteryouthmoroccounpolandglobalwarmingnojusticenodealdohasdgsindiaadp2warsawgccunited1o5cd12forestsethiopiakpauspolcoy8latviacop18gocop21sustainabilitycop18volunteerscdnpolirenewableclimatefinanceipccqatarfoundationaction2015climateactiondtff12coalclimate2014greenpartofthesolutionindcroadtodohaactonclimateearthdayclimatechangefitncop21climatecleanenergydemdebatenhpolitics50by30gotvbigblockofcheesedayparisagreementcop21globalthinklandscapeearthhourearthislamclimatechangepariscoy8ngosocialmediaopportunitiesdayenvironmentrecyclewalessavemenaawarenessdofinancefoodoceansearthdaytreesarabgeorgiacommunityworldbankworldinvestorqgbcunfcccdubaiajclimategreenbuildingqfsustainableeconomydcef2013myqatarinstagrampeopleenergyroad2dohaestedamaunfaowastemanagementsportagricultureqnfspqatarbestsustainabledevelopmentqatarpost2015investingentrepreneursmiddleeastresearchglfcop21foghealthnhpoliticsinnovationelectronicsspecialistjobsar5businessnatureecoagreementwaterjobchinaqfrecruitmentnewscitieschangeenergyefficiencyprotectbiodiversitylifestylehealthyomanfloridaqsciencerecognizedplaceactionsb42foodsecuritypollutionlovecoloradogarbageegyptrecyclingafricatarsheedfeaturedweathercarbonfsdlconferenceplasticconservationcaliforniacop18indianawasterenewablesgogreenactonclimatelifeclimatelegacyparis2015worldenvironmentdaywildlifeestedamaqatarenvironmentaljordansolarcop19qeeriinyteftobama50by30animalsscienceworktreeunusyouthmoroccofutureemissionsukhotnewsglobalwarmingbeautifuldohatalkdirtyqaindiacleantechadp2thepearlqataruaewed2015bagsgcceducationclimatepeacegreenillinoisasmfecotipeventscsrdevelopmentoilmangrovessustainabilitycdnpoliwarmingbuildingipccqatarfoundationsummitministry_of_environmentpartofthesolutionindcinstagramqatarwindgivesyouwingssummerdohariyadhsandstormsdrivesafelyinterstellardesertlivingwindystreetwestbaydohasandstormhellmoi_qatardohanewsuaealsaddsandyqatarduststormgulfzchizzerzsandmiddleeasthothealthiphonestormalkhorksadust_stormbahrainqatarlivinghealthy_tipsfridayabudhabigccrainzerovisibilityaspireqatarsandstormdustnewsweekenddesertipaddubaisamsungdustycarinstadohadrivemorningsandstormbadweathersaudiseemydohawindviewdohacop21energyindiabiomassdcef2013qatarco2climatechangeinvestingcdnpolifrackingmiddleeastresearchenvironmentconservationrenewablecop18waleswasterenewablesmenaeuropeoilpowerfoodwatersustainabilitysolarcarboninyteftinvestorchangearabclimategreenemissionsukqfsusdevdohaqfcop21energythinklandscapefeaturednewsuaetransportqatarclimatechangeinvestingenvironmentbiogashealthcop18waleswasteeconomyoilfinancebusinessnatureparis2015stealthrangerdubaiwatersustainabilityadvancedglobaltradingchinasolarcop19carboncitiesinvestorchangeunfcccwbgyouthsummitclimateioniphone5greenukhotnewsagtpollutionecologydohacop21energybiomassislamqatarclimatechangemiddleeastenvironmentconservationcop18saverenewablescsrnatureecowildlifetreeswatersustainabilitysolarcarbonarabclimatebiodiversitytreegreenqfwaterqatardohaenergyoceansclimatemiddleeastenvironmentsustainabilitygreenseasummerclimateclimatechangewinterdohanaturecop21arcticglobalqatarairenvironmentsustainabilityweathergreenwarmingchange the conference, suggesting that organizing big events is not
enough to raise any lasting public awareness toward climate
change. The backbone graph analysis on the other side, has
shown that hashtag co-occurrence is not random and leads
in most of the cases to meaningful connected communities
(clusters of hashtags).
The findings of this paper did not just highlight interesting
answers on public's opinion about climate change in Qatar,
but also raise questions for future research to explore. First,
we would like to alleviate the limitation of keyword based
filtering by designing a better classifier that accounts for se-
mantics of tweets. Second, we aspire to complement this
work with a longitudinal analysis of causality between dif-
ferent climate change topics using transfer of entropy tech-
nique. Finally, we would like to extend this study to other
countries in the region and examine the geographical influ-
ences on public's opinion about climate change.
References
[1] Alhorr, Y.; Eliskandarani, E.; and Elsarrag, E. 2014. Ap-
proaches to reducing carbon dioxide emissions in the built envi-
ronment: Low carbon cities. International Journal of Sustain-
able Built Environment 3(2):167 – 178.
[2] An, X.; Ganguly, A. R.; Fang, Y.; Scyphers, S. B.; Hunter,
A. M.; and Dy, J. G. 2014. Tracking climate change opin-
ions from twitter data. In Workshop on Data Science for Social
Good.
[3] Authors. 2016. Qatar Environment Taxonomy. http://
bit.ly/1Qms5eM. [Online; accessed 15-Feb-2016].
[4] Cody, E. M.; Reagan, A. J.; Mitchell, L.; Dodds, P. S.; and
Danforth, C. M. 2015. Climate change sentiment on twitter: An
unsolicited public opinion poll. PloS one 10(8):e0136092.
[5] Hart, P. S., and Feldman, L. 2014. Threat without efficacy?
climate change on us network news. Science Communication
36(3):325–351.
[6] Jang, S. M., and Hart, P. S. 2015. Polarized frames on cli-
mate change and global warming across countries and states:
Evidence from twitter big data. Global Environmental Change
32:11–17.
[7] Kirilenko, A. P., and Stepchenkova, S. O. 2014. Public mi-
croblogging on climate change: One year of twitter worldwide.
Global Environmental Change 26:171 – 182.
[8] Nielsen, R. C. 2016. RE:UN global pulse taxonomy. Message
to author. 22 Feb, 2016. [Email].
[9] Olteanu, A.; Castillo, C.; Diakopoulos, N.; and Aberer, K.
2015. Comparing events coverage in online news and social
media: The case of climate change. In Proceedings of the Ninth
International AAAI Conference on Web and Social Media, num-
ber EPFL-CONF-211214.
[10] QNB.
2015.
Qatar National Bank, Qatar Eco-
http://www.qnb.com/cs/
(p 12).
nomic Insight
Satellite?blobcol=urldata&blobheader=
application%2Fpdf&blobkey=
id&blobtable=MungoBlobs&blobwhere=
1355526969220&ssbinary=true.
15-Feb-2016].
[Online; accessed
[11] Radcliffe, D.
2015.
Middle East:
//damianradcliffe.files.wordpress.
2014.
of
The
Story
Social Media
the
in
https:
Figure 7: Matrix of pairwise correlations between different
climate change and weather variables. The highest p-value
score observed in this matrix is 4.64e− 07 which means that
all reported correlations are statistically significant.
as politics is tightly related to the two aforementioned topics.
Overall, one could easily see that the topics General, Negoti-
ations&Summits, Politics&Opinion, Economy, and Energy
constitute a well correlated group. Expending our analysis to
include the weather variables, we find that Sandstorm topic
showed some significant positive dependencies with mean
wind speed (ρ = 0.21) and a negative correlation with visi-
bility (ρ = −0.14). Thus, the lower the visibility, the more
likely a sandstorm is going on, which means more people are
tweeting about it. The other low yet statistically significant
correlation is observed between Negotiations&Summits and
mean temperature (ρ = −0.1) which is simply explained
by the fact that the most important summits captured in our
collection (COP18, COP21) took place in winter time.
Conclusions and Future Work
In our study, we start by monitoring, collecting, filtering, and
analyzing a large dataset of 109 millions tweets posted by
98K distinct users, in a course of 9 years in order to capture
the public interest in climate change. We use a taxonomy of
climate change to classify a significant amount of relevant
tweets into different topics covering a wide range of issues
such as politics, economy, and air quality. We perform dif-
ferent types of analysis to understand and reveal the timely
distribution of public interest toward different topics.
In our examination of the relevant sub set of tweets, the
results revealed that people's interest is mainly driven by
widely covered events (e.g., Paris COP21), or local events
that have a direct impact on users' daily life (e.g., sand-
storms, Doha COP18). The number of users who engage
in climate change discourse does not seem to be on a in-
creasing trend. Our analysis revealed that the burst in users'
activity observed during COP18 has quickly fade out after
generalrisk_disasterpolitics_opinioneconomynegotiations_summitsweatherenergyagriculture_forestryocean_watercampaignsarcticair_qualitysand_stormmeanvismprecipmhumiditymeantempmmeanwindspdmmeanwindspdmmeantempmhumidityprecipmmeanvismsand_stormair_qualityarcticcampaignsocean_wateragriculture_forestryenergyweathernegotiations_summitseconomypolitics_opinionrisk_disastergeneral 0.052 0.029 0.00026 0.0093 0.0140.00250.00080.03 0.018 0.066 0.0230.00140.21 0.0620.031 0.380.0191 0.022 0.068 0.079 0.025 0.1 0.008 0.0780.0110.0230.0440.04 0.0310.0087 0.16 0.079 0.5310.0190.0230.0340.0550.0180.0830.0360.046 0.047 0.02 0.0210.0110.024 0.087 0.0830.111 0.53 0.380.0690.080.0690.0190.0190.0590.0460.011 0.013 0.012 0.0160.043 0.0015 0.01910.11 0.0790.0310.0820.0540.0550.0410.00550.0180.0910.00250.0520.059 0.00330.036 0.141 0.019 0.083 0.16 0.0620.017 0.0130.0240.0056 0.0160.00740.0180.0290.0032 0.0083 0.00580.00941 0.14 0.0015 0.0870.00870.210.0450.00490.120.069 0.00480.050.040.0440.0590.0480.01310.00940.0360.0430.024 0.0310.00140.110.0520.110.120.0940.0510.0840.0420.0640.1410.013 0.0058 0.0033 0.0160.0110.04 0.0230.330.240.320.310.240.0930.330.180.06810.140.048 0.00830.059 0.012 0.0210.044 0.0660.140.0760.210.220.130.0910.230.1410.0680.0640.0590.00320.052 0.013 0.020.023 0.0180.210.190.390.40.20.0950.3210.140.180.0420.0440.0290.00250.011 0.0470.0110.030.380.270.610.570.490.1610.320.230.330.0840.040.0180.0910.0460.046 0.0780.00080.140.110.360.160.1310.160.0950.0910.0930.0510.050.00740.0180.0590.036 0.0080.00250.560.440.730.5210.130.490.20.130.240.094 0.0048 0.0160.00550.0190.083 0.1 0.0140.430.320.6210.520.160.570.40.220.310.120.0690.00560.0410.0190.018 0.025 0.00930.540.4510.620.730.360.610.390.210.320.110.120.0240.0550.0690.055 0.079 0.000260.5610.450.320.440.110.270.190.0760.240.0520.0049 0.0130.0540.080.034 0.068 0.02910.560.540.430.560.140.380.210.140.330.110.0450.0170.0820.0690.023 0.022 0.0520.60.30.00.30.6 com/2015/01/damian-radcliffe-social-\
media-in-the-middle-east-the-story-of-2014.
pdf. [Online; accessed 15-Aug-2016].
[12] Serrano, M. ´A.; Bogun´a, M.; and Vespignani, A. 2009.
Extracting the multiscale backbone of complex weighted net-
works. National academy of sciences 106(16):6483–6488.
[13] Snoj, J. 2013. Qatar's Population - by Nationality. BQ Mag-
https://www.bq-magazine.com/economy/
azine.
2013/12/population-qatar. [Online; accessed 15-Feb-
2016].
[14] Underground, W.
2016. History API. http://www.
wunderground.com/weather/api/. [Online; accessed
15-Feb-2016].
[15] UNFCCC. 2012. United Nations Framework Convention
on Climate Change, 2012. http://unfccc.int/2860.
php#decisions. [Online; accessed 15-Feb-2016].
[16] 2014. United Nations Global Pulse Taxonomy for climate
change. unglobalpulse.net/climate/taxonomy/.
[17] Veltri, G. A., and Atanasova, D. 2015. Climate change on
twitter: Content, media ecology and information sharing be-
haviour. Public Understanding of Science 0963662515613702.
[18] Wang, S.; Paul, M. J.; and Dredze, M. 2015. Social media as
a sensor of air quality and public response in china. Journal of
medical Internet research 17(3).
[19] WB. 2015. CO2 emissions (metric tons per capita). http:
//data.worldbank.org/indicator/EN.ATM.
CO2E.PC?order=wbapi_data_value_2011+wbapi_
data_value+wbapi_data_value-first&sort=
desc. [Online; accessed 15-Feb-2016].
[20] Weber, E. U., and Stern, P. C. 2011. Public understanding
of climate change in the united states. American Psychologist
66(4):315.
|
1610.06703 | 1 | 1610 | 2016-10-21T08:47:38 | Revisiting Link Prediction: Evolving Models and Real Data Findings | [
"cs.SI",
"physics.soc-ph"
] | The explosive growth of Web 2.0, which was characterized by the creation of online social networks, has reignited the study of factors that could help us understand the growth and dynamism of these networks. Various generative network models have been proposed, including the Barabasi-Albert and Watts-Strogatz models. In this study, we revisit the problem from a perspective that seeks to compare results obtained from these generative models with those from real networks. To this end, we consider the dating network Skout Inc. An analysis is performed on the topological characteristics of the network that could explain the creation of new network links. Afterwards, the results are contrasted with those obtained from the Barabasi-Albert and Watts-Strogatz generative models. We conclude that a key factor that could explain the creation of links originates in its cluster structure, where link recommendations are more precise in Watts-Strogatz segmented networks than in Barabasi-Albert hierarchical networks. This result reinforces the need to establish more and better network segmentation algorithms that are capable of clustering large networks precisely and efficiently. | cs.SI | cs |
Revisiting Link Prediction: Evolving Models and Real
Data Findings
Marcelo Mendozaa,∗, Mat´ıas Estradab
aUniversidad T´ecnica Federico Santa Mar´ıa, Santiago, Chile
bSkout Inc., Santiago, Chile
Abstract
The explosive growth of Web 2.0, which was characterized by the creation of
online social networks, has reignited the study of factors that could help us
understand the growth and dynamism of these networks. Various generative
network models have been proposed, including the Barab´asi-Albert and Watts-
Strogatz models. In this study, we revisit the problem from a perspective that
seeks to compare results obtained from these generative models with those from
real networks. To this end, we consider the dating network Skout Inc. An
analysis is performed on the topological characteristics of the network that
could explain the creation of new network links. Afterwards, the results are
contrasted with those obtained from the Barab´asi-Albert and Watts-Strogatz
generative models. We conclude that a key factor that could explain the cre-
ation of links originates in its cluster structure, where link recommendations are
more precise in Watts-Strogatz segmented networks than in Barab´asi-Albert hi-
erarchical networks. This result reinforces the need to establish more and better
network segmentation algorithms that are capable of clustering large networks
precisely and efficiently.
Keywords: Link Prediction, Network Evolving Models, Discrete Dynamics in
Networks
∗Corresponding author
Email addresses: [email protected] (Marcelo Mendoza),
[email protected] (Mat´ıas Estrada)
Preprint submitted to ArXiV
June 17, 2021
1. Introduction
The explosive growth of Web 2.0, which is characterized by its consolidation
of online social networks, has brought millions of users into participating in
these networks, sharing their life experiences and establishing new relationships
through these platforms. The relationships created in these networks can be
modeled as large graphs. Their evolution and dynamism represent an interesting
challenge for the complex network community; this community is dedicated
to proposing and discussing the significance of generative models that could
precisely explain the growth of these networks.
The most widely discussed generative models are those developed by Barab´asi-
Albert [2] and Watts-Strogatz [23]; each of these models has their own strengths
and weaknesses in their ability to explain topological characteristics of real net-
works. Specifically, the Watts-Strogatz model is able to model large but small-
diameter networks, which are properties that can be effectively observed in real
networks, whereas the Barab´asi-Albert networks follow an evolving model of
preferential attachment with a tendency to generate networks with few vertices
that contain a high number of links, where the majority of vertices have low con-
nectivity. This approach has been used as a very effective model for biological
networks [17].
A growing interest in establishing better methods for link predictions is using
a microscale approximation of the network evolution problem. This approxima-
tion has driven investigations by the complex network community in their efforts
to not only find models that effectively fit the dynamic laws of social networks
but can also provide the possibility of link recommendation, which is also known
as link prediction and is particular useful in online social networks.
In this study, we perform analyses on link prediction methods based on the
topological characteristics of real data from the dating network Skout. Subse-
quently, we contrast the results obtained with those observed from the Barab´asi-
Albert and Watts-Strogatz generative models. A number of experimental config-
urations were evaluated, including intra-cluster recommendations in segmented
2
networks. The contribution of this study is the evaluation of the effectiveness of
these models in light of the results observed from data of a real network. The
problem is analyzed from a purely topological standpoint, which contrasts the
classical standpoint of FOAF (friend-of-a-friend) recommendations with those
based on segmented networks.
This article is organized as follows. In Section 2 we discuss related work.
Background is summarized in Section 3. The case study is presented in Sec-
tion 4. Results from synthetic data are presented in Section 5 and discussed in
Section 6. Finally, we conclude in Section 7 with a brief discussion about some
open questions and future work.
2. Related Work
Link prediction is mainly addressed using transitivity properties of the graph
also known as friend-of-a-friend approaches [20]. The rationale of this kind of
approaches is the following: If a user A is friend of a user B and a user B is friend
of a user C, then A and C probably will be friends [9]. However, the success of
link prediction approaches is limited because real online social networks tend to
be very sparse [4], meaning that the total amount of potential links to be created
is much greater than the links that are actually created. Sparsity turns the link
prediction problem into a very dificult problem returning success improvements
over a random predictor from 3% to 54% [10].
A number of different types of features are available for this problem [11].
By exploiting locality we do not require the full graph from being stored. One of
the simplest locality features is the Common Neighboors index [16]. This index
considers the amount of common neighboors that two nodes have. When the
number of common neighboors is high, then is more likely that those nodes will
create a link in the future [9]. However, this feature is biased by the number of
neighboors that a given node registers. A more robust feature is the Jaccard in-
dex [5] which compares the common neighboors cardinality with the cardinality
of the union of both neighboors, getting a value that represents the proportion
3
between the cardinality of these two sets. Other features used are the Salton
index [19] and the Sorensen index [21] that are mainly used in ecological net-
works [12]. Finally, another well known locality measure is the Adamic & Adar
index [1].
Graph-based features considers the whole network to be calculated. Sim-
Rank index [6] is calculated using a random walk process which is propagated
through the graph with a decay factor. It is an expensive feature because it
needs several passes through the whole network. Then, unexpensive and con-
ventional graph-based features as degree or transitivity are more recommended
for link prediction [12]. Finally, the combination of global/local scope is also
explored for this task, using approaches as HITS (Hypertext Induced Topic
Search) which comes from the information retrieval community, proposed by
Kleinberg [8] for web page ranking. An undirected version of HITS can be ex-
plored for link prediction purposes in undirected graphs, giving us an authority
feature for each node of the graph.
There are more information sources that are useful for link prediction tasks.
For instance, by recovering text from user messages, it is possible to create more
sophisticated user descriptors [18]. It is also possible to characterize neighboors
by recovering text and then using this information for user description [26].
Transactional registers has been also explored for link prediction [24]. These
methods tend to involve high computational costs due to the necessity of main-
tain content-based indexes or log-repositories. These methods are out of the
scope of this study.
Regarding network segmentation methods, there has been recent interest in
the exploration of spectral clustering methods for link prediction. [22] explored
the use of segmented networks for link prediction, proving in synthetic networks
that this approach is feasible. The study is close in aim to our article but it
does not consider an evaluation regarding evolving graph models. In addition
the comparison is constrained only to the top-1 recommended node.
Link prediction has been explored in several domains, showing that is still a
relevant task. For instance, different algorithms has been developed to predict
4
new links in protein networks [13], energy grid networks [15], medical co-author
networks [25], disease networks [7], among others. Further details about this
topic can be read in the survey of L´u [14].
3. Background
Link prediction is the problem of inferring whether potential edges between
pairs of vertices in a graph will be present or absent in the near future. In a
link prediction task, we first assume the existence of an original observed graph,
with a completely known set of vertices and a partially known set of edges.
Accordingly, the link prediction task is to infer the rest of the edges.
Let G = (V, E) be an undirected graph. At a given point in time t0, we
assume that all vertices v ∈ V are known but only a subset of E is known. Let
E obs be the subset of edges e ∈ E that is known at t0, and let Emiss be the
subset of unobserved edges at t0, such that Emiss = E \ E obs. Link prediction
is to predict the edges in Emiss from G = (V, E obs).
A natural way to address link prediction is to use a ranking-based strategy.
A ranking-based strategy considers two steps to rank links. For a given vertex
u, a vertex retrieval step is conducted, using a locality constraint to recover
a list of proximal vertices to u. Then, the list is sorted according to a given
scoring measure.
A number of graph-based measures can be used for scoring. We explore the
perfomance of three of them: Normalized degree, authority, and transitivity.
These features are query independent, that is to say, these measures define a
collection of pointwise estimations at vertex level. The value of a measure of this
collection is the same for the whole graph. The definitions of these measures
are given below.
Degree coefficient. Let Γ(u) be the neighborhood of a vertex u and let Γ(u)
be the cardinality of this set, also known as the degree of u. We define a degree
coefficient by normalizing Γ(u) with the maximum degree of G. Formally:
5
Degree(u) =
Γ(u)
Maxv∈G Γ(v)
.
Transitivity coefficient. Let Γ(u) be the neighborhood of a vertex u. The tran-
sitivity coefficient Transitivity(u) (a.k.a. clustering coefficient) is the ratio
between the number of links in Γ(u) and the maximum number of such links.
If Γ(u) has eu links, we have:
Transitivity(u) =
eu
Γ(u)·(Γ(u)−1)
2
.
Authority. HITS (Hypertext Induced Topic Search) coefficients come from the
information retrieval community and were proposed by [8] for web page rank-
ing. The idea is that pages that have many links pointing to them are called
authorities and pages that have many outgoing links are called hubs. Good hubs
point to good authority pages, and vice-versa. Lets hub(u) and auth(u) be hub
and authority coefficients for a vertex u. The following equations can be solved
through an iterative algorithm that addresses the fixed point problem defined
by:
hub(u) = X
auth(v),
v∈Gu→v
auth(u) = X
hub(v).
v∈Gv→u
In undirected graphs, both coefficients have the same value. In this study,
we denote this feature as authority, given that a dating network is an undirected
graph.
4. Link prediction in a real network: the Skout Inc. case
We explore properties of the problem of link prediction in a real network
provided directly by Skout Inc.. After recording each link created until January
1st 2014, we generated a network with 3,855,389 links among 1,920,015 active
6
users. We define an active user as one that, after creating their account, has
added at least one friend during the period.
Afterwards, we recorded links created during the first 25 days of January
2014 between active users. A total of 582,119 new links were recorded solely
from users who were active during 2013. A total of 76,848 users added at least
one new friend during the observation period. The distribution of new friends
per user is shown in Figure 1.
s
r
e
s
U
f
o
r
e
b
m
u
N
100000
10000
1000
100
10
1
1
New Friends Distribution
New Friends Bins
10
100
1000
New Friends
Figure 1: Number of links created per active user during the observation period.
As shown in Figure 1, the distributions of new friends per user follow the law
of the rich get richer: only a few users added a high number of new friends; the
majority of users added only a few new friends, confirming that link prediction
is a very difficult task due to the imbalance between potential friends and actual
friends.
The collection of links that could have potentially been created during this
period corresponds to the group of links that was not created until January
1st 2014 among active users. Then, we analyzed the properties of separability
that exist between the group of potential links created versus the group of not-
created links; the logic of the methodology used in this study is explained in
the scheme shown in Figure 2.
7
1/1
t
0
1/25
t
1
Links at t
0
Links created
between t
0
and t
1
Potential links
at t
1
Nodes at t0
Figure 2: Case study methodology. Blue links and nodes depict the graph at
t0. Red links depict links created during the observation period. Links depicted
with dashed lines indicates potential unobserved links. In our strategy, dashed
links are labeled as false link instances and red links as real link instances.
Each link in the group of potential links is classified as either real (created)
or not (not created), where the authority, degree, and transitivity scores for
both the target user and the candidate user are calculated. In this experiment,
the target user is the user who is looking for a new friend, and the candidate
user is the potential new friend that could be recommended to the target user.
Table 1 shows the results of information gain from separating the two classes.
As shown in Table 1, the most relevant characteristic for this problem is
the authority of the candidate user. Because the first vertex corresponds to the
target user (the one who accepts or declines the new relationship), the authority
of the candidate user is a measure of the visibility (and therefore popularity) of
the user for the rest of the network. However, the characteristics of the target
user are not extremely relevant, which indicates that the present connectivity of
the user is not necessarily an indicator of how connected the user will be in the
8
Table 1: Information Gain values for each feature considered in our data set.
Target users features are depicted with a subindex equals to one. Candidate
users are depicted with a subindex equals to two.
Feature
Authority 2
Authority 1
Transitivity 2
Degree 2
Transitivity 1
Degree 1
Information Gain
0.9155
0.5367
0.1192
0.1054
0.0392
0.0227
future. We should note that an algorithm that predicts links should be capable
of making recommendations to newly target users in the network, i.e., users
with few neighbors and a short user history; this challenge is also referred to as
a cold start. Because of this, the characteristics of the target user are not as
relevant as those of the candidate user. In this strategy, we always consider that
it is the user with fewer friends who gets the most recommendations to equalize
the size of the network neighborhoods. This focus is paralleled with information
retrieval, where the query (the target user) is considered as a document with
few descriptors, and therefore, the descriptors of the candidate documents are
primarily exploited on being recommended in the list of answers.
5. Results from synthetic data
Networks with 10,000 vertices were generated based on three generative link
models: random-based graphs models, also known as Erd´os-R´enyi graphs [3],
Watts-Strogatz [23], with a rewired probability equal to 0.1, and Barab´asi-
Albert [2]. For each of these models, we performed a pruning process according
to a factor taken as an experimental parameter that could control the ratio
9
of links eliminated to all the links in the network. Later, link prediction was
performed on the eliminated links, where a recommended and eliminated link
was considered a success, and a recommended but kept (not-eliminated) link is
considered a failure. In this way, the pruning factor controls the ratio of real
solutions to the total candidates. A low pruning factor, which corresponds to
a smaller ratio of links to be predicted over the total candidates, makes the
problem more difficult.
Each of the networks was segmented with a directed k-clustering algorithm
(k-means for list of arcs), with k values ranging between 3 and 7. Then, using
one of the three candidate ranking functions performed on the Skout network
(authority, degree, and transitivity), an intra-cluster ranking process was per-
formed. We were thus able to evaluate the impact of network segmentation on
the quality of the recommendations.
Tables 2 and 3 show the average precision and recall results for the first
ten recommendations, which are micro-averages, i.e., the results correspond
to averages for all the nodes in a network. Table 2 shows the results for a
pruning factor of 0.1, and Table 3 shows the results for a pruning factor of 0.25.
More configurations were examined in these experiments, which incorporated a
greater number of small networks and different pruning factors. The P@N and
R@N curves, with N values ranging between 1 and 10 for all configurations, can
be seen in greater detail for some small networks (up to 1,000 nodes) at the
following url [http://104.236.107.4:17264] 1.
6. Discussion
In this study, we addressed the problem of link prediction, which segments
networks to later analyze each resulting partition in terms of degree, authority,
and transitivity. Several of the results from the experiments using model data
demonstrated certain relevant characteristics. One characteristic is related to
1We will replace the ip number by a .org url in the camera ready version
10
1
1
k
Erd´os-R´enyi
Barab´asi-Albert Watts-Strogatz
Erd´os-R´enyi
Barab´asi-Albert Watts-Strogatz
y
t
i
r
o
h
t
u
A
e
e
r
g
e
D
y
t
i
v
i
t
i
s
n
a
r
T
3
4
5
6
7
3
4
5
6
7
3
4
5
6
7
P@10
0.0126
0.0121
0.0121
0.0110
0.0104
0.0110
0.0110
0.0110
0.0112
0.0112
0.0050
0.0120
0.0102
0.0126
0.0114
P@10
0.0604
0.0462
0.0467
0.0328
0.0424
0.0612
0.0464
0.0465
0.0326
0.0412
0.0406
0.0412
0.0405
0.0333
0.0384
P@10
0.0516
0.0798
0.1114
0.1310
0.1552
0.0312
0.0513
0.0400
0.0746
0.1066
0.0236
0.0311
0.0400
0.0441
0.0472
R@10
0.0001
0.0005
0.0005
0.0005
0.0008
0.0001
0.0004
0.0005
0.0005
0.0008
0.0001
0.0005
0.0005
0.0005
0.0008
R@10
0.0022
0.0015
0.0010
0.0007
0.0010
0.0026
0.0012
0.0010
0.0007
0.0010
0.0026
0.0014
0.0010
0.0007
0.0010
R@10
0.0081
0.0169
0.0174
0.0232
0.0298
0.0072
0.0143
0.0143
0.0212
0.0268
0.0074
0.0143
0.0122
0.0204
0.0212
Table 2. Precision and recall results with pruning rate = 0.10.
1
2
k
Erd´os-R´enyi
Barab´asi-Albert Watts-Strogatz
Erd´os-R´enyi
Barab´asi-Albert Watts-Strogatz
y
t
i
r
o
h
t
u
A
e
e
r
g
e
D
y
t
i
v
i
t
i
s
n
a
r
T
3
4
5
6
7
3
4
5
6
7
3
4
5
6
7
P@10
0.0272
0.0241
0.0280
0.0282
0.0276
0.0264
0.0224
0.0264
0.0270
0.0272
0.0244
0.0212
0.0270
0.0292
0.0264
P@10
0.0712
0.0688
0.0612
0.0575
0.0510
0.0791
0.0689
0.0616
0.0579
0.0520
0.0512
0.0522
0.0614
0.0507
0.0472
P@10
0.0988
0.1548
0.2028
0.2616
0.2901
0.0744
0.1135
0.1486
0.2235
0.2735
0.0485
0.0603
0.0850
0.1015
0.1177
R@10
0.0018
0.0021
0.0022
0.0014
0.0017
0.0018
0.0020
0.0022
0.0014
0.0017
0.0016
0.0022
0.0022
0.0014
0.0017
R@10
0.0033
0.0025
0.0019
0.0018
0.0013
0.0033
0.0024
0.0018
0.0018
0.0013
0.0033
0.0026
0.0019
0.0018
0.0013
R@10
0.0144
0.0254
0.0255
0.0339
0.0399
0.0122
0.0178
0.0231
0.0336
0.0377
0.0102
0.0124
0.0146
0.0147
0.0141
Table 3. Precision and recall results with pruning rate = 0.25.
the link pruning parameter because it is also related to the network evolution
time. That is, a low pruning factor emulates a short period of observation and
thus, a more difficult problem because worse results would be obtained compared
with that of higher pruning factors. This result is due to predictability improv-
ing as the limits of observation increase for prediction problems. Specifically for
our experiments, increasing the pruning ratio raises the number of possible links
that are candidates for selection; each of the candidate links added corresponds
to occurrences of links already existing in the original network, and thus, it is
expected that the results would improve when they are selected.
The results demonstrate that the process of segmenting the networks is key
to link prediction because dissimilar results are obtained from different segmen-
tation routines. Such is the case with the Watts-Strogatz generative model,
which achieved a precision of approximately 0.1 with three clusters and approx-
imately 0.3 with seven clusters. The optimum number of clusters depends on the
nature of the network, the number of vertices, and the density of the network,
among other factors.
Most of the methods of link prediction use locality from a FOAF (friend-of-
a-friend) standpoint. However, the results obtained in this research based on
transitivity exhibited worse results than those based on authority and degree. It
is possible that when using locality, transitivity restricts the predictions to the
graph of nearest neighbors, whereas clustering uses global network characteris-
tics, which are only partially approximated by a FOAF standpoint. Barab´asi-
Albert-type networks give worse results from segmentation, where prediction
results worsen as the number of clusters increases. Barab´asi-Albert networks
are essentially hierarchical, and a flat clustering segmentation method is there-
fore not extremely suitable for this type of structure.
The measures used to characterize and select candidates that behave best
depend on the network type.
In Barab´asi-Albert-type networks, predictions
based on degree are slightly better than those based on authority. The opposite
is observed in Watts-Strogatz-type networks; predictions based on authority are
somewhat better than those based on degree. In both network types, the worst
13
precision is achieved using transitivity.
When comparing these results to those obtained from Skout Inc., we should
remember that in the latter case, the best properties of prediction were shown
in the authority, degree, and transitivity scores (in that order). These results
are similar to those obtained from a Watts-Strogatz-type generative model,
which suggests that this model could approximate the structure of the Skout
network, provide a means to perform link prediction on a segmented network,
and complement FOAF strategies with ranking strategies based on intra-cluster
authority.
7. Conclusions
This study revisited the problem of link prediction from a perspective that
aims to unify results obtained from real networks and those from generative
models. We conclude that the recommendations provided from FOAF are over-
valued because nearest-neighbor structures are unable to explain the growth
and dynamism of a network on their own. Instead, the use of network segmen-
tation methods can explain the creation of links between users that might not
be among the nearest-neighbors, which highlights how the dynamism and evolu-
tion of a network is explained via a combination of local (nearest neighbor) and
global (clusters) factors. These results emphasize the need for more and better
network segmentation algorithms capable of working on large data collections.
There is also a need to consolidate link prediction models capable of effectively
combining local and global factors.
Acknowledgment
Marcelo Mendoza was supported by project FONDECYT 11121435 and by
FB0821 Centro Cient´ıfico Tecnol´ogico de Valpara´ıso.
14
References
[1] Adamic, L. A., Adar, E. (2003) Friends and neighbors on the Web, Social
Networks, 25(3), 211-230.
[2] Barab´asi, A. L., Albert, R. (1999) Emergence of scaling in random
networks. Science, 286, 509-512.
[3] Erd´os, P., R´enyi, A. (1960) On the evolution of random graphs, Publica-
tions of the Mathematical Institute of the Hungarian Academy of Sciences,
5, 17-61.
[4] Genio, C. I. Del, Gross, T., Bassler, K. E. (2011) All scale-free networks
are sparse, Physical Review Letters, 107, 178701, 1-4.
[5] Jaccard, P. (1912) The distribution of the flora in the alphine zone, The
New Phytologist, XI, 37-50.
[6] Jeh, G., Widom, J. (2002) SimRank: a measure of structural-context
similarity, In Proceedings of the ACM International Conference on Infor-
mation and Knowledge Management, KDD, 1-11.
[7] Kaya, B., Poyraz, M. (2014) Supervised link prediction in symptom
networks with evolving case, Measurement, 56, 231-238.
[8] Kleinberg, J. (1998) Authoritative Sources in a Hyperlinked Environment,
in Proceedings of the Symposium on Discrete Algorithms, SODA, 668-677.
[9] Kossinets, G. (2006) Effects of missing data in social networks, Social Net-
works, 28(3), 247-268.
[10] Liben-Nowell, D., Kleinberg, J. (2003) The link prediction problem for
social networks, In Proceedings of the ACM Conference on Information
Knowledge and Management, CIKM, 556-559.
[11] Lin, D. (1998) An Information-Theoretic Definition of Similarity, in Pro-
ceedings of the International Conference on Machine Learning, ICML, 296-
304.
15
[12] Linyuan, L. (2011) Link Prediction in Complex Networks : A Survey. Phys-
ica A: Statistical Mechanics and its Applications, 390(6), 1150-1170.
[13] L´u, L., Jin, C.-H., Zhou, T. (2009) Similarity index based on local paths
for link prediction of complex networks, Physical Review. E, Statistical,
Nonlinear, and Soft Matter Physics, 80, 046122.
[14] L´u, L., Zhou, T. (2011) Link prediction in complex networks: A survey,
Physica A: Statistical Mechanics and its Applications, 390(6), 1150-1170.
[15] Murata, T., Moriyasu, S. (2008) Link Prediction based on Structural
Properties of Online Social Networks, New Generation Computing, 26(3),
245-257.
[16] Newman, M. E. (2001) Clustering and preferential attachment in growing
networks, Physical Review. E, Statistical, Nonlinear, and Soft Matter
Physics, 64, 025102.
[17] Ravasz, E., Somera, A. L., Mongru, D. A., Oltvai, Z. N., Barab´asi, A.
L. (2002) Hierarchical organization of modularity in metabolic networks.
Science, 297, 1551-1555.
[18] Roth, M., Flysher, G., Matias, Y., Leichtberg, A., Merom, R. (2010) Sug-
gesting Friends Using the Implicit Social Graph Categories and Subject De-
scriptors. In Proceedings of the ACM International Conference on Knowl-
edge Discovery and Data Mining, KDD, 233-242.
[19] Salton, G., McGill, M. J. (1983) Introduction to Modern Information
Retrieval, New York (Vol. 22, p. xv, 448 p.).
[20] Silva, N. B., Tsang, I.-R., Cavalcanti, G. D. C., Tsang, I.-J. (2010) A
graph-based friend recommendation system using Genetic Algorithm, In
Proceedings of the IEEE Congress on Evolutionary Computation, 1-7.
[21] Sorensen, T. (1948) A method of establishing groups of equal amplitude in
plant sociology based on similarity of species content and its application to
16
analyses of the vegetation on Danish commons, Kongelige Danske Viden-
skabernes Selskab, 5(4), 1-34.
[22] Symeonidis, P., Iakovidou, N., Mantas, N., Manolopoulos, Y. (2013) From
biological to social networks: Link prediction based on multi-way spectral
clustering, Data & Knowledge Engineering, 87, 226-242.
[23] Watts, D., Strogatz, S. (1998) Collective dynamics of
'small-world'
networks, Nature, 393, 6684, 440-442.
[24] Xiang, R., Neville, J., Rogati, M. (2010) Modeling relationship strength
in online social networks, In Proceedings of the ACM World Wide Web
Conference, WWW, 981-990.
[25] Yu Q., Long C., Lv Y., Shao H., He P. (2014) Predicting Co-Author
Relationship in Medical Co-Authorship Networks, PLoS ONE 9(7):
e101214.
[26] Zhou, T., L, L., Zhang, Y.-C. (2009) Predicting missing links via local
information. The European Physical Journal B, 71(4), 623-630.
17
|
1602.01107 | 1 | 1602 | 2016-02-02T21:01:04 | Do Cascades Recur? | [
"cs.SI",
"physics.soc-ph",
"stat.ML"
] | Cascades of information-sharing are a primary mechanism by which content reaches its audience on social media, and an active line of research has studied how such cascades, which form as content is reshared from person to person, develop and subside. In this paper, we perform a large-scale analysis of cascades on Facebook over significantly longer time scales, and find that a more complex picture emerges, in which many large cascades recur, exhibiting multiple bursts of popularity with periods of quiescence in between. We characterize recurrence by measuring the time elapsed between bursts, their overlap and proximity in the social network, and the diversity in the demographics of individuals participating in each peak. We discover that content virality, as revealed by its initial popularity, is a main driver of recurrence, with the availability of multiple copies of that content helping to spark new bursts. Still, beyond a certain popularity of content, the rate of recurrence drops as cascades start exhausting the population of interested individuals. We reproduce these observed patterns in a simple model of content recurrence simulated on a real social network. Using only characteristics of a cascade's initial burst, we demonstrate strong performance in predicting whether it will recur in the future. | cs.SI | cs | Do Cascades Recur?
Justin Cheng1, Lada A Adamic2, Jon Kleinberg3, Jure Leskovec4
1,4Stanford University, 2Facebook, 3Cornell University
1,4{jcccf, jure}@cs.stanford.edu, [email protected], [email protected]
6
1
0
2
b
e
F
2
]
I
S
.
s
c
[
1
v
7
0
1
1
0
.
2
0
6
1
:
v
i
X
r
a
ABSTRACT
Cascades of information-sharing are a primary mechanism by which
content reaches its audience on social media, and an active line of
research has studied how such cascades, which form as content is
reshared from person to person, develop and subside. In this paper,
we perform a large-scale analysis of cascades on Facebook over
significantly longer time scales, and find that a more complex pic-
ture emerges, in which many large cascades recur, exhibiting mul-
tiple bursts of popularity with periods of quiescence in between.
We characterize recurrence by measuring the time elapsed between
bursts, their overlap and proximity in the social network, and the
diversity in the demographics of individuals participating in each
peak. We discover that content virality, as revealed by its initial
popularity, is a main driver of recurrence, with the availability of
multiple copies of that content helping to spark new bursts. Still,
beyond a certain popularity of content, the rate of recurrence drops
as cascades start exhausting the population of interested individu-
als. We reproduce these observed patterns in a simple model of
content recurrence simulated on a real social network. Using only
characteristics of a cascade's initial burst, we demonstrate strong
performance in predicting whether it will recur in the future.
Keywords: Cascade prediction; content recurrence; information
diffusion; memes; virality.
1.
INTRODUCTION
In many online social networks, people share content in the form
of photos, videos, and links with one another. As others reshare
this content with their friends or followers in turn, cascades of re-
sharing can develop [14]. Substantial previous work has studied
the formation of such information cascades with the aim of charac-
terizing and predicting their growth [7, 23, 47]. Cascades tend to
be bursty, with a spike of activity occurring within a few days of
the content's introduction into the network [34, 37]. This property
forms the backdrop to a line of temporal analyses that focus on the
basic rising-and-falling pattern that characterizes the initial onset
of a cascade [2, 10, 36, 48].
However, the temporal patterns exhibited by cascades over sig-
nificantly longer time scales is largely unexplored. Do successful
cascades display a long monotonic decline after their initial peak,
.
Figure 1: An example of a image meme that has recurred, or resur-
faced in popularity multiple times, sometimes as a continuation of
the same copy, and sometimes as a new copy of the same meme (ex-
ample copies are shown as thumbnails). This recurrence appears as
multiple peaks in the plot of reshares as a function of time.
or do they exhibit more complex behavior in which they can re-
cur, experiencing renewed bursts of popularity long after their ini-
tial introduction? Anecdotally, many of us have experienced déjà
vu when a friend shared content we had seen weeks or months ago,
but it is not clear whether these are isolated occurrences or glimpses
into a robust phenomenon. Resolving these basic distinctions in the
long-time-scale behavior of cascades is crucial to understanding the
longevity of content beyond its initial popularity, and points toward
a more holistic view of how content spreads in a network.
The present work: Cascade recurrence. We perform a year-
long large-scale analysis of cascades of public content on Face-
book, measuring them over significantly longer time scales than
previously investigated. Our first main finding is that recurrence
is widespread in the temporal dynamics of large cascades. Among
large cascades appearing in 2014, over half come back in one or
more subsequent bursts. While reshare activity does peak and then
drop to very low or even zero levels relatively soon after introduc-
tion, the same content can recur after a short or extended lull.
The prevalence of recurrence prompts several questions about
how and why content recurs. Is more broadly or narrowly appealing
content more likely to recur? Does a larger initial burst indicate a
greater likelihood of recurrence, or does it inhibit subsequent bursts
by exposing and thus satiating many people in the initial wave? Do
different bursts of the same content spread in different parts of the
network? Is the second burst a continuation of the initial cascade,
or a fresh re-introduction of the content into the network? Does the
media type of the reshared content matter - for example, whether
2.5K5K7.5K10KFebMarAprMayJunJulAugSepOctNovDec# ResharesHow to be skinny1. Notice that your body is coveredin skin2. Say "Wow I'm skinny"Congratulations you are now skinnyImage Meme (a)
(b)
(c)
(d)
Figure 2: (a) The diffusion cascade of the example meme from Figure 1 as it spreads over time, colored from red (early) to blue (late). Only
reshares that prompted subsequent reshares are shown. (b) The cascade is made up of separately introduced copies of the same content; in
this drawing of the cascade from (a), each copy is represented in a different color. (c) Sometimes, individual copies experience a resurgence
in popularity; again we draw the cascade from (a), but now highlight a single resurgent copy in red with the spread of all other copies depicted
in black. (d) A different network on the same set of users who took part in the cascade, showing friendship edges rather than reshare edges.
These edges span reshares across copies and time, showing that multiple copies of the meme are not well-separated in the friendship network.
it is a photo or a video? Finally, how well can one combine such
features to predict whether a piece of widely reshared content is
likely to experience additional bursts in popularity later on?
We motivate our discussion with an example of content recur-
rence. Figure 1 shows an image meme that first became popular
on Facebook at the end of February 2014, and it depicts how the
number of reshares of that meme changed over time. Here, while
an initial burst in resharing activity is followed by a gradual de-
crease, this meme recurred, experiencing multiple resurgences in
popularity - first in mid-March, then several times over the next
few months. Perhaps surprisingly, there is little to no resharing
between consecutive bursts. Additionally, multiple near-identical
copies of this image meme, represented in different colors, are
shared in the network. This distinction between different copies of
the same content will prove important in our later analyses: when
a user reshares content through the reshare mechanism provided by
the site, the content continues onward as the same copy; in contrast,
when a user reposts or re-uploads the same content and thus shares
it afresh, this is a new copy.
Figure 2 sketches the diffusion cascade of this meme, or its prop-
agation over edges in the social network. As shown in (a), bursts
in activity are connected through the same large long-lived cascade
and can be traced through the network, from the initial bursts in
March (shown in red), to the smaller bursts nearer the end of 2014
(shown in blue). In (b), where the same network is now colored ac-
cording to the copy of the image being reshared, different copies of
the same content appear at different times, sometimes correspond-
ing to when bursts occur, suggesting that recurrence sometimes oc-
curs from the introduction of new copies. However, recurrence may
also occur as a continuation of a previous copy:
the copy high-
lighted in red in (c) experiences an initial burst in March, but then
resurfaces in popularity later in the year. Further, we see in (d) that
friendship ties exist between even the earliest and latest reshares -
the meme appears to be diffusing rapidly, but also revisits parts of
the network through which it had earlier diffused.
While the meme in our example recurred several times, are such
memes the exception or the norm? And if such memes are in fact
typical, what are the bases for such robust patterns of recurrence?
To answer these questions, we use a dataset of reshare activity of
publicly viewable photos and videos on Facebook in 2014.
Characterizing recurrence. First, we develop a simple definition
of a burst, corresponding informally to a spike in the number of
reshares over time, that we can use to quantify when recurrence
occurs (via multiple observed bursts), and when it does not (a single
burst). We show that a significant volume of popularly reshared
content recurs (59% of image memes and 33% of videos), and that
recurring bursts tend to take place over a month apart from each
other. Recurrence is itself relatively bursty - rarely do we observe
long sustained periods of resharing.
Studying the temporal patterns of recurrence, user characteristics
of the resharing population, and the network structure of cascades,
we find that the recurrence of a piece of content is moderated to a
large extent by its virality, or broadness of appeal: cascades with
initial bursts of activity that are larger, last longer, and have a more
diverse population of resharers are more likely to recur. Nonethe-
less, it is not the cascades that start out the largest or most viral
that recur, but those that are moderately appealing. Specifically, a
moderate number of initial reshares, as well as a moderate amount
of homophily (or diversity) in the initial resharing population is
correlated with higher rates of recurrence. This lies in contrast to
more appealing (or popular) content, where one is likely to see a
single large outbreak which results in a large single burst, as well
as less appealing content, where one is likely to only see a sin-
gle small outbreak and thus a smaller single burst. In the former
case, we show evidence that a large initial burst inhibits subsequent
recurrence by effectively "immunizing" a large proportion of the
susceptible population.
While individual copies of content already recur in the network
(18% for image memes and 30% for videos), the presence of mul-
tiple copies catalyzes recurrence, allowing that content to spread
rapidly to different parts of the network, significantly boosting the
rate of recurrence. To a smaller extent, the principle of homophily,
suggesting that people are more likely to share content received
from users similar to themselves, also plays a role in recurrence,
with user similarity positively correlated with the rate of spreading.
Modeling recurrence. Motivated by the above picture of recur-
rence, and inspired by classic epidemiological models of diffusion
[39] and disease recurrence [3, 25, 40], we present a simple model
of cascading behavior that is primarily driven by content virality
and the availability of multiple copies, and is able to reproduce the
observed recurrence features. A simulation of this model, which
introduces multiple copies of the same content into the network,
can cause independent cascades that peak at different times and in
aggregate are observed as recurring. As the virality of the content
increases, the shape of a plot of overall reshares in the network over
time transforms from a shorter independent single burst, to multiple
bursts of differing sizes, to a single large burst of a longer duration.
Replicating our previous findings, increasing virality increases re-
currence, up to a point: once a meme has exposed a large part of
the network, further recurrence is inhibited.
Predicting recurrence. Finally, we show how temporal, network,
demographic, and multiple-copy features may be used to predict
whether a cascade will recur, if the recurrence will be smaller or
larger than the original burst, and when the recurrence occurs. We
demonstrate strong performance in predicting whether the same
content will recur after observing its initial burst of popularity (ROC
AUC=0.89 for image memes), as well as in predicting the relative
size of the resulting burst (0.78). The time of recurrence, on the
other hand, appears to be more unpredictable (0.58). Features re-
lating to content virality and multiple copies perform best. Though
multiple-copy features account for significant performance in pre-
dicting the recurrence of content, we obtain similarly strong perfor-
mance (0.88) when predicting the recurrence of an individual copy
of a piece of content.
Together, these results not only provide the first large-scale study
of content recurrence in social media, but also begin to suggest
some of the factors that underpin the process of recurrence.
2. TECHNICAL PRELIMINARIES
Studying cascade recurrence requires both sufficiently rich data
that accurately measures activity throughout a network over long
periods of time, as well as a robust definition of what recurrence is.
2.1 Dataset Description
In this paper, we use over a year of sharing data from Facebook.
All data was de-identified and analyzed in aggregate. Facebook
presents a particularly rich ecosystem of users and pages (entities
that can represent organizations or brands) sharing a large amount
of content over long periods of time.
Reliably measuring the spread of content in a network over time
is challenging because multiple copies of the same content may
exist at any time. As we will later show, the presence of multi-
ple copies in a cascade is an important catalyst for recurrence. On
Facebook, users and pages may introduce a new copy of the same
content by re-posting or re-uploading it; resharing an existing copy
instead creates an attribution back to that same copy. Content may
be reintroduced, instead of reshared, for various reasons - multi-
ple users may have independently discovered the same content, or
downloaded and then re-uploaded an image.
To construct a dataset of popularly shared content, we initially
selected a seed set of reshared content uploaded to Facebook in
March 2014. We selected the top 200,000 most reshared images,
which were publicly viewable, counting only reshares within the
180 days since the image was uploaded, then used a neural net-
work classifier [28] to identify images with overlaid text (i.e., im-
age memes). One advantage of studying image memes in par-
ticular is that the information that these memes transmit is un-
likely to change, as opposed to unembellished images which may
be used differently (e.g., if the same photo is used to support sepa-
rate causes).
Next, we tried to identify other copies of content that exist in
this seed set. Beyond exact copies of the same image, many near-
identical images, which have slightly different dimensions or intro-
duce compression artifacts or borders, also exist (as seen in Figure
1). As such, a binary k-means algorithm [21] was used to iden-
tify clusters of near-identical images to which each of these can-
didates belonged, including images beyond the original set. For
each cluster, we then obtained all reshares of images in that clus-
ter that were made in 2014. To verify the quality of the clustering,
we manually examined the top 100 most-reshared copies in each of
Figure 3: Recurrence occurs when we observe multiple peaks (p0,
p1, red crosses) in the number of reshares over time. Bursts (b0, b1)
capture the activity around each peak.
Figure 4: Examples of time series of recurring and non-recurring
cascades over a year, colored by copy. Identified peaks are marked
with red crosses; the number of reshares is normalized per cascade.
100 randomly sampled clusters. In 94 clusters, all 100 copies were
near-identical. The remaining clusters mainly comprised the same
image overlaid with different text.
This sample of resharing activity in 2014 that we use consists of
395,240,736 users and pages that made 5,167,835,292 reshares of
105,198,380 images. These images were aggregated into 76,301
clusters. Repeating the process above for videos shared on Face-
book, we obtain a sample comprising 323,361,625 users and pages
that made 2,187,047,135 reshares of 6,748,622 videos, aggregated
into 156,145 clusters. Images, videos, users, and pages that were
deleted were excluded from analysis. On average, each image clus-
ter is made up of 1379 copies of the same content. Video clusters
were smaller, with 43 copies in each cluster on average.
As we only measured reshares for a year, we may only be ob-
serving part of a cascade's spread if it began prior to 2014. Thus,
we also considered subsets of each dataset containing only clus-
ters that began in 2014. We identified these subsets by additionally
measuring reshares of content in the three months prior to 2014
(October to December 2013) and excluding clusters where activity
was observed during this period.
Though we mainly analyze recurrence at the cluster level, we
also investigate the recurrence of individual copies by studying the
top 100,000 individually most reshared copies in each dataset.
2.2 Defining Recurrence
In this work, we define recurrence relative to peaks and bursts
in popularity over time. In practice, almost all popular content on
t# Reshareshwrp0p1b0b1Recurring CascadesNon-Recurring Image Memes
Videos
# Clusters
51,415 (76,793)
149,253 (156,145)
Copies/Cluster
523 (1378)
13 (43)
Prop. Recurrence
0.40 (0.59)
0.30 (0.33)
# Peaks
2.3 (4.6)
1.6 (2.0)
Days Observed
202 (280)
170 (182)
Days betw. 1st/2nd Burst
31 (32)
47 (44)
Table 1: Recurrence occurs in a large proportion of popular image memes and videos shared on Facebook. We note in parentheses statistics
computed on all cascades, as opposed to the cascades that began in 2014 whose initial spread we can observe.
Facebook experiences at least one peak in popularity. If content
peaks in popularity more than once, we say that it recurs.
To identify these peaks, and thus whether a cascade recurs, we
measure the number of reshares of content over time. Figure 4
shows several examples of recurring memes. Empirically, reshare
activity is varied across different content but is generally bursty,
with long periods of inactivity between peaks. As recurrence oc-
curs over a long amount of time, we discretize time into days.
Intuitively, a recurrence occurs when a peak is observed in the
time series. Not only should these peaks be relative outliers on a
timeline, but they should also last for a significant amount of time.
Further, we should be able to tell these peaks apart from each other.
Motivated by this intuition, suppose we observe a meme for t days.
Let ri, i ∈ {1, 2, ..., t} be the number of reshares observed on day
i. We parameterize recurrence using four variables - h0, m, and
w place constraints on identified peaks, and v places a constraint on
the "valley" between peaks. Specifically, the height h of each peak
must be at least h0 and at least m times the mean reshares per day
¯r (Figure 3). Additionally, a peak day must be a local maximum
within ±w days. Finally, between any two adjacent peaks pi and
pi+1, the number of reshares must drop below v · min{rpi , rpi+1}.
We call the area around the peak a burst (b0, b1 respectively for p0,
p1 in Figure 3), whose duration or width w, is defined as the sum of
the number of days the number of reshares is increasing before pi
and falling after pi, while remaining above ¯r. There is a one-to-one
correspondence between peaks and bursts.
In practice, we set h0=10, m=2, w=7, and v=0.5 so that each
burst is relatively well-defined. The red crosses in Figure 4 show
the identified peaks under this regime. While this definition does
not strictly minimize activity between bursts, empirically, activity
does drop significantly (and in many cases, falls to zero) in be-
tween bursts. Stricter definitions that reduce the number of iden-
tified peaks (e.g., requiring a well-defined "valley floor" between
two peaks, or increasing h0 or m) also resulted in qualitatively
similar findings. The approach we take is fairly rudimentary; fu-
ture work may involve developing more specific definitions of re-
currence which take into account the shape of resulting bursts.
3. CHARACTERIZING RECURRENCE
moderate number of reshares, are more likely to recur.
We first introduce recurrence at a high level, showing that it is
both common and bursty, with the same content sometimes resur-
facing multiple times. We then discuss four important classes of
observations that we later draw on to model and predict recurrence:
• Temporal patterns: cascades with longer initial bursts, but a
• Sharer characteristics: recurring and non-recurring cascades
differ in demographic makeup, and moderate diversity in the
initial sharing population encourages recurrence. Further,
changes in homophily in the network affect the speed at which
content spreads, and hence burstiness.
• Network structure: bursts in a cascade occur in different, but
nonetheless connected parts of the network. Also, large ini-
tial bursts tend to exhaust the supply of susceptible users, po-
tentially accounting for why moderate, but not high cascade
volume or diversity results in greater recurrence.
• Catalysts of recurrence: the availability of multiple copies
in the network may catalyze recurrence. Still, neither does
the presence of multiple copies suggest that recurrence is en-
tirely an externally-driven phenomenon, nor is it a necessary
condition for recurrence.
In the remainder of this paper, we report results primarily on
image memes, and note any salient differences with videos. All
differences reported are significant at p<10-10 using a t-test unless
otherwise noted.
3.1 Recurrence is common
Once introduced on Facebook, popular content continues spread-
ing for a long time. On average, the maximum time between re-
shares of the same content is 280 days. But rather than being shared
at a constant rate (among popularly reshared content, less than 1%
of memes have no discernible peak), resharing tends to be bursty,
with bursts typically separated by substantial periods of relative in-
activity. A mean of 32 days separates the initial and subsequent
bursts for image memes (Figure 5b).
Previously, we defined recurrence as observing multiple peaks
in the number of reshares observed over time, and non-recurrence
as observing only a single peak. Over these long periods of time,
59% of popular image memes recur. In fact, a significant propor-
tion of these cascades experience resurgences in popularity (Figure
5a), and may even have experienced bursts prior to our observation
window. If we limit the sample to the set of image memes which
began spreading in 2014, 40% of these memes recur (Table 1).
3.2 Temporal patterns
Cascades with larger initial bursts of activity that last longer are
more likely to recur, suggesting that more viral, or appealing cas-
cades are more likely to recur. However, it is not the most popular
cascades that recur the most, but those that are only moderately
popular - while recurrence initially increases with the size of the
initial peak, it subsequently decreases.
Recurring cascades have larger, longer-lived initial bursts. The
initial burst of a cascade is already indicative of recurrence. Recur-
ring cascades start out larger (15,547) and initially last longer (9.3
days) than non-recurring cascades (6128 reshares, 6.9 days), spend-
ing more time "building up" (Figure 5c) and "winding down". The
greater initial popularity of recurring cascades suggests that more
viral cascades are more likely to recur, but is this the case?
Recurring content is moderately popular. Plotting the total num-
ber of reshares in the initial burst against the subsequent number
of bursts observed, rather than the number of reshares monotoni-
cally increasing or decreasing the rate of recurrence, we observe a
striking interior maximum at approximately 105 reshares for both
image memes and videos (Figure 7a). Neither the initially best-
performing (or most viral), nor poorest-performing (or least viral)
cascades tend to resurface. In the former case, a single large burst
tends to dominate with smaller bursts after; in the latter case, a
small number of small bursts is typically observed.
They keep coming back! While most of our analyses focus on the
initial burst and subsequent recurrence, several general trends arise
as more recurrence is observed:
(a) Number of Peaks
(b) Days Between Bursts
(c) Cascade Duration
Figure 5: (a) 40% of cascades that began in 2014 came back, and (b) over 30% of recurring cascades only resurfaced after a month or more.
(c) Further, the initial burst of a recurring cascade tends to last longer than that of a non-recurring cascade.
The latter observation corroborates previous work that showed a
correlation with eventual cascade size [14].
(b) # Reshares in Each Burst
(a) Probability of Subsequent Bursts
Figure 6: (a) The probability of subsequent recurrences increases
after the initial recurrence. (b) Cascades that recur less tend to have
bursts that diminish in size over time, while those that recur more
tend to have a stable burst size.
• Once a cascade has recurred, it is more likely to resurface
again. The probability of recurrence jumps from 0.40 ini-
tially, to 0.60 for subsequent recurrences before gradually
decreasing (Figure 6a). This observation parallels prior work
showing that the prior popularity of Youtube videos predicts
their future popularity [11].
In fact, for 26% of all image
meme cascades, we observe resharing activity on the first and
last day of our observation period. These image memes may
be "evergreen", tending to continuously recur.
• For cascades that recur less, subsequent bursts tend to be
smaller; for cascades that recur more, subsequent bursts are
more similar in size (Figure 6b), suggesting that they depend
less on external factors (e.g., breaking news) to spread.
• Subsequent recurrences are briefer than their predecessors.
Burst duration monotonically decreases from a mean of 7.6
days for the first burst to 6.3 for the tenth.
• On average, the lull between recurrences is substantial, with
bursts happening an 28 to 32 days apart for image memes,
and 30 to 44 days apart for videos. Again, these long periods
between bursts suggest that recurrence can only be observed
over substantial periods of time.
3.3 Sharer characteristics
People who participate in recurring cascades differ significantly
from those who participate in non-recurring cascades. While a di-
verse user population encourages recurrence, moderately diverse
cascades recur the most. Homophily, the concept that similar peo-
ple are likely to share the same content, also affects how quickly
content spreads, suggesting that it modulates recurrence.
Demographics vary with recurrence. For recurring cascades, the
average age of people participating in the initial burst is lower (40
vs. 42), but the proportion of women is higher (65% vs. 58%).
Demographics also change across bursts. In the case of image
meme cascades, the mean age changes by 2.7 years, and the pro-
portion of women by 6.1 percentage points (in absolute terms). The
same content may become popular in different parts of the world at
different times, resulting in recurrence: 13% of the time, the major-
ity of people in the initial two bursts come from different countries.
Diversity encourages recurrence. We now turn our attention to
the diversity (or homophily) of people who take part in a cascade.
We quantify homophily in the network by measuring the entropy
of the distribution of demographic characteristics. A low entropy
in the distribution of countries users are from (or country-entropy)
corresponds to high homophily, suggesting that a majority of shar-
ers belong to a small number of countries. On the other hand, a
high country-entropy suggests that the countries sharers belong to
are more diverse and distributed more evenly.
It is not a priori clear whether homophily encourages or inhibits
recurrence. Homophily within a community, meaning that con-
nected users are receptive to sharing the same content, may help a
cascade gain the initial traction it needs to spread, but may also re-
sult in the content getting "trapped" in a local part of the network.
In contrast, diversity in the users sharing that content suggests it
has wider appeal and might come back, but may also result in only
a single burst if the initial spread overwhelms the network.
We find that diversity in the country distribution is predictive
of recurrence. Controlling for the duration (w), peak height (h),
and the number of reshares in the initial bursts of recurring and
non-recurring cascades [42], a Wilcoxon Signed-rank test shows
that a higher country-entropy is indicative of recurrence (W>108,
p<10-10, effect size r=0.19). Thus, if the initial burst of a cascade
occurs in more countries, it is more likely to recur. Higher gender-
entropy (i.e., greater gender-balance) also predicts recurrence, but
its effect is weaker (W>108, p<10-2, r=0.02). The effect of age is
inconsistent across image memes and videos.
Recurring content is moderately diverse. Again, it is not the
most diverse populations that bring about recurrence: a moderate
country-entropy of approximately 3.0 in the initial burst of a cas-
cade results in the most recurrence (Figure 7b). An interior maxi-
mum can also be observed with respect to the gender-entropy of the
initial burst (Figure 7c). These results, combined with the previous
observation of a similar interior maximum with respect to the ini-
tial number of reshares, suggests that the virality of content plays a
significant role in recurrence.
Cascades spread quickly in pockets of homophily. The viral-
ity of a cascade and homophily in the network are closely related,
lllllllllllllll0.000.250.500.751.00051015# BurstsEmpirical CCDF0.000.250.500.751.00050100150200Days Between 1st/2nd BurstEmpirical CCDFllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll0.000.250.500.751.000102030Duration of Initial Burst (Days)Empirical CCDFllNon−RecurringRecurringlllllllllllllll0.30.40.50.60.70.8051015# BurstsProbability of Subsequent Burstlllllllllllllllllllllllllllllllllllllllllllllllllllllll10K15K20K25K0246810Burst Index# Reshares# Burstsllllllllll12345678910 (a) # Bursts vs. # Reshares
(b) # Bursts vs. Country-Entropy
(c) # Bursts vs. Gender-Entropy
Figure 7: (a) A moderate number of reshares results in more recurrence. (b), (c) Similarly, recurrence is more likely when the entropy of the
distribution of users across countries, as well as gender, is moderate.
ities of 0.02 (i.e., 2% of people share the same content in both
bursts) and 0.03 mean that bursts have very little direct overlap.
and perhaps represent two perspectives on the spread of content.
Greater virality enables content to appeal to a larger population;
more homophily suggests that receptive users are closer in the net-
work. In fact, homophily in the network modulates the speed of
resharing (and thus bursts in a cascade). If we measure the aver-
age country-entropy of a sliding window of 100 reshares ordered
in time and the time elapsed, and then compute the average corre-
lation, we find a slight positive correlation between the two (0.08),
suggesting that homophily among those sharing results in faster re-
sharing, and hence burstiness. Gender-entropy and age-entropy are
also positively, but more weakly correlated with burstiness (0.06
and 0.04 respectively). One potential cause of this is that pages,
with their substantial and homophilous followings, are driving re-
sharing. The more pages that share content, the more homophilous
the set of potentially reachable people, and thus the quicker content
is reshared. However, as the proportion of reshares attributable to
pages increases, the entropy in these demographic characteristics
instead increases, an effect opposite to what we observe.
3.4 Network structure
The initial bursts of recurring cascades tend to be better con-
nected. Further, successful recurrence tends to occur in different,
but not disconnected parts of the network. Considering the peo-
ple potentially exposed in each burst beyond the resharers, large
initial cascades may exhaust the population of susceptible people
in the network, a fact that will subsequently become important in
explaining the mechanism of recurrence.
Bursts of recurring cascades are internally more connected.
More people and pages share in the initial burst of recurring cas-
cades than non-recurring cascades (15,050 users and 59 pages, and
5855 users and 24 pages respectively). To measure connectivity
within a burst, we used the induced subgraph G0 of the Facebook
network made up of the people and pages resharing in the initial
burst. The subgraph includes two kinds of edges: friend edges be-
tween people, and follow edges between people and the pages they
like. On this subgraph, people in the initial burst of recurring cas-
cades have an average of 3.4 connections to other people and pages
in the same burst, relative to 3.1 connections in non-recurring cas-
cades, suggesting that the initial bursts of recurring cascades are
slightly more connected.
Subsequent bursts happen in different parts of the network.
Bursts of a cascade are separate in time, but may overlap in terms
of sharers, or be connected via friend and follow edges.
To start, the sharer overlap between bursts is small. For recurring
cascades, an average of 15,050 people and 59 pages make up the
initial burst, and 8892 people and 28 pages the subsequent burst.
Comparing people and pages across these bursts, Jaccard similar-
We also find evidence of community structure within bursts by
considering whether the second burst is proximate in the network to
the first. Even if individuals in the second burst do not repost con-
tent a second time, they may not be far removed in the social net-
work from someone in the first burst. Here, we instead consider the
induced subgraph G0+1 of the individuals and pages who reshared
in the initial two bursts for each cascade. If these bursts correspond
to communities within the network, we would expect more edges
within bursts than between them. An average of 17,457 friend and
17,242 follower edges exist in the first burst; 10,094 friend and
6,310 follower edges exist in the second. Note that these commu-
nities are very sparse. A person within a burst has a connection to
an average of 3.2 friends or pages within the same burst, indicat-
ing that memes tend to diffuse out through the network rather than
stay within a narrow community. Still, across bursts, we observe
an average of 8,273 friend and 4,755 follower edges, resulting in
an average of 1.4 connections to friends and pages in a different
burst, indicating that these bursts are somewhat separated.
Large initial bursts exhaust the supply of susceptible people. As
noted above, the second burst in a cascade has fewer sharers. Intu-
itively, people may tire of content they have seen before, but is this
the case? Studying overlap a third way, we look at the susceptible
populations of the initial and subsequent bursts of a cascade. We
approximate these populations by considering people who could
have been exposed through their connections (friend and follow
edges) to those who shared content in these two bursts.
On average, 6.4 million unique individuals are potentially ex-
posed in the initial burst, with recurring cascades, having a greater
number of initial resharers, having greater reach (8.2 million vs.
4.0 million for non-recurring cascades). For recurring cascades,
the potential reach of the second burst is smaller, but still sizable at
6.8 million. Comparing the sets of individuals exposed in the first
two bursts of recurring cascades, we obtain a Jaccard similarity of
0.15, indicating this second burst is mostly reaching a different set
of people, but where some people will have seen the same content
twice, creating a sense of déjà vu for many.
In particular, 28% of people who are potentially exposed in the
second peak would have also been exposed in the first. This high
overlap could be a contributing factor to the second peak being
smaller. Further, the proportion of exposed individuals in the sec-
ond peak is positively correlated (ρ=0.39) with the size of the first
peak, meaning that the larger the initial peak, the more likely that
those exposed in the second peak have previously seen the content.
In other words, in the case of large initial bursts, subsequent bursts
are likely reaching a similar part of the network.
lllllllll1.52.02.53.03.54.0103104105106# Reshares in Initial Burst# Burstslllllllllll2.02.53.03.5012345Country Entropy in Initial Burst# Burstsllllllllll1.52.02.50.20.40.60.81.0Gender Entropy in Initial Burst# Bursts 3.5 Catalyzing recurrence
As shown in Figures 1 and 2, cascades are made up of reshares
of multiple copies of the same content, and the presence of these
copies can help catalyze recurrence. Still, neither are copies the
only cause of recurrence (recurrence is substantial even with a sin-
gle copy), nor must they be independently or externally introduced
(many later copies are attributable to previously seen copies).
Cascades whose reshares are divided across multiple copies tend
to recur. Recurring cascades are made up of more copies than non-
recurring cascades (2277 vs. 93). Reshares are also more spread
out across multiple copies in the former case (841 and 3445 re-
shares per copy for recurring and non-recurring cascades respec-
tively), suggesting recurrence may be characterized by multiple
smaller outbreaks. The most reshared copy accounts for 72% of
reshares in the initial burst for recurring cascades, and 93% for
non-recurring cascades. Altogether, the substantial differences here
suggest the strong predictive power of these characteristics.
The appearance of new copies correlates with recurrence. Fur-
ther, the introduction of new copies and the number of reshares over
time is significantly correlated (Pearson's r=0.66), suggesting that
the appearance of new copies causes bursts, and thus recurrence.
On a related note, prior work showed that reposting content helps
make it popular [44].
Copies are not the only cause of recurrence. Nonetheless, not all
copies burst (only 6% are reshared at least 10 times on any single
day), and not all bursts are caused by new copies, as we will later
show. And while correlations between the number of copies and
other characteristics such as duration and country-entropy also ex-
ist, when we control for the number of copies in the initial bursts of
recurring and non-recurring cascades [42], all previously observed
differences in the temporal, sharer, and network characteristics of
these cascades still hold (W>108, p<10-10, mean effect size r=0.08).
Comparing recurring and non-recurring cascades with similar num-
bers of copies in their initial bursts, the initial bursts of recurring
cascades are still larger, longer-lived, and more diverse. In all, this
suggests that recurrence is not simply caused by distinct copies of
the same content spreading through the network, but is a result of a
more complex phenomenon which we explain in Section 4.
A majority of copies are internal to the network. Still, where do
these copies come from, and are they internal or external to the net-
work? By using the network to identify friends and pages who may
have previously shared a different copy of some content, we can
attribute 75% of newly uploaded copies to previously seen copies
in the network (this approach roughly estimates content-copying
that occurs within Facebook, as users who share a new copy may
not have seen a friend's shared copy). This suggests a nuanced ap-
proach to studying recurrence - external sources may drive some
of the introduction of new copies to a social system, but a large pro-
portion of activity, which we can study, occurs within the network.
Pages may also catalyze recurrence. Pages are responsible for a
large proportion of highly-reshared copies (over 70% of reshares
are attributable to page-created copies in the second burst of re-
curring cascades). In recurring cascades, pages tend to re-upload,
rather than reshare content, doing so 50% of the time, as opposed
to 2% for users. Further, the most popular copy in the second burst
is likely to have been created by a page (70%). Given the relatively
higher degree of pages, which tend to have tens of thousands of
followers, as opposed to users who typically only have hundreds of
friends, pages may spark recurrence by posting a new copy of the
same content, rapidly exposing a number of followers to it.
Individual copies recur too! Recurrence of the individually most
popular copies in our datasets, while lower than when copies are
studied in clusters, is still substantial (18%). These individual copies
last a significant amount of time (261 days), with bursts further
apart (41 days). Like cascades of multiple copies, the initial bursts
of recurring individual-copy cascades are larger and longer-lived
than those of non-recurring cascades, with later bursts occurring in
different parts of the network. Recurrence of the same copy can
also be observed within clusters - 22% of the time, the most re-
shared copy in a burst was also most reshared in a previous burst.
4. MODELING RECURRENCE
Tying our observations together, we present an overall picture of
the mechanisms of recurrence, then suggest a model of recurrence
which we evaluate through simulations on a real social network.
4.1 Why do cascades recur?
Our findings as a whole suggest a model of recurrence where
virality is a primary factor, and where the availability of multiple
copies can help spark recurrence.
Virality plays a primary role in recurrence. Virality, or broad-
ness of appeal, affects recurrence: cascades with initial bursts that
are larger, last longer, and are more demographically diverse are
more likely to recur. Specifically, moderately popular and diverse
cascades are most likely to recur. While recurrence typically oc-
curs in different parts of the network, the larger the initial burst
of a cascade, the larger the proportion of the potentially exposed
population in the subsequent burst that was already previously ex-
posed. This observation, coupled with the fact that users tend not
to reshare the same content multiple times, suggests that large ini-
tial bursts inactivate a significant portion of the network, inhibiting
a cascade's future spread. Our subsequent simulations show more
clearly that this may indeed happen as the initial burst grows large.
Multiple copies in the network help spark recurrence. Bursts
in a cascade are separated by relatively long periods of inactivity.
By studying the availability of multiple copies of the same con-
tent, we find that these copies can act as catalysts for recurrence
in different parts of the network. Indeed, multiple introductions of
the same content correlate with recurrence. However, while more
copies initially increases the chance of recurrence, they are not the
only cause of it; recurring and non-recurring cascades with similar
numbers of copies differ significantly in virality. Moreover, mul-
tiple copies do not explain the substantial recurrence of individual
copies. To a lesser extent, we also discover that homophily in the
network affects the speed of the spread of a cascade in a network.
Together, moderate content virality and the presence of multiple
copies results in recurrence. While the likelihood of recurrence
does increase with the number of copies (or potential "sparks"), we
can still observe an interior maximum in how recurrence varies with
the number of reshares after fixing the number of copies, where a
moderate number of reshares results in the most recurrence.
4.2 A simple model of recurrence
Motivated by these findings, we suggest a simple model of cas-
cading behavior where recurrence depends on content virality:
• If the virality of a cascade is low, it may only appeal to a
small group of people, and is thus unable to spread far in
the network. Thus, a single, small peak results, with many
attempts to propagate in the network failing (Figure 8a).
• As virality increases, the cascade is able to spread substan-
tially further in the network, and may occasionally even jump
to other local communities in the network, spreading faster
When the virality of the contagion is high (pc
of infections (or reshares) in each time step, and identify bursts as
defined in Section 2.2.
Results. Within certain ranges of virality (6×10-4≤ pc
0 <8×10-4),
we can consistently reproduce recurring cascades. Figure 9a shows
several examples of the time series of these simulations. In aggre-
gate, we can obtain a distribution of number of peaks similar in
shape to Figure 5. Plotting the number of peaks against number
0), we observe an
of reshares in the initial burst (or alternatively, pc
interior maximum - a moderate amount of virality results in the
most recurrence (Figure 9b), replicating our previous findings.
0≥8×10-4), a large
fraction of the highly connected portion of the graph becomes in-
fected by a single copy in the initial burst, suppressing subsequent
bursts as many nodes are now resistant. To show this happening,
we consider for each simulation, in addition to our original model,
an alternate-universe setting where the resistances of nodes are re-
set following the initial burst. We can then measure how much
the initial burst inhibited the second by observing the likelihood
of a second burst in the alternate case, as well as the overlap of
nodes infected in the second burst with nodes in the initial burst.
A significant difference in the total number of peaks when virality
is high (1.0 vs. 2.0, t=92, p<10-10), but not when virality is low
0≤7×10-4, n.s.) suggests that the supply of susceptible nodes
(pc
is indeed being used up in the former case, but not the latter. A
significant positive correlation of initial peak size with the size of
the overlap of the second peak in the alternate setting (0.76) further
supports this hypothesis and our prior observations.
Likewise, the connectivity of graph deteriorates significantly af-
0≥8×10-4). Here, we measure the al-
ter a large initial burst (pc
gebraic connectivity [18] of the graph if all the nodes involved in
the initial burst are removed, and compare this to a baseline that
removes the same number of nodes at random. Connectivity is sig-
nificantly lower in the former case (579 vs. 1065, t>17, p<10-10),
especially in comparison to the graph's initial connectivity (1105).
These results together suggest that under such a model of recur-
rence, a large initial burst does indeed inhibit subsequent bursts,
as we previously hypothesized (Figure 8c). Also in support of our
prior observations, increasing the number of introduced copies m
monotonically increases recurrence.
Limitations and alternatives.
Importantly, our model assumes
that recurrence is sparked primarily by independent copies intro-
duced to the network. However, the reality of recurrence is subtler:
individual copies recur significantly in the network, and homophily
may also moderate recurrence. Allowing virality to vary with time
[24] or having nodes wait according to a power-law distribution
[17, 35] may also reproduce recurrence with only a single copy.
Decision-based queuing processes [8] may also help model the long
periods of inactivity between bursts.
5. PREDICTING RECURRENCE
Is it possible to predict if a cascade will resurface in the future?
Observing just the initial burst of a cascade, we use features re-
lated to the temporality, network structure, user demographics, and
presence of multiple copies to determine a) whether recurrence oc-
curs, b) if the recurrence will be relatively smaller or larger, and
c) when the recurrence occurs. Overall, we find that cascades with
longer initial bursts that consist of multiple small outbreaks tend to
recur, supporting the hypothesis that content virality and multiple
copies play a significant role in recurrence. Nonetheless, we obtain
similarly strong performance predicting recurrence for individual
copies of content. Predicting recurrence may enable us to better
forecast content longevity in a network.
Figure 8: When virality is low, only a small number of attempts at
infection succeed. When virality is moderate, more attempts suc-
ceed, which aggregate into observable recurrence. When virality
is high, rather than a large number of bursts aggregating to form a
single large peak, the first successful burst infects a large portion of
the network, making it difficult for other copies to spread.
within them. As several bursts occur in the network, they
may be observed as recurring in aggregate (Figure 8b).
• However, as virality increases beyond some threshold, any
individual burst is likely to spread through a large portion
of the susceptible population, inhibiting the transmission of
subsequent copies (Figure 8c). This last point lies in contrast
to the trivial hypothesis that more independent copies leads
to more independent bursts that aggregate to form a single
large burst, which does not appear to be the case, as most
reshares in initial bursts can be attributed to a single copy.
4.3 Simulating recurrence
To see if such a model of recurrence can reproduce characteris-
tics of recurrence observed in the data, we now simulate recurrence
on a real social network. Our observations and model suggest the
use of an SIR model, where nodes in a network are initially sus-
ceptible (S) to a contagion, and then may become infected (I) when
exposed. Infected nodes subsequently recover (R) and become re-
sistant to the contagion. These models have been used to study the
spread of disease [6] and information [9, 13, 39] in a network.
Setup. Our simulation thus consists of an SIR model with multiple
outbreaks introduced at different times, and with resistant nodes re-
infectable at a lower rate. We parameterize our model as follows:
For a given contagion c, its virality, or equivalently, the susceptibil-
0. In other words, if exposed
ity of every node in the network, is pc
to the contagion, the probability that the node will be infected is
0. Infected nodes attempt to infect all neighbors in the subsequent
pc
time step, and then become resistant. As users sometimes share the
same content multiple times, resistant nodes have a constant lower
probability pc
0 of being re-infected. The introduction of each
copy of a contagion is normally distributed in time (N (µ, σ)).
1 < pc
Here, we make a simplifying assumption that independent copies
of the same content are introduced into the network at different
points in time. Following the intuition that more connected entities
(e.g., pages) are likely to start outbreaks, the target nodes to infect
are sampled, with replacement, proportional to the node's degree.
m copies are introduced in total.
We simulate this model for 1000 discrete time steps with µ=500,
0 between 5×10-4 and 10-3 and where
σ=250, and m=50, varying pc
1=0.5·pc
1. We run our simulation on the network of a country with
pc
approximately 1.4 million nodes and 160 million friendship edges,
repeating the simulation 5000 times. We measure the total number
# Resharesa. Low Viralityb. Moderate Viralityc. High ViralityOverallIndividual Burstst cascades (N=40,912 for image memes, 89,368 for videos) so that
guessing results in a baseline accuracy of 0.5. Given the non-linear
relation of several features to recurrence (e.g., that a moderate num-
ber of reshares results in the most recurrence), we use a random
forest classifier. In all cases, we perform 10-fold cross-validation
and report the classification accuracy, F1 score, and area under the
ROC curve (AUC).
Results. Overall, we find strong performance in predicting recur-
rence (Accuracy=0.82, F1=0.81, AUC=0.89). A logistic regression
classifier results in slightly worse performance (AUC=0.78). Ta-
ble 2 shows how performance improves as features are added to
the model, as well as individual feature set performance. While
multiple-copy features perform best, temporal and network fea-
tures, and to a lesser extent demographic features, also individually
exhibit robust performance, suggesting that each significantly con-
tributes to recurrence. In the absence of strong multiple-copy fea-
tures (fewer copies of any one video exist), we obtain worse perfor-
mance in predicting the recurrence of videos (Acc=0.69, F1=0.66,
AUC=0.76), with temporal features instead performing best.
For image meme cascades, the most predictive features of re-
currence relate to cascades having multiple small outbreaks (fewer
reshares per copy (0.78) and a higher entropy in the distribution
of reshares across copies (0.72)), and longer initial bursts (more
days before (0.63) and after (0.63) the peak). These features re-
main important for video cascades. Mirroring the dual importance
of multiple-copy and temporal features, just the number of reshares
per copy and the average gradient of the initial burst after its peak
alone achieve strong performance (0.81). Though the initial burst
of a recurring cascade is on average significantly larger, size-related
features are weaker signals of recurrence (≤0.59).
5.3 Will the recurrence be smaller/larger?
Prediction task. Assuming that we know that a cascade will recur,
how much smaller or larger will the second burst be? Knowing the
relative size of the next recurrence can differentiate bursty cascades
that are rising or falling in popularity. Given the initial burst of a
cascade, we aim to predict if the relative size of the second burst, or
the ratio of the size of the second burst to that of the first, is above or
below the median (0.28). As the median evenly divides the dataset,
we again have a balanced binary classification task with a random
guessing baseline accuracy of 0.5.
Results. We also find strong performance in predicting the rela-
tive size of the subsequent burst (Acc=0.72, F1=0.69, AUC=0.78
for image memes, AUC=0.85 for videos). Temporal features here
outperform all other feature sets, with the most predictive features
relating to the cascade having a long initial burst.
5.4 When does it recur?
Prediction task.
If a cascade will recur, when will we observe
the next burst? With a cascade's initial burst, can we predict if the
duration between bursts will be greater than the median (14 days)?
Results. We find that the timing of recurrence is far less predictable
(Acc=0.56, F1=0.51, AUC=0.58 for image memes, AUC=0.60 for
videos). Nevertheless, longer initial bursts are most indicative of
recurrence happening earlier.
5.5 Predicting recurrence for individual copies
Given the correlation of the appearance of multiple copies with
bursts, multiple-copy features perform strongest in predicting re-
currence. But what if we want to predict recurrence of a single
instance of some content, where multiple copies do not exist by
(a) Simulated Time Series
(b) # Peaks vs. Reshares, Simulated
Figure 9: (a) By varying content virality, a model of recurrence that
assumes independent introductions of copies of the same content
can simulate recurrence. (b) It also replicates the observation that a
moderate number of reshares results in more recurrence.
AUC on Feature Sets
Temporal
+ Demographic
+ Network
+ Multiple-Copy
0.74
0.78 (0.63)
0.81 (0.72)
0.89 (0.82)
0.76
0.76 (0.58)
0.77 (0.66)
0.78 (0.70)
0.55
0.56 (0.52)
0.57 (0.53)
0.58 (0.54)
Table 2: We obtain strong performance in predicting whether recur-
rence occurs and if the subsequent burst will be smaller or larger,
but not in predicting when recurrence occurs. Individual feature set
performance is in parentheses. The column headers refer to Sec-
tions 5.2, 5.3, and 5.4 respectively.
5.1 Factors driving recurrence
Based on our observations, we develop several features that help
predict recurrence, and group them into four categories:
Temporal features (7). Initially longer-lived bursts are suggestive
of recurrence, motivating the importance of the number of days be-
fore and after the peak is reached, as well as the number of reshares
before and after, and the height of the initial peak. The average
gradient of the initial burst before and after the peak further char-
acterize the shape of the initial burst.
Demographic features (5). The differences in user characteris-
tics and diversity we previously observed suggest the importance
of age, gender, as well as the entropy in the distribution of age,
gender and country of the initial burst.
Network features (6). Recurring cascades appear to be more con-
nected in their initial bursts, having more friendship and follower
edges, in addition to having a larger potentially exposed popula-
tion. The number of users, pages, and proportion of pages in the
initial burst also vary.
Multiple-copy features (8). The availability of multiple copies
plays a significant role in recurrence, motivating the use of the num-
ber of copies observed in the initial peak, the entropy in the distri-
bution of reshares of each copy, the mean reshares per copy, and
the proportion of reshares attributable to the most popular copy.
Pages also play a role in recurrence, suggesting that the proportion
of copies created by pages, the proportion of all reshares made by
pages or attributable to page-created copies, and whether the most
popular copy was created by a page are useful features.
5.2 Does it recur?
Prediction task. We formulate our prediction task as a binary clas-
sification problem: given only the initial burst of a cascade, we aim
to predict if a second burst will be observed (i.e., if the cascade will
recur). We use a balanced dataset of recurring and non-recurring
Recurring CascadesNon-Recurringllllllll1.01.52.02.53.0101102103104# Reshares in Initial Burst# Bursts definition? Surprisingly, we obtain similarly strong performance in
predicting the recurrence of individual copies (N=28,454, Acc=0.80,
F1=0.79, AUC=0.88 for image memes, AUC=0.82 for videos). Net-
work features are strongest (AUC=0.84), with fewer edges between
users and pages (0.68) in the initial peak the most predictive of
recurrence. As individual copies have a single point of origin,
fewer edges between pages and users and more edges between
users (0.61) suggests that the burst may have resulted more from
users sharing content from other users than high-degree pages shar-
ing that content with their followers. This observation, together
with the fact that longer initial bursts continue to be strongly pre-
dictive of recurrence (>0.65), suggests the continued significance
of virality with respect to individual copies.
The relative size of the subsequent burst is similarly predictable
for individual copies (0.83 for image memes, 0.84 for videos), but
interestingly, the time of recurrence is more predictable (0.68 and
0.63 respectively), which may be because any recurrence must be
a continuation of the initial copy, as opposed to possibly being
sparked by a new, less related copy.
6. RELATED WORK
Significant prior work has studied information diffusion in on-
line social media [7, 14, 37] - with respect to memes, work has
demonstrated the effect of meme similarity [16] and competition
for limited attention on subsequent popularity [46]. Most relevant
is previous work that looked at the temporal dynamics of diffusion
and developed epidemiological models of recurrence.
Among work that aims to predict the future popularity of on-
line content [12, 32, 44], one relevant line of research has involved
modeling the temporal patterns of the diffusion of information in
social media [2, 36, 49] or using these patterns to predict future
popularity or forecast trends [4, 9, 10, 14, 15, 24]. Perhaps driven
by the general shape of the initial burst of a cascade, many of these
models implicitly assume that the temporal shape of a cascade con-
sists primarily of a rising and falling period, and focus on modeling
the initial activity around a peak [2, 36, 48] or the overall popular-
ity discounting subsequent spikes [10]. Beyond the initial burst of
activity, we studied the long-term temporal dynamics of content on
Facebook over a year.
In prior work, when multiple bursts are observed in a time se-
ries, they tend to be of a topic or hashtag rather than an individual
piece of content, and are commonly attributed to external stimuli
[23, 29, 33, 38] (e.g., news related to that topic). While knowing
about external events can help forecast the temporal pattern of the
resulting spike [36], there has been little work in predicting if new
spikes will appear in the future lacking such knowledge. In partic-
ular, rumor recurrence is bursty, with or without external stimuli,
and sometimes with embellishments and other mutations [1, 31,
19], but there is little understanding of this phenomenon. Patterns
of human activity can also explain periodicity in popularity [5, 22,
34], but the vast majority of recurrence we observe in this paper
is aperiodic. While external stimuli explains some instances of re-
currence, we discover other factors that influence recurrence. In
contrast to most work that has observed multiple bursts in topics,
we observed recurrence even at the level of an individual copy.
Finally, substantial work has studied how bursts in streams or
time series can be detected [26, 27, 41]. In this paper, we adopted
a simple definition of burstiness, parameterizing peaks and bursts
relative to the mean activity observed.
Recurrence has also been studied in the context of epidemiology,
though primarily from a modeling perspective. Many base their
analysis on SIR models [39], simulating recurrence through intro-
ducing dormant periods [25], seasonality effects [3], or changes in
contagion fitness [20], which may be periodic [40]. More recently,
some work studied content popularity using these models, while ac-
counting for user login dynamics and content aging [13]. The struc-
ture of the network can also cause periodicity in epidemics [30,
45]. Many focus on modeling specific types of recurrence (e.g.,
historical disease epidemics [3]). In contrast, many recurrences we
observe are aperiodic, and findings on synthetic networks may not
easily generalize. Inspired by this line of work, we adapted an SIR
model assuming multiple points of infection on a real social net-
work, and show that key characteristics of recurrence we observed
can be reproduced.
7. DISCUSSION AND CONCLUSION
Our results start to shed light on the mechanism of content re-
currence - studying a large dataset of popularly reshared content,
we find that recurrence is common, and that content can come back
not just once, but several times. Strikingly, content may nearly
cease to circulate for days, weeks or even months, prior to experi-
encing another surge in popularity. Such a phenomenon may seem
highly unpredictable, but we find trends in how recurring cascades
behave, and can predict whether content will come back. The vi-
rality, or appeal of a cascade plays a role in recurrence: cascades
whose initial bursts are long-lasting, moderately popular, and mod-
erately diverse are most likely to recur. The presence of multiple
copies of the same content sparks recurrence, though homophily in
the network may also influence recurrence.
One limitation of our work is that we only analyze content within
a single network. Though most copies of the same content were
made within the network, a minority appeared without a prior path.
Analyzing the transfer of content between different social networks
may reveal different mechanisms of recurrence. Separately, while
the appearance of multiple copies correlates with recurrence, this
does not hold in the case of individual-copy recurrence. Under-
standing recurrence in the absence of multiple copies (e.g., through
studying homophily in more detail) remains future work.
Based on our observations, we presented a simple model that
exhibits some features of recurrence (e.g., pronounced bursts with
little activity in-between, and an internal maximum in the number
of bursts as a function of the number of reshares). Future work
could extend such models to account for homophily and commu-
nity structure in the network.
While the temporal shape, network structure, and user attributes
are already highly predictive of resharing behavior, other factors
may improve prediction accuracy further: sentimentality or humor
may make content evergreen, while content tied to current events
may have an expiration date. Seasonality effects may also cause pe-
riodic recurrence: we did observe an instance of a daylight-savings
image meme which appeared, as expected, exactly at the two points
during the year when people needed to adjust their clocks. Also,
other types of content may exhibit different properties of recur-
rence (e.g., link sharing may be more externally driven); the in-
teractions of users with shared content (e.g., comments) may also
reveal the reasons why some content came back; the societal con-
text of memes, as well as their interactions (or competition) with
other content, may also reveal more insight into their popularity
[43]. Perhaps most suggestive that much remains to be studied is
that while we can predict if recurrence will happens, it remains a
significant challenge to predict when recurrence will happen.
Acknowledgments. This work was supported in part by a Mi-
crosoft Research PhD Fellowship, a Simons Investigator Award,
NSF Grants CNS-1010921 and IIS-1149837, and the Stanford Data
Science Initiative.
8. REFERENCES
[1] L. A. Adamic, T. M. Lento, E. Adar, and P. C. Ng.
Information evolution in social networks. WSDM, 2016.
[2] M. Ahmed, S. Spagna, F. Huici, and S. Niccolini. A peek
into the future: Predicting the evolution of popularity in user
generated content. WSDM, 2013.
[3] S. Altizer, A. Dobson, P. Hosseini, P. Hudson, M. Pascual,
and P. Rohani. Seasonality and the dynamics of infectious
diseases. Ecol. Lett., 2006.
[4] S. Asur, B. Huberman, et al. Predicting the future with social
media. WI-IAT, 2010.
[5] S. Asur, B. Huberman, G. Szabo, and C. Wang. Trends in
social media: Persistence and decay. ICWSM, 2011.
[6] N. T. Bailey et al. The mathematical theory of infectious
diseases and its applications. 1975.
[7] E. Bakshy, I. Rosenn, C. Marlow, and L. A. Adamic. The role
of social networks in information diffusion. WWW, 2012.
[8] A.-L. Barabasi. The origin of bursts and heavy tails in human
dynamics. Nature, 2005.
[9] C. Bauckhage, F. Hadiji, and K. Kersting. How viral are viral
videos? ICWSM, 2015.
[10] C. Bauckhage, K. Kersting, and F. Hadiji. Mathematical
models of fads explain the temporal dynamics of internet
memes. ICWSM, 2013.
[11] Y. Borghol, S. Ardon, N. Carlsson, D. Eager, and
A. Mahanti. The untold story of the clones: content-agnostic
factors that impact Youtube video popularity. KDD, 2012.
[12] Y. Borghol, S. Mitra, S. Ardon, N. Carlsson, D. Eager, and
A. Mahanti. Characterizing and modelling popularity of
user-generated videos. Perform. Eval., 2011.
[13] M. Cha, F. Benevenuto, Y.-Y. Ahn, and K. P. Gummadi.
Delayed information cascades in Flickr: Measurement,
analysis, and modeling. Computer Networks, 2012.
[14] J. Cheng, L. A. Adamic, P. A. Dow, J. M. Kleinberg, and
J. Leskovec. Can cascades be predicted? WWW, 2014.
[15] H. Choi and H. Varian. Predicting the present with Google
Trends. Econ. Rec., 2012.
[16] M. Coscia. Average is boring: How similarity kills a meme's
success. Sci. Rep., 2014.
[17] R. Crane and D. Sornette. Robust dynamic classes revealed
by measuring the response function of a social system.
PNAS, 2008.
[18] M. Fiedler. Algebraic connectivity of graphs. Czech. Math.
J., 1973.
[19] A. Friggeri, L. A. Adamic, D. Eckles, and J. Cheng. Rumor
cascades. ICWSM, 2014.
[20] M. Girvan, D. S. Callaway, M. E. Newman, and S. H.
Strogatz. Simple model of epidemics with pathogen
mutation. Phys. Rev. E, 2002.
[21] Y. Gong and S. Lazebnik. Iterative quantization: A
procrustean approach to learning binary codes. CVPR, 2011.
[22] N. Grinberg, M. Naaman, B. Shaw, and G. Lotan. Extracting
diurnal patterns of real world activity from social media.
ICWSM, 2013.
[23] D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins.
Information diffusion through blogspace. WWW, 2004.
[24] A. Guille and H. Hacid. A predictive model for the temporal
dynamics of information diffusion in online social networks.
WWW Companion, 2012.
[25] A. Johansen. A simple model of recurrent epidemics. J.
Theor. Biol., 1996.
[26] D. Kifer, S. Ben-David, and J. Gehrke. Detecting change in
data streams. VLDB, 2004.
[27] J. Kleinberg. Bursty and hierarchical structure in streams.
Data Min. Knowl. Discov., 2003.
[28] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks.
NIPS, 2012.
[29] R. Kumar, J. Novak, P. Raghavan, and A. Tomkins. On the
bursty evolution of blogspace. WWW, 2005.
[30] M. Kuperman and G. Abramson. Small world effect in an
epidemiological model. Phys. Rev. Lett., 2001.
[31] S. Kwon, M. Cha, K. Jung, W. Chen, and Y. Wang.
Prominent features of rumor propagation in online social
media. ICDM, 2013.
[32] H. Lakkaraju, J. J. McAuley, and J. Leskovec. What's in a
name? understanding the interplay between titles, content,
and communities in social media. ICWSM, 2013.
[33] J. Leskovec, L. Backstrom, and J. Kleinberg. Meme-tracking
and the dynamics of the news cycle. KDD, 2009.
[34] J. Leskovec, M. McGlohon, C. Faloutsos, N. S. Glance, and
M. Hurst. Patterns of cascading behavior in large blog
graphs. SDM, 2007.
[35] D. Liben-Nowell and J. Kleinberg. Tracing information flow
on a global scale using internet chain-letter data. PNAS,
2008.
[36] Y. Matsubara, Y. Sakurai, B. A. Prakash, L. Li, and
C. Faloutsos. Rise and fall patterns of information diffusion:
model and implications. KDD, 2012.
[37] S. A. Myers and J. Leskovec. The bursty dynamics of the
twitter information network. WWW, 2014.
[38] S. A. Myers, C. Zhu, and J. Leskovec. Information diffusion
and external influence in networks. KDD, 2012.
[39] M. E. Newman. Spread of epidemic disease on networks.
Phys. Rev. E, 2002.
[40] L. F. Olsen, G. L. Truty, and W. M. Schaffer. Oscillations and
chaos in epidemics: a nonlinear dynamic study of six
childhood diseases in Copenhagen, Denmark. Theor. Popul.
Biol., 1988.
[41] G. Palshikar et al. Simple algorithms for peak detection in
time-series. ICADABAI, 2009.
[42] P. R. Rosenbaum and D. B. Rubin. The central role of the
propensity score in observational studies for causal effects.
Biometrika, 1983.
[43] B. H. Spitzberg. Toward a model of meme diffusion (M3D).
Communication Theory, 2014.
[44] G. Stoddard. Popularity dynamics and intrinsic quality in
reddit and hacker news. ICWSM, 2015.
[45] J. Verdasca, M. T. Da Gama, A. Nunes, N. Bernardino,
J. Pacheco, and M. Gomes. Recurrent epidemics in small
world networks. J. Theor. Biol., 2005.
[46] L. Weng, A. Flammini, A. Vespignani, and F. Menczer.
Competition among memes in a world with limited attention.
Sci. Rep., 2012.
[47] J. Yang and S. Counts. Predicting the speed, scale, and range
of information diffusion in twitter. ICWSM, 2010.
[48] J. Yang and J. Leskovec. Modeling information diffusion in
implicit networks. ICDM, 2010.
[49] J. Yang and J. Leskovec. Patterns of temporal variation in
online media. WSDM, 2011.
|
1706.09297 | 5 | 1706 | 2018-08-10T01:49:10 | Optimal Investment Strategies for Competing Camps in a Social Network: A Broad Framework | [
"cs.SI",
"physics.soc-ph"
] | We study the problem of optimally investing in nodes of a social network in a competitive setting, wherein two camps aim to drive the average opinion of the population in their own favor. Using a well-established model of opinion dynamics, we formulate the problem as a zero-sum game with its players being the two camps. We derive optimal investment strategies for both camps, and show that a random investment strategy is optimal when the underlying network follows a popular class of weight distributions. We study a broad framework, where we consider various well-motivated settings of the problem, namely, when the influence of a camp on a node is a concave function of its investment on that node, when a camp aims at maximizing competitor's investment or deviation from its desired investment, and when one of the camps has uncertain information about the values of the model parameters. We also study a Stackelberg variant of this game under common coupled constraints on the combined investments by the camps and derive their equilibrium strategies, and hence quantify the first-mover advantage. For a quantitative and illustrative study, we conduct simulations on real-world datasets and provide results and insights. | cs.SI | cs | Optimal Investment Strategies for Competing
Camps in a Social Network: A Broad Framework
Swapnil Dhamal, Walid Ben-Ameur, Tijani Chahed, and Eitan Altman
1
8
1
0
2
g
u
A
0
1
]
I
S
.
s
c
[
5
v
7
9
2
9
0
.
6
0
7
1
:
v
i
X
r
a
Abstract -- We study the problem of optimally investing in nodes
of a social network in a competitive setting, wherein two camps
aim to drive the average opinion of the population in their own
favor. Using a well-established model of opinion dynamics, we
formulate the problem as a zero-sum game with its players
being the two camps. We derive optimal investment strategies
for both camps, and show that a random investment strategy
is optimal when the underlying network follows a popular class
of weight distributions. We study a broad framework, where we
consider various well-motivated settings of the problem, namely,
when the influence of a camp on a node is a concave function of
its investment on that node, when a camp aims at maximizing
competitor's investment or deviation from its desired investment,
and when one of the camps has uncertain information about the
values of the model parameters. We also study a Stackelberg
variant of this game under common coupled constraints on the
combined investments by the camps and derive their equilibrium
strategies, and hence quantify the first-mover advantage. For a
quantitative and illustrative study, we conduct simulations on
real-world datasets and provide results and insights.
Index Terms -- Social networks, opinion dynamics, election,
zero-sum games, common coupled constraints, decision under
uncertainty, Stackelberg game.
I. INTRODUCTION
Opinion dynamics is a natural phenomenon in a system of
cognitive agents, and is a well-studied topic across several
disciplines. It is highly relevant to applications such as elec-
tions, viral marketing, propagation of ideas and behaviors, etc.
In this paper, we consider two competing camps who aim to
maximize the adoption of their respective opinions in a social
network. In particular, we consider a strict competition setting
where the opinion value of one camp is denoted by +1 and
that of the other camp by −1; we refer to these camps as good
and bad camps respectively. Opinion adoption by a population
can be quantified in a variety of ways; here we consider a well-
accepted way, namely, the average or equivalently, the sum of
opinion values of the nodes in the network [1], [2]. Hence the
good camp's objective would be to maximize this sum, while
the bad camp would aim to minimize it.
The average or sum of opinion values of the nodes or
individuals is of relevance in several applications. In a fund
collection scenario, for instance, the magnitude of the opinion
value of an individual can be viewed as the amount of funds
and its sign as the camp towards which he or she is willing to
S. Dhamal is with INRIA Sophia Antipolis, France, working at Laboratoire
Informatique d'Avignon; he was with Samovar, T´el´ecom SudParis, CNRS,
Universit´e Paris-Saclay, when most of this work was done.
W. Ben-Ameur and T. Chahed are with Samovar, T´el´ecom SudParis, CNRS,
E. Altman is with INRIA Sophia Antipolis and Laboratoire Informatique
Universit´e Paris-Saclay.
d'Avignon.
contribute. Another example is that of a group of sensors or
reporting agents, who are assigned the job of reporting their
individual measurements of a particular parameter or event;
the resulting measurement would be obtained by averaging
the individual values. In this case, two competitors may aim
to manipulate the resulting average (one perhaps for a good
cause of avoiding panic, and another for elevating it).
While the opinion values can be unbounded in the above
examples, there are scenarios which can be modeled aptly
by bounded opinion values. In elections, for instance, an
individual can vote at most once. Here one could view bounded
opinion value of an individual as a proxy for the probability
with which the individual would vote for a camp. For instance,
an opinion value of v ∈ [−1, +1] could imply that
the
probability of voting for the good camp is (1 + v)/2 and
that of voting for the bad camp is (1 − v)/2. Hence the
good (or respectively bad) camp would want to maximize (or
respectively minimize) the sum of opinion values, since this
sum would indicate the expected number of votes in favor of
the good camp. Product adoption is another example where
bounded opinion values are well justified; the opinion value
of an individual would indicate its probability of purchasing
the product from the company that corresponds to good camp.
Social networks play a prime role in determining the opin-
ions, preferences, behaviors, etc. of the constituent individuals
[3]. There have been efforts to develop models which could
determine how the individuals update their opinions based on
the opinions of their connections, and hence study the dynam-
ics of opinions in the network [4]. With such an underlying
model of opinion dynamics, a camp would aim to maximize
the adoption of its opinion in a social network, in presence
of a competitor. A camp could act on achieving this objective
by strategically investing on selected individuals in a social
network who could adopt its opinion; these individuals would
in turn influence the opinions of their connections, who would
then influence the opinions of their respective connections, and
so on. Based on the underlying application, this investment
could be in the form of money, free products or discounts,
attention, convincing discussions, etc. Given that both camps
have certain budget constraints, the strategy of the good camp
hence comprises of how much to invest on each node in the
network, so as to maximize the sum of opinion values of the
nodes, while that of the bad camp comprises of how much to
invest on each node, so as to minimize this sum.
This setup results in a game, and since we consider a
strict competition setting with constraints such as budget (and
other constraints as we shall encounter), the setup fits into the
framework of constrained zero-sum games [5].
A. Motivation
There have been studies to identify influential nodes and
the amounts to be invested on them, specific to analytically
tractable models of opinion dynamics (such as DeGroot) [6],
[7], [2]. Such studies are important to complement the empiri-
cal and experimental studies, since they provide more concrete
results and rigorous reasonings behind them. However, most of
the studies are based in a very preliminary setting and a limited
framework. This paper aims to consider a broader framework
by motivating and analyzing a variety of settings, which could
open interesting future directions for a broader analytical study
of opinion dynamics.
Throughout the paper, we study settings wherein the invest-
ment per node by a camp could be unbounded or bounded.
Bounded investments could be viewed as discounts which
cannot exceed 100%, attention capacity or time constraint of a
voter to receive convincing arguments, company policy to limit
the number of free samples that can be given to a customer,
government policy of limiting the monetary investment by a
camp on a voter, etc. As we will see, bounded investments
in our model would result in bounded opinion values, which
as explained earlier, could be transformed into probability
of voting for a party or adopting a product, and hence the
expected number of votes or sales in the favor of each camp.
We first study in Section III, the cases of unbounded and
bounded investment in a fundamental setting where a camp's
influence on a node is a linear function of its investment.
While the linear influence function is consistent with the
well-established Friedkin-Johnsen model, the influence of a
camp on a node might not increase linearly with the corre-
sponding investment. In fact, several social and economic set-
tings follow law of diminishing marginal returns, which says
that for higher investments, the marginal returns (influence in
this context) are lower for a marginal increase in investment.
An example of this law is when we watch a particular product
advertisement on television; as we watch the advertisement
more number of times, its marginal influence on us tends to
get lower. A concave influence function naturally captures this
law. We study such an influence function in the settings of both
unbounded and bounded investment per node, and relate it to
the skewness of investment in optimal strategies as well as
user perception of fairness. We study this in Section IV.
There are scenarios where a camp may want to maximize
the total investment of the competing camp, so as to upset the
latter's broad budget allocation, which might lead to reduction
in its available budget for future investments or for other
channels such as mass media advertisement. The latter may
also be forced to implement unappealing actions such as
increasing the product cost or seek further monetary sources
in order to compensate for its investments. Alternatively, the
camps may have been instructed a desired investment strategy
by a mediator such as government or a central authority,
and deviating from this strategy would incur a penalty. For
instance,
the mediator itself would have its own broader
optimization problem (which could be for the benefit for it or
the society), whose optimal solution would require the camps
to devise their investment strategies in a particular desired
2
way. The mediator would then instruct the camps to follow
the corresponding desired investment strategies, and in case
of violation, the mediator could impose a penalty so as to
compensate for the suboptimal outcome of its own optimiza-
tion problem. For similar reasons as mentioned before, a camp
may want to maximize the penalty incurred by the competing
camp. We study these settings which capture the adversarial
behavior of a camp towards another camp, in Section V.
For all of the aforementioned settings, we show in this
paper that it does not matter whether the camps strategize
simultaneously or sequentially. We use Nash equilibrium as
the equilibrium notion to analyze the game in these settings.
However,
there could be settings where a sequential play
would be more natural than a simultaneous one, which would
result in a Stackelberg game. The sequence may be determined
by a mediator or central authority which, for example, may
be responsible for giving permissions for campaigning or
scheduling product advertisements to be presented to a node.
We use subgame perfect Nash equilibrium as the equilibrium
notion for the game in such sequential play settings. Moreover,
since we are concerned with a zero-sum game, we express the
equilibrium in terms of maxmin or minmax. Assuming the
good camp plays first (without loss of analytical generality),
the bad camp would choose a strategy that minimizes the sum
of opinion values as a best response to the good camp's strat-
egy. Knowing this, the good camp would want to maximize
this minimum value. We motivate two such settings.
It would often be the case that the total attention capacity
of a node or the time it could allot for receiving campaigning
from both camps combined, is bounded. This leads us to
study the game under common coupled constraints (CCC)
that the sum of investments by the camps on any node is
bounded. These are called common coupled constraints since
the constraints of one camp are satisfied if and only if the
constraints of the other camp are satisfied, for every strategy
profile. We study this setting in Section VI.
Another sequential setting is one that results in uncertainty
of information, where the good camp (which plays first) may
not have exact information regarding the network parameters.
However, the bad camp (which plays second) would have
perfect
information regarding these parameters, which are
either revealed over time or deduced based on the effect of
the good camp's investment. Forecasting the optimal strategy
of bad camp, we derive a robust strategy for the good camp
which would give it a good payoff even in the worst case. We
study this setting in Section VII.
It can be noted that the common coupled constraints setting
captures the first mover advantage, while the uncertainty
setting captures the first mover disadvantage.
B. Related Work
A principal part of opinion dynamics in a population is how
nodes update their opinions over time. One of the most well-
accepted and well-studied approaches of updating a node's
opinion is based on imitation, where each node adopts the
opinion of some of its neighbors with a certain probability.
One such well-established variant is DeGroot model [8] where
Acemoglu and Ozdaglar [4] review several other models of
opinion dynamics. Lorenz [11] surveys modeling frameworks
concerning continuous opinion dynamics under bounded con-
fidence, wherein nodes pay more attention to beliefs that do
not differ too much from their own. Xia, Wang, and Xuan
[12] give a multidisciplinary review of the field of opinion
dynamics as a combination of the social processes which are
conventionally studied in social sciences, and the analytical
and computational tools developed in mathematics, physics
and complex system studies. Das, Gollapudi, and Munagala
[13] show that the widely studied theoretical models of opinion
dynamics do not explain their experimental observations, and
hence propose a new model as a combination of the DeGroot
model and the Voter model [14], [15]. Parsegov et al. [16]
develop a multidimensional extension of Friedkin-Johnsen
model, describing the evolution of the nodes' opinions on
several interdependent topics, and analyze its convergence.
Ghaderi and Srikant [17] consider a setting where a node
iteratively updates its opinion as a myopic best response to
the opinions of its own and its neighbors, and hence study
how the equilibrium and convergence to it depend on the
network structure, initial opinions of the nodes, the location
of stubborn agents (forceful nodes with unchanging opinions)
and the extent of their stubbornness. Ben-Ameur, Bianchi, and
Jakubowicz [18] analyze the convergence of some widespread
gossip algorithms in the presence of stubborn agents and
show that the network is driven to a state which exclusively
depends on the stubborn agents. Jia et al. [19] propose an
empirical model combining the DeGroot and Friedkin models,
and hence study the evolution of self-appraisal, social power,
and interpersonal influences for a group of nodes who discuss
and form opinions. Halu et al. [20] consider the case of
two interacting social networks, and hence study the case of
political elections using simulations.
each node updates its opinion using a weighted convex com-
bination of its neighbors' opinions. The model developed by
Friedkin and Johnsen [9], [10] considers that, in addition to
its neighbors' opinions, a node also gives certain weightage
to its initial biased opinion.
Yildiz, Ozdaglar, and Acemoglu [21] study the problem of
optimal placement of stubborn agents in the discrete binary
opinions setting with the objective of maximizing influence,
given the location of competing stubborn agents. Gionis, Terzi,
and Tsaparas [1] study from an algorithmic and experimental
perspective, the problem of identifying a set of target nodes
whose positive opinions about an information item would
maximize the overall positive opinion for the item in the
network. Ballester, Calv´o-Armengol, and Zenou [22] study
optimal targeting by analyzing a noncooperative network game
with local payoff complementarities. Sobehy et al. [23] pro-
pose strategies to win an election using a Mixed Integer Linear
Programming approach.
The basic model we study is similar to that considered by
Grabisch et al. [2], that is, a zero-sum game with two camps
holding distinct binary opinion values, aiming to select a set
of nodes to invest on, so as to influence the average opinion
that eventually emerges in the network. Their study, however,
considers non-negative matrices and focuses on the existence
3
and the characterization of equilibria in a preliminary setting,
where the influence and cost functions are linear, camps have
network information with certainty, and there is no bound on
combined investment by the camps per node. Dubey, Garg,
and De Meyer [6] study existence and uniqueness of Nash
equilibrium, while also considering convex cost functions. The
study, however, does not consider the possibility of bounded
investment on a node, and the implications on the extent of
skewness of investment and user perception of fairness owing
to the convexity of cost functions. Bimpikis, Ozdaglar, and
Yildiz [7] provide a sharp characterization of the optimal
targeted advertizing strategies and highlight their dependence
on the underlying social network structure, in a preliminary
setting. Their study emphasizes the effect of absoption cen-
trality, which is encountered in our study as well.
The problem of maximizing information diffusion in social
networks under popular models such as Independent Cas-
cade and Linear Threshold, has been extensively studied [3],
[24], [25]. The competitive setting has resulted in several
game theoretic studies of this problem [26], [27], [28]. There
have been preliminary studies addressing interaction among
different informations, where the spread of one information
influences the spread of the others [29], [30].
There have been studies on games with constraints. A
notable study by Rosen [31] shows existence of equilibrium
in a constrained game, and its uniqueness in a strictly concave
game. Altman and Solan [32] study constrained games, where
the strategy set available to a player depends on the choice
of strategies made by other players. The authors show that,
in constrained zero-sum games, the value of the game need
not exist (that is, maxmin and minmax values need not be the
same) and contrary to general functions, maxmin value could
be larger than minmax.
The topic of decision under uncertainty has been of interest
to the game theory and optimization communities. An estab-
lished way of analyzing decision under uncertainty is using
robust optimization tools. Ben-Tal, El Ghaoui, and Nemirovski
[33] present a thorough review of such tools.
C. Contributions of the Paper
A primary goal of this work is to provide a broad framework
for optimal investment strategies for competing camps in a
social network, and propose and explore several aspects of
the problem. In particular, we study several well-motivated
variants of a constrained zero-sum game where two competing
camps aim to maximize the adoption of their respective
opinions, under the well-established Friedkin-Johnsen model
of opinion dynamics. Following are our specific contributions:
• We show that a random investment strategy is optimal
when the underlying network follows a particular popular
class of weight distributions. (Section III-C)
• We investigate when a camp's influence on a node is
a concave function of its investment on that node, for
the cases of unbounded and bounded investment per
node. We hence provide implications for the skewness
of optimal investment strategies and user perception of
fairness. (Section IV)
• We look at the complementary problem where a camp
acts as an adversary to the competing camp by aiming
to maximize the latter's investment. We also look at the
problem where a camp aims to maximize the deviation
from the desired investment of the competing camp.
(Section V)
• We study the Stackelberg variant under common coupled
constraints, that the combined investment by the good and
bad camps on any given node cannot exceed a certain
limit. We study the maxmin and minmax values and
present some interesting implications. (Section VI)
• We analyze a setting where one of the camps would need
to make decision under uncertainty. (Section VII)
• Using simulations, we illustrate our analytically derived
results on real-world social networks, and present further
insights based on our observations. (Section VIII)
v0
i
w0
ii
wig
wib
wij
xi
yi
kg
kb
vi
II. MODEL
where
Consider a social network with N as its set of nodes and
E as its set of weighted, directed edges. Two competing
camps (good and bad) aim to maximize the adoption of
their respective opinions in the social network. We consider a
strict competition setting where the opinion value of the good
camp is denoted by +1 and that of the bad camp by −1.
In this section, we present the parameters of the considered
model of opinion dynamics, and the update rule along with
its convergence result. We first provide an introduction to
the well-established Friedkin-Johnsen model, followed by our
proposed extension.
A. Friedkin-Johnsen Model
As per Friedkin-Johnsen model [9], [10], prior to the process
of opinion dynamics, every node holds a bias in opinion which
could have been formed owing to various factors such as the
node's fundamental views, its experiences, past information
from news and other sources, opinion dynamics in the past,
etc. We denote this opinion bias of a node i by v0
i and the
weightage that the node attributes to it by w0
ii.
connections is (cid:80)
The network effect is captured by how much a node is
influenced by each of its friends or connections, that is, how
much weightage is attributed by a node to the opinion of each
of its connections. Let vj be the opinion held by node j and
wij be the weightage attributed by node i to the opinion of
node j. The influence on node i owing to node j is given
by wijvj, thus the net influence on i owing to all of its
j∈N wijvj (where wij (cid:54)= 0 only if j is a
connection of i). It is to be noted that we do not make any
assumptions regarding the sign of the edge weights, that is,
they could be negative as well (as justified in [34], [35]). A
negative edge weight wij can be interpreted as some form of
distrust that node i holds on node j, that is, i would be driven
towards adopting an opinion that is opposite to that held or
suggested by j.
Since in Friedkin-Johnsen model, each node updates its
opinion using a weighted convex combination of its neighbors'
opinions, the update rule is given by
∀i ∈ N : vi ← w0
iiv0
i +
(cid:88)
j∈N
4
TABLE I
NOTATION TABLE
the initial biased opinion of node i
weightage given to the initial opinion by node i
weightage given by node i to the good camp's opinion
weightage given by node i to the bad camp's opinion
weightage given by node i to the opinion of node j
investment made by good camp to directly influence node i
investment made by bad camp to directly influence node i
budget of the good camp
budget of the bad camp
the resulting opinion of node i
∀i ∈ N : w0
ii +
(cid:88)
j∈N
wij ≤ 1
B. Our Extended Model
We extend Friedkin-Johnsen model
to incorporate the
camps' investments and the weightage attributed by nodes
to the camps' opinions. The good and bad camps attempt to
directly influence the nodes so that their opinions are driven
towards being positive and negative, respectively. This direct
influence depends on the investment or effort made by the
camps, and also on how much a node weighs the camps'
opinions. A given amount of investment may have different
influence on different nodes based on how much these nodes
weigh the camps' recommendations. We denote the investment
made by the good and bad camps on node i by xi and yi
respectively, and the weightage that node i attributes to them
by wig and wib respectively. Since the influence of good camp
on node i would be an increasing function of both xi and wig,
we assume the influence to be wigxi so as to maintain the
multilinearity of Friedkin-Johnsen model. Similarly, wibyi is
the influence of bad camp on node i. Also note that since the
good and bad camps hold the opinions +1 and −1 respectively,
the net influence owing to the direct recommendations from
these camps is (wigxi − wibyi).
The camps have budget constraints stating that the good
camp can invest a total amount of kg across all the nodes,
while the bad camp can invest a total amount of kb.
Table I presents the required notation. Consistent with the
standard opinion dynamics models, we have the condition
on the influence weights on any node i that they sum to
at maximum 1 (since a node updates its opinions using a
weighted 'convex' combination of the influencing factors).
∀i ∈ N :
w0
ii +
wij + wig + wib ≤ 1
A standard assumption for guaranteeing convergence of the
dynamics is
wij < 1
(cid:88)
(cid:88)
j∈N
j∈N
wijvj
This assumption is actually well suited for our model where
we would generally have non-zero weights attributed to the
influence outside of the network, namely, the influence due to
bias (w0
ii) and campaigning (wig, wib).
Nodes update their opinions in discrete time steps starting
with time step 0. With the aforementioned factors into consid-
eration, each node i updates its opinion at each step, using the
following update rule (an extension of the Friedkin-Johnsen
update rule):
∀i ∈ N : vi ← w0
(cid:104)τ(cid:105)
i
iiv0
i +
wijvj + wigxi − wibyi
(cid:104)0(cid:105)
i .
i = v0
be the opinion of node i at time step τ, and v
Let v
The update rule can hence be written as
(cid:104)τ−1(cid:105)
wijv
j
∀i ∈ N : v
(cid:104)τ(cid:105)
i = w0
(cid:88)
iiv0
i +
+ wigxi − wibyi
(1)
(cid:88)
j∈N
j∈N
i (weighed
For any node i, the static components are xi, yi, v0
by wig, wib, w0
ii), while the dynamic components are vj's
(weighed by wij's). The static components remain unchanged
while the dynamic ones get updated in every time step.
i and w0
Let w be the matrix consisting of the elements wij for each
pair (i, j) (note that w contains only the network weights
and not wig, wib, w0
ii). Let v be the vector consisting of the
opinions vi, v0 and w0 be the vectors consisting of the
elements v0
ii respectively, x and y be the vectors
consisting of the investments xi and yi respectively, wg and
wb be the vectors consisting of the weights wig and wib
respectively. Let the operation ◦ denote Hadamard product
(elementwise product) of vectors, that is, (a ◦ b)i = aibi.
Let Hadamard power be expressed as (a◦p)i = ap
i .
Assuming v(cid:104)τ(cid:105) to be the vector consisting of the opinions
(cid:104)τ(cid:105)
, the update rule (1) can be written in matrix form as
i
v
(cid:104)τ(cid:105)
v
= wv
(cid:104)τ−1(cid:105)
+ w0 ◦ v0 + wg ◦ x − wb ◦ y
(2)
Proposition 1. The dynamics defined by the update rule in
(2) converges to v = (I− w)−1(w0 ◦ v0 + wg ◦ x− wb ◦ y).
Proof. The recursion in (2) can be simplified as
(cid:16)
(cid:104)τ(cid:105)
v
= wv
(cid:104)τ−1(cid:105)
= w
wv
+ w0 ◦ v0 + wg ◦ x − wb ◦ y
(cid:104)τ−2(cid:105)
+ w0 ◦ v0 + wg ◦ x − wb ◦ y
(cid:17)
+ w0 ◦ v0 + wg ◦ x − wb ◦ y
= w2v
= w3v
(cid:104)τ−2(cid:105)
(cid:104)τ−3(cid:105)
+ (I + w)(w0 ◦ v0 + wg ◦ x − wb ◦ y)
+ (I + w + w2)(w0 ◦ v0 + wg ◦ x − wb ◦ y)
= wτ v
(cid:104)0(cid:105)
+
wη
(w0 ◦ v0 + wg ◦ x − wb ◦ y)
η=0
substochastic matrix, since ∀i ∈ N : (cid:80)
limτ→∞ wτ v0 = 0. Furthermore, limτ→∞(cid:80)τ−1
Now, the initial opinion: v(cid:104)0(cid:105) = v0. Also, w is a strictly
j∈N wij < 1;
its spectral radius is hence less than 1. So when τ → ∞,
we have limτ→∞wτ = 0. Since v0 is a constant, we have
η=0 wη = (I −
w)−1, an established matrix identity [36]. This implicitly
means that (I − w) is invertible. Hence,
(cid:104)τ(cid:105)
= (I − w)
−1(w0 ◦ v0 + wg ◦ x − wb ◦ y)
lim
τ→∞ v
which is a constant vector, that is, the dynamics converges to
this steady state of opinion values.
(cid:32)τ−1(cid:88)
(cid:33)
5
III. THE FUNDAMENTAL PROBLEM
We now present the fundamental problem of competitive
opinion dynamics under the Friedkin-Johnsen model.
A. Introduction of the Fundamental Problem
The problem of maximizing opinion adoption can be mod-
eled as an optimization problem. In particular, considering
perfect competition, this problem can be modeled as a maxmin
problem as we now present. Here our objective is to determine
the strategies of the good and bad camps (the values of xi and
yi such that they satisfy certain constraints), so that the good
camp aims to maximize the sum of opinion values of the nodes
while the bad camp aims to minimize it. Considering linear
constraints for setting the problem in the linear programming
framework, we represent these constraints by Ax ≤ b and
Cy ≤ d, respectively, where A, C are matrices and b, d are
vectors, in general.
Owing to xi and yi being investments, we have the natural
constraints: xi, yi ≥ 0,∀i ∈ N. We can hence write the
maxmin optimization problem in its general form as
max
Ax≤b
x≥0
s.t. ∀i ∈ N : vi = w0
iiv0
i +
min
Cy≤d
y≥0
i∈N
wijvj + wigxi − wibyi
(cid:88)
(cid:88)
j∈N
vi
From Proposition 1, we have
v = (I − w)
−1(w0 ◦ v0 + wg ◦ x − wb ◦ y)
Let
=⇒ 1T v = 1T (I − w)
−1(w0 ◦ v0 + wg ◦ x − wb ◦ y)
the constant 1T (I − w)−1 = (cid:0)((I − w)−1)T 1(cid:1)T
(cid:0)(I − wT )−11(cid:1)T
(cid:88)
(cid:88)
(cid:88)
(cid:88)
riwigxi −(cid:88)
riwigxi −(cid:88)
(cid:88)
= rT . So the above is equivalent to
So our objective function becomes
riwibyi
riwibyi
riw0
riw0
iiv0
vi =
i∈N
i∈N
iiv0
i +
i∈N
i∈N
i +
max
Ax≤b
x≥0
min
Cy≤d
y≥0
i∈N
i∈N
i∈N
=
(3)
This can be achieved by solving two independent optimization
problems, namely,
max
Ax≤b
x≥0
riwigxi
and min
Cy≤d
y≥0
riwibyi
which can be easily solved.
1) The Specific Case: Overall Budget Constraints: For
studying the problem in a broader framework, we consider
the case, specific to our model that we introduced in Section
II. This case that considers overall budget constraints kg and
kb for the good and bad camps respectively, corresponds
i∈N yi ≤ kb. That is, we have
A = C = 1T , b = kg, d = kb. It is clear that the solution
to this specific optimization problem is
i∈N xi ≤ kg and (cid:80)
to (cid:80)
(cid:88)
i = 0 , ∀i /∈ arg max
∗
x
∗
x
i = kg ,
i∈N
if max
i∈N
riwig
riwig > 0
(4)
i∈arg maxi riwig
x∗
i ≥0
(cid:88)
i∈N
(cid:88)
i∈N
and
(cid:88)
i∈arg maxi riwib
y∗
i ≥0
i = 0 , ∀i /∈ arg max
∗
y
∗
i = kb ,
i∈N
if max
i∈N
y
riwib
riwib > 0
(5)
Note that if maxi∈N riwig ≤ 0, then x∗
i = 0,∀i ∈ N.
maxi∈N riwib ≤ 0, then y∗
i = 0,∀i ∈ N and if
Equations (4) and (5) lead to the following result.
Proposition 2. In Setting III-A1, it is optimal for the good
and bad camps to invest their entire budgets in node i with
maximum value of riwig and riwib respectively, subject to the
value being positive.
Insight 1. Parameter ri could be interpreted as the influencing
power of node i on the network, while wig and wib are
respectively the influencing powers of the good and bad camps
on node i. So it is clear why these parameters factor into the
result. Furthermore, the strategies of the camps are mutually
independent, which arises from the sum of steady state values
of nodes as derived in (3). The multilinearity of the model
and unconstrained investment on nodes allow the camps to
exhaust their budgets by concentrating their entire investments
on a node possessing the highest value of riwig or riwib
respectively. Also, the camps' strategies are independent of the
initial opinions, since they aim to optimize the sum of opinion
values without considering their relative values.
1 =(cid:0)I − αAT(cid:1)−1
(cid:16)(cid:0)I − αAT(cid:1)−1− I
(cid:17)
r =(cid:0)I − wT(cid:1)−1
Actually, ri can be viewed as a variant of Katz centrality
[37] in that, Katz centrality of node i measures its relative
influence in a social network (say having adjacency matrix
A) with all edges having the same weight (say α), while ri
measures its influence in a general weighted social network.
Katz centrality of node i is defined as the ith element of vector
1 − 1, for 0 < α < 1ρ
where ρ is the largest eigenvalue of A. In our case where
1, A is replaced by the weighted adjacency
matrix w, for which ρ < 1 (since w is strictly substochastic),
and we have α = 1. The subtraction of vector 1 is common
for all nodes, so its relative effect can be ignored. ri can also
be viewed as a variant of absorption centrality of node i [7],
which captures the expected number of visits to node i in
a random walk starting at a node other than i uniformly at
random, with transition probability matrix w (assuming all
elements of w to be non-negative).
Furthermore, recall that
rT = 1T (I − w)
−1 = 1T(cid:16)
I +
wη(cid:17)
∞(cid:88)
η=1
So if we have wij ≥ 0 for all pairs of nodes (i, j), we will
have that all elements of vector r are at least 1. That is, wij ≥
0,∀(i, j) =⇒ ri ≥ 1,∀i ∈ N.
2) The Case of Bounded Investment Per Node: This setting,
includes an additional bound on the
as motivated earlier,
investment per node by a camp. We assume this bound to be
1 unit without loss of generality, that is, xi, yi ≤ 1,∀i ∈ N.
With respect to the generic constraints Ax ≤ b and Cy ≤ d,
(cid:18) kb
this case corresponds to A = C =
, d =
(cid:18) 1T
(cid:18) kg
, b =
(cid:19)
(cid:19)
(cid:19)
1
I
.
1
6
From Equation (3), an optimal x can be obtained as follows.
Let Iriwig>0 = 1 if riwig > 0, and 0 otherwise. Let
ω1, ω2, . . . , ωn be the ordering of nodes in decreasing values
of riwig with any tie-breaking rule. So (3) is maximized with
respect to x when
xi = 1 · Iriwig >0, for i = ω1, . . . , ω(cid:98)kg(cid:99)
xi = (kg − (cid:98)kg(cid:99)) · Iriwig >0, for i = ω(cid:98)kg(cid:99)+1
xi = 0, for i = ω(cid:98)kg(cid:99)+2, . . . , ωn
An optimal y is analogous, hence the following result.
Proposition 3. In Setting III-A2, it is optimal for the good
camp to invest in nodes one at a time, subject to a maximum
investment of 1 unit per node, in decreasing order of values
of riwig until either the budget kg is exhausted or we reach a
node with a non-positive value of riwig. The optimal strategy
of the bad camp is analogous.
Also, from Proposition 1, if a camp's investment per node is
bounded by 1 unit, the opinion value of every node would be
bounded between −1 and +1. As stated earlier, such bounded
opinion value is relevant to elections and product adoption
scenarios, where the bounded opinion value of a node could
be translated into the probability of the node voting for a camp
or adopting a particular product.
B. Maxmin versus Minmax Values
With no bounds on investment per node, it is clear that the
maxmin and minmax values are the same, since the strategies
of the camps are mutually independent, that is,
max
x≥0
min
y≥0
vi = min
y≥0
max
x≥0
vi
(6)
The equality would hold even with mutually independent
bounds on the camps' investment on a node, that is,
max
0≤x≤1
min
0≤y≤1
vi = min
0≤y≤1
max
0≤x≤1
vi
(7)
seen using Equation (3); the value of(cid:80)
would stay the same while the value of(cid:80)
It is to be noted that we cannot compare the values in (6) and
(7), in general. For instance, if all i's have equal values of
riwib and only one i has good value of riwig, then for kg > 1,
the value in (6) would be greater than that in (7). This can be
i − wibyi)
i∈N riwigxi would
be higher in (6) than in (7). On the other hand, if all i's have
equal values of riwig and only one i has good value of riwib,
then for kb > 1, the value in (7) would be greater than that in
(6).
i∈N ri(w0
iiv0
C. Result for a Popular Class of Weight Distributions
We now present a result concerning a class of distribution
of edge weights in a network, which includes the popular
weighted cascade (WC) model.
Proposition 4. Let Ni = {j : wij (cid:54)= 0}, di = Ni, and
j ∈ Ni ⇐⇒ i ∈ Nj. If ∀i ∈ N , wig = wib = w0
ii = 1
=
wij,∀j ∈ Ni, where α > 0, then riwig = riwib = 1
α ,∀i ∈ N.
α+di
(cid:88)
i∈N
(cid:88)
i∈N
(cid:88)
i∈N
(cid:88)
i∈N
Proof. We know that
(cid:0)I − wT(cid:1)r = 1
⇐⇒ r = 1 + wT r
⇐⇒ ∀i ∈ N : ri = 1 +
(cid:88)
j∈Ni
wjirj = 1 +
(cid:18) 1
α + dj
(cid:19)
rj
(cid:88)
j∈Ni
Let us assume ri = γ(α + di), where γ is some constant. If
this satisfies the above equation, the uniqueness of ri ensures
that it is the only solution. Hence we have
∀i ∈ N : γ(α + di) = 1 +
⇐⇒ γ =
1
α
(cid:88)
j∈Ni
γ = 1 + γdi
∴ ∀i ∈ N : riwig = riwib =
α + di
α
·
1
α + di
=
1
α
ii = 1
The above result implies that models which assign weights
= wij,∀j ∈ Ni,
for all i such that wig = wib = w0
are suitable for the use of a random strategy, since the decision
parameter for either camp (riwig, riwib) holds the same value
for all nodes. That is, in these models, a random strategy that
exhausts the entire budget is optimal. This class of models
includes the popular weighted cascade model, which would
assign the weights with α = 3.
α+di
IV. EFFECT OF CONCAVE INFLUENCE FUNCTION
The linear influence function (1) without any bound on
investment per node, leads to an optimal strategy that con-
centrates the investment on a single node (Proposition 2). As
motivated in Section I-A, several social and economic settings
follow law of diminishing marginal returns, which says that
for higher investments, the marginal returns (influence in our
context) are lower for a marginal increase in investment. A
concave influence function would account for such diminish-
ing marginal influence of a camp with increasing investment
on a node which, as we shall see, would advise against
concentrated investment on a single node. For the purpose of
our analysis so as to arrive at precise closed-form expressions
and specific insights, we consider a particular form of concave
functions: x1/t
i when the investment is xi. It is to be noted,
however, that it can be extended to other concave functions
since we use a common framework of convex optimization,
however the analysis could turn out to be more complicated
or intractable.
A. The Case of Unbounded Investment per Node
min(cid:80)
i yi≤kb
yi≥0
i − wiby1/t
wijvj + wigx1/t
s.t. ∀i ∈ N : vi = w0
max(cid:80)
i xi≤kg
xi≥0
(cid:88)
iiv0
i∈N
i +
vi
i
(cid:88)
j∈N
t
Proposition 5. In Setting IV-A, for t > 1, it is optimal for
the good and bad camps to invest in node i proportional to
t−1 , subject to positivity of riwig and
(riwig)
riwib respectively.
A proof of Proposition 5 is provided in Appendix A.
t−1 and (riwib)
t
7
Remark 1 (Skewness of investment). When we compare
the results for lower and higher values of t, the investment
made by the good camp has an exaggerated correlation with
the value of riwig for lower values of t. In particular, the
investment made is very skewed towards nodes with high
values of riwig when t is very low, while it is proportional
to riwig when t is very high. Note that t = 1 corresponds
to the linear case in Setting III-A1 where the investment is
extremely skewed with each camp investing its entire budget
on only one node.
Remark 2 (User perception of fairness). The skewness can
be linked to user perception of fairness [38]. Suppose a node
p is such that rpwpg = maxi riwig, and it is the unique
node with this maximum value. Suppose a node q is such that
rqwqg = rpwpg − , where is positive and infinitesimal.
From the perspective of node q, the strategy would be fair
if the investment in q is not much less than that in p, since
they are almost equally valuable. However, t = 1 leads to a
highly skewed investment where p receives kg and q receives
0, which can be perceived as unfair by q. As t increases, the
investment becomes less skewed; in particular, t → ∞ leads
to investment on a node i to be proportional to riwig, which
could be perceived as fair by the nodes.
B. The Case of Bounded Investment Per Node
With the additional constraints xi ≤ 1 and yi ≤ 1,∀i ∈ N,
the optimal investment strategies are given by Proposition 6.
We provide its proof in Appendix B.
Proposition 6. In Setting IV-B, if the number of nodes with
riwig > 0 is less than kg, it is optimal for good camp to invest
1 unit on each node i with riwig > 0 and 0 on all other nodes.
If the number of nodes with riwig > 0 is at least kg, let γ > 0
be the solution of(cid:88)
(cid:18) riwig
(cid:19) t
t−1
(cid:88)
i:riwig∈(0,tγ]
tγ
+
1 = kg
i:riwig >tγ
It can be shown that γ exists and is unique; it is then optimal
for the good camp to follow the investment strategy:
x
x
i = 0, if riwig ≤ 0
∗
∗
i = 1, if riwig > tγ
kg − (cid:88)
∗
i =
x
1
i:riwig >tγ
(cid:80)
,
t
t−1
(riwig)
t
t−1
i:riwig∈(0,tγ](riwig)
if riwig ∈ (0, tγ]
(cid:16) riwig
(cid:17) t
The optimal strategy of the bad camp is analogous.
tγ
Note that for riwig ∈ (0, tγ], we can alternatively write
x∗
t−1 , which would be between 0 and 1. So the
i =
nodes with positive values of riwig should be classified into
two sets, one containing nodes with riwig ∈ (0, tγ] (for which
i ∈ (0, 1]) and the other containing nodes with riwig > tγ
x∗
(x∗
i forcefully limited to 1). So we can effectively start with
all nodes in the former set (meaning tγ ≥ maxi∈N riwig) and
then transfer nodes to the latter set as per descending values
of riwig (as we reduce tγ), until we have two sets, one with
(cid:33)
(cid:16) riwig
(cid:17) t
(cid:32)
(cid:80)
t
t
tγ
=
t−1
t−1
t−1
(riwig)
≤ 1 and the
i:riwig∈(0,tγ](riwig)
i forcefully limited to 1.
x∗
i =
other with x∗
Insight 2. The solution suggests that the optimal strategy
can be obtained using a trial-and-error iterative process. A
camp could use the optimal strategy for the unbounded case
suggested in Proposition 5. If we get x∗
i > 1 for any node, we
assign x∗
i = 1 to node i with the highest value of riwig, and
use Proposition 5 again by excluding node i and decrementing
the available budget by 1. This process would be repeated until
i ≤ 1,∀i ∈ N.
x∗
V. ACTING AS COMPETITOR'S ADVERSARY
In this setting, a camp explicitly acts to maximize the
competitor's investment or deviation from its desired invest-
ment, that is required to drive the sign of the average opinion
value of the population in the latter's favor. Without loss of
analytical generality, we consider that the good camp acts as
the adversary.
A. The Case of Unbounded Investment per Node
max(cid:80)
s.t. (cid:88)
i xi≤kg
xi≥0
min
yi≥0
i∈N
vi ≤ 0
i∈N
yi
(cid:88)
i ) −(cid:88)
i∈N
ri(wigxi + w0
iiv0
riwibyi
(cid:88)
i∈N
vi =
(cid:88)
i∈N
8
(cid:88)
i∈N
(cid:88)
(yi − ¯yi)2
(cid:80)
i(xi−¯xi)2≤kg
max
xi≥0
min
yi≥0
s.t. (cid:88)
i∈N
vi ≤ 0
(cid:88)
(cid:18) riwig
(cid:88)
i∈N
i∈N
(cid:19)2
i ) −(cid:88)
(cid:88)
i∈N
+
(¯xi)2 = kg
2γ
i:riwig
≥−2γ ¯xi
i:riwig
<−2γ ¯xi
vi =
ri(wigxi + w0
iiv0
riwibyi
Let γ > 0 be the solution of
2
2
(¯xi)2
(cid:32)
(riwig)2
i:riwig
<−2γ ¯xi
1
= ¯xi + sgn (riwig)
kg − (cid:88)
Then the good camp's optimal strategy is the following:
i = 0, if riwig < −2γ ¯xi
∗
x
(cid:33) 1
(cid:80)
If there does not exist a γ > 0 (because(cid:80)
(cid:33)
(riwig)2
i:riwig
≥−2γ ¯xi
if riwig ≥ −2γ ¯xi
i:riwig<0(¯xi)2 < kg
and no node with riwig > 0), we invest 0 on any node with
riwig < 0 and ¯xi on any node with riwig = 0.
This can be proved on similar lines as Proposition 6. Here,
kg − (cid:88)
i − ¯xi)2 =
∗
(riwig)2
(cid:32)
(¯xi)2
(x
(cid:80)
i:riwig
<−2γ ¯xi
(riwig)2
i:riwig
≥−2γ ¯xi
and the optimal square root is determined by sgn (riwig)
(since a positive riwig would mean a higher optimal invest-
ment as opposed to a negative riwig). Here, it is possible that a
node i is invested on by the good camp even if it has negative
riwig, so as to have the investment close to ¯xi.
B. The Case of Bounded Investment per Node
The optimal strategies of the camps can be easily obtained
for this setting on similar lines as Proposition 3.
Proposition 8. In Setting V-B, it is optimal for the good camp
to invest in nodes one at a time, subject to a maximum invest-
ment of 1 unit per node, in decreasing order of values of riwig
until either the budget kg is exhausted or we reach a node
with a non-positive value of riwig. Say the so derived optimal
investment on node i is x∗
i . The optimal strategy of the bad
camp is to invest in nodes one at a time, subject to a maximum
investment of 1 unit per node, in decreasing order of values
of riwib until(cid:80)
(cid:80)
(cid:80)
Note that
j∈N rj(wjgx∗
j + w0
i∈N vi ≤ 0, when x∗
i∈N riwibyi ≥(cid:80)
the terminating condition (cid:80)
j∈N rj(wjgx∗
j + w0
i∈N riwibyi ≥
j ) is same as the required condition
jjv0
j is the optimal investment by the good
j ).
jjv0
camp on node j.
VI. COMMON COUPLED CONSTRAINTS RELATING
BOUNDS ON COMBINED INVESTMENT PER NODE
As motivated in Section I-A, a sequential play would be
more natural than a simultaneous one in certain scenarios, for
instance, in presence of a mediator or central authority which
Proposition 7. In Setting V-A, it is optimal for the good camp
to invest its budget in node i with the maximum value of riwig,
subject to it being positive. For the bad camp, it is optimal
to invest in node i with maximum value of riwib, subject to
its positivity. (If there does not exist any node i with positive
value of riwib, it is optimal for the bad camp to not invest at
all). The optimal amount of investment made by the bad camp
is
(cid:40)
(cid:32)
(cid:26)
max
1
maxj∈N wjbrj
kg max
max
i∈N
riwig, 0
+
riw0
iiv0
i
, 0
(cid:27)
(cid:88)
i∈N
(cid:33)
(cid:41)
A proof of Proposition 7 is provided in Appendix C.
iiv0
It is to be noted that, contrary to the previous settings, the
amount of investment made by the bad camp in this setting is
dependent on the good camp's parameters (riwig and kg) as
well as the opinion bias parameters (w0
i ). This is because
in the previous settings, the bad camp's objective was to mini-
mize the sum of opinion values without considering the actual
value of this sum, while the current setting necessitates the
bad camp to ensure that this sum is non-positive; this requires
taking into account the effects of good camp's influence and
the initial biases on this sum.
Remark 3 (Maximizing competitor's deviation). Let the de-
sired investments for the good and bad camps be ¯xi and ¯yi,
respectively. Thus the optimization problem is
may be responsible for giving permissions for campaigning
or scheduling product advertisements to be presented to an
individual. We hence consider two sequential play settings,
which result in Stackelberg variants of the considered game.
We use subgame perfect Nash equilibrium as the equilibrium
notion; also since it is a zero-sum game, we refer to the
equilibrium as either maxmin or minmax, based on which
camp plays first. Without loss of analytical generality, we
conduct our analysis while assuming the good camp plays first.
The bad camp would hence choose a strategy that minimizes
the sum of opinion values as a best response to the good
camp's strategy. Foreseeing this, the good camp would want
to maximize this minimum value; this is popularly known
as the backward induction approach. We hence derive the
subgame perfect Nash equilibrium strategy profile and the
corresponding maxmin value. The minmax profile and value
can be obtained symmetrically.
In this section, we consider a setting in which the combined
investment on a node by both camps is bounded by a certain
limit. Without loss of generality, we assume this limit to be
1 unit. This leads to the introduction of common coupled
(cid:88)
constraints (CCC): xi + yi ≤ 1,∀i ∈ N.
max(cid:80)
i xi≤kg
xi≥0
vi
i∈N
min(cid:80)
i yi≤kb
0≤yi≤(1−xi)
(cid:88)
j∈N
s.t. ∀i ∈ N : vi = w0
iiv0
i +
wijvj + wigxi − wibyi
(cid:88)
The inner term is
(cid:88)
yi ≤ kb or −(cid:88)
i∈N
min
vi
s.t. y ≥ 0
yi ≥ −kb ← α
∀i ∈ N : vi −(cid:88)
i∈N
wijvj + wibyi = wigxi + w0
i ← zi
∀i ∈ N : xi + yi ≤ 1 or − yi ≥ −(1 − xi) ← γi
iiv0
j∈N
i∈N
Its dual problem can be written as
max −αkb +
(cid:88)
(cid:0)zi(wigxi + w0
s.t. ∀i ∈ N : zi −(cid:88)
i∈N
i ) − γi(1 − xi)(cid:1)
iiv0
wjizj = 1 ← vi
∀i ∈ N : wibzi − γi − α ≤ 0 ← yi
α ≥ 0
∀i ∈ N : zi ∈ R, γi ≥ 0
j∈N
As earlier, from (9), we have zi =(cid:0)(I − wT )−11(cid:1)
satisfying Constraint (10), it is required that
(8)
(9)
(10)
i = ri. For
∀i ∈ N : riwib − γi − α ≤ 0 or γi ≥ riwib − α
To maximize objective function (8), it is required that γi
should be as low as possible (knowing that 1 − xi ≥ 0). So
the above condition γi ≥ riwib − α along with γi ≥ 0 gives
(11)
∀i ∈ N : γi = max{riwib − α, 0}
9
So we need to maximize the objective function with respect
to γi, xi,∀i ∈ N and α. For this purpose, let us define a set
with respect to α, namely,
Jα = {j : rjwjb − α ≥ 0}
So the objective function to be maximized is
(rjwjb − α)(1 − xj) +
ri(wigxi + w0
iiv0
i ) (12)
−αkb − (cid:88)
(cid:16)(cid:88)
j∈Jα
which is equal to
(1−xj) − kb
α
j∈Jα
(cid:17) −(cid:88)
j∈Jα
(cid:88)
i∈N
(cid:88)
i∈N
(1−xj)rjwjb +
ri(wigxi +w0
iiv0
i )
(13)
to search the values of α ∈
It
is sufficient
Claim 1.
{rjwjb}j:rj wjb>0 ∪ {0} to find an optimal solution.
Proof. Since α ≥ 0, we have α (cid:54)= rjwjb for any rjwjb < 0.
Consider a range [rlwlb, ruwub] for a consecutive pair of
distinct values of rjwjb. If a range has both these values
negative, we do not search for α in that range, since α ≥ 0.
If a range has rlwlb ≤ 0 and ruwub > 0, we search for α in
[0, ruwub]. We will now determine an optimal value of α in
the valid searchable subset of [rlwlb, ruwub], for a given x.
We have Jα = {j : rjwjb ≥ rlwlb}. The value of the objective
function (12) becomes
Case 1: If α = rlwlb (where rlwlb ≥ 0):
(cid:88)
(rjwjb − rlwlb)(1 − xj) +
ri(wigxi + w0
iiv0
i )
i∈N
j∈Jα
−rlwlbkb − (cid:88)
Case 2a: If (cid:80)
Case 2b: Instead, if(cid:80)
j∈Jα
Case 2: If α ∈ (rlwlb, ruwub]:
We have Jα = {j : rjwjb ≥ ruwub}.
j∈Jα
(1 − xj) − kb ≥ 0, we have an optimal
α = ruwub (from (13)).
(1−xj)−kb < 0, we have optimal
α → rlwlb if rlwlb ≥ 0, and the optimal value is the same as
that for α = rlwlb (Case 1). Note here that if rlwlb < 0 and
ruwub ≥ 0, we would have optimal α = 0.
Case 3: If rlwlb = maxi∈N riwib, that is, when we are look-
ing for α ≥ maxi∈N riwib. For α = maxi∈N riwib, we have
(rjwjb −
α)(1− xj) in (12) vanishes. For α > maxi∈N riwib, we have
(rjwjb − α)(1 − xj) in
(12) vanishes in this case too. So the objective function to
be maximized becomes
−αkb +
Jα = {arg maxi∈N riwib} and so the term (cid:80)
Jα = {} and so the term (cid:80)
(cid:88)
ri(wigxi + w0
j∈Jα
j∈Jα
iiv0
i )
i∈N
for which the optimal α = maxi∈N riwib (the lowest value of
α such that α ≥ maxi∈N riwib).
The above cases show that it is sufficient to search the values
of α ∈ {rjwjb}j:rj wjb>0 ∪ {0} to determine an optimal value
of the objective function.
Now that we have established that the only possible values
of optimal α are {rjwjb}j:rj wjb>0 ∪ {0}, we can assume
optimal α = rjwjb for j ∈ {j : rjwjb > 0} ∪ {d}, where
the dummy node d is such that rdwdb = 0.
Recalling the objective function in (13),
ri(wigxi +w0
iiv0
i ) + α
(1−xj) − kb
(1−xj)rjwjb
α = rjwjb, objective function (12) becomes
i∈N
(cid:88)
(cid:88)
(cid:88)
(cid:34)(cid:88)
i∈N
=
=
(cid:16)(cid:88)
(cid:34)(cid:88)
j∈Jα
j∈Jα
j∈Jα
(cid:17) −(cid:88)
(cid:16)
kb −(cid:88)
(cid:35)
j∈Jα
(cid:17)(cid:35)
The term(cid:80)
(cid:88)
max
x,j
i∈N
(cid:104)
⇐⇒ max
j
ri(wigxi +w0
iiv0
i ) −
(1−xj)rjwjb + α
(1−xj)
ri(wigxi +w0
iiv0
i )
j∈Jα
(1−xj)rjwjb +
i∈N
−
Let Iα = {j : rjwjb > α}, Pα = {j : rjwjb = α = rjwjb}.
(cid:88)
So the objective function is
rjwjb
ri(wigxi +w0
iiv0
i ) −
(cid:17)
j∈Jα
(1−xj)
(cid:16)
kb − (cid:88)
(cid:34) (cid:88)
(1−xj)rjwjb − (cid:88)
(cid:17)
(cid:16)
kb − (cid:88)
(cid:16)
kb − (cid:88)
(1−xj)
(1−xj)
j∈Pα
(cid:17)
j∈Iα
j∈Pα
rjwjb
rjwjb
j∈Iα
(cid:35)
i∈N
=
+
(cid:88)
(cid:88)
(cid:34)(cid:88)
i∈N
−
j∈Iα
(1−xj)rjwjb +
j∈Iα
ri(wigxi +w0
iiv0
i )
(1−xj)rjwjb +
(1−xj)rjwjb
(cid:35)
ii, v0
j∈Iα
j∈Iα
Comparing this with generic objective function (3) and since
it should hold for any ri, wig, wib, w0
i , it is necessary that
the coefficients of non-zero values of riwib are the same in
both forms of the objective function. This along with the fact
yj =
j∈Iα
yj = kb −
yj =
j /∈Jα
that ∀j ∈ Iα : rjwjb > 0 (since α ≥ 0), gives (cid:80)
(1 − xj). Also if rjwjb > 0, then (cid:80)
(cid:80)
(1− xj). And for all other terms, we have(cid:80)
(cid:80)
yj = kb −(cid:88)
(cid:88)
0. Since ∀j ∈ N : 0 ≤ yj ≤ 1 − xj, these are equivalent to
∀j ∈ Iα : yj = 1−xj ,∀j /∈ Jα : yj = 0 ,
(1−xj)
(14)
To check for the consistency of budget of the bad camp,
yj ≤ kb. This gives the constraint
j∈Pα
j∈Pα
j∈Iα
j∈Iα
j∈Iα
j∈Iα
xj ≥ Iα − kb
(1 − xj) ≤ kb or equivalently,
it is necessary that (cid:80)
(cid:80)
(cid:88)
necessary that (cid:80)
yj ≤ (cid:80)
kb −(cid:80)
(1 − xj) ≤(cid:80)
(cid:88)
(cid:88)
is necessary that(cid:80)
xj ≤ kg and(cid:80)
j∈Pα
j∈Iα
j∈Pα
j∈Iα
xj +
Also if rjwjb > 0, for the consistency of investment on
the nodes in Pα (that is, ∀j ∈ Pα : xj + yj ≤ 1), it is
(1 − xj) or equivalently,
j∈Pα
(1 − xj) or equivalently,
j∈Pα
xj ≤ Iα + Pα − kb
(16)
To check for the consistency of budget of the good camp, it
xj ≥
j∈Iα
0. These along with Inequalities (15) and (16) give Iα−kb ≤
kg and Iα + Pα − kb ≥ 0, or equivalently,
j∈Pα
j∈Iα
Iα ≤ kg + kb
and
Iα + Pα ≥ kb
(17)
The sets Iα and Pα depend only on j. So let the set of j's
that satisfy the constraints in (17) be denoted by J, that is,
J = {j : Iα ≤ kg + kb and Iα + Pα ≥ kb}
xj +(cid:80)
(15)
i∈N riw0
iiv0
i being a constant, and substituting
(cid:88)
xj(rjwjb − rjwjb)
riwigxi +
− (cid:88)
(cid:88)
−(cid:88)
i∈N
x
i∈N
max
j:rj wjb≥rj wjb
j:rj wjb≥rj wjb
(rjwjb − rjwjb) − rjwjbkb
xi(riwig + max{riwib − rjwjb, 0})
(cid:105)
max{riwib − rjwjb, 0} − rjwjbkb
10
(18)
Hence the good camp's optimal strategy can be obtained
by maximizing (18) with respect to x and j ∈ J, subject to
Constraints (15) and (16), and xi∈ [0, 1],∀i ∈ N.
A Greedy Approach for Determining Optimal Strategy
(cid:80)
For a given j, it can be seen from (18) that the optimal
strategy of the good camp is to determine x which maximizes
i∈N xi(riwig + max{riwib − rjwjb, 0}). Since Constraint
(15) should be satisfied, the minimum total investment by the
good camp on nodes belonging to set Iα should be Iα −
kb. Hence it should invest in nodes belonging to Iα one at a
time (subject to a maximum investment of 1 unit per node) in
decreasing order of values of (riwig +max{riwib−rjwjb, 0}),
until a total investment of Iα − kb is made. Let xi be the
good camp's investment on node i after this step; its remaining
budget is kg − (Iα − kb) and the maximum amount that it
could henceforth invest on a node i is 1− xi (since each node
has an investment capacity of 1 unit).
Now since Constraint (16) should also be satisfied, the max-
imum total investment by the good camp on nodes belonging
to set Iα ∪ Pα should be Iα + Pα − kb. Hence it should
now invest in nodes one at a time (maximum investment of
1 − xi in node i) in decreasing order of values of (riwig +
max{riwib − rjwjb, 0}) until one of the following occurs: (a)
the remaining budget (kg − Iα + kb) is exhausted or (b) a
node with a negative value of (riwig +max{riwib−rjwjb, 0})
is reached or (c) the investment made on nodes belonging to
Iα∪ Pα reaches Iα +Pα− kb. If condition (a) or (b) is met,
the so obtained strategy x∗
) is the optimal x for the
j
given j. However, if condition (c) is met, good camp should
continue investing the remaining available amount on nodes
belonging to N\(Iα∪Pα) one at a time (subject to a maximum
investment of 1 unit per node) in decreasing order of values
of (riwig + max{riwib − rjwjb, 0}). The so obtained strategy
x∗
= (x∗
ji
j
The absolute optimal strategy of the good camp can now
be computed by iterating over all j ∈ J and taking the one
that maximizes (from Expression (18))
) would hence be the optimal x for the given j.
= (x∗
ji
max
j∈ J
ji(riwig + max{riwib − rjwjb, 0})
∗
x
max{riwib − rjwjb, 0} − rjwjbkb
(19)
For the bad camp's optimal strategy, recall that
vi =
riw0
iiv0
i +
riwibyi
(cid:88)
i∈N
riwigxi −(cid:88)
i∈N
(cid:88)
i∈N
i∈N
(cid:88)
−(cid:88)
(cid:88)
i∈N
i∈N
Since yi ∈ [0, 1 − xi],∀i ∈ N, the optimal strategy of the
bad camp is to invest in nodes one at a time (subject to a
maximum investment of 1 − xi per node) in decreasing order
of values of riwib until either its budget kb is exhausted or
we reach a node with a negative value of riwib.
It can also be seen that if kg and kb are integers, it is an
optimal investment strategy of the good and bad camps to
invest one unit or not invest at all in a node.
Insight 3. Assuming a j, the strategy of the good camp is
to choose nodes with good values of (riwig + max{riwib −
rjwjb, 0}). That is, it chooses nodes with not only good values
of riwig, but also good values of riwib. This is expected since
the budget constraint per node allows the good camp (which
plays first) to block those nodes on which the bad camp would
have preferred to invest. Also, based on (14) and the definitions
of Jα, Iα, Pα, node j can be viewed as a boundary for the bad
camp's investment, that is, the bad camp would not invest in
any node i such that riwib < rjwjb.
Time Complexity of the Greedy Approach
For a given j, the above greedy approach would require
the good camp to select a total of O(kg) nodes to invest on.
This could be done by either (a) iteratively choosing a node
with the maximum value of (riwig + max{riwib − rjwjb, 0})
according to the greedy approach or (b) presorting the nodes
as per decreasing values of (riwig + max{riwib − rjwjb, 0})
and then choosing nodes according to the greedy approach.
The time complexity of (a) would be O(nkg) and that of (b)
would be O(n log n + kg) = O(n log n). Hence, following (a)
would be more efficient if kg << log n, while (b) would be
better if kg >> log n (else, asymptotically indifferent between
the two). So its time complexity is O(n · min{kg, log n}).
Now, the absolute optimal strategy is computed by iterating
over all j ∈ J. Based on (14) and the definitions of Jα, Iα, Pα,
node j can be viewed as a boundary for the bad camp's
investment since it would not invest in any node i such that
riwib < rjwjb. So if the nodes are ordered in decreasing
order of riwib values, such a node j would be in a position
no later than (cid:100)kg + kb(cid:101) (this limiting case is met if the good
camp invests kg in nodes with the highest values of riwib, and
then the bad camp invests kb in nodes with the highest values
of riwib which are not exhaustively invested on by the good
camp). So the possible j's are at most the top (cid:100)kg + kb(cid:101) nodes
in the sorted order of riwib, hence determining the possible
j's requires O(min{nkb, n log n}) time and the number of
possible j's is O(kg + kb).
Since the absolute optimal strategy of the good camp is
computed by iterating over all j ∈ J,
time
complexity of the greedy approach is O(min{nkb, n log n} +
(kg + kb) · n · min{kg, log n}), which is same as O(n(kg +
kb) · min{kg, log n}).
Maxmin versus Minmax Values
the overall
Here, we compare the maxmin and minmax values of the
game in the fundamental setting (Section III-B) with that in the
common coupled constraints setting. The introduction of the
total budget constraints per node disturbs the equality between
(cid:88)
(cid:88)
i∈N
i∈N
(cid:88)
i∈N
(cid:88)
(cid:88)
i∈N
i∈N
vi
vi
(cid:88)
i∈N
11
maxmin and minmax, as we show now. Let (x(cid:48), y(cid:48)) be an
optimal maxmin strategy profile in (7). Adding the constraint
0 ≤ y ≤ 1−x restricts the set of feasible strategies for the bad
camp, and this set of feasible strategies and hence its optimal
strategy now depends on x. So we have
max
0≤x≤1
min
0≤y≤1
Similarly,
min
0≤y≤1
max
0≤x≤1
vi ≤ max
0≤x≤1
min
0≤y≤1−x
vi ≥ min
0≤y≤1
max
0≤x≤1−y
These two inequalities, along with (7), result in the following
inequality,
max
0≤x≤1
min
0≤y≤1−x
vi ≥ min
0≤y≤1
max
0≤x≤1−y
vi
(20)
This result, which is contrary to general functions (for which
maxmin is less than or equal to minmax), has also been derived
in [32]. In our problem, this is a direct consequence of the first
mover advantage, which restricts the strategy set of the second
mover. In the maxmin case as analyzed earlier, the good camp
invests in nodes with good values of (riwig + max{riwib −
rjwjb, 0}) (assuming a j). That is, it is likely to invest in nodes
with good values of riwib which are the preferred investees
of the bad camp. Owing to total investment limit per node,
the bad camp may not be able to invest in its preferred nodes
(those with high values of riwib). It can be shown on similar
lines that, in the minmax case where the bad camp plays first,
it would play symmetrically opposite, thus limiting the ability
of good camp to invest in nodes with good values of riwig.
Remark 4 (CCC under simultaneous play). If
instead of
sequential play, the two camps play simultaneously under CCC
setting, it can be seen that the uniqueness of Nash equilibrium
is not guaranteed (an immediate example is that the maxmin
and minmax values could be different). We address this precise
question in [39] for general resource allocation games, albeit
assuming strict preference ordering of the camps over nodes.
Therein, however, we do not derive an equilibrium strategy
profile since there could be infinite number of Nash equilibria.
In order to derive a precise strategy profile which would be of
practical and conceptual interest, we considered a sequen-
tial play in this paper and computed the subgame perfect
Nash equilibrium. For the case where the Nash equilibrium
is unique, the sequence of play would not matter (that is,
the maxmin and minmax values would be the same), and
our derived subgame perfect Nash equilibrium would be the
unique pure strategy Nash equilibrium.
VII. DECISION UNDER UNCERTAINTY
In this section, we look at another sequential play setting
which considers the possibility that the good camp, which
plays first, may not have complete or exact
information
regarding the extrinsic weights, namely, wig, wib, w0
ii. The bad
camp, however, which plays second, has perfect information
regarding the values of these parameters, and hence it is known
that it would act optimally. Forecasting the optimal strategy
of the bad camp, the good camp aims at choosing a robust
u1
...
w0
. That is, u =
, where ui =
(cid:1)T .
strategy which would give it a good payoff even in the worst
case.
(cid:0)w0
un
ii
wig
wib
nn wng wnb
Let u =
11 w1g w1b ··· w0
Let U be a polytope defined by Eu ≤ f (that is, u ∈ U). It
can be viewed as the uncertainty set, which in this case, is a
convex set. The polytope would be based on the application at
hand and could be deduced from observations, predictions, etc.
We use the framework of robust optimization [33] for solving
this problem.
ii ≥ 0,∀i ∈ N.
(cid:88)
For the purpose of this section, let us assume that all the
elements of u are non-negative. This is to ensure bounded
values of the parameters. For instance, if we have a constraint
ii ≤ θi, the
in the linear program such as wig + wib + w0
ii can be unbounded. So for this
individual values wig, wib, w0
current setting (under uncertain parameters), we will assume
wig, wib, w0
min
Eu≤f
riwigxi +
(cid:88)
min(cid:80)
i yi≤kb
yi≥0
i −(cid:88)
Since the good camp aims to optimize in the worst case of
parameter values, while the bad camp has knowledge of these
values with certainty, the optimization problem is
max(cid:80)
i xi≤kg
xi≥0
one feasible node to invest on, then we have(cid:80)
kb maxj∈N rjwjb, else we have (cid:80)
If maxj∈N rjwjb > 0, that is, the bad camp has at least
i∈N riwibyi =
such that rdwdb = 0. So now we have (cid:80)
i∈N riwibyi = 0. For
let d be a dummy node
i∈N riwibyi =
max(cid:80) xi≤kg
kb maxj∈N∪{d} rjwjb. The optimization problem thus is
arriving at a concise solution,
i − kb max
j∈N∪{d} rjwjb
(cid:88)
(cid:88)
riwigxi +
riw0
riw0
iiv0
iiv0
i∈N
i∈N
min
u
i∈N
riwibyi
xi≥0
i∈N
i∈N
s.t. Eu ≤ f
Note that
there are in general n + 1 possibilities for
maxj∈N∪{d} rjwjb. We could write a linear program for
each possibility of i0 = arg maxj∈N∪{d} rjwjb. For a fixed
i0 ∈ N ∪ {d}, the inner term is
riw0
(cid:88)
(cid:88)
riwigxi +
min
i∈N
iiv0
i − kbri0 wi0b
s.t. Eu ≤ f ← αi0
∀i ∈ N : riwib ≤ ri0 wi0b ← βi0i
u
i∈N
For this problem to be feasible, the constraint set should be
non-empty. Let Nf be the subset of N ∪ {d} consisting of
nodes i0 such that the constraint set satisfying Eu ≤ f and
∀i ∈ N : riwib ≤ ri0wi0b is non-empty. Its dual is:
s.t. ∀i ∈ N \ {i0} : −αT
if i0 (cid:54)= d : −αT
i0 Ei0b +
max−αT
i0 f
i(cid:54)=i0
(cid:88)
i0 Eib − βi0iri ≤ 0 ← wib
βi0iri0 ≤ −kbri0 ← wi0b
i0 Eig ≤ rixi ← wig
i0 Eii ≤ riv0
i ← w0
αi0 ≥ 0
∀i ∈ N : βi0i ≥ 0
ii
∀i ∈ N : −αT
∀i ∈ N : −αT
12
We need to find a common x for all possibilities of i0 ∈ Nf .
So we have a constraint on the value of the dual, say ρ,
f,∀i0. We hence obtain a solution to the
namely, ρ ≤ −αT
optimization problem by solving the following LP.
i0
s.t. (cid:88)
max ρ
xi ≤ kg
∀i ∈ N : xi ≥ 0
i∈N
∀i0 ∈ Nf :
∀i ∈ N \ {i0} : −αT
if i0 (cid:54)= d : −αT
i0 Ei0b +
i(cid:54)=i0
∀i ∈ N : −αT
∀i ∈ N : −αT
ρ + αT
i0 f ≤ 0
(cid:88)
i0 Eib − βi0iri ≤ 0
βi0iri0 ≤ −kbri0
i0 Eig − rixi ≤ 0
i0 Eii ≤ riv0
αi0 ≥ 0
∀i ∈ N : βi0i ≥ 0
i
We solve the above LP for a specific example in our
simulation study, so as to derive insights on the effect of
uncertainty on the optimal strategy of the good camp.
VIII. SIMULATIONS AND RESULTS
Throughout this paper, we analytically derived the optimal
investment strategies of competing camps in a social network
for driving the opinion of the population in their favor. We
hence presented either closed-form expressions or algorithms
with polynomial running time. With the aim of determining
implications of the analytically derived results on real-world
networks and obtaining further insights, we conducted a sim-
ulation study on two popular network datasets. In this section,
we present the setup and observed results, and provide insights
behind them.
A. Simulation Setup
We consider an academic collaboration network obtained
from co-authorships in the "High Energy Physics - Theory"
papers published on the e-print arXiv from 1991 to 2003. It
contains 15,233 nodes and 31,376 links among them, and is
popularly denoted as NetHEPT. This network exhibits many
structural features of large-scale social networks and is widely
used for experimental justifications, for example, in [24], [40],
[41]. For the purpose of graphical illustration, we use the
popular Zachary's Karate club dataset consisting of 34 nodes
and 78 links among them [42].
Our analyses throughout the paper are valid for any distribu-
tion of edge weights satisfying the general constraints in Sec-
tion II-B. However, in order to concretize our simulation study,
we need to consider a particular distribution of edge weights.
Proposition 4 showed that some popular models of distributing
edge weight in a graph would result in random strategy being
optimal, and so are not suited for our simulations. Hence in
order to transform an undirected unweighted network dataset
into a weighted directed one for our simulations, we consider
that for any node i (having di number of connections), the
tuple (w0
ii, wig, wib) is randomly generated such that
13
(a) The value of wig
(b) The value of wib
(d) The value of ri
Fig. 1. Details about the Karate club dataset used in our simulations. The size and color saturation of a node i represent the value of the parameter.
(a) τ = 1 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i =−0.0342)
(b) τ = 2 (maxxminy
(cid:80)
(cid:104)τ(cid:105)
i = 0.2058)
i v
(d) τ = 4 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.1566)
Fig. 2. Progression of opinion values for Karate club dataset with kg = kb = 5 under linear influence function with unbounded investment per node. Opinions
of nodes are signified by their shapes, sizes, color saturations (circular blue nodes: positive opinions, square red nodes: negative opinions)
∀i ∈ N : w0
ii + wig + wib)
1−(w0
ii + wig + wib = 0.5, and
wij =
di
, if there is an edge between i and j.
A primary reasoning for considering w0
ii + wig + wib = 0.5
is to have a natural first guess that nodes give equal weightage
to intra-network influencing factors ({wij}j∈N ) and extra-
network influencing factors (w0
ii, wig, wib). We provide results
for the extreme cases in Appendix D, namely, when the value
of w0
ii+wig+wib is 0.1 or 0.9 (with the values of the individual
parameters scaled proportionally). We also highlight some key
effects of this value on the obtained results, throughout this
section.
Figure 1 presents the values of the parameters used for our
experimentation on the Karate club dataset, considering our
weight distribution with w0
ii + wig + wib = 0.5. The size
and color saturation of a node i represent the value of the
parameter mentioned in the corresponding caption (bigger size
and higher saturation implies higher value). Unless otherwise
specified, we consider kg = kb = 5 for this dataset. Also,
unless otherwise specified, we start with an unbiased network,
that is, v0
i = 0,∀i ∈ N.
Progression with time: Throughout this paper, our analyses
were based on the steady state opinion values. However for
the purpose of completion, we now provide a brief note on
the progression of opinion values with time, which occurs
according to our update rule given by (1).
vestment per node, where (cid:80)
Figure 2 illustrates the progression of opinion values with
time under the linear influence function and unbounded in-
is the sum of opinion
(cid:104)τ(cid:105)
i
i∈N v
values of nodes in time step τ. The network starts with
i = 0,∀i ∈ N, and then at τ = 1, the good and bad camps
v0
invest their entire budgets on their respective target nodes
having maximum values of riwig and riwib, respectively.
Hence the opinion values of these nodes change to being
highly positive and negative respectively, while other nodes
still hold an opinion value of 0 (Figure 2(a)). At τ = 2, nodes
which are directly connected to these target nodes, update their
opinions; as seen from Figure 2(b), nodes in the left region
hold positive opinions while the ones in the right region hold
negative opinions. Few nodes like the ones on the top, center,
and extreme right regions, still hold an opinion value of 0. By
τ = 3, all nodes hold a non-zero opinion value and at τ = 4,
the individual opinion values and hence the sum of opinion
values almost reach the convergent value. The sum of opinion
values at τ = 4 is 0.1566 (Figure 2(c)), while the convergent
sum is 0.1564 (Table II). In general, assuming the threshold
of convergence to be 10−4, the convergence is reached in 8-10
time steps for the Karate club dataset, and in 12-15 time steps
for the NetHEPT dataset.
An important insight is that any significant changes in the
opinion value of a node occur in the earlier time steps. For in-
stance, a node which is geodesically closer to the target nodes
receive influence from them in the earlier time steps, and also
the influence is strong since the entries of the substochastic
weight matrix wτ are significantly higher for lower values of
τ. Hence, owing to the substochastic nature of wτ , the change
in a node's opinion value is insignificant at a later time step.
We could also observe that the rate of convergence depends
on the investment strategies, for instance, the convergence
RESULTS FOR KARATE CLUB AND NETHEPT DATASETS
TABLE II
14
Karate club
NetHEPT
maxx miny
100 maxx miny
100
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
kb
100
-
-
100
400
100
400
kg
100
200
100
100
100
100
400
100
400
100
400
100
400
100 maxx miny
400 maxx miny
100 maxx miny
400 maxx miny
100
100
maxx miny
miny maxx
Setting
Aspect
Case
Section
Fundamental
Adversary
Concave (t = 2)
Concave (t = 10)
Unbounded
III-A1
Bounded
III-A2
Unbounded
Bounded
V-A
V-B
Unbounded
IV-A
Bounded
IV-B
Unbounded
IV-A
Bounded
IV-B
CCC
Bounded
VI
kg
5
10
5
5
5
5
20
5
20
5
20
5
20
5
kb
5
5
5
-
-
5
20
5
20
5
20
5
20
5
Results
(cid:80)
(cid:80)
i∈N vi = 0.1564
(cid:80)
i∈N vi = 5.8811
(cid:80)
i∈N vi = −0.0538
(cid:80)
i∈N yi = 5.1404
(cid:80)
i∈N yi = 4.8936
(cid:80)
i∈N vi = 0.4581
(cid:80)
i∈N vi = 0.9163
(cid:80)
i∈N vi = 0.4612
(cid:80)
i∈N vi = 1.2653
(cid:80)
i∈N vi = 1.1180
(cid:80)
i∈N vi = 1.2842
(cid:80)
i∈N vi = 1.1180
(cid:80)
i∈N vi = 1.3104
(cid:80)
i∈N vi = 1.5399
i∈N vi = −0.8900
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
miny maxx
Results
(cid:80)
(cid:80)
i∈N vi = 73.2539
(cid:80)
i∈N vi = 347.0770
(cid:80)
i∈N vi = 2.8513
(cid:80)
i∈N yi = 136.5231
(cid:80)
i∈N yi = 102.7266
(cid:80)
i∈N vi = −0.8446
(cid:80)
i∈N vi = −1.6892
(cid:80)
i∈N vi = −0.8446
(cid:80)
i∈N vi = −1.7117
(cid:80)
i∈N vi = −12.4212
(cid:80)
i∈N vi = −14.2682
(cid:80)
i∈N vi = −12.4212
(cid:80)
i∈N vi = −14.2682
(cid:80)
i∈N vi = 7.4843
i∈N vi = −3.6795
is almost
immediate under the concave influence function
setting where the investment is already distributed over nodes
(see Figures 11-12 in Appendix D); since this investment is
made in the earliest time step, it plays a significant role in
determining the nodes' convergent opinion values. Also, it is
usually observed that the individual opinion values as well as
sum of opinion values are, more often than not, monotone
increasing or decreasing with time (see Figures 11-14 in
Appendix D).
B. Simulation Results
Table II presents the quantitative results of our simulations
on Karate club and NetHEPT datasets. (The results for w0
ii +
wig + wib = 0.1 and 0.9 are provided in Tables III and IV of
Appendix D.)
1) The fundamental setting: For both the datasets,
the
overall opinion value is positive in the fundamental unbounded
setting (III-A1) when both camps had the same budget (first
row of Table II), which implies that the maximum value of
riwig in the network was higher than the maximum value
of riwib (as clear from Equation (3) and Proposition 2). The
results of the good camp doubling its budget can also be seen
(second row of Table II). When both camps had the same
budget, their overall influences tend to nullify each other to
a great extent and so the sum of opinion values is neither
exceedingly positive nor exceedingly negative. However, with
the good camp doubling its budget, this additional budget
could be used to have a large surplus of positive influence
in the network. Actually, owing to close competition, even
a slight imbalance in the camps' budgets would result in
significant skewness in the overall opinion of the network.
The effect of bounded investment (Setting III-A2) can also
be seen (third row of Table II); for these particular datasets,
the maxmin value decreases implying that the value of riwig is
probably concentrated on one node, while that of riwib is well
distributed, thus giving the bad camp an advantage in bounded
(and hence distributed) investment. This can be seen from
Proposition 3. Figure 4(a) illustrates the case with bounded
investment per node (xi, yi ≤ 1,∀i ∈ N) for the Karate club
dataset. Here, the label 'c' means that, that node is invested
on by both the camps with 1 unit, while 'g' or 'b' mean that
the node is invested on by the good or bad camp, respectively.
2) The adversary setting: Considering the adversary setting
of maximizing the competitor's (bad camp's) investment, we
could see how much budget the bad camp required to draw the
overall opinion in its favor. As expected from the results of the
above setting where the overall opinion value turns out to be
positive when the camps have the same budget, the investment
required in the unbounded case of Setting V-A (5.1404 for
Karate club and 136.5231 for NetHEPT) is more than the
budget available in Setting III-A1 (5 for Karate club and 100
for NetHEPT). In the bounded setting (V-B), for Karate club
dataset, we can see that the bad camp could have driven
the overall opinion in its favor by expending 4.8936 instead
of its entire budget of 5. For NetHEPT dataset, as expected
from the results of the fundamental setting where the overall
opinion in the bounded setting (III-A2) was less than that in
the unbounded setting (III-A1), the investment required by the
bad camp in the bounded setting (V-B), which is 102.7266, is
less than that required in the unbounded setting (V-A), which
is 136.5231.
3) Concave influence function: Results under the concave
influence function are presented for both unbounded (IV-A)
and bounded (IV-B) cases for t = 2 and t = 10, in Table II.
Figure 3 shows the effect of the value of t on the distribution
of investment and final opinion values for the Karate club
dataset, in the unbounded case. In Figures 3(a),(b),(d),(e), the
size and color saturation of a node represent the amount of
15
(a) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 2
(b) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 2
(c) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 2
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
(d) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 10
Fig. 3. Simulation results for the Karate club dataset with kg = kb = 5 when the influence function is concave (Setting IV-A)
(e) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 10
(f) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 10
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
investment on it by the camp mentioned in the corresponding
caption. With careful observation,
it can be seen that for
t = 2, the investments are more skewed, while for t = 10,
the investments by the good and bad camps on a node i are
close to being proportional to the values of riwig and riwib,
respectively (as suggested in Proposition 5 and Remark 1). In
Figures 3(c) and 3(f), the shape and color of a node represent
its opinion sign (blue circle implies positive, red square implies
negative), while its size and color saturation represent
its
opinion magnitude.
In the unbounded case, for some values of budgets, there
exist nodes for which either xi or yi or both exceed 1 unit.
So for the bounded case, the camps are directed to have
different investment strategies than in the unbounded case
(as can be understood from Insight 2). The effects can be
seen in Table II where the values are different in Settings
IV-A and IV-B for the same values of budgets. In some
scenarios such as NetHEPT with kg = kb = 100, however, the
investment strategies inadvertently assured xi, yi ≤ 1,∀i ∈ N
even for the unbounded case; so the investment strategies
remain the same in both settings and hence resulted in the
same overall opinion value (−0.8446 for t = 2 and −12.4212
for t = 10). A careful analysis of the values would indicate
that the constraints xi, yi ≤ 1 are likely to come into picture
for some nodes, for lower values of t and higher values of
budgets. Lower values of t lead to skewed investment and so
a higher likelihood of some nodes having investment more
than 1 unit in the unbounded case. Similarly, higher values of
budgets scale up the investments on the nodes, resulting in a
higher likelihood of some nodes having investment more than
1 unit in the unbounded case. This can also be inferred from
our analytically derived investment strategies.
For w0
ii + wig + wib = 0.1 (see Figure 9 of Appendix D),
ii, wig, wib = 0.1)
the values of extra-network parameters (w0
get scaled down proportionally, while the intra-network ones
(wij's) scale up; that is, the network influence plays a stronger
role than the camps' recommendations. Because of the weak
impact of the camps' recommendations, that is, the nodes
being unwilling to accept opinions that are external to the net-
work, and since all nodes started with a zero initial opinion, the
magnitudes of their opinion values are very low. Furthermore,
because of the strong impact of the intra-network influence,
the camps invest greatly on the most influential nodes (and
not so much on the lesser influential ones), and allow the
network to spread its influence. Owing to the low magnitudes
of opinion values and a strong network effect which aids in
distributing the influence evenly, the opinion values in the
network are less skewed. For w0
ii + wig + wib = 0.9 (see
Figure 10 of Appendix D), the above reasoning gets reversed,
and so the magnitudes of individual opinion values are high,
the camps' investments are more evenly distributed, and the
opinion values are highly skewed.
4) Common coupled constraints: Figures 4 (b-c) illustrate
the effect of common coupled constraints xi +yi ≤ 1,∀i ∈ N,
for the Karate club dataset. The advantage of playing first is
clearly visible from the overall opinion value as well as the
16
(maxx miny
xi, yi ≤ 1,∀i
(cid:80)
(a) Bound on investment per node
i vi = −0.0538)
(cid:80)
(c) Common coupled constraints
i vi = −0.8900)
Fig. 4. Results in presence of additional constraints for the Karate club dataset with kg = kb = 5 ; The nodes are labeled 'g/b/c' to signify if invested on
by good/bad/both camps respectively. The sign of the opinion value of a node is signified by its shape and color (circle and blue for good, square and red for
bad), while the absolute value of its opinion is signified by its size and color saturation.
(b) Common coupled constraints
xi + yi ≤ 1,∀i
xi + yi ≤ 1,∀i
i vi = 1.5399)
(miny maxx
(maxx miny
(cid:80)
For w0
(cid:80)
(cid:80)
distribution of opinion values. Specifically, the overall opinion
value is the highest in Figure 4(b) with a healthy distribution
of positive opinion values, followed by the value in Figure
4(a), followed by that in Figure 4(c) which is dominated by
negative opinions.
ii + wig + wib = 0.9,∀i ∈ N on NetHEPT
dataset, we observed that maxx miny≤1−x
i∈N vi =
i∈N vi = 1.5930 (see Table IV of Ap-
miny maxx≤1−y
pendix D). So the camps did not have the first mover advan-
tage; and it would not matter if the camps played sequentially
or simultaneously (Remark 4). Hence the camp which plays
first, could devise its investment strategy without having to
consider the best response investment strategy of the camp
which plays second. Also, as reasoned earlier, the opinion
values are less skewed for w0
ii + wig + wib = 0.1, while highly
skewed for w0
ii + wig + wib = 0.9.
5) Decision under uncertainty: For studying the effects of
decision under uncertainty as analyzed in Section VII, we
consider that the good camp (first mover) is uncertain about
parameters wig and wib up to a certain limit. In particular,
there is a fractional uncertainty of l regarding the values of
these parameters, while there is a fractional uncertainty of o
regarding the values of the sums of these parameters over the
entire network. l can be hence viewed as local uncertainty and
o as global uncertainty. Let wig and wib be the underlying
ground truth values for a node i (the actual values destined to
be realized).
(cid:88)
(cid:88)
i∈N
i∈N
wig
wib
wib ≤ (1 + o)
(1 − l) wig ≤ wig ≤ (1 + l) wig
(1 − l) wib ≤ wib ≤ (1 + l) wib
wig ≤ (1 + o)
(cid:88)
(cid:88)
i∈N
i∈N
wig ≤(cid:88)
wib ≤(cid:88)
i∈N
i∈N
(1 − o)
(1 − o)
It is clear that the latter two constraints would come into
picture only if o < l (this would usually be the case since,
though there may be significant relative deviation for the
individual parameters, the relative deviation of their sum is
usually low owing to significant balancing of positive and
negative deviations of the individual parameters). For different
values of l and o, Figure 5 presents the maxmin values:
(a) as computed by the good camp (first mover) as its worst
Fig. 5. The worst case and realized values of the maxmin value for different
values of l and o for the Karate club dataset with kg = kb = 5
Fig. 6. Investment made by good camp on nodes signified by their sizes and
color saturations for the Karate club dataset with kg = kb = 5 (under high
uncertainty)
case value using our robust optimization approach and (b) as
realized based on the ground truth. We assume the ground
truth values as depicted in Figure 1. It can be seen that for a
large enough range of values of l and o, though the good
camp computes the worst case maxmin value to be very low,
the corresponding realized value is the same as when the good
camp is certain about parameter values. The uncertainty factor
starts affecting the good camp only for very high l and o.
As clear from our linear program formulation in Sec-
the good camp under
the optimal strategy of
tion VII,
uncertainty is a pessimistic one, that is, ensuring that the
good camp performs well even in the worst case. Hence it
is fundamentally different from the optimal strategy under
certainty, where there is no requirement of being pessimistic
or optimistic since all the information required to solve the
optimization problem is known with certainty. As pointed
above, the uncertainty factor plays a role only for very high
values of l and o. Specifically, for values of l upto 0.6, the
good camp's investment is concentrated on one node, like in
the certainty setting (Proposition 2). For l in excess of 0.6,
the distribution of investment showed a very similar nature
across different values of l and o; Figure 6 illustrates how
the good camp distributes its investment over nodes. Thus,
the investment is distributed under high levels of uncertainty,
while it is concentrated on one node under certainty as well
as under low levels of uncertainty. Furthermore, one could
notice a clear correlation between the values of wig (Figure
1(a)) and the investment amounts on the corresponding nodes
under high uncertainty (Figure 6).
IX. CONCLUSION
Using a variant of Friedkin-Johnsen model for opinion
dynamics, we studied a zero-sum game framework for optimal
investment strategies for two competing camps in a social
network. We derived closed form expressions and efficient al-
gorithms for a number of well-motivated settings. We showed
all the results quantitatively as well as illustratively using
simulations on network datasets.
Our analysis arrived at a decision parameter analogous
to Katz centrality. We also showed that for some of the
popular models of weighing edges, a random strategy is indeed
optimal. We further looked at a setting where the influence of
a camp on a node is a concave function of the amount of
investment, and derived that a more concave function results
in a less skewed investment strategy, which could be perceived
as a fairer strategy by the nodes. We studied an adversarial
problem where a camp aims to maximize its competitor's
investment required to drive the average opinion value of the
population in its favor, and saw that the optimal strategies
fundamentally remain the same, albeit with different forms of
the exact optimal strategies.
We studied Stackelberg variant of the game, under common
coupled constraints stating the bound on combined investment
by the two camps on any node. We analytically derived the
subgame perfect Nash equilibrium strategies of the camps, and
hence quantified the first-mover advantage. We derived a linear
program for obtaining optimal strategy for a camp to whom the
parameters' values are uncertain, while playing against a camp
having exact information. In our simulations, we observed that
a camp is likely to get affected only under considerably high
levels of uncertainty. Here, the optimal strategy is to have a
distributed investment over nodes, as against the concentrated
investment in the exact information setting.
Future Work
This work has several interesting directions for future work;
we mention a few. The two camps setting can be extended to
multiple camps where each camp would attempt to drive the
17
opinion of the population towards its own. With more than
two camps, we would need to have the camps hold opinions
in a multidimensional plane rather than on the real number
line (+1/− 1). We explain one way in which this could be
done. Let each camp (say h) have a vector associated with its
opinion (say (cid:126)ch). Let its investment on node i be denoted by
zhi and the weightage attributed by i to the camp's opinion
be wih. Analogous to Equation (3), the vector-sum of nodes'
opinions could be written as
(cid:126)vi =
riw0
ii(cid:126)v 0
i +
riwihzhi(cid:126)ch
(cid:88)
i∈N
(cid:88)
i∈N
(cid:88)
(cid:88)
i∈N
h
and its own opinion vector (that is, (cid:126)ch ·(cid:80)
A camp's objective would be to drive this vector-sum towards
the direction of its own opinion vector, that is, to maximize
the inner product between the vector-sum of nodes' opinions
i∈N (cid:126)vi). This paper
studied the special case where we have two camps: h = g, b
with (cid:126)cg = +1, (cid:126)cb = −1 and zgi = xi, zbi = yi. Since the
optimization terms for different camps get decoupled under
the settings in Sections 3 and 4, the optimization strategies of
all camps would be analogous to those derived in this paper.
However, our set of equations will not be directly adaptable
for the CCC and uncertainty settings, and is an interesting
direction for future work.
One could study other models of opinion dynamics with
respect to optimal investment strategies of competing camps
and investigate if it is possible to arrive at closed form expres-
sions under them. One such type of models could be where
the investment parameters get coupled in the optimization
problem, and the sum of opinion values is no longer a multilin-
ear function in the investment parameters {xi}i∈N ,{yi}i∈N .
Another possible future direction is to study the setting of
common coupled constraints under more complex constraints,
while maintaining analytical tractability; our analysis in this
paper could act as a conceptual base for solving such a
problem. In general, it would be interesting to consider hybrid
settings which combine the individually studied settings in this
paper, for instance, a concave influence function with the cost
function accounting for deviation from the desired investment.
It would also be interesting to study the tradeoff between the
advantages and disadvantages of playing first, since playing
first would allow a camp to block certain investments of the
competing camp (like in common coupled constraints) but it
may also force it to take decision under uncertainty.
REFERENCES
[1] A. Gionis, E. Terzi, and P. Tsaparas, "Opinion maximization in social
networks," in Proceedings of the 2013 SIAM International Conference
on Data Mining. SIAM, 2013, pp. 387 -- 395.
[2] M. Grabisch, A. Mandel, A. Rusinowska, and E. Tanimura, "Strategic
influence in social networks," Mathematics of Operations Research,
2017.
[3] D. Easley and J. Kleinberg, "Networks, crowds, and markets," Cam-
bridge Univ Press, vol. 6, no. 1, pp. 6 -- 1, 2010.
[4] D. Acemoglu and A. Ozdaglar, "Opinion dynamics and learning in social
networks," Dynamic Games and Applications, vol. 1, no. 1, pp. 3 -- 49,
2011.
[5] A. Charnes, "Constrained games and linear programming," Proceedings
of the National Academy of Sciences, vol. 39, no. 7, pp. 639 -- 641, 1953.
[6] P. Dubey, R. Garg, and B. De Meyer, "Competing for customers in a
social network: The quasi-linear case," in WINE. Springer, 2006, pp.
162 -- 173.
18
[33] A. Ben-Tal, L. El Ghaoui, and A. Nemirovski, Robust optimization.
Princeton University Press, 2009.
[34] C. Altafini, "Consensus problems on networks with antagonistic inter-
actions," IEEE Transactions on Automatic Control, vol. 58, no. 4, pp.
935 -- 946, 2013.
[35] A. V. Proskurnikov, A. S. Matveev, and M. Cao, "Opinion dynamics in
social networks with hostile camps: Consensus vs. polarization," IEEE
Transactions on Automatic Control, vol. 61, no. 6, pp. 1524 -- 1536, 2016.
[36] J. H. Hubbard and B. B. Hubbard, Vector calculus, linear algebra, and
differential forms: a unified approach. Matrix Editions, 2015.
[37] L. Katz, "A new status index derived from sociometric analysis,"
Psychometrika, vol. 18, no. 1, pp. 39 -- 43, 1953.
[38] R. Jain, D.-M. Chiu, and W. R. Hawe, A quantitative measure of fairness
and discrimination for resource allocation in shared computer system,
1984, vol. 38.
[39] S. Dhamal, W. Ben-Ameur, T. Chahed, and E. Altman, "Resource
allocation polytope games: Uniqueness of equilibrium, price of stability,
and price of anarchy," in Proceedings of the 32nd AAAI Conference on
Artificial Intelligence. AAAI, 2018, pp. 997 -- 1006.
[40] W. Chen, Y. Wang, and S. Yang, "Efficient influence maximization in
social networks," in Proceedings of the 15th ACM SIGKDD Interna-
tional Conference on Knowledge Discovery and Data Mining. ACM,
2009, pp. 199 -- 208.
[41] 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. ACM, 2010, pp. 1029 -- 1038.
[42] 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.
[7] K. Bimpikis, A. Ozdaglar, and E. Yildiz, "Competitive targeted adver-
tising over networks," Operations Research, vol. 64, no. 3, pp. 705 -- 720,
2016.
[8] M. H. DeGroot, "Reaching a consensus," Journal of
the American
Statistical Association, vol. 69, no. 345, pp. 118 -- 121, 1974.
[9] N. E. Friedkin and E. C. Johnsen, "Social influence and opinions,"
Journal of Mathematical Sociology, vol. 15, no. 3-4, pp. 193 -- 206, 1990.
[10] -- -- , "Social positions in influence networks," Social Networks, vol. 19,
no. 3, pp. 209 -- 222, 1997.
[11] J. Lorenz, "Continuous opinion dynamics under bounded confidence: A
survey," International Journal of Modern Physics C, vol. 18, no. 12, pp.
1819 -- 1838, 2007.
[12] H. Xia, H. Wang, and Z. Xuan, "Opinion dynamics: A multidisciplinary
review and perspective on future research," in Multidisciplinary Studies
in Knowledge and Systems Science.
IGI Global, 2013, pp. 311 -- 332.
[13] A. Das, S. Gollapudi, and K. Munagala, "Modeling opinion dynamics
in social networks," in Proceedings of
the 7th ACM International
Conference on Web Search and Data Mining. ACM, 2014, pp. 403 -- 412.
[14] P. Clifford and A. Sudbury, "A model for spatial conflict," Biometrika,
vol. 60, no. 3, pp. 581 -- 588, 1973.
[15] R. A. Holley and T. M. Liggett, "Ergodic theorems for weakly interact-
ing infinite systems and the voter model," The annals of probability, pp.
643 -- 663, 1975.
[16] S. E. Parsegov, A. V. Proskurnikov, R. Tempo, and N. E. Friedkin,
"Novel multidimensional models of opinion dynamics in social net-
works," IEEE Transactions on Automatic Control, vol. 62, no. 5, pp.
2270 -- 2285, 2017.
[17] J. Ghaderi and R. Srikant, "Opinion dynamics in social networks: A local
interaction game with stubborn agents," in American Control Conference
(ACC), 2013.
IEEE, 2013, pp. 1982 -- 1987.
[18] W. Ben-Ameur, P. Bianchi, and J. Jakubowicz, "Robust distributed con-
sensus using total variation," IEEE Transactions on Automatic Control,
vol. 61, no. 6, pp. 1550 -- 1564, 2016.
[19] P. Jia, A. MirTabatabaei, N. E. Friedkin, and F. Bullo, "Opinion
dynamics and the evolution of social power in influence networks," SIAM
review, vol. 57, no. 3, pp. 367 -- 397, 2015.
[20] A. Halu, K. Zhao, A. Baronchelli, and G. Bianconi, "Connect and win:
The role of social networks in political elections," EPL (Europhysics
Letters), vol. 102, no. 1, p. 16002, 2013.
[21] E. Yildiz, A. Ozdaglar, D. Acemoglu, A. Saberi, and A. Scaglione,
"Binary opinion dynamics with stubborn agents," ACM Transactions
on Economics and Computation, vol. 1, no. 4, p. 19, 2013.
[22] C. Ballester, A. Calv´o-Armengol, and Y. Zenou, "Who's who in net-
works. wanted: The key player," Econometrica, vol. 74, no. 5, pp. 1403 --
1417, 2006.
[23] A. Sobehy, W. Ben-Ameur, H. Afifi, and A. Bradai, "How to win
elections," in International Conference on Collaborative Computing:
Networking, Applications and Worksharing. Springer, 2016.
[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. ACM, 2003, pp. 137 -- 146.
[25] A. Guille, H. Hacid, C. Favre, and D. A. Zighed, "Information diffusion
in online social networks: A survey," ACM SIGMOD Record, vol. 42,
no. 1, pp. 17 -- 28, 2013.
[26] S. Bharathi, D. Kempe, and M. Salek, "Competitive influence maximiza-
tion in social networks," in International Workshop on Web and Internet
Economics. Springer, 2007, pp. 306 -- 311.
[27] S. Goyal, H. Heidari, and M. Kearns, "Competitive contagion in net-
works," Games and Economic Behavior, 2014.
[28] S. R. Etesami and T. Bas¸ar, "Complexity of equilibrium in competitive
diffusion games on social networks," Automatica, vol. 68, pp. 100 -- 110,
2016.
[29] S. A. Myers and J. Leskovec, "Clash of the contagions: Cooperation
and competition in information diffusion," in Proceedings of the Twelfth
International Conference on Data Mining (ICDM).
IEEE, 2012, pp.
539 -- 548.
[30] Y. Su, X. Zhang, L. Liu, S. Song, and B. Fang, "Understanding
information interactions in diffusion: an evolutionary game-theoretic
perspective," Frontiers of Computer Science, vol. 10, no. 3, pp. 518 --
531, 2016.
[31] J. B. Rosen, "Existence and uniqueness of equilibrium points for concave
n-person games," Econometrica: Journal of the Econometric Society, pp.
520 -- 534, 1965.
[32] E. Altman and E. Solan, "Constrained games: The impact of the attitude
to adversary's constraints," IEEE Transactions on Automatic Control,
vol. 54, no. 10, pp. 2435 -- 2440, 2009.
APPENDIX A
PROOF OF PROPOSITION 5
s.t. ∀i ∈ N : vi −(cid:88)
(cid:88)
(cid:88)
j∈N
Proof. We make the following one-to-one transformation:
yi = Y t
i . So the above optimization problem
is equivalent to
i and xi = X t
(cid:88)
max(cid:80)
i ≤kg
i X t
Xi≥0
min(cid:80)
i ≤kb
i Y t
Yi≥0
i∈N
wijvj + wibYi = wigXi + w0
vi
iiv0
i
i ) −(cid:88)
Similar to the derivation of Equation (3) in main text, we get
riwibYi
ri(wigXi + w0
i∈N
i∈N
i∈N
vi =
iiv0
Owing to the mutual
(21)
independence between ¯X and ¯Y
(where ¯X and ¯Y are the vectors with components Xi and
Yi respectively), we can write
max(cid:80)
(cid:88)
(cid:88)
i ≤kg
i X t
Xi≥0
(cid:88)
(cid:88)
min(cid:80)
i ≤kb
i Y t
Yi≥0
i + max(cid:80)
iiv0
i ≤kg
i X t
Xi≥0
riwigXi − max(cid:80)
i ≤kb
i Y t
Yi≥0
riwibYi
riw0
i∈N
i∈N
i∈N
i∈N
vi
=
We now solve for the first optimization term with respect to
¯X , which can be written as the following convex optimization
problem
min −(cid:88)
s.t. (cid:88)
riwigXi
i∈N
X t
i − kg ≤ 0 ← γ
i∈N
∀i ∈ N : −Xi ≤ 0 ← βi
The modified objective function with the Lagrangian mul-
tipliers is
g = −(cid:88)
i∈N
(cid:32)(cid:88)
i∈N
(cid:33)
−(cid:88)
i∈N
riwigXi + γ
X t
i − kg
βiXi
The KKT conditions with the modified objective function g
give the following additional constraints
i − βi = 0
∀i ∈ N :
∂g
∂Xi
= −riwig + tγX t−1
γ ≥ 0
∀i ∈ N : βi ≥ 0
X t
i − kg
(cid:33)
(cid:32)(cid:88)
γ
= 0
∀i ∈ N : βiXi = 0
i
Constraint (22) gives
Xi =
(cid:18) βi + riwig
(cid:19) 1
t−1
tγ
(22)
(23)
(24)
(25)
(26)
(27)
Note that if γ = 0, Constraint (22) gives βi + riwig =
0,∀i ∈ N. Since βi ≥ 0, this can hold only when riwig ≤
0,∀i ∈ N. For nodes with riwig = 0, the value of riwigXi
stays 0 for any Xi and so Xi = 0 is an optimal solution. For
nodes with riwig < 0, we must have βi > 0 and hence Xi = 0
(from (26)).
19
Henceforth we assume γ > 0 and so Equation (27) is valid.
X t
i = kg
i∈N
This, with (25) and (23), gives(cid:88)
(cid:18) βi + riwig
From (27), we get(cid:88)
(cid:19) 1
(cid:18) 1
(cid:16)(cid:80)
This, with (27) gives
X ∗
i = k
⇐⇒
i∈N
t−1
tγ
tγ
1
t
g
=
(cid:16)(cid:80)
(cid:19) t
t−1
= kg
1
t
g
k
i∈N (βi + riwig)
t
t−1
t
(cid:17) 1
(βi + riwig)
1
t−1
j∈N (βj + rjwjg)
t
t−1
(cid:17) 1
t
(28)
Now, the following cases are possible depending on the sign
of riwig.
Case 1:
Case 2: If riwig = 0, we must have X ∗
constraint X ∗
i ≥ 0). This with (26) gives X ∗
If riwig < 0, we must have βi > 0 (from the
i = βi = 0 (from
i = 0.
(26) and (27)).
Furthermore in both Case 1 and Case 2, from (27), we have
βi + riwig = 0. So the i-terms for which riwig ≤ 0 vanish
from the denominator of (28) and it transforms into
X ∗
i = k
1
t
g
(cid:16)(cid:80)
(βi + riwig)
1
t−1
j:rj wjg >0 (βj + rjwjg)
t
t−1
(cid:17) 1
t
, if riwig ≥ 0
and X ∗
i = 0, if riwig < 0
(29)
Case 3: If riwig > 0, we get X ∗
i > 0 (from (24)). This
with (26) gives βi = 0, and hence from (29), we get
(cid:16)(cid:80)
X ∗
i = k
1
t
g
(riwig)
j:rj wjg >0 (rjwjg)
The above cases can be concisely written as
X ∗
i = k
1
t
g
⇐⇒ x
∗
i = kg
(riwig)
1
t−1
j:rj wjg >0 (rjwjg)
(riwig)
t
t−1
t
t−1
j:rj wjg >0(rjwjg)
t
t−1
t
t
t−1
1
t−1
(cid:17) 1
· Iriwig >0
(cid:17) 1
· Iriwig >0
· Iriwib>0
t
(30)
(31)
(cid:16)(cid:80)
(cid:80)
(cid:80)
where Iriwig>0 = 1 if riwig > 0, and 0 otherwise.
Similarly, it can be shown that
∗
y
i = kb
(riwib)
t
t−1
j:rj wjb>0(rjwjb)
t
t−1
From the optimization problem, it can be directly seen that,
i = 0,∀i ∈ N. Similarly, if
if ∀i ∈ N , riwig ≤ 0, then x∗
∀i ∈ N , riwib ≤ 0, then y∗
i = 0,∀i ∈ N.
APPENDIX B
PROOF OF PROPOSITION 6
Proof. This proof goes as an extension to the proof of Propo-
sition 5, with the additional constraint
∀i ∈ N : Xi ≤ 1 ← ξi
Equation (22) changes to
∀i ∈ N : −riwig + tγX t−1
i − βi + ξi = 0
and Equation (27) changes to
(cid:18) βi − ξi + riwig
(cid:19) 1
t−1
tγ
Xi =
(32)
(33)
(34)
The complementary slackness conditions for the additional
constraints are
∀i ∈ N : ξi(Xi − 1) = 0
(35)
Further note that the constraints Xi ≥ 0 and Xi ≤ 1 cannot
both be tight, and so at least one of βi or ξi must be 0. That
is,
∀i ∈ N : βiξi = 0
(36)
When γ = 0, Equation (33) transforms into βi−ξi+riwig =
0,∀i ∈ N. Now if riwig = 0, we must have βi − ξi = 0 and
hence βi = ξi = 0 (from Equation (36)). It is further clear
from the objective function that if riwig = 0, it is optimal to
have Xi = 0 and hence x∗
i = 0. If riwig < 0, we must have
βi > 0 (since βi − ξi + riwig = 0) and so Xi = x∗
i = 0 (from
Equation (26)). If riwig > 0, we must have ξi > 0 and so
Xi = x∗
i = 1 (from Equation (35)). That is, when γ = 0, we
invest an amount of 1 on all nodes i with positive values of
riwig and 0 on all other nodes.
i = 1 · Iriwig >0
γ = 0 =⇒ x
∗
(37)
(cid:88)
(cid:18) 1
i∈Jγ
tγ
⇐⇒
(cid:18) riwig
(cid:19) t
t−1
tγ
=
+
t−1
(cid:19) t
kg −(cid:80)
(cid:80)
20
(cid:88)
i /∈Jγ
1 = kg
(38)
i:riwig >0
i /∈Jγ ,i:riwig >0 1
(riwig)
t−1
t
i∈Jγ
Note that if i ∈ Jγ, then any j for which rjwjg < riwig
belongs to Jγ. So Jγ and hence γ can be determined by adding
nodes to Jγ, one at a time in increasing order of riwig, subject
to riwig > 0. Let γ be the value of γ so obtained. It can be
seen that as γ decreases, the left hand side of (38) increases.
Since the right hand side is a constant, we would obtain a
unique γ satisfying the equality. Furthermore, for γ > 0, this
budget constraint is tight and so we are ensured the existence
of γ. Once γ and hence Jγ are obtained, an optimal solution
for the good camp can be expressed as follows (recall that
xi = X t
i ):
i = 0.
i = 1.
If riwig ≤ 0, x∗
If riwig > tγ, x∗
If 0 < riwig ≤ tγ,
(cid:32)
kg − (cid:88)
∗
i =
x
(cid:33) (riwig)
(cid:80)
1
t
t−1
i∈Jγ
(riwig)
t
t−1
i /∈Jγ
i:riwig >0
(cid:88)
(cid:18) riwig
(cid:19) t
t−1
i∈Jγ
tγ
(cid:88)
+
1 < kg
i /∈Jγ
i:riwig >0
However, if Equation (38) is not satisfied for any γ, that is,
when
hence (cid:80)
even for the lowest value of γ > 0, we have Jγ = {} and
i:riwig>0 1 < kg. Here, the number of nodes with
riwig > 0 is less than kg, meaning that the budget constraint
is not tight and so γ = 0 (from Equation (25)). The investment
is thus as per Equation (37).
For γ > 0, we have(cid:88)
⇐⇒ (cid:88)
i∈N
X t
i = kg
(cid:18) βi − ξi + riwig
(cid:19) t
t−1
= kg
Proof.
i∈N
tγ
(cid:18)−ξi + riwig
When βi > 0, we have ξi = 0 (from (36)) and Xi = 0 (from
(26)). From the above equation, this corresponds to riwig < 0.
Further note that for riwig = 0, Xi = 0 is an optimal solution.
So we have (cid:88)
(cid:26)(cid:16) riwig
We can have Xi = 1 or if not, we should have ξi =
0. So for any i such that riwig > 0, we have Xi =
. Let Jγ = {i : 0 < riwig ≤ tγ}. So
min
the above equation results in
(cid:17) 1
(cid:19) t
(cid:27)
i:riwig >0
= kg
t−1
t−1
, 1
tγ
tγ
APPENDIX C
PROOF OF PROPOSITION 7
max(cid:80)
s.t. (cid:88)
i xi≤kg
xi≥0
min
yi≥0
i∈N
vi ≤ 0
i∈N
yi
(cid:88)
i ) −(cid:88)
(cid:88)
i∈N
min
ri(wigxi + w0
iiv0
riwibyi
(cid:88)
i∈N
vi =
(cid:88)
i∈N
The inner term of this optimization problem is
s.t. (cid:88)
i∈N
riwibyi ≥(cid:88)
if (cid:80)
i∈N
yi
i∈N
ri(wigxi + w0
iiv0
∀i ∈ N : yi ≥ 0
i ) ← π
It
is clear that
is
optimal for the bad camp to have yi = 0,∀i ∈ N, that is,
i∈N ri(wigxi + w0
iiv0
it
i ) ≤ 0,
RESULTS FOR KARATE CLUB DATASET FOR DIFFERENT VALUES OF w0
ii + wig + wib
TABLE III
21
Setting
Aspect
Case
Section
Fundamental
Unbounded
3.1.1
Bounded
3.1.2
Adversary
Concave (t = 2)
Concave (t = 10)
Unbounded
Bounded
Unbounded
Bounded
5.1
5.2
4.1
4.2
Unbounded
4.1
Bounded
4.2
CCC
Bounded
6
kg
5
10
5
5
5
5
20
5
20
5
20
5
20
5
Value of
w0
ii + wig + wib
0.1
0.5
0.9
0.3545
5.0768
0.1141
7.6575
0.7089
0.3931
4.8542
0.1564
0.9163
0.4581
4.8936
i∈N vi
i∈N vi
5.8811
i∈N vi −0.0678 −0.0538
i∈N yi
5.1404
i∈N yi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
1.5399
i∈N vi −1.8655 −0.8900
0.4612
1.2653
1.1180
1.2842
1.1180
1.3104
1.4950
1.2512
1.4373
1.2512
1.5027
2.8105
0.1723
3.5673
0.3309
5.2674
5.6844
0.5435
1.0870
0.5435
1.0714
1.0120
1.1625
1.0120
1.1633
0.3951
0.2581
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
kb
5
5
5
-
-
maxx miny
maxx miny
maxx miny
maxx miny
maxx miny
5
maxx miny
20 maxx miny
5
maxx miny
20 maxx miny
5
maxx miny
20 maxx miny
5
maxx miny
20 maxx miny
5
maxx miny
miny maxx
(cid:80)
(cid:80)
(cid:80)
(cid:80)
i∈N yi = 0. So we need to only consider the case where
i ) > 0. In this case, for the constraint
i∈N ri(wigxi + w0
i∈N vi ≤ 0 to be satisfied, it is necessary that there exists
The dual problem of the above optimization problem can
a node j with positive value of rjwjb.
iiv0
be written as
max π
ri(wigxi + w0
iiv0
i )
s.t. ∀i ∈ N : πriwib ≤ 1 ← yi
π ≥ 0
From the dual constraints, we have π ≤
(recall
that maxj∈N rjwjb > 0 in this considered case). Since we aim
to maximize the dual objective function, we have
maxj∈N rj wjb
1
(cid:88)
i∈N
π =
1
maxj∈N wjbrj
The optimization problem now becomes
max(cid:80) xi≤kg
xi≥0
= π
π
(cid:88)
max(cid:80) xi≤kg
i∈N
xi≥0
1
(cid:88)
(cid:32)
i∈N
ri(wigxi + w0
iiv0
i )
riwigxi +
(cid:27)
riw0
iiv0
i
(cid:88)
(cid:26)
i∈N
(cid:33)
riw0
iiv0
i
(cid:88)
i∈N
=
maxj∈N wjbrj
kg max
max
i∈N
riwig, 0
+
(cid:80)
The above expression thus gives the total investment made by
the bad camp in the case where (kg max{maxi∈N riwig, 0} +
iiv0
i ) > 0.
i∈N riw0
Accounting for the case that it optimal for the bad camp
to have yi = 0,∀i ∈ N when (kg max{maxi∈N riwig, 0} +
i∈N riw0
iiv0
camp is
(cid:40)
i ) ≤ 0, the total investment made by the bad
(cid:41)
(cid:33)
(cid:32)
1
max
maxj∈N wjbrj
kg max{max
i∈N
riwig, 0} +
riw0
iiv0
i
, 0
(cid:88)
i
APPENDIX D
FURTHER SIMULATION RESULTS
The following pages present simulation results not included
in the main text.
• The quantitative results for Karate club and NetHEPT
ii + wig + wib are
datasets for different values of w0
presented in Tables III and IV, respectively.
• The illustrative results for Karate club dataset under ad-
ditional constraints such as bounded investment per node
and common coupled constraints, when w0
ii +wig +wib =
0.1 and 0.9 are presented in Figures 7 and 8, respectively.
• The illustrative results for Karate club dataset under
concave influence function when w0
ii + wig + wib = 0.1
and 0.9 are presented in Figures 9 and 10, respectively.
• The illustrations for progression of opinion values in
Karate club dataset under different settings are presented
in Figures 11-14.
RESULTS FOR NETHEPT DATASET FOR DIFFERENT VALUES OF w0
ii + wig + wib
TABLE IV
22
Setting
Aspect
Case
Section
kg
kb
Value of
w0
ii + wig + wib
0.1
0.5
0.9
Fundamental
Unbounded
3.1.1
Bounded
3.1.2
Adversary
Concave (t = 2)
Concave (t = 10)
Unbounded
Bounded
Unbounded
Bounded
5.1
5.2
4.1
4.2
Unbounded
4.1
Bounded
4.2
100
200
100
100
100
100
400
100
400
100
400
100
400
100 maxx miny
100 maxx miny
100 maxx miny
-
-
maxx miny
maxx miny
100 maxx miny
400 maxx miny
100 maxx miny
400 maxx miny
100 maxx miny
400 maxx miny
100 maxx miny
400 maxx miny
CCC
Bounded
6
100
100
maxx miny
miny maxx
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
2.8513
5.7026
7.6865
105.1987
143.6201
626.5203
145.9562
i∈N vi
i∈N vi
i∈N vi
i∈N yi
i∈N yi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
i∈N vi
28.5077
i∈N vi −19.2724
2.3493
−7.3610
−8.4556
−7.3610
−8.4556
2.8513
73.2539
13.5988
347.0770
131.4902
2.8513
1.5930
136.5231
113.0391
101.9037
102.7266
−1.7849
−0.8446
−3.5698
−1.6892
−1.7849
−0.8446
−1.7117
−3.5698
−12.4212 −13.9045
−14.2682 −15.9721
−12.4212 −13.9045
−14.2682 −15.9721
7.4843
−3.6795
1.5930
1.5930
(maxx miny
xi, yi ≤ 1,∀i
(cid:80)
(a) Bound on investment per node
i vi = −0.0678)
(cid:80)
(c) Common coupled constraints
i vi = −1.8655)
Fig. 7. Results in presence of additional constraints for the Karate club dataset with kg = kb = 5 ; The nodes are labeled 'g/b/c' to signify if invested on
by good/bad/both camps respectively. The sign of the opinion value of a node is signified by its shape and color (circle and blue for good, square and red for
bad), while the absolute value of its opinion is signified by its size and color saturation. (w0
(b) Common coupled constraints
xi + yi ≤ 1,∀i
xi + yi ≤ 1,∀i
i vi = 2.8105)
(miny maxx
(maxx miny
(cid:80)
ii + wig + wib = 0.1)
(a) Bound on investment per node
(b) Common coupled constraints
(maxx miny
i vi = 0.3309)
(maxx miny
i vi = 0.3951)
xi, yi ≤ 1,∀i
(cid:80)
xi + yi ≤ 1,∀i
(cid:80)
(c) Common coupled constraints
xi + yi ≤ 1,∀i
(miny maxx
i vi = 0.2581)
(cid:80)
Fig. 8. Results in presence of additional constraints for the Karate club dataset with kg = kb = 5 ; The nodes are labeled 'g/b/c' to signify if invested on
by good/bad/both camps respectively. The sign of the opinion value of a node is signified by its shape and color (circle and blue for good, square and red for
bad), while the absolute value of its opinion is signified by its size and color saturation. (w0
ii + wig + wib = 0.9)
23
(a) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 2
(b) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 2
(c) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 2
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
(d) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 10
(e) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 10
(f) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 10
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
Fig. 9. Simulation results for the Karate club dataset with kg = kb = 5 when the influence function is concave (w0
ii + wig + wib = 0.1)
(a) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 2
(b) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 2
(c) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 2
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
(d) Investment made by good camp
on nodes signified by their sizes
and color saturations for t = 10
(e) Investment made by bad camp
on nodes signified by their sizes
and color saturations for t = 10
(f) Final opinion values of nodes signified
by shapes, sizes, color saturations for t = 10
(Circular blue nodes: positive opinions,
square red nodes: negative opinions)
Fig. 10. Simulation results for the Karate club dataset with kg = kb = 5 when the influence function is concave (w0
ii + wig + wib = 0.9)
24
(a) τ = 1 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.3037)
(b) τ = 2 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.3681)
(d) τ = 4 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.4353)
Fig. 11. Progression of opinion values for the Karate club dataset with kg = kb = 5 when the influence function is concave (t = 2)
(a) τ = 1 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.4956)
(b) τ = 2 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 0.7986)
(d) τ = 4 (maxxminy
(cid:80)
i v
(cid:104)τ(cid:105)
i = 1.0362)
Fig. 12. Progression of opinion values for the Karate club dataset with kg = kb = 5 when the influence function is concave (t = 10)
(a) τ = 1 (maxxminy
(d) τ = 4 (maxxminy
Fig. 13. Progression of opinion values for the Karate club dataset with kg = kb = 5 under common coupled constraints (maxmin value)
(b) τ = 2 (maxxminy
i v
i v
(cid:104)τ(cid:105)
i =−0.2375)
(cid:104)τ(cid:105)
i = 0.8117)
(cid:80)
i v
(cid:104)τ(cid:105)
i = 1.3420)
(a) τ = 1 (minymaxx
(d) τ = 4 (minymaxx
Fig. 14. Progression of opinion values for the Karate club dataset with kg = kb = 5 under common coupled constraints (minmax value)
(b) τ = 2 (minymaxx
i v
i v
(cid:104)τ(cid:105)
i =−0.4655)
(cid:104)τ(cid:105)
i = 0.1775)
(cid:104)τ(cid:105)
i =−0.7709)
i v
(cid:80)
(cid:80)
(cid:80)
(cid:80)
(cid:80)
|
1905.09945 | 1 | 1905 | 2019-05-23T21:50:55 | Multifaceted Privacy: How to Express Your Online Persona without Revealing Your Sensitive Attributes | [
"cs.SI",
"cs.CR"
] | Recent works in social network stream analysis show that a user's online persona attributes (e.g., gender, ethnicity, political interest, location, etc.) can be accurately inferred from the topics the user writes about or engages with. Attribute and preference inferences have been widely used to serve personalized recommendations, directed ads, and to enhance the user experience in social networks. However, revealing a user's sensitive attributes could represent a privacy threat to some individuals. Microtargeting (e.g.,Cambridge Analytica scandal), surveillance, and discriminating ads are examples of threats to user privacy caused by sensitive attribute inference. In this paper, we propose Multifaceted privacy, a novel privacy model that aims to obfuscate a user's sensitive attributes while publicly preserving the user's public persona. To achieve multifaceted privacy, we build Aegis, a prototype client-centric social network stream processing system that helps preserve multifaceted privacy, and thus allowing social network users to freely express their online personas without revealing their sensitive attributes of choice. Aegis allows social network users to control which persona attributes should be publicly revealed and which ones should be kept private. For this, Aegis continuously suggests topics and hashtags to social network users to post in order to obfuscate their sensitive attributes and hence confuse content-based sensitive attribute inferences. The suggested topics are carefully chosen to preserve the user's publicly revealed persona attributes while hiding their private sensitive persona attributes. Our experiments show that adding as few as 0 to 4 obfuscation posts (depending on how revealing the original post is) successfully hides the user specified sensitive attributes without changing the user's public persona attributes. | cs.SI | cs |
MULTIFACETED PRIVACY: HOW TO EXPRESS YOUR ONLINE
PERSONA WITHOUT REVEALING YOUR SENSITIVE ATTRIBUTES
A PREPRINT
Victor Zakhary
Ishani Gupta
Rey Tang
Department of Computer Science
Department of Computer Science
Department of Computer Science
UC Santa Barbara
Santa Barbara, CA 93106
[email protected]
UC Santa Barbara
Santa Barbara, CA 93106
[email protected]
UC Santa Barbara
Santa Barbara, CA 93106
[email protected]
Amr El Abbadi
Department of Computer Science
UC Santa Barbara
Santa Barbara, CA 93106
[email protected]
May 27, 2019
ABSTRACT
Recent works in social network stream analysis show that a user's online persona attributes (e.g.,
gender, ethnicity, political interest, location, etc.) can be accurately inferred from the topics the
user writes about or engages with. Attribute and preference inferences have been widely used to
serve personalized recommendations, directed ads, and to enhance the user experience in social
networks. However, revealing a user's sensitive attributes could represent a privacy threat to some
individuals. Microtargeting (e.g., Cambridge Analytica scandal), surveillance, and discriminating
ads are examples of threats to user privacy caused by sensitive attribute inference. In this paper,
we propose Multifaceted privacy, a novel privacy model that aims to obfuscate a user's sensitive
attributes while publicly preserving the user's public persona. To achieve multifaceted privacy, we
build Aegis, a prototype client-centric social network stream processing system that helps preserve
multifaceted privacy, and thus allowing social network users to freely express their online personas
without revealing their sensitive attributes of choice. Aegis allows social network users to control
which persona attributes should be publicly revealed and which ones should be kept private. For
this, Aegis continuously suggests topics and hashtags to social network users to post in order to
obfuscate their sensitive attributes and hence confuse content-based sensitive attribute inferences.
The suggested topics are carefully chosen to preserve the user's publicly revealed persona attributes
while hiding their private sensitive persona attributes. Our experiments show that adding as few as 0
to 4 obfuscation posts (depending on how revealing the original post is) successfully hides the user
specified sensitive attributes without changing the user's public persona attributes
Keywords Attribute Privacy, Online-Persona, Content-Based Inference
1
Introduction
Over the past decade, social network platforms such as Facebook, Twitter, and Instagram have attracted hundreds of
millions of users [11 -- 13]. These platforms are widely and pervasively used to communicate, create online communi-
ties [26], and socialize. Social media users develop, over time, online persona [45] that reflect their overall interests,
activism, and diverse orientations. Users have numerous followers that are specifically interested in their personas and
their postings which are aligned with these personas. However, due to the rise of machine learning and deep learning
A PREPRINT - MAY 27, 2019
techniques, user posts and social network interactions can be used to accurately and automatically infer many user
persona attributes such as gender, ethnicity, age, political interest, and location [33, 38, 46, 47]. Recent work shows that
it is possible to predict an individual user's location solely using content-based analysis of the user's posts [19, 20].
Zhang et al. [46] show that hashtags in user posts can alone be used to precisely infer a user's location with accuracy of
70% to 76%. Also, Facebook likes analysis was successfully used to distinguish between Democrats and Republicans
with 85% accuracy [33].
Social network giants have widely used attribute inference to serve personalized trending topics, to suggest pages to like
and accounts to follow, and to notify users about hyper-local events. In addition, social networks such as Facebook use
tracking [9] and inference techniques to classify users into categories (e.g. Expats, Away from hometown, Politically
Liberal, etc.). These categories are used by advertisers and small businesses to enhance directed advertising campaigns.
However, recent news about the Cambridge Analytica scandal [6] and similar data breaches [14] suggest that users
cannot depend on the social network providers to preserve their privacy. User sensitive attributes such as gender,
ethnicity, and location have been widely misused in illegally discriminating ads, microtargeting, and surveillance. A
recent ACLU report [5] shows that Facebook illegally allowed employers to exclude women from receiving their job
ads on Facebook. Also, several reports have shown that Facebook allows discrimination against some ethnic groups in
housing ads [10]. News about the Russian-linked Facebook Ads during the 2016 election suggests that the campaign
targeted voters in swing states [2] and specifically in Michigan and Wisconsin [3]. In addition, location data collected
from Facebook, Twitter, and Instagram has been used to target activists of color [1].
An online-persona can be thought of as the set of user attributes that can be inferred about a user from their online
postings and interactions. These attributes fall into two categories: public and private persona attributes. Users should
decide which attributes fall in each category. Some attributes (e.g., political orientation and ethnicity) should be publicly
revealed as a user's followers might follow her because of her public persona attributes. Other attributes (e.g., gender
and location) are private and sensitive, and the user would not like them to be revealed. However, with the above
mentioned inference methods, the social media providers, as well as any adversary receiving the user posting can reveal
a user's sensitive attributes.
To remedy this situation, in this paper, we propose multifaceted privacy, a novel privacy model that aims to obfuscate
a user's sensitive attributes while revealing the user's public persona attributes. Multifaceted privacy allows users to
freely express their online public personas without revealing any sensitive attributes of their choice. For example, a
#BlackLivesMatter activist might want to hide her location from the police and from discriminating advertisers while
continuing to post about topics specifically related to her political movement. This activist can try to hide her location
by disabling the geo-tagging feature of her posts and hiding her IP address using an IP obfuscation browser like Tor [4].
However, recent works have shown that content-based location inferences can successfully and accurately predict a
user's location solely based on the content of her posts [19, 20, 46]. If this activist frequently posts about topics that
discuss BLM events in Montpelier, Vermont, she is most probably a resident of Montpelier (Montpelier has a low
African American population).
To achieve multifaceted privacy, we build Aegis1, a prototype client-centric social network stream processing system that
enables social network users to take charge of protecting their own privacy, instead of depending on the social network
providers. Our philosophy in building Aegis is that social network users need to introduce some noisy interactions
and obfuscation posts to confuse content based attribute inferences. This idea is inspired from Rivest's chaffing
and winnowing privacy model in [39]. Unlike in [39] where the sender and receiver exchange a secret that allows
the receiver to easily distinguish the chaff from the wheat, in social networks, a user (sender) posts to an open world
of followers (receivers) and it is not feasible to exchange a secret with every recipient. In addition, a subset of the
recipients could be adversaries who try to infer the user's sensitive attributes from their postings. Choosing this noise
introduces a challenging dichotomy and tension between the utility of the user persona and her privacy. Similar notions
of dichotomy between sensitive and non sensitive personal attributes have been explored in sociology and are referred
to as contextual integrity [37]. Obfuscation posts need to be carefully chosen to achieve obfuscation of private attributes
without damaging the user's public persona. For example, a #NoBanNoWall activist loses persona utility if she writes
about #BuildTheWall to hide her location. Multifaceted privacy represents a continuum between privacy and persona
utility. Figure 1 captures this continuum, where the x-axis represents the persona attributes that need to be obfuscated or
kept private while the y-axis represents the persona attributes that should be publicly preserved or revealed. Figure 1
shows that privacy is the reciprocal of the persona utility. Any attribute that needs to be kept private cannot be preserved
in the public persona. As illustrated, the more attributes are kept private, the more obfuscation overhead is needed to
achieve their privacy. A user who chooses to publicly reveal all her persona attributes achieves no attribute privacy and
hence requires no obfuscation posting overhead.
1Aegis: a shield in the Greek mythology.
2
A PREPRINT - MAY 27, 2019
Figure 1: The dichotomy of multifaceted privacy: persona vs. privacy.
As illustrated in Figure 1, a user chooses a point on the diagonal line of the multifaceted privacy that determines which
persona attributes should be publicly revealed (e.g., political interest, ethnicity, etc) and which ones should be kept
private (e.g., location, gender, etc). Note that users can reorder the attributes on the axes of Figure 1 in order to achieve
their intended public/private attribute separation. Unlike previous approaches that require users to change their posts
and hashtags [46] to hide their sensitive attributes, Aegis allows users to publish their original posts without changing
their content. Our experiments show that adding obfuscation posts successfully preserves multifaceted privacy. Aegis
considers the added noise as the cost to pay for achieving multifaceted privacy. Therefore, Aegis targets users who are
willing to write additional posts to hide their sensitive attributes.
Aegis is user-centric, as we believe that users need to take control of their own privacy concerns and cannot depend on
the social media providers. This is challenging as it requires direct user engagement and certain sacrifices. However,
we believe Aegis will help better understand the complexity of privacy as well as the role for individual engagement
and responsibility. Aegis represents a first step in the long path to better understand the tensions between user privacy,
the utility of social media, and trust in public social media providers. This is an overdue discussion that needs to be
discussed by the scientific community, and we believe Aegis will facilitate the medium for this discussion.
Our contributions are summarized as follows:
• We propose multifaceted privacy, a novel privacy model that represents a continuum between the privacy of
sensitive private attributes and public persona attributes.
• We build Aegis, a prototype user-centric social network stream processing system that preserves multifaceted
privacy. Aegis continuously analyzes social media streams to suggest topics to post that are aligned with the
user's public persona but hide their sensitive attributes.
• We conduct an extensive experimental study to show that Aegis can successfully achieve multifaceted privacy.
The rest of the paper is organized as follows. We explain the models of user, topic, and security in Section 2. Topic
classification algorithms and data structures that achieve multifaceted privacy are described in Section 3 and Aegis's
system design is explained in Section 4. Afterwards, an experimental evaluation is conducted in Section 5 to evaluate
the effectiveness of Aegis in achieving the multifaceted privacy. The related work is presented in Section 6 and future
extensions are presented in Section 7. The paper is concluded in Section 8.
2 Models
In this section, we present the user, topic, and security models. The user and topic models explain how users and topics
are represented in the system. The security model presents both the privacy and the adversary models.
3
2.1 User Model
A PREPRINT - MAY 27, 2019
Our user model is similar to the user model presented in [27]. The set U is the set of social network users where
U = {u1, u2, ...}. A user ui is represented by a vector of attributes Vui (e.g., gender Vui[g], ethnicity Vui[e], age Vui [a],
political interest Vui[p], location Vui[l], etc). Each attribute a has a domain a.d and the attribute values are picked
from this domain. For example, the gender attribute g has domain g.d = {male, female}2 and ∀ui∈U Vui[g] ∈ g.d. An
example user ux is represented by the vector Vux where Vux = (g: female, e: African American, a: 23, p: Democrat, l:
New York). Attribute domains can form a hierarchy (e.g., location: city → county → state → country) and an attribute
can be generalized by climbing up this hierarchy. A user who lives in Los Angeles is also a resident of Orange County,
California, and the United States. Other attributes can form trivial hierarchies (e.g., gender: male or female → * (no
knowledge)).
The user attribute vector Vui is divided into two main categories: 1) the set of public persona attributes V p
set of private sensitive attributes V s
hiding all sensitive attributes in V s
are the complement of the attributes in V s
ui. Multifaceted privacy aims to publicly reveal all persona attributes in V p
ui. Each user defines her V s
ui and 2) the
ui while
a priori. As shown in Figure 1, attributes in V p
ui
ui. Therefore, each attribute either belongs to V p
ui and V p
ui
ui.
ui or V s
2.2 Topic Model
i , t2
i , ..., tn
i = {t1
i ⊂ T represents the
The set T represents the set of all topics that are discussed by all the social network users in U. T τ
i }. Social network topics are characterized
set of all the topics posted by user ui's till time τ where T τ
by the attributes of the users who post about these topics. Unlike user attributes, which are discrete values, topic
attributes are represented as distributions. For example, an analysis of the ethnicity of the users who post about the
topic #BlackLivesMatter can result in the distribution 10% Asian, 25% White, 15% Hispanic, and 50% Black. This
distribution means that Asians, Whites, Hispanics, and Blacks post about the topic #BlackLivesMatter and 50% of the
users who post about this topic are Black. A topic ti is represented by a vector of attribute distributions Vti where Vti[g],
Vti[e], Vti[p], and Vti[l] are respectively the gender, the ethnicity, the political party, and the location distributions of
the users who post about ti.
To extract the gender, ethnicity, and political interest attribute distributions of different topics, we use the language
models introduced in [26, 41]. However, any other available model could be used to infer user attributes. The location
distribution of a topic is inferred using the geo-tagged public posts about this topic, where the location of the publisher
is explicitly attached to the post.
2.3 Security Model
An approach that is commonly used for attribute obfuscation is generalization. The idea behind attribute generalization
is to report a generalized value of a user's sensitive attribute in order to hide the actual attribute value within. Consider
location as a sensitive attribute example. Many works [36, 46] have used location generalization in different contexts.
Mokbel et al. [36] use location generalization to hide a user's exact location from Location Based Services (LBS). A
query that asks "what is the nearest gas station to my exact location in Stanford, CA?" should be altered to "list all gas
stations in California". Notice that the returned result of the altered query has to be filtered at the client side to find the
answer of the original query. Similarly, Andres et al. [17] propose geo-indistinguishability, a location privacy model
that uses differential privacy to hide a user's exact location in a circle of radius r from LBS providers. The wider the
generalization range, the more privacy achieved, and the more network and processing overhead are added at the client
side. Similarly, in the context of social networks, Zhang et al. [46] require Twitter users to generalize their location
revealing hashtags in order to hide their exact location. For example, a user whose post includes "#WillisTower" should
be generalized to "#Chicago" to hide a user's exact location. Notice that generalization requires users to alter their
original posts or queries.
To overcome these limitations and to allow users to write their posts using their own words, we adopt the notion of
k-attribute-indistinguishability privacy that is defined as follows. For every sensitive attribute s ∈ V s
ui, the user defines
an indistinguishability parameter ks. ks determines the number of attribute values among which the real value of
parameter s is hidden. For example, a user who lives in CA sets kl = 3 in order to hide her original state location, CA,
among 3 different states (e.g., CA, IL, and NY). This means that a content-based inference attack should not be able to
distinguish the user's real location among the set {CA, IL, NY}. As explained in 2.1, attribute domains either form
multi-level hierarchies (e.g., location) or trivial hierarchies (e.g., gender and ethnicity). Unlike in attribute generalization
where a user's attribute value is generalized by climbing up the attribute hierarchy, k-attribute-indistinguishability
2Due to the limitation of the inference models, the gender attribute is considered only binary. However, better models can be used
to infer non binary gender attribute values.
4
A PREPRINT - MAY 27, 2019
achieves the privacy of an attribute value by hiding it among k − 1 attribute values chosen from the siblings of the actual
attribute value in the same hierarchy level (e.g., a user's state level location is hidden among k − 1 other states instead
of generalizing it to the entire country). The following inference attack explains when k-attribute-indistinguishability is
achieved or violated.
The adversary assumptions: the adversary model and the inference attacks are similar to the ones presented in [46].
However, unlike in [46], our inference attack is not only limit to the location attribute but also can be extended to infer
every user sensitive attribute in V s
ui. The adversary has access to the set of all topics T and all the public posts related
to each topic. This assumption covers any adversary who can crawl or get access to the public posts of every topic
in the social network. As proposed in Section 1, the target user ui does not reveal her sensitive attribute values to the
public (e.g., a user who wants to hide her location must obfuscates her IP address and disable the geo-tagging feature
for her posts). Therefore, the adversary can only see the content of the public posts published by ui. The adversary uses
their knowledge about the set of all topics T and the set of topics T τ
i discussed by ui to infer her sensitive attributes.
Multifaceted privacy protects users against an adversary who only performs content-based inference attacks. Therefore,
multifaceted privacy assumes that the adversary does not have any side channel knowledge that can be used to reveal a
user's sensitive attribute value (e.g., another online profile that is directly linked to the user ui where sensitive attributes
such as gender, ethnicity, or location are revealed).
Inference attack: the adversary's ultimate goal is to reveal or at least have high confidence in the knowledge of the
sensitive attribute values of the target user ui. For this, the adversary runs a content-based attack as follows. First, the
that user ui wrote about. For each topic, the adversary infers the demographics of
adversary crawls the set of topics T τ
i
the users who wrote about this topic. Then, the adversary aggregates the demographics of all the topics in T τ
i . The
adversary uses the aggregated demographics to estimate the sensitive attributes of user ui. The details of the inference
attack is explained as follows.
i , an adversary crawls the set of posts Ptj that discusses topic tj and for each post pi ∈ Ptj , the
For each topic tj ∈ T τ
adversary uses some models to infer the gender, ethnicity, political interest, and location of the user who wrote this
post. Then, the adversary uses the inferred attributes of each post in topic tj to populate tj's distribution vector Vtj .
For example, Vtj [g] is the gender distribution of all users who wrote about topic tj. Similarly, Vtj [e], Vtj [p], and Vtj [l]
are the ethnicity, political interest, and location distributions of the users who posted about topic tj. We define V ∗
ui as
a vector of attribute distributions that is used to estimate the attributes of user ui. V ∗
ui is the result of aggregating the
normalized Vtj for every topic tj ∈ T τ
i as shown in Equation 1.
(cid:80)
V ∗
ui
=
i
tj∈T τ
T τ
i
Vtj
Ptj
(1)
ui
i . V ∗
ui
i divided by the number of topics in T τ
ui is the summation of the normalized Vtj for every topic tj ∈ T τ
Equation 1 shows that the topic's attribute distribution vector Vtj is first normalized by dividing Vtj by the number of
posts in topic tj. This normalization equalizes the effect of every topic tj ∈ T τ
i on the user's attribute estimations in
V ∗
ui. V ∗
[a]
is the distribution of attribute a for user ui. For example, a user might have a gender distribution V ∗
[g] = {female:0.8,
male:0.2}. This means that the inference attack using ui's posted topics suggests that the probability ui is a female is
80% while ui is a male is only 20%. For every attribute a, an attacker uses the maximum attribute value max(V ∗
[a])
as an estimation of the actual value Vui[a]. In the previous example, an attacker would infer V ∗
[g] = female as an
estimate of the gender of user ui. An inference attack succeeds in estimating an attribute a if the attacker can have
sufficient confidence in estimating the actual value Vui[a]. This confidence is achieved if the difference between the
maximum estimated attribute value of a and the top-kth estimated attribute value of a is greater than ∆a. For example,
if ∆g = 0.1 and kg = 2 (assuming gender is a binary attribute and it needs to be hidden among the 2 gender attribute
values), then an attacker successfully estimates ui's gender if the max(V ∗
[g]) is distinguishable from the 2nd highest
value in V ∗
[g] = {female:0.8, male:0.2}, an attacker succeeds
to estimate ui's gender = female as the difference between V ∗
Figure 2 show an example of a successful inference attack and another of a failed inference attack on attribute a of user
ui. As shown in Figure 2.a, the maximum estimated attribute value V ∗
[a]1 is distinguishable from the top-kth (ka for
attribute a) attribute values in V ∗
[a] by more than ∆a. In this scenario, an attacker can conclude with high confidence
that V ∗
[a]1 is indistinguishable from the top-kth attribute
values in V ∗
The parameter k is used to determine the number of attribute values within which the user's actual attribute value is
hidden. The bigger the k, the less the attacker's confidence about the user's actual attribute value. As a result, increasing
k introduces uncertainty in the attacker's inference and hence boosts the adversary cost to micro-target users who hide
[a]. In this scenario, the attack is marked failed and k-attribute-indistinguishability is achieved.
[a]1 is a good estimate for Vui[a]. However, in Figure 2.b, V ∗
[g] by more than 10%. In the previous example where V ∗
ui
ui
ui
[g].f emale − V ∗
[g].male ≥ ∆g.
ui
ui
ui
ui
ui
ui
ui
ui
ui
5
A PREPRINT - MAY 27, 2019
Figure 2: Illustrating k-attribute-indistinguishability
their actual attribute values among k different attribute values. For example, an adversary who wants to target a user
in location CA has to pay 3 times the advertisement cost to reach the same user if the user equally hides her location
among 3 other locations (e.g., CA, IL, and NY).
We understand that the requirement to determine the sensitive attributes and an indistinguishability parameter value ks
for every sensitive attribute s ∈ V s
ui could be challenging for many users. Users might not have a sense of the number
of attribute values to obfuscate the actual value of a particular attribute. One possible solution to address this usability
challenge is to design a questionnaire for Aegis's first time users. This questionnaire helps Aegis understand which
persona attributes are sensitive and how critical the privacy of every sensitive attribute is to every user. This allows
Aegis to auto-configure ks of every sensitive attribute s for every user. The details of such an approach is out of the
ui and for every attribute s ∈ V s
scope of this paper. In this paper, we assume that Aegis is preconfigured with the set V s
ui,
the value of ks is determined.
3 Multifaceted Privacy
Multifaceted privacy aims to obfuscate a user's sensitive attributes for every attribute in V s
ui. This has to be done
while publicly revealing every attribute in the user's public persona in V p
ui. Various approaches have been used to
obfuscate specific sensitive attributes, in particular, Tagvisor [46] protects users against content-based inference attacks
by requiring users to alter their posts by changing or replacing hashtags that reveal their sensitive attributes, in their
case location. Our approach is different, as it is paramount to not only preserving the privacy of the sensitive attributes,
but also to preserve the on-line persona of the user, and hence reveal their public attributes. It is critical for a user to
post their posting in their own words that reflect their persona. We therefore preserve multifaceted privacy by hiding a
specific post among other obfuscation posts. Our approach needs to suggest posts that are aligned with the user's public
persona but linked to alternative attribute values of their sensitive attributes in order to obfuscate them. This requires
a topic classification model that simplifies the process of suggesting obfuscation posts. For example, consider State
level location as a sensitive attribute. To achieve k-location indistinguishability, a user's exact State should be hidden
among k − 1 other States. This requires suggesting obfuscation postings about topics that are linked to these other
k − 1 States. Users in NY state can use topics that are mainly discussed in IL to obfuscate their location among NY
and IL. To discover such potential topics, all topics that are discussed on a social network need to be classified by the
sensitive attributes that need to be obfuscated, State level location in this example. A topic is linked to some State if the
maximum estimated State location of this topic, max(Vti[l]), is distinguishable from other State location estimates in
Vti[l] by more than ∆l. For example, if ∆l = 10%, a topic that has a State location distribution of {NY=0.6, IL=0.2,
CA=0.1, Others=0.1} is linked to NY State while a topic that does not have a distinguishable State location inference
by more than ∆l is not linked to any State. In this section, we first explain a simple but incorrect topic classification
model that successfully suggests obfuscation posts that hide a user's sensitive attributes but does not preserve her public
persona. Then, we explain how to modify the topic classification model to suggest obfuscation posts that do achieve
multifaceted privacy.
A simple incorrect proposal:
in this proposal, topics are classified independently by every sensitive attribute. As
shown in Figure 3, each attribute forms an independent hierarchy. The root of the hierarchy has the topics that are not
linked to a specific attribute value. Topics that are strongly linked to some attribute value fall down in the hierarchy
node that represents this attribute value. For example, State level location attribute forms a hierarchy of two levels. The
first level, the root, has all the topics that do not belong to a specific State. A topic like #Trump is widely discussed in all
the States and therefore it resides on the root of the location attribute. However, #cowboy is mainly discussed in TX and
6
A PREPRINT - MAY 27, 2019
Figure 3: Each attribute independently forms a hierarchy
therefore it falls down in the hierarchy to the TX node. To obfuscate a user's State location, topics need to be selected
from the sibling nodes of the user's State in the State level location hierarchy. These topics belong to other locations and
can be used to achieve k-location-indistinguishability privacy. Although this proposal successfully achieves location
privacy, the suggested posts are not necessarily aligned with the user's public online persona. For example, this
obfuscation technique could suggest the topic #BuildTheWall (from TX) to an activist (from NY) who frequently posts
about #NoBanNoWall in order to hide her location. This misalignment between the suggested obfuscation posts and the
user's public persona discourages users from seeking privacy fearing the damage to their public online persona.
Figure 4: A dependent topic tree where public attributes are at the top while private attributes are at the bottom
A persona preserving proposal: To overcome the independent classification shortcomings, obfuscation postings need
to be suggested from a tree hierarchy that captures relevant dependencies among all user attributes including both public
persona and private sensitive attributes. Public attributes need to reside on the upper levels of the classification tree while
private ones reside on the bottom levels of the tree as shown in Figure 4. To achieve k-attribute-indistinguishability,
sibling values of the sensitive attributes are used to hide the actual value of these sensitive attributes. By placing
the public attributes higher up on the tree, we ensure that the suggest topics adhere to the public persona. Finally,
multifaceted privacy only requires all the public attributes, regardless of their order, to reside on the upper levels of the
tree while all the private attributes, regardless of their order, to reside on the lower levels of the tree.
Social network topics are dependently classified in the tree by the attribute domain values at each level. For example,
if the top most level of the tree is the political party attribute, topics that are mainly discussed by Democrats are placed
in the left green child while topics that are mainly discussed by Republicans are place in the right green child. Note
that topics that have no inference reside in the root of the hierarchy. Now, if the second public persona attribute is
ethnicity (shown as blue nodes in Figure 4), topics in both the Democratic and Republican nodes are classified by the
ethnicity domain attribute values. For example, topics that are mainly discussed by White Democrats are placed under
7
A PREPRINT - MAY 27, 2019
the Democratic node in the White ethnicity node while topics that are mainly discussed by Asian Republicans are
put under the Republican node in the Asian ethnicity node. This classification is applied at every tree level for every
attribute. Now, assume a user is White, Female, Democrat, who lives in CA and wants to hide her location (shown as
red nodes in Figure 4) while publicly revealing her ethnicity, gender, and political party. In this case, topics that reside in
the sibling nodes of the leaf of her persona path, e.g., topics that are mainly discussed by White Female Democrats who
live in locations other than CA (e.g., NY and IL) can be suggested as obfuscation topics. This dependent classification
guarantees that the suggested topics are aligned with the user's public persona but belong to other sensitive attribute
values (different locations in this example). Note that this technique is general enough to obfuscate any attribute and
any number of attributes. Each user defines her sensitive attribute(s) and the classification tree would be constructed
with these attributes to the bottom thus guaranteeing that the suggested posts do not violate multifaceted privacy.
4 Aegis System Design
This section presents Aegis, a prototype social network stream processing system that implements multifaceted privacy
and overcomes the adversarial content-based attribute inference attacks discussed in Section 2.3. Aegis achieves
k-attribute-indistinguishability by suggesting topics to post that are aligned with the social network user's public persona
while hiding the user's sensitive attributes among their other domain values. To achieve this, Aegis uses the classification
and suggestion models discussed in Section 3. Aegis is designed in a user-centric manner which is configured on the
user's local machine. In fact, Aegis can be developed as a browser extension where all the user interactions with the
social network are handled through this extension. Every local deployment of Aegis only needs to construct a partition
of the attribute-based topic classification hierarchy developed in Figure 4. This partition (or sub-hierarchy) include the
user's attribute path from the root to a leaf in addition to the sibling nodes of the user's sensitive attribute nodes. For
example, a user whose attributes are Female, White, Democrat, and CA and wants to obfuscate her State location only
requires Aegis to construct the Female, White, Democrat, CA path in addition to k − 1 other paths with the shared
prefix Female, White, Democrat but linked to k − 1 other States. These k − 1 States are used to hide the user's true state
in order to achieve k-location-indistinguishability. Note that if another user's public persona is specifically associated
with their location while they consider their ethnicity to be sensitive, then the hierarchy needs to be reordered to reflect
this criteria.
Although Aegis can be integrated with different online social network platforms, our prototype implementation of Aegis
only supports Twitter to illustrate Aegis's functionality. Twitter provides developers with several public APIs [8] that
allow them to stream tweets that discuss certain topics. In addition, Twitter streaming APIs allow developers to sample
1% of all the tweets posted on Twitter. In Twitter, a topic is represented by either a hashtag or a keyword. Aegis is built
to work for new Twitter profiles in order to continuously confuse an adversary about a profile's true sensitive attribute
values from the genesis of this profile. Aegis is not designed to work with existing old profiles as an adversary could
have already used their existing posts to reveal their sensitive attribute values. Even though Aegis suggests obfuscation
posts, an adversary can distinguish the old original posts from the newly added original posts accompanied by their
obfuscation posts and hence reveals the user's sensitive attributes true values.
Aegis is designed to achieve the following goals:
1. to automate the process of streaming and classifying Twitter topics according to their attributes,
2. to construct and continuously maintain the topic classification sub-hierarchy,
3. and finally to use the topic classification sub-hierarchy to suggest topics to the user that achieve multifaceted
privacy.
To achieve these goals, Aegis consists of two main processes:
• a Twitter analyzer Process T P and
• a topic Suggestion Process SP .
T P continuously analyzes the topics that are being discussed on Twitter and for each topic ti, T P uses the topic
attribute inference models to infer ti's attribute distribution vector Vti. The accuracy of the topic attribute inference
increases as the number of posts that discuss topic tj, Ptj, increases. T P uses a local key-value store as a topic
repository where the key is the topic id tj ∈ T and the value is the topic attribute distribution vector Vti. In addition,
T P constructs and continuously maintains the topic classification sub-hierarchy that classifies the topics based on
their attributes. This topic classification sub-hierarchy is used for suggesting obfuscation topics. For this, the user
provides T P with both their attribute vector values Vui and their sensitive attribute vector V s
ui determine
ui. Figure 5 shows the interactions
the sub-hierarchy that T P needs to maintain in order to obfuscate the attributes in V s
ui. Vui and V s
8
A PREPRINT - MAY 27, 2019
Figure 5: Aegis System Design and User Interaction Flow
s inferred attribute value by more than 10%.
among the user, Aegis, and the social network. As shown, step 0 represents the continuous Twitter stream analysis
performed by T P . As T P analyzes Twitter streams, it continuously updates the topic repository and the classification
sub-hierarchy.
The topic suggestion process SP mainly handles user interactions with Twitter. SP uses the topic classification
sub-hierarchy constructed and maintained by T P to suggest obfuscation topics. For every sensitive attribute s ∈ V s
ui,
the user provides the indistinguishability parameter ks that determines how many attribute values from the domain of s
should be used to hide the true value of s. Aegis allows users to configure the privacy parameter ∆s for every attribute
s. However, to enhance usability, Aegis maintains a default value for the privacy parameter ∆s = 10%. This means
that the privacy of attribute s is achieved if the inference attack cannot distinguish the maximum inferred attribute value
from the kth
SP uses ks, and ∆s for every attribute s ∈ V s
ui to generate the topic suggestion set Si. Note that SP is locally deployed
at the user's machine. Therefore, the user does not have to trust any service outside of her machine. Aegis is designed
to transfer user trust from the social network providers to the local machine. For every sensitive attribute s, SP selects
a fix set of ks − 1 attribute domain values. These ks − 1 attribute values are used to obfuscate the true value of attribute
s, Vui[s].
As shown in Figure 5, in Step 1, the user writes a post to publish on Twitter. SP receives this post and queries T P
about the attributes of all the topics mentioned in this post. SP uses these topic attributes to simulate the adversarial
attack. If T P 's topic inference indicates that k-attribute-indistinguishability is violated for any attribute s ∈ V s
ui, SP
ui but linked to the other attribute values of s in the set of ks − 1 attribute
queries T P for topics with public persona V p
values. For every returned topic, SP ensures that writing about this topic enhances the aggregated inference of the
original post and the obfuscation posts towards ks-attribute-indistinguishability. SP adds these topics to the set Si and
returns them to the user (Step 2). The user selects a few topics from Si to post in Step 3 and submits the posts to SP .
Note that users are required to write the obfuscation posts using their personal writing styles to ensure that the original
posts and the obfuscation posts are indistinguishable [16, 30]. Afterwards, SP ensures that the aggregated inference of
submitted obfuscation posts in addition to the original post lead to k-attribute-indistinguishability. Otherwise, SP keeps
suggesting more topics. As every original post along with its obfuscation posts achieve k-attribute-indistinguishability,
the aggregated inference over the whole user's posts achieve k-attribute-indistinguishability. In Step 4, SP queues the
original and the obfuscation posts and publishes them on the user's behalf in random order and intervals to prevent
timing attacks (Step 5). An adversary can perform a timing attack if the original posts and the obfuscation posts are
distinguishable. Queuing and randomly publishing the posts prevents the adversary from distinguishing original posts
from the obfuscation posts and hence prevents timing attacks.
We understand that the obfuscation writing overhead might alienate users from Aegis. As a future extension, Aegis
can exploit deep neural network language models to learn the user's writing style [21]. Aegis can use such a model
to generate [31] full posts instead of hashtags and users can either directly publish these posts or edit them before
publishing.
5 Experimental Evaluation
In this section, we experimentally evaluate the effectiveness of Aegis in achieving multifaceted privacy. We first present
the experimental setup and analyze some properties of the used dataset in Section 5.1. Then, Sections 5.2 and 5.3
present illustrative inference and obfuscation examples that show the functionality of Aegis using real Twitter topics.
We experimentally show how Aegis can be used to hide user location in Section 5.4 and measure the effect of changing
the indistinguishability parameter k on the obfuscation overhead in Section 5.5. Finally, Section 5.6 illustrates the
efficiency of Aegis on hiding the user gender while preserving other persona attributes.
9
5.1 Experimental Setup
A PREPRINT - MAY 27, 2019
(a) #GiveAway (Negligible Connection)
(b) #2a (Weak Connection)
(c) #Disney (Strong Connection)
Figure 6: Examples of negligible, weak, and strong connection distribution topics for top-20 personas
Although Aegis can be integrated with different social network platform, our prototype implementation is integrated
only with the Twitter social network. Twitter provides developers with several public APIs [8] including the streaming
API that allows developers to crawl 1% of all the postings on Twitter. For our experiments, we use the 1% random
sampling of the Twitter stream during the year 2017. The attributes gender, ethnicity, and location are used to build a
three level topic classification hierarchy that classifies all topics according to their attribute distribution. For simplicity
and without loss of generality, we use the language models in [41] to infer both gender and ethnicity attributes of a post
writer. In addition, we infer the location distribution of different topics using the explicitly geo-tagged posts about these
topics. In the 1% of Twitter's 2017 postings, our models were able to extract 2,126,791 unique topics. Our classification
hierarchy suggests that 66% of the dataset tweets are posted by males. This analysis is consistent with the statistics
published in [7]. In addition, the dataset has 6,864,300 geo-tagged posts, 15% of which originated in California. Finally,
the predominant ethnicity extracted from the dataset is White.
As the classification hierarchy is built using only gender, ethnicity, and state location attributes, this results in a hierarchy
of 500 different paths from the root to a leaf of the hierarchy. These 500 paths result from all the possible combinations
of gender (male, female), ethnicity (White, Black, Asian, Hispanic, Native American), and state location (50 States).
The 500 paths represent the different 500 personas considered in our experiments. Our topic classification hierarchy
suggests that topics vary significantly in their connection to a specific persona path (a gender, ethnicity, and location
combination). For example, #GiveAway is widely discussed among the 500 personas across the 50 States with very
little skew towards specific personas over others. For topics that are widely discussed across all different persona, their
skew is usually proportional to the population density of different States. For example, the top five highly populated
states (CA, TX, FL, NY, and PA) usually appear as the top locations where widely discussed topics are posted. On
the other hand, other topics show strong connection to specific personas. For example, 33.93% of the personas who
write about #Disney are Male, Asian, and live in Florida where Disney World is located. Figure 6 shows 3 examples of
topics that have trivial, weak, and strong connection to specific personas. In Figure 6, the x-axis represents the top-20
personas who post about a topic and the y-axis represents the percentage of postings for each persona. As shown in
Figure 6a, #GiveAway has slight skew (negligible connection) towards some personas over other personas. Also, the
top most persona who post about #GiveAway represent only 3.94% of the overall postings about the topic. On the other
hand, Figure 6b shows that the persona distribution for #2a (refers to the second amendment) has more skew (weak
connection) towards some personas over others. As shown, the top posting persona on the topic #2a contributes 14.03%
of the overall postings of this topic. Finally, a topic like #Disney has remarkable skew (strong connection) towards
some personas over others. As shown in Figure 6c, the top posting persona on the topic #Disney contributes 33.93% of
the overall postings of this topic.
Strength Minimum δ Maximum δ
Negligible
Weak
Mild
Strong
10%
20%
30%
100%
0%
10%
20%
30%
Table 1: Topic to persona connection strength categories and their corresponding δ ranges
We define a topic to persona connection strength parameter δ. δ is defined as the difference in posting percentage
between the top-1 posting persona and the top-k posting persona. For #2a and k = 3, δ = 14.03 − 3.45 = 10.6 while
for #Disney and for k=3, δ = 33.93 − 6.19 = 27.74. We categorize topics into 4 categories according to their δ value.
As shown in Table 1, a topic to persona connection that ranges from 0% to 10% represents a negligible connection and
hence this topic does not reveal the persona attributes of the users who write about it. As the topic to persona connection
10
0 5 10 15 20 25 30 351234567891011121314151617181920Topic postings per persona (%)Top 20 Personas#GiveAway Distribution 0 5 10 15 20 25 30 351234567891011121314151617181920Topic postings per persona (%)Top 20 Personas#2a Distribution 0 5 10 15 20 25 30 351234567891011121314151617181920Topic postings per persona (%)Top 20 Personas#Disney Distribution A PREPRINT - MAY 27, 2019
Topic
#teen
#hot
#etsy
#diy
#actor
#cowboys
Freq M W Ca F W Ca M W Tx F W Tx
7094
7478
2739
1987
725
797
15
13
27
11
9
1
Table 2: Topic Analysis By Persona
4
3
1
1
11
2
7
5
6
3
1
5
1
1
5
7
2
16
increases, the potential of revealing the attributes of the users who write about this topic increases. In our experiments,
we measure the overhead of obfuscation for three different distributions with weak, mild and strong connections having
a topic to persona strength connection ranges that are shown in Table 1
5.2
Illustrative Inference Example
Using our tree data structure we can infer interesting information about different topics on Twitter in our dataset,
specifically regarding correlations between persona and topics. In Table 2 we analyze 7 topics and their connection to 4
of the most prominent personas in our dataset (Male-White-CA, Female-White-CA, Male-White-TX, Female-White-
TX). Frequency denotes the number of times the topic was observed and the number under a persona for a particular
topic denotes the order or rank in which a persona discusses this topic most. For example, among all the persona we
analyze in our dataset, #actor is most discussed by White Male Californians (rank 1) closely followed by White Female
also from California (rank 2). This is followed by other persona out of the focus of Table 2, until White Female Texans
are reached at rank 9 and White Male Texans at rank 11. Also, Table 2 shows that both #teen and #hot are discussed the
most by Female White Californians and that overall Females (in both CA and TX) who discuss this topic are more than
Males in both CA and TX. We can also observe correlations across topics that have semantic connections like #etsy and
#diy. Etsy is an online platform for users to sell DIY (Do It Yourself) projects. Female White Texans are much more
interested in such DIY specific topics than any other of the personas. Lastly, high correlations of certain topics can be
observed with specific locations such as #actor with California and #cowboys with Texas. As Table 2 reveals, the topics
you post on social media significantly reveal your attributes, even private sensitive attributes you are unwilling to share.
As such, we need a tool like Aegis to prevent adversaries from inferring private attributes while preserving others.
5.3
Illustrative Obfuscation Example
Figure 7: Illustrative Gender Obfuscation Example.
In this section, we provide an illustrative obfuscation example that shows how Aegis achieves multifaceted privacy.
This example begins with a newly created Twitter profile of a Male, White user who lives in California. The user
wants to obfuscate his gender among the gender domain values {male, female} achieving 2-gender-indistinguishably.
Indistinguishability holds if the privacy parameter ∆g = 0.1 is achieved. ∆g is set to 0.1 (or 10%) to ensure that users
who write topics with only negligible topic to persona strength connection do not have to add any obfuscation posts to
their timelines as the topics they post do not reveal their sensitive attributes. In addition, the user wants to preserve
his ethnicity and location attributes as his public persona. Now, assume that the user tweets #gowarriors to show his
support for his favorite Californian basketball team, the Golden State Warriors.
11
0 10 20 30 40 50 60 70 80#gowarriors#womenintech#organicfood#bodybuildingGender Inference (%)The Tweeted HashtagsMaleFemaleδ43.0019.0011.007.00 A PREPRINT - MAY 27, 2019
(a) Location (Weak)
(b) Location (Mild)
(c) Location (Strong)
(d) Persona (Weak)
(e) Persona (Mild)
(f) Persona (Strong)
Figure 8: Effect of obfuscation posts on location and user persona given weak, mild and strong connected topics to
locations
Unfortunately, as shown in Figure 7, #gowarriors has a strong connection to the male gender attribute value. In
Figure 7, the x-axis represents the posted hashtags one after the other and the y-axis represents the aggregated
gender inferences for both male and female attribute values over all the posted hashtags. In addition, δ represents the
difference of the aggregated gender inference between male and female attribute values. As shown, initially, δ = 43%
which indicates a strong link between the user's gender and the male attribute value. As δ ≥ ∆g, this indicates that
2-gender-indistinguishably is violated.
Therefore, Aegis suggests to post topics that are mainly discussed by White people who live in California but linked
to the female gender attribute value. Figure 7 shows the effect of posting subsequent topics on the aggregated gender
inference. The topic #womenintech helps to reduce the aggregate inference difference to 19%. #organicfood brings
the difference down to 11% and #bodybuilding reduces it to 7%. Notice that δ = 7% achieves δ ≤ ∆g and hence
2-gender-indistinguishably is achieved. Notice that the same example holds if the user's true gender is female. The goal
of Aegis is not to invert the gender attribute value but to achieve inference indistinguishability among kg = 2 different
gender attribute values.
5.4 User Location Obfuscation
The number and the topic to persona connection strength of the obfuscation topics largely depend on the topic to
persona connection strength of the original posts. In this experiment, we show how Aegis is used to hide user location
while preserving their gender and ethnicity. In this experiment, we set kl = 3 where user location is hidden among
three locations. ∆l is set to 0.1 to indicate that 3-location-indistinguishability is achieved if the difference between
the highest (top-1) aggregated location inference and the 3rd (top-3) aggregated location inference is less than 10%.
The number of obfuscation posts needed and their effect on the user persona are reported. This experiment runs using
4707 weak topics, 1984 mild topics, and 1106 strong topics collected from Twitter over several weeks. This experiment
assumes a newly created twitter profile simulated with one of the personas in Table 2. First, a post with some topic to
location connection strength (weak, mild, or strong) is added to the user profile. Then, we add the suggested obfuscation
posts one at a time to the user's timeline. After every added obfuscation post, the location and persona inference are
reported. The reported numbers are aggregated and averaged for every topic to location connection strength category.
Figure 8 shows the effect of adding obfuscation posts on both the location inference and the persona inference for
weak, mild, and strong topics. As shown in Figure 8a, the 4707 weak topics on average need just one obfuscation
suggestion to achieve 3-location-indistinguishability when ∆l = 0.1. Note that adding more obfuscation posts achieves
3-location-indistinguishability for smaller ∆ls (e.g., ∆l = 0.05 (4 suggestions), ∆l = 0.04 (5 suggestions)). Also,
adding one suggestion post achieves 26.5% reduction in δ. The same experiment is repeated for mild and strong topics
and the results are reported in Figures 8b and 8c respectively. Notice that strong topics requires four suggestions on
average to achieve 3-location-indistinguishability when ∆ = 0.1. Also, in Figures 8c, adding a single suggestion post
achieves 49.8% reduction in δ.
12
0 10 20 30 40 50012345Aggregated Location Inference (%)Number of Suggested TopicsTop-1 LocationTop-3 Locationδ12.689.326.935.894.623.76 0 10 20 30 40 50012345Aggregated Location Inference (%)Number of Suggested TopicsTop-1 LocationTop-3 Locationδ23.8612.359.758.227.206.20 0 10 20 30 40 50012345Aggregated Location Inference (%)Number of Suggested TopicsTop-1 LocationTop-3 Locationδ45.8218.4213.1010.909.218.03 65 70 75 80 85 90 95012345Aggregated Attribute Inference (%)Number of Suggested TopicsGenderEthnicity 65 70 75 80 85 90 95012345Aggregated Attribute Inference (%)Number of Suggested TopicsGenderEthnicity 65 70 75 80 85 90 95012345Aggregated Attribute Inference (%)Number of Suggested TopicsGenderEthnicity A PREPRINT - MAY 27, 2019
Figures 8d, 8e, and 8f show the effect on user persona after adding obfuscation posts to weak, mild, and strong posts
respectively. User persona is represented by gender and ethnicity. As obfuscation posts are carefully chosen to align
with the user persona, we observe negligible changes on the average gender and ethnicity inferences after adding
obfuscation posts.
5.5 The Effect of Changing kl
Figure 9: Change in δ as obfuscation posts are added
In this experiment, we measure the effect of changing the parameter kl on the number of obfuscation posts required to
achieve k-location-indistinguishability. kl determines the number of locations within which user ui wants to hide her
true location. Increasing kl increases the achieved privacy and boosts the required obfuscation overhead to achieve
k-location-indistinguishability. Assume users in Texas hide their State level location among 3 States: Texas, Alabama,
and Arizona. A malicious advertiser who wants to target users in Texas is uncertain about their location and now has to
pay 3 times the cost of the original advertisement campaign to reach the same target audience. Therefore, increasing k
inflates the cost of micro-targeting.
This experiment uses 621 strong topics collected from Twitter over several weeks. The average δ = Vui[l]1 − Vui[l]kl
is reported for all topics. In addition, the effect of adding obfuscation posts on the average δ for different values of
kl = 3, 5, and 7 is reported. Figure 9 shows the effect of adding obfuscation posts on the aggregated location inference
for different values of kl. The privacy parameter ∆l is set to ∆l = 0.1.
As shown in Figure 9, achieving 3-location indistinguishability for strong topics requires 4 obfuscation posts on the
average for ∆l = 0.1. On the other hand, 7-location indistinguishability requires more than 5 obfuscation posts for the
same value of ∆l. This result highlights the trade-off between privacy and obfuscation overhead. Achieving higher
privacy levels by increasing kl or lowering ∆l requires more obfuscation posts and hence more overhead. Obfuscation
posts are carefully chosen to align with the user persona. Therefore, we observe negligible changes on the average
gender and ethnicity inferences after adding obfuscation posts for different values of kl.
5.6 User Gender Obfuscation
Figure 10: The effect of obfuscation posts on gender inference for strong connected topics to gender.
13
5 10 15 20 25 30 35 40 45 50012345Aggregated Location Inference (%)Number of Suggested Topicsδ, kl = 3δ, kl = 5δ, kl = 7 0 10 20 30 40 50 60 70 80012345Aggregated Gender Inference (%)Number of Suggested TopicsMaleFemaleδ57.9134.6316.358.414.522.34 A PREPRINT - MAY 27, 2019
This experiment shows how Aegis is used to hide user gender while preserving their ethnicity and location. In this
experiment, we set kg = 2 where user gender should be hidden among male and female gender domain values. ∆g is
set to ∆g = 0.1 to indicate that 2-gender-indistinguishability is achieved if the difference between the highest (top-1)
aggregated gender inference and the 2nd (top-2) aggregated gender inference is less than 10%. This experiment runs
over 40 gender strongly connected topics. The aggregated gender inference is reported when adding the original strong
post to the user's timeline and after adding every obfuscation post one at a time. In addition, δ, the difference between the
male gender inference and the female gender inference is reported. 2-gender-indistinguishability is achieved if δ ≤ 10%.
As shown in Figure 10, gender strongly connected topics result in high δ that violates the 2-gender-indistinguishability
privacy target. Therefore, Aegis suggest obfuscation posts that results in δ reduction. Figure 10 shows that strong topics
need on the average 3 obfuscation posts to achieve the gender privacy target. This result is quite consistent with the
location obfuscation experiments in Section 5.4. This shows Aegis's obfuscation mechanism is quite generic and can
be efficiently used to hide different user sensitive attributes. Finally, as the obfuscation posts are carefully chosen to
align with the user's public persona attributes, we observed negligible changes on the average location and ethnicity
inferences after adding obfuscation posts.
6 Related Work
The problem of sensitive attribute privacy of social network users has been extensively studied in the literature from
different angles. k-anonymity [40, 42, 43] and its successors l-diversity [35] and t-closeness [34] are well-known and
widely used privacy models in publishing dataset to hide user information among a set of indistinguishable users in
the dataset. Also, differential privacy [22 -- 24] has been widely used in the context of dataset publishing to hide the
identity of a user in a published dataset. These models focus on hiding user identity among other users in the published
dataset. Another variation of differential privacy is pan-privacy [25]. Pan-privacy is designed to work for data streams
and hence it is more suitable for social network streams privacy. However, these models are service centric and assume
trusted service providers. In this paper, we tackle the privacy problem from the end-user angle where the user identity is
known and all their online social network postings are public and connected to their identity. Our goal is to confuse
content-based sensitive attribute inference attacks by hiding the user's original public posts among other obfuscation
posts. Our multifaceted privacy achieves the privacy of user sensitive attributes without altering their public persona.
In the context of social networks privacy, earlier works focus on sensitive attribute inferences due to the structure of
social networks. In [47], Zhelava and Getoor attempt to infer the user's sensitive attributes using public and private user
profiles. However, the authors do not provide a solution to prevent such inference attacks. Georgiou et al. [26, 27]
study the inference of sensitive attributes in the presence of community-aware trending topic reports. An attacker can
increase their inference confidence by consuming these reports and the corresponding community characteristics of the
involved users. In [27], a mechanism is proposed to prevent social network services from publishing trending topics
that reveal information about individual users. However, this mechanism is service centric and it is not suitable for
hiding a user's sensitive attributes against content-based inference attacks. Ahmad et al. [15] introduce a client-centered
obfuscation solution for protecting user privacy in personalized web searches. The privacy of a search query is achieved
by hiding it among other obfuscation search queries. Although this work is client-centered, it is not suitable for social
networks privacy where the user online persona has to be preserved.
Recent works have focused on the privacy of some sensitive attributes such as location of social network users. Ghufran
et al. [28] show that social graph analysis can reveal user location from friends and followers locations. Although, it
is important to protect user sensitive attributes like location against this attack, Aegis focuses only on content-based
inference attacks. Yakout et al. [44] proposed a system called Privometer, which measures how much privacy leaks
from certain user actions (or from their friends' actions) and creates a set of suggestions that could reduce the risk of a
sensitive attribute being successfully inferred. Similar to Privometer, [32] proposes sanitation techniques to the structure
of the social graph by introducing noise, and obfuscating edges in the social graphs to prevent sensitive information
inference. Andres et al. [17] propose geo-indistinguishability, a location privacy model that uses differential privacy
to hide a user's exact location in a circle of radius r from locaction based service providers. In a recent work, Zhang
et al. [46] introduce Tagvisor, a system to protect users against content-based inference attacks. However, Tagvisor
requires users to alter their posts by changing or replacing hashtags that reveal their location. Other works [18, 29]
depend on user collaboration to hide an individual's exact location among the location of the collaboration group. This
approach requires group members to collaborate and synchronously change their identities to confuse adversaries.
However, these techniques are prone to content-based inference attacks and collaboration between users might be hard
to achieve in the social network context. For location based services, Mokbel et al. [36] use location generalization and
k-anonymity to hide the exact location of a query. These works do not preserve the user online persona while achieving
location privacy. In addition, these works do not provide a generic mechanism to hide other sensitive attributes such as
user gender and ethnicity.
14
A PREPRINT - MAY 27, 2019
This paper presents Aegis, the first persona friendly system that enables users to hide their sensitive attributes while
preserving their online persona. Aegis is a client-centric solution that can be used to hide any user specified sensitive
attribute. Aegis does not require users to alter their original posts or topics. Instead, Aegis hides the user's original
posts among other obfuscation posts that are aligned with their persona but linked to other sensitive attribute values
achieving k-attribute-indistinguishability.
7 Future Extensions
Research in the social network privacy has focused on dataset publishing and obfuscating user information among other
users. Such focus is service centric and assumes that service providers are trusted. However, Aegis is user centric and
aims to give the users control over their own privacy. This control comes with a cost represented by the obfuscation
posts that need to be posted by users. The role of Aegis is to automate the topic suggestion process and to ensure that
k-attribute-indistinguishability holds against content-base inference attacks. There are several directions where these
obfuscation and privacy models can evolve.
Obfuscation Post Generation. Aegis suggests topics as keywords or hashtags and requires users to write the obfuscation
posts using their personal writing styles to ensure that the original posts and the obfuscation posts are indistinguishable.
However, the obfuscation writing overhead might alienate users from Aegis. Instead, Aegis can exploit deep neural
network language models to learn the user's writing style. Aegis can use such a model to suggest full posts instead of
hashtags and users can either directly publish these posts or edit them before publishing. This extension aims to reduce
the overhead on the users by automating the obfuscation post generation.
Social Graph Attack Prevention. Aegis is mainly designed to obfuscate the user sensitive attributes against content-based
inference attacks. An orthogonal attack is to use the attribute values of friends and followers to infer a user's real
sensitive attribute value. Take location as a sensitive attribute example. Ghufran et al., [28] show that user location can
be inferred from the locations of followers and friends. A user whose friends are mostly from NYC is highly probable
to be from NYC. Aegis can be extended to prevent this attack. Users of similar persona but different locations can
create an indistinguishability network where users in this network have followers and friends from different locations.
Similar to the obfuscation topics, Aegis could suggest users to follow with similar persona but different locations.
8 Conclusion
In this paper, we propose multifaceted privacy, a novel privacy model that obfuscates a user's sensitive attributes
while publicly revealing their public online persona. To achieve the multifaceted privacy, we build Aegis, a prototype
client-centric social network stream processing system that achieves multifaceted privacy. Aegis is user-centric and
allows social network users to control which persona attributes should be publicly revealed and which should be kept
private. Aegis is designed to transfer user trust from the social network providers to her local machine. For this, Aegis
continuously suggests topics and hashtags to social network users to post in order to obfuscate their sensitive attributes
and hence confuse content-based sensitive attribute inferences. The suggested topics are carefully chosen to preserve
the user's publicly revealed persona attributes while hiding their private sensitive persona attributes. Our experiments
show that Aegis is able to achieve sensitive attributes privacy such as location and gender. Adding as few as 0 to 4
obfuscation posts (depending on how strongly connected the original post is to a persona) successfully hides the user
specified sensitive attributes without altering the user's public persona attributes.
References
[1] Facebook, instagram, and twitter provided data access for a surveillance product marketed to target activists of
color. https://www.aclunc.org/blog/facebook-instagram-and-twitter-provided-data-access-
surveillance-product-marketed-target, 2016.
[2] Russia-linked facebook ads reportedly aimed for swing states. http://fortune.com/2017/10/04/trump-
russia-facebook-ads-michigan-wisconsin-swing/, 2017.
[3] Russian-linked facebook ads targeted michigan and wisconsin.
https://www.cnn.com/2017/10/03/
politics/russian-facebook-ads-michigan-wisconsin/index.html, 2017.
[4] Tor Browser. https://www.torproject.org/, 2017.
[5] Aclu and workers take on facebook for gender discrimination in job ads. https://www.aclu.org/news/aclu-
and-workers-take-facebook-gender-discrimination-job-ads, 2018.
15
A PREPRINT - MAY 27, 2019
[6] Cambridge analytica data breach.
http://www.businessinsider.com/cambridge-analytica-data-
breach-harvested-private-facebook-messages-2018-4, 2018.
[7] Distribution of twitter users worldwide as of october 2018, by gender.
https://www.statista.com/
statistics/828092/distribution-of-users-on-twitter-worldwide-gender/, 2018.
[8] Docs -- twitter developers. https://developer.twitter.com/en/docs.html, 2018.
[9] Facebook pixel. https://www.facebook.com/business/learn/facebook-ads-pixel, 2018.
[10] Facebook will
https://www.
theverge.com/2018/8/21/17764480/facebook-ad-targeting-options-removal-housing-racial-
discrimination, 2018.
remove 5,000 ad targeting categories
to prevent discrimination.
[11] Monthly active Facebook users. https://www.statista.com/statistics/264810/number-of-monthly-
active-facebook-users-worldwide/, 2018.
[12] Monthly active instagram users. https://www.statista.com/statistics/253577/number-of-monthly-
active-instagram-users/, 2018.
[13] Monthly active Twitter users. https://www.statista.com/statistics/282087/number-of-monthly-
active-twitter-users/, 2018.
[14] Other data breaches like cambridge analytica. https://www.cnbc.com/2018/04/08/cubeyou-cambridge-
like-app-collected-data-on-millions-from-facebook.html, 2018.
[15] W. U. Ahmad, K.-W. Chang, and H. Wang. Intent-aware query obfuscation for privacy protection in personalized
web search. In The 41st International ACM SIGIR Conference on Research & Development in Information
Retrieval, pages 285 -- 294. ACM, 2018.
[16] M. Anderson. Researchers take author recognition to neural networks. https://thestack.com/world/2016/
06/07/researchers-take-author-recognition-to-neural-networks/, 2016.
[17] M. E. Andrés, N. E. Bordenabe, K. Chatzikokolakis, and C. Palamidessi. Geo-indistinguishability: Differential
privacy for location-based systems. In Proc. of ACM SIGSAC, CCS '13, pages 901 -- 914. ACM, 2013.
[18] J. W. Brown, O. Ohrimenko, and R. Tamassia. Haze: Privacy-preserving real-time traffic statistics. In Proc. of
ACM SIGSPATIAL, pages 540 -- 543. ACM, 2013.
[19] S. Chandra, L. Khan, and F. B. Muhaya. Estimating twitter user location using social interactions -- a content based
approach. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Social
Computing (SocialCom), 2011 IEEE Third International Conference on, pages 838 -- 843. IEEE, 2011.
[20] 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,
pages 759 -- 768. ACM, 2010.
[21] S. H. Ding, B. C. Fung, F. Iqbal, and W. K. Cheung. Learning stylometric representations for authorship analysis.
IEEE Transactions on Cybernetics, 2017.
[22] C. Dwork. Automata, Languages and Programming: ICALP 2006, Proceedings, Part II, chapter Differential
Privacy, pages 1 -- 12. Springer Berlin Heidelberg, 2006.
[23] C. Dwork. Differential privacy: A survey of results. In International Conference on Theory and Applications of
Models of Computation, pages 1 -- 19. Springer, 2008.
[24] C. Dwork. Differential privacy. In Encyclopedia of Cryptography and Security, pages 338 -- 340. Springer, 2011.
[25] C. Dwork, M. Naor, T. Pitassi, G. N. Rothblum, and S. Yekhanin. Pan-private streaming algorithms. In Proceedings
of ICS, 2010.
[26] T. Georgiou, A. El Abbadi, and X. Yan. Extracting topics with focused communities for social content recom-
mendation. In Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social
Computing, 2017.
[27] T. Georgiou, A. El Abbadi, and X. Yan. Privacy-preserving community-aware trending topic detection in online
In IFIP Annual Conference on Data and Applications Security and Privacy, pages 205 -- 224.
social media.
Springer, 2017.
[28] M. Ghufran, G. Quercini, and N. Bennacer. Toponym disambiguation in online social network profiles. In
Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems,
page 6. ACM, 2015.
16
A PREPRINT - MAY 27, 2019
[29] X. Gong, X. Chen, K. Xing, D.-H. Shin, M. Zhang, and J. Zhang. Personalized location privacy in mobile
networks: A social group utility approach. In Computer Communications (INFOCOM), 2015 IEEE Conference
on, pages 1008 -- 1016. IEEE, 2015.
[30] T. Grant. How text-messaging slips can help catch murderers. https://www.independent.co.uk/voices/
commentators/dr-tim-grant-how-text-messaging-slips-can-help-catch-murderers-923503.
html, 2008.
[31] A. Graves. Generating sequences with recurrent neural networks. arXiv preprint arXiv:1308.0850, 2013.
[32] R. Heatherly, M. Kantarcioglu, and B. Thuraisingham. Preventing private information inference attacks on social
networks. IEEE TKDE, 25(8):1849 -- 1862, 2013.
[33] M. Kosinski, D. Stillwell, and T. 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.
[34] N. Li, T. Li, and S. Venkatasubramanian. t-closeness: Privacy beyond k-anonymity and l-diversity. In IEEE ICDE,
pages 106 -- 115, April 2007.
[35] A. Machanavajjhala, D. Kifer, J. Gehrke, and M. Venkitasubramaniam. L-diversity: Privacy beyond k-anonymity.
ACM Trans. Knowl. Discov. Data, 1(1), Mar. 2007.
[36] M. F. Mokbel, C.-Y. Chow, and W. G. Aref. The new casper: Query processing for location services without
compromising privacy. In Proc. of VLDB, pages 763 -- 774, 2006.
[37] H. Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119, 2004.
[38] M. J. Paul and M. Dredze. You are what you tweet: Analyzing twitter for public health. Icwsm, 20:265 -- 272, 2011.
[39] R. L. Rivest et al. Chaffing and winnowing: Confidentiality without encryption. CryptoBytes (RSA laboratories),
4(1):12 -- 17, 1998.
[40] P. Samarati. Protecting respondents identities in microdata release. IEEE TKDE, 13(6):1010 -- 1027, 2001.
[41] H. A. Schwartz, J. C. Eichstaedt, M. L. Kern, L. Dziurzynski, S. M. Ramones, M. Agrawal, A. Shah, M. Kosinski,
D. Stillwell, M. E. Seligman, et al. Personality, gender, and age in the language of social media: The open-
vocabulary approach. PloS one, 8(9):e73791, 2013.
[42] L. Sweeney. Achieving k-anonymity privacy protection using generalization and suppression. International
Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 10(05):571 -- 588, 2002.
[43] L. Sweeney. K-anonymity: A model for protecting privacy. Int. J. Uncertain. Fuzziness Knowl.-Based Syst.,
10(5):557 -- 570, Oct. 2002.
[44] M. Yakout, M. Ouzzani, H. Elmeleegy, N. Talukder, and A. K. Elmagarmid. Privometer: Privacy protection in
social networks. IEEE ICDEW, 00:266 -- 269, 2010.
[45] V. Zakhary, C. Sahin, T. Georgiou, and A. El Abbadi. Locborg: Hiding social media user location while
maintaining online persona. In Proceedings of the 25th ACM SIGSPATIAL International Conference on Advances
in Geographic Information Systems, page 12. ACM, 2017.
[46] Y. Zhang, M. Humbert, T. Rahman, C.-T. Li, J. Pang, and M. Backes. Tagvisor: A privacy advisor for sharing
hashtags. arXiv preprint arXiv:1802.04122, 2018.
[47] E. Zheleva and L. Getoor. To join or not to join: the illusion of privacy in social networks with mixed public and
private user profiles. In Proc. of WWW, pages 531 -- 540. ACM, 2009.
17
|
1801.09946 | 2 | 1801 | 2018-04-20T10:48:17 | "23andMe confirms: I'm super white" -- Analyzing Twitter Discourse On Genetic Testing | [
"cs.SI",
"cs.CY"
] | Recent progress in genomics is bringing genetic testing to the masses. Participatory public initiatives are underway to sequence the genome of millions of volunteers, and a new market is booming with a number of companies like 23andMe and AncestryDNA offering affordable tests directly to consumers. Consequently, news, experiences, and views on genetic testing are increasingly shared and discussed online and on social networks like Twitter. In this paper, we present a large-scale analysis of Twitter discourse on genetic testing. We collect 302K tweets from 113K users, posted over 2.5 years, by using thirteen keywords related to genetic testing companies and public initiatives as search keywords. We study both the tweets and the users posting them along several axes, aiming to understand who tweets about genetic testing, what they talk about, and how they use Twitter for that. Among other things, we find that tweets about genetic testing originate from accounts that overall appear to be interested in digital health and technology. Also, marketing efforts as well as announcements, such as the FDA's suspension of 23andMe's health reports, influence the type and the nature of user engagement.Finally, we report on users who share screenshots of their results, and raise a few ethical and societal questions as we find evidence of groups associating genetic testing to racist ideologies. | cs.SI | cs | "23andMe confirms: I'm super white"
Analyzing Twitter Discourse On Genetic Testing
Alexandros Mittos1, Jeremy Blackburn2, and Emiliano De Cristofaro1
1University College London 2University of Alabama at Birmingham
{a.mittos, e.decristofaro}@ucl.ac.uk, [email protected]
8
1
0
2
r
p
A
0
2
]
I
S
.
s
c
[
2
v
6
4
9
9
0
.
1
0
8
1
:
v
i
X
r
a
Abstract
Recent progress in genomics is bringing genetic testing to
the masses. Participatory public initiatives are underway to se-
quence the genome of millions of volunteers, and a new mar-
ket is booming with a number of companies like 23andMe and
AncestryDNA offering affordable tests directly to consumers.
Consequently, news, experiences, and views on genetic testing
are increasingly shared and discussed online and on social net-
works like Twitter. In this paper, we present a large-scale anal-
ysis of Twitter discourse on genetic testing. We collect 302K
tweets from 113K users, posted over 2.5 years, by using thir-
teen keywords related to genetic testing companies and public
initiatives as search keywords. We study both the tweets and
the users posting them along several axes, aiming to under-
stand who tweets about genetic testing, what they talk about,
and how they use Twitter for that. Among other things, we find
that tweets about genetic testing originate from accounts that
overall appear to be interested in digital health and technology.
Also, marketing efforts as well as announcements, such as the
FDA's suspension of 23andMe's health reports, influence the
type and the nature of user engagement. Finally, we report on
users who share screenshots of their results, and raise a few
ethical and societal questions as we find evidence of groups
associating genetic testing to racist ideologies.
1
Introduction
In 1990, the Human Genome Project was kicked off with
the goal of producing the first complete sequence of a human
genome; at a cost of almost $3 billion, it was completed 13
years later [23]. Since then, costs have dropped at a stagger-
ing rate: by 2006, high-quality sequencing of a human genome
cost $14 million, and, by 2016, private individuals could have
their genomes sequenced for about $1,500 [27]. This rapid
progress is paving the way to personalized medicine, a con-
cept advocating for diagnosis and treatment to be tailored to
patients' genetic features, aiming to make healthcare more
preventive and effective [2]. It also enables public initiatives
to sequence large numbers of genomes and build large bio-
repositories for research purposes; for instance, the Precision
Medicine research program in US (now called All Of Us)
or the Genomics England project in UK are sequencing the
genomes of, respectively, 1M and 100K volunteers.
Moreover, a number of companies have entered the flour-
ishing market of direct-to-consumer (DTC) genetic testing.
Rather than visiting a clinic, customers purchase a collection
kit for a few hundred dollars or less, deposit a saliva sample,
and mail it back; after a few days, they receive a report with
information about genetic health risks (e.g., susceptibility to
Alzheimer's), wellness (e.g., lactose intolerance), and/or an-
cestry and genealogy information. Today, there are hundreds
of DTC companies – naturally, some more reputable than oth-
ers [30] – including 23andMe (which provides reports on car-
rier status, health, and ancestry) and AncestryDNA (which fo-
cuses on genealogy and ancestry). As of Jan 2018, 23andMe
has tested 2M and AncestryDNA 5M customers.1
Traditionally, health-related issues were communicated to
patients primarily by doctors and clinicians-the advent of
direct-to-consumer genetic testing changes this substantially.
Individuals can now learn potentially life-changing results
with a few clicks of the mouse, without contacting a medical
professional. Also, as results are delivered electronically, they
are more easily shared with others. Affordable DTC products
and participatory sequencing initiatives make genetic testing
increasingly more accessible and available to the general pop-
ulation. Like with other aspects of digital health, this leads to
social media attracting discussions, sharing of experiences, and
molding of perceptions around genetic testing, thus becoming
a key platform for related news and marketing efforts. How-
ever, while the research community has analyzed in great detail
the interlinked relationship between health and social networks
such as Twitter, to the best of our knowledge, genetic testing
discourse on social media has not been adequately studied.
To this end, in this paper, we set to address a few open ques-
tions: 1) What are the tweets related to genetic testing really
about? 2) Which accounts are particularly active in tweeting
and what do they talk about? 3) Is the discussion about genetic
testing dominated by certain keywords, themes, or companies?
4) What is the overall sentiment and what topics relate to more
negative sentiment? (We focus on Twitter due to its popularity
and the relatively ease of collecting data).
Aiming to answer these questions, we present an ex-
ploratory, large-scale analysis of Twitter discourse related to
genetic testing. Starting from 10 keywords related to DTC ge-
netics companies and 3 to genomics initiatives, we search and
crawl all available tweets containing these keywords that were
posted between Jan 1, 2015 and Jul 31, 2017. We collect 302K
tweets from 113K users, and analyze them along several axes,
1See http://for.tn/2eYVJuD and http://ancstry.me/2zH3HBN
1
seeking to understand who tweets about genetic testing, what
they talk about, and how they use Twitter for that. Specifically,
after presenting a general characterization of our dataset, we
analyze the tweets content-wise, studying the most common
hashtags/URLs and measuring sentiment. Next, we perform a
user-based analysis, looking at the profiles and their location,
and assessing whether they are likely to be bots. We also se-
lect a random sample of 15K users and analyze their latest 1K
tweets to study their interests. Note that, as a substantial chunk
of tweets turns out to be about DTC companies 23andMe and
AncestryDNA, we present a few case studies focused on them.
Finally, we examine the most negative tweets in our dataset,
finding a number of tweets related to racism and hate-speech,
as well as fears of privacy and data misuse, and look for in-
stances of users sharing screenshots of their test results.
Overall, our study leads to a few interesting observations:
1. Users tweeting about genetic testing seem overall inter-
ested in digital health and technology, although the con-
versation is often dominated by those with a vested in-
terest in its success, e.g., specialist journalists, medical
professionals, entrepreneurs, etc.
2. The two most popular DTC companies, 23andMe and An-
cestryDNA, also generate the most tweets. However, al-
though 23andMe has half the customers, it produces al-
most 5 times more tweets, which is also due to contro-
versy around their failure to get FDA approval in 2015.
3. Sentiment around initiatives is positive, with cover-
age boosted by mainstream news and announcements
(e.g. President Obama's) and neutral for DTC companies,
although with a few strongly opinionated users.
4. There is a clear distinction in the marketing efforts under-
taken by different companies, which naturally influence
the type and the nature of users' engagement; e.g., we
find the the promotional hashtag #sweepstakes in 1 out 8
tweets containing the keyword AncestryDNA.
5. We find a limited presence of social bots, with some key-
words attracting a different degree of automated publish-
ing, as some topics seem to be more popular among indi-
viduals than others.
6. We find evidence of groups using genetic testing to push
racist and anti-semitic agendas, and of users expressing
concerns about privacy and data protection
7. A non-negligible amount of users share and discuss
screenshots of their ancestry test results, despite the pos-
sible privacy implications.
Paper organization. The rest of the paper is organized as fol-
lows. In Section 2 we present related work focusing on the
following three themes: i) user perspectives on genetic testing,
ii) health in social networks, and iii) analysis of social media
discourse. In Section 3 we present our methodology for col-
lecting the dataset. In Section 4 we present a general charac-
terization of the tweets in our dataset. In Section 5 we study the
most common hashtags and URLs, and we measure the over-
all sentiment of tweets. In Section 6 we study the profiles of
people who tweet about genetic testing. Finally, in Section 7
we examine several cases where genetic testing is being used
to promote racist agendas, instances of people expressing con-
cerns about privacy, and of people sharing screenshots of their
ancestry test results online.
2 Related Work
User perspectives on genetic testing. A few qualitative stud-
ies have analyzed users' perspectives on genetic testing. Gold-
smith et al. [17] review 17 studies conducted in 6 different
countries, finding that, although participants appear to be inter-
ested in the health-related aspects of testing, they also express
concerns about privacy and reliability. Covolo et al. [12] re-
view 118 articles, highlighting that users are drawn to genetic
testing by the potential to monitor and improve their health.
Caulfield et al. [7] analyze the controversy around Myriad Ge-
netics following their attempt to patent the BRCA gene which
is associated with predisposition to breast cancer.
Closer to our work are quantitative studies using social me-
dia, however, to the best of our knowledge, the only relevant
work is by Chow-White et al. [10], who look at one week's
worth of tweets containing '23andMe', performing a simple
sentiment analysis, finding that positive tweets outnumber neg-
ative ones and that people tend to be enthusiastic about it.
Their analysis only studies one company and only sentiment,
whereas, we focus on 10 companies and 3 initiatives and con-
duct deeper statistical, content, and user-based analyses. Also,
they rely on a much smaller dataset, 2K vs 324K tweets, col-
lected over 1 week vs 2.5 years.
Health in social networks. Social networks like Twitter have
been used extensively to study health and health-related is-
sues, e.g., to measure and predict depression. De Choudhury
et al. [13] identify 476 users self-reporting depression, collect
their tweets, and study their engagement, emotion, and use of
depressive language. By comparing to a control group, they
extract significant differences, and build a classifier to predict
the likelihood of an individual's depression. Coppersmith et
al. [11] study tweets related to various mental disorders, while
Paul et al. [29] gather public health information from Twitter,
discovering statistically significant correlations between Twit-
ter and official health statistics. Abbar et al. [1] analyze the
nutritional behavior of US citizens: they collect 892K tweets
by 400K US users using food-related keywords and find that
foods match obesity and diabetes statistics, and that Twitter
friends tend to share the same preferences in food consump-
tion. Prasetyo et al. [31] study how social media can effect
awareness in health campaigns. Focusing on the Movember
charity campaign, they collect more than 1M tweets, using the
keyword 'Movember', and uncover correlations between the
visitors of the Movember website and popular Twitter users,
but none between tweets and donations.
Analyzing social media discourse. Cavazos-Rehg et al. [8]
study drinking behaviors on Twitter: using keywords related
to drinking (e.g., drunk, alcohol, wasted), they collect 10M
2
tweets and identify the most common themes related to pro-
drinking and anti-drinking behavior. Lerman et al. [24] con-
duct an emotion analysis on tweets from Los Angeles: using
public demographic data, they find that users with lower in-
come and education levels, and who engage with less diverse
social contacts, express more negative emotions, while peo-
ple with higher income and education levels post more posi-
tive messages. Chatzakou et al. [9] study the GamerGate con-
troversy2 on Twitter, collecting a dataset of tweets containing
keywords indicating abusive behavior. They compare the char-
acteristics of the related Twitter profiles to a baseline, find-
ing that users tweeting about Gamergate are more technologi-
cally savvy and active, and that their tweets are more negative.
Burnap et al. [6] study Twitter response to a terrorist attack
occurred in Woolwich in 2013. Using 'Woolwich' as a key-
word search, they collect 427K tweets, finding that opinions
and emotional factors are predictive of size and survival of in-
formation flows.
3 Dataset
We now present the methodology used to gather the tweets
studied in this paper. We build a dataset with tweets containing
keywords related to (1) direct-to-consumer (DTC) genetic test-
ing companies, and (2) public genome sequencing initiatives,
using these keywords as search queries and crawling all tweets
posted from Jan 1, 2015 to Jul 31, 2017 returned as results.
DTC genetic testing companies. We start from a list of 36
DTC genetic testing companies compiled by the International
Society of Genetic Genealogy3, which provides a good sam-
ple of the DTC ecosystem. We use each company's name as
a search keyword; if the search returns less than 1,000 tweets,
we discard it. In the end, we collect tweets for 10 companies:
23andMe, AncestryDNA, Counsyl, DNAFit, FamilyTreeDNA,
FitnessGenes, MapMyGenome, PathwayGenomics, Ubiome,
and VeritasGenetics. We opt for keywords not separated by
spaces (e.g., VeritasGenetics) rather than quoted search (e.g.,
"Veritas Genetics") since we notice that companies are primar-
ily discussed via hashtags or mentions, and because Twitter's
search engine does not provide exact results with quotes.
Genomics initiatives. Besides tweets related to for-profit com-
panies, we also want to study discourse related to public se-
quencing initiatives and related concepts. Thus, we select three
more keywords: PrecisionMedicine, PersonalizedMedicine,
and GenomicsEngland. Personalized Medicine aims to make
diagnosis, treatment, and care of patients tailored and opti-
mized to their specific genetic makeup. Precision Medicine
conveys a similar concept, but also refers to the initiative se-
quencing the genome of 1M individuals announced by Presi-
dent Obama in 2015to understand how a person's genetics, en-
vironment, and lifestyle can help determine the best approach
to prevent or treat disease.4 Genomics England is a similar UK
initiative with 100K volunteers, primarily focusing on cancer
and rare disease research. Once again, we search for keywords
not separated by spaces (e.g., PrecisionMedicine) since these
2https://en.wikipedia.org/wiki/Gamergate controversy
3https://isogg.org/wiki/List of DNA testing companies
4https://ghr.nlm.nih.gov/primer/precisionmedicine/initiative
Figure 1: Number of tweets per user account.
concepts are mostly discussed via hashtags and because of the
incorrectness of the search engine.
Crawl. We use a custom Python script to collect all posted
tweets from Jan 1, 2015 to Jul 31, 2017 returned as search re-
sults using the 10 DTC keywords and the 3 keywords related to
genomics initiatives. The crawler, run with self-imposed throt-
tling to avoid issues for the site operators over in Fall 2017,
collects, for each tweet, its content, the username, date and
time, the number of retweets and likes, as well as the URL
of the tweet. It also visits the profile of the users posting each
tweet, collecting their location (if any), the number of follow-
ers, following, tweets, and likes. Overall, we collect a total of
191K tweets from 94K users for the 10 DTC companies and
111K from 19K users for the 3 initiatives.
Baseline. We also collect a set of 163,260 random English
tweets, from the same Jan 2015 to Jul 2017 period (approx.
170 per day), which serves as a baseline set for comparisons.
Remarks. The keyword search also returns accounts that
match that keyword, e.g.,
including 23andMe,
#23andMe, or @23andMe, but also those posted by the
@23andMe account. For consistency, we discard the latter, an-
alyzing them separately when relevant. Note that our dataset
includes tweets from users who discuss their opinions on ge-
netic testing, but also blog posts, ads, news articles, etc. As our
goal is to discover how genetic testing is reflected through the
lens of Twitter, we choose not to discard any of the above sub-
sets in an attempt to "clean" the dataset, or to focus only on
certain kinds of profiles. Nevertheless, in Sec. 6, we do shed
light on the users tweeting about genetic testing, as well as
those who publish their genetic tests results.
tweets
4 General Characterization
We now present a general characterization of the tweets in
our dataset. Simple statistics of our keyword-based dataset are
reported in Table 1. From left to right, the table lists the total
number of tweets, unique users, retweets, and likes for each
of the 13 keywords and the random baseline. We also quantify
the percentage of tweets made by the official accounts of each
company or initiative, as well as the percentage of tweets in-
cluding media (images and videos), quoted tweets, hashtags,
and URLs, and how many of them are in the Alexa Top 1M.
3
Top 1M
75.40%
23andMe
49.68%
AncestryDNA
74.97%
Counsyl
79.18%
DNAFit
69.21%
FamilyTreeDNA
71.28%
FitnessGenes
64.30%
MapMyGenome
68.12%
PathwayGenomics
64.19%
Ubiome
71.95%
VeritasGenetics
48.82%
Genomics England
71.98%
Personalized Medicine
77.16%
Precision Medicine
Total
71.80%
Baseline
89.57%
Table 1: Our keyword dataset, with all tweets from January 1, 2015 to July 31, 2017 containing keywords related to genetic testing.
Quotes Hashtags
Likes Official Media
3.49%
6.14%
1.31%
2.69%
8.79%
7.08%
4.64%
3.49%
6.98%
5.37%
15.34% 18.74%
6.62%
4.36% 19.97%
9.40%
16.29% 18.47%
4.99%
15.30% 13.13%
6.11%
2.13% 18.51%
2.71%
2.85%
4.37%
6.65% 17.07% 17.07%
19.68% 17.80% 11.58%
6.93%
7.55%
8.56% 10.41%
7.75%
5.92%
41.20% 12.07%
URLs
27.23% 68.68%
54.29% 75.50%
44.01% 83.94%
57.22% 78.94%
34.18% 36.47%
44.53% 56.76%
53.63% 80.35%
61.01% 76.55%
27.85% 73.28%
46.13% 58.28%
61.19% 69.18%
99.93% 87.42%
99.88% 83.39%
56.54% 74.77%
23.48% 45.49%
149,897
47,249
4,255
2,508
3,111
2,809
3,726
2,521
13,991
2,526
18,756
15,514
128,303
395,166
486,960,753
Tweets
132,597
29,071
3,862
2,118
2,794
2,142
1,568
1,544
14,420
1,292
7,009
20,302
83,329
302,048
163,260
Users
64,014
16,905
1,834
844
1,205
773
704
579
6,762
497
1,863
4,631
13,012
113,624
131,712
RTs
72,848
16,266
2,716
1,336
1,196
908
4,488
1,968
9,223
1,443
19,772
19,085
118,043
269,292
282,063,006
–
–
2.26%
–
tinct users); AncestryDNA is a distant second (30K tweets
from 16.9K users). Given their large customer bases, this
should not come as a surprise. However, it is surprising that
23andMe has 4.6 times as many tweets as AncenstryDNA
even though AncestryDNA has over twice the customers as
23andMe. The least popular companies are MapMyGenome,
PathwayGenomics, and VeritasGenetics, with less than 2K
tweets each over our 2.5 year collection period. Among the ini-
tiatives, Precision Medicine generates a relative high number
of tweets (83K from 13K users), much more so than Personal-
ized Medicine (20K tweets).
Tweets per user. For each keyword, we also measure the num-
ber of tweets per user(see Figure 1). We find that the me-
dian for every keyword is 1; i.e., 50% of users tweet about
a given DTC company or initiative only once. However, we do
find differences in the outliers for different keywords. For in-
stance, there are several highly engaged users tweeting about
Personalized Medicine and Precision Medicine. Manual exam-
ination of these users indicates that most of them are medical
researchers and companies actively promoting the initiatives
as hashtags. The presence of these heavily "invested" users be-
comes more apparent when we look at the number of tweets
as a function of the number of unique keywords a user posts
about, as plotted in Figure 2: 95% of them post about only one
keyword, and those that post in more than one tend to post sub-
stantially more tweets about genetic testing in general; in some
cases, orders of magnitude more tweets.
Retweets and Likes. The total number of retweets and likes
per tweet in the baseline is substantially higher than for
tweets related to genetic testing due to outliers, i.e., viral
tweets or tweets posted by famous accounts (e.g., a tweet by
@POTUS44 on January 11, 2017 has 875,844 retweets and
1,862,249 likes). However, the median for retweets and likes
in the baseline dataset mirrors that of tweets in our keywords
dataset, with values between 0 and 1.
Note that, although the number of retweets and likes per
tweet could be influenced by how old the tweets are, this is
not really the case in our dataset. We collect tweets posted
up to July 2017 starting in late-August 2017, allowing ample
time for retweets and likes to occur, considering that previ-
ous work [22] indicates that 75% of retweets happen within 24
Figure 2: Number of tweets per user as a function of the number of
unique keywords they tweeted about.
DTC vs Initiatives. Overall, we find differences between
tweets about DTC genetic testing companies and those about
genomics initiatives. The majority of the latter come from a
smaller set of users compared to the former, i.e., a few very
dedicated users drive the discussion about genomics initiatives.
This is evident from Figure 1, which plots the number of tweets
per user for each keyword: Personalized Medicine and Preci-
sion Medicine have more outliers than most of the DTC genetic
companies (although the median for all keywords is 1). We
also find these tweets are more likely to contain URLs (87%
and 83% of tweets, respectively) than most companies, and
even more so when compared to the baseline (45%). This sug-
gests that tweets about these topics often include links to news
and/or other external resources.
Only around 50% of URLs linked from tweets related to
Genomics England or AncestryDNA are in the Alexa top 1M,
compared to 60–75% for other keywords. For Genomics Eng-
land, this is due to many URLs pointing to genomicsengland.
co.uk itself. For AncenstryDNA, whose official site at ancestry.
com is in the top 1M, it appears to be due to a very large num-
ber of marketing URLs tweeted along with the keyword; we
discuss this further later on.
Number of tweets. 23andMe is by far the most popular key-
word, with one order of magnitude more tweets than any
other company (130K in total, around 140/day, from 64K dis-
4
promotion strategies (see Sec. 5). We also find a number of
outliers for 23andMe, mostly around Feb 20 and Oct 19, 2015,
and Apr 6, 2017, which are key dates related to 23andMe's
failure to get FDA approval for their health reports in 2015,
then obtained in 2017. In fact, 20K/132K 23andMe tweets are
posted around those dates. As for Personalized and Precision
Medicine (Figure 3(b)), the volume of tweets stays relatively
flat. There are outliers for Precision Medicine too, e.g., 2,628
tweets on February 25, 2016, when the White House hosted
the Precision Medicine Initiative summit [18].
Discussion. Overall, our characterization shows that highly
engaged users drive the discussion around public genomics
initiatives, which is particularly influenced, at least in terms
of volumes, by important announcements such as the one
made by President Obama. As for direct-to-consumer (DTC)
genetic testing, the conversation is, as expected, dominated
by the two most popular companies: 23andMe and Ances-
tryDNA. However, it is interesting that the former generates
4 times more tweets even though the latter has more than twice
the customers. Some of this "popularity" seems to be due to
23andMe's controversy around FDA approval. We also find a
non-negligible use of hashtags, possibly used for promotion
and marketing efforts, and that a lot of tweets include URLs to
popular domains, indicating that they are used to disseminate
news and links to external resources. This warrants further ex-
ploration, thus, we perform hashtag and URL analysis in the
next section.
5 Content Analysis
Next, we analyze the content of the tweets related to genetic
testing, studying hashtags and URLs included in them and per-
forming a simple sentiment analysis. We also performed Latent
Dirichlet Allocation (LDA) [5] and Term Frequency-Inverse
Document Frequency (TF-IDF) [33] analysis, however, we
omit the results since they do not yield any actionable findings.
5.1 Hashtag Analysis
In Table 2, we report the top three hashtags for every key-
word, while differentiating between tweets made by regular
users and those by official accounts. We also quantify the
percentage of tweets with at least one hashtag (WH) and
that of tweets including the keyword as a hashtag (KH), e.g.,
#23andMe.
We find a few unexpected hashtags among the DTC
tweets, e.g., #sweepstakes (AncestryDNA), #startup (Fit-
nessgenes), #vote (Ubiome), #shechat and #appguesswho
(MapMyGenome). AncestryDNA's top hashtag, #sweepstakes
(12%), is related to a marketing campaign promoting a TV
series, "America: Promised Land." There are 3.5K tweets,
from distinct users, with the very same content (most likely
due to a "share" button): "I believe I've discovered my
@ancestry! Discover yours for the chance to win an An-
cestryDNA Kit. #sweepstakes journeythroughhistorysweeps.
com." We also find hashtags like #feistyfrugal and #holiday-
giftguide in the AncestryDNA top 10 hashtags, which confirms
how AncestryDNA uses Twitter for relatively aggressive mar-
keting campaigns. Moreover, in the Fitnessgenes tweets, we
(a) 23andMe and AncestryDNA
(b) Personalized/Precision Medicine
Figure 3: Number of tweets per day. Note the log scale in y-axis.
hours and 85% happen within a month.
Official accounts tweets. We also look at the tweets including
a given keyword (e.g., Ubiome) made by the corresponding of-
ficial account (e.g., @Ubiome). There are no official accounts
for Personalized and Precision Medicine, however, the Preci-
sion Medicine initiative is now called All Of Us and has a Twit-
ter account (created in February 2017) that has posted only a
few tweets (224 as of April 4, 2018), so we do not consider
it. Tweets made by the official accounts of most companies in-
cluding the name of the company as keyword is unsurprisingly
very (e.g., 1% for 23andMe). However, it is higher for oth-
ers (e.g., 15% for DNAfit, Fitnessgenes, and MapMyGenome),
due to the fact that these companies actually add their names
in their tweets as a hashtag (e.g., #AncestryDNA).
Hashtags, media, and quotes. Table 1 shows that around
a quarter of 23andMe's and Ubiome's tweets have hashtags
(slightly more than 23% for the baseline); for most other
keywords, it is above 40%. For Personalized and Precision
Medicine, we find that almost all tweets have the keyword in
the form of hashtag (99%). For Genomics England, this only
happens 61% of the time, since a lot of tweets include @Ge-
nomicsEngland. We perform a more detailed hashtag analysis
in the relevant section.
We then find that the percentage of tweets with media vary
from 4% in Ubiome to almost 20% in FamilyTreeDNA. Anec-
dotally, we see that images often contain text, i.e., are used to
overcome the character limit and comment on issues related to
the company (e.g., [28]). We also look at "quotes", i.e., tweets
including the URL of another tweet: for most keywords, per-
centages are lower than the baseline, except for VeritasGenet-
ics (mostly due to the official account), although less so for the
initiatives. Possibly, users tweeting about genomics initiatives
tend to be discuss more with each other, by commenting on
relevant tweets.
Temporal analysis. Finally, we analyze how the volume
of tweets changes over time. In Figure 3, we plot
the
number of tweets per day in our dataset (between Jan
1, 2015–July 31, 2017) for the two most popular compa-
nies (23andMe/AncestryDNA) and the two most popular ge-
nomics initiatives (Personalized/Precision Medicine). On av-
erage, there are 145 and 30 tweets per day for 23andMe and
AncestryDNA keywords, respectively. While the former is
relatively constant, the latter increases steadily in 2017(Fig-
ure 3(a)). This may be the result of AncestryDNA's aggressive
5
WH Top 3 Hashtags
– Without Official Accounts –
–
–
Keyword
23andMe
AncestryDNA
Counsyl
DNAFit
FamilyTreeDNA
FitnessGenes
MapMyGenome
PathwayGenomics
Ubiome
VeritasGenetics
Genomics England
Personalized Medicine
Precision Medicine
Table 2: Top 3 hashtags for each keyword, along with the percentage of tweets with at least a hashtag (WH) as well as that of of "keyword
hashtags" (KH), e.g., #23andMe.
27.09% dna (3.58%), genetics (2.07%), tech (1.96%)
75.48% sweepstakes (12.38%), dna (4.90%), genealogy (4.86%)
45.24% getaheadofcancer (2.64%), cap (1.93%), medical (1.94%)
55.30% diet (4.19%), fitness (3.72%), crossfit (3.54%)
29.31% dna (14.24%), genealogy (13.42%), ancestryhour (3.18%)
72.19% startup (5.93%), london (5.73%), job (5.59%)
54.98% shechat (7.94%), appguesswho (5.32%), genomepatri (4.22%)
55.85% coloncancer (6.91%), genetictesting (3.29%), cancer (2.85%)
28.57% microbiome (13.23%), tech (2.14%), vote (2.07%)
57.16% brca (3.92%), genome (3.62%), genomics (3.32%)
62.05% genomes100k (14.84%), genomics (7.72%), raredisease (5.24%)
precisionmedicine (22.74%), genomics (9.77%), pmcon (8.37%)
genomics (6.70%), personalizedmedicine (5.49%), cancer (4.89%)
KH
9.74%
46.88%
0%
40.37%
56.66%
45.29%
0%
19.35%
6.12%
0%
0%
–
–
– Only Official Accounts –
Top 3 Hashtags
23andMestory (6.67%), genetics (6.35%), video (5.19%)
dna (11.74%), ancestry (5.92%), familyhistory (5.07%)
acog17 (6.18%), womenshealthweek (5.15%), teamcounsyl (5.15%)
dna (5.33%), fitness (3.71%), generictogenetic (3.48%)
geneticgenealogy (5.55%), ftdnasuccess (4.44%), ftdna (3.33%)
fitness (5.85%), dna (4.32%), gtsfit (2.79%)
genomepatri (7.28%), knowyourself (4.04%), genetics (2.02%)
dnaday16 (9.67%), ashg15 (9.67%), health (3.22%)
microbiome (24.48%), bacteria (4.76%), meowcrobiome (2.72%)
brca (11.82%), liveintheknow (11.82%), wholegenome (10.75%)
genomes100k (32.45%), raredisease (19.49%), genomics (18.71%)
–
–
KH
12.46%
25.94%
3.08%
22.91%
10.86%
18.22%
15.80%
3.34%
6.61%
4.22%
1.77%
–
–
Keyword
23andMe
AncestryDNA
Counsyl
DNAFit
FamilyTreeDNA
FitnessGenes
MapMyGenome
PathwayGenomics
Ubiome
VeritasGenetics
Genomics England
Without Official Accounts
23andMe.com (7.33%), techcrunch.com (3.09%), fb.me (2.48%)
journeythroughhistorysweeps.com (15.18%), ancestry.com (13.94%),
ancstry.me (6.67%)
techcrunch.com (8.42%), businesswire.com (5.30%), bioportfolio.com (4.46%)
fb.me (15.81%), instagram.com (14.65%), dnafit.com (2.99%)
familytreedna.com (11.31%), myfamilydnatest.com (4.28%), fb.me (4.17%)
instagram.com (14.77%), fitnessgenes.com (8.48%), workinstartups.com (6.29%)
yourstory.com (11.84%), owler.us (11.44%), mapmygenome.in (9.18%)
paper.li (11.96%), atjo.es (10.82%), pathway.com (3.31%)
techcrunch.com (9.30%), bioportfolio.com (4.83%), ubiomeblog.com (4.21%)
veritasgenetics.com (10.97%), technologyreview.com (5.01%), buff.ly (2.30%)
genomicsengland.co.uk (33.85%), youtube.com (1.98%), buff.ly (1.64%)
Personalized Medicine
Precision Medicine
Baseline
instagram.com (8.78%), myriad.com (2.54%), buff.ly (2.32%)
buff.ly (2.92%), instagram.com (2.27%), nih.gov (1.87%)
instagram.com (4.18%), fb.me (3.44%), youtu.be (2.72%)
Only Official Accounts
23me.co (50.88%), 23andMe.com (21.13%), instagram.com (5.40%)
ancstry.me (74.11%), youtube.com (3.27%), ancestry.com.au (2.88%)
businesswire.com (14.78%), counsyl.com (13.91%), medium.com (5.21%)
fb.me (11.74%), dnafit.com (10.52%), dnafit.gr (2.83%)
familytreedna.com (76.56%), abcn.ws (3.12%), instagram.com (1.56%)
fitnessgenes.com (31.11%), instagram.com (4.44%), pinterest.com (4.44%)
mapmygenome.in (42.12%), youtu.be (14.35%), indiatimes.com (3.70%)
pathway.com (23.07%), nxtbook.com (3.84%), drhoffman.com (3.84%)
ubiomeblog.com (34.32%), igg.me (26.07%), ubiome.com (6.60%)
veritasgenetics.com (75.67%), biospace.com (1.35%), statnews.com (1.35%)
genomicsengland.co.uk (98.03%), peoplehr.net (0.58%),
campaign-archive1.com (0.21%)
–
–
–
Table 3: The top 3 domains per keyword, without official accounts and only considering the official accounts.
find hashtags like #startup, #london, and #job due to a num-
ber of tweets advertising jobs for Fitnessgenes, while #shechat
appears in tweets linking to an article related to women in busi-
ness about MapMyGenome's founder.
By contrast, top hashtags for official accounts' tweets are
closer to their main expertise/business. Similarly, those for ge-
nomics initiatives are pretty much always related to genetic
testing, and this is actually consistent besides top 3. (The
top 10 hashtags include, e.g., #digitalhealth, #genetics, and
#lifestylemedicine). Finally, the percentage of tweets with the
keyword appearing as a hashtag (KH), range from 12% for
23andMe to 25% for AncestryDNA even when excluding of-
ficial accounts, which might be the by-product of promotion
campaigns. When looking at tweets by official accounts KH
values go up for some companies, e.g., AncestryDNA heavily
promotes their brand using hashtags (46% KH).
5.2 URL Analysis
Next, we analyze the URLs contained in the tweets of our
dataset. Recall that the ratio of tweets containing URLs, as well
as the percentage of those in the Alexa top 1M domains, are re-
ported in Table 1. Once again, we distinguish between tweets
from the official accounts and report the top 3 (top-level) do-
mains per keyword in Table 3. Since there are several URL
shorteners in our dataset (e.g., bit.ly), so we first extract the top
10 domains for each keyword and identify those only provid-
ing URL shortening services, then, we "unshorten" the URLs
and use them in our analysis instead.
Among the top URLs shared by the official accounts, we
find, unsurprisingly, their websites, as well as others leading to
other domains owned by them, e.g., 23me.co, ancestry.com.au,
and ancstry.me. A few companies also promote news articles
about them or related topics, e.g., top domains for Counsyl and
MapMyGenome include businesswire.com and indiatimes.
com, while DNAfit seems more focused on social media with
its top domain being Facebook. As discussed previously, the
domain journeythroughhistorysweeps.com appears frequently
in AncestryDNA tweets. Then, note that techcrunch.com, a
blog about technology, appears several times, as it often covers
news and stories about genetic testing. We also highlight the
presence of owler.us, an analytics/marketing provider some-
times labeled as potentially harmful by Twitter, as one of the
top domains for MapMyGenome.
Finally, for genomics initiatives, we notice buff.ly, a social
media manager, suggesting that interested users appear to be
extensively scheduling posts, thus potentially being more tech-
savvy. We also find myriad.com, the domain of Myriad Ge-
netics, which discovered the BRCA1 gene and tried to patent
it [7], as their account is quite active in posting tweets with
Precision/Personalized Medicine keywords.
5.3 Sentiment Analysis
We perform sentiment analysis using SentiStrength [34],
which is designed to work on short texts. The tool outputs two
scores, one positive, in [1, 5], and one negative, in [−1,−5].
We calculate the sum value of the positive+negative scores for
every tweet, then, collect all tweets with that keyword from the
same user, and output the mean sentiment score.
In Figure 4, we report the distribution of sentiment across
the different keywords. The vast majority of tweets have neu-
tral sentiment, ranging from 0 to 1 scores. We run pair-wise
two-sample Kolmogorov-Smirnov tests on the distributions,
and in most cases reject the null hypothesis that they come
from a common distribution at α = 0.05. However, we are
6
understand their interests.
6.1 User Profiles
We start by analyzing the profiles tweeting about genetic
testing: in Figure 5, we plot the distribution of the number of
their followers, following, likes, and tweets.
Followers. Accounts tweeting about genomics initiatives have
a median number of followers similar to baseline, while for the
DTC companies the median is always lower, except for Coun-
syl, MapMyGenome, PathwayGenomics, and VeritasGenetics
(see Figure 5(a)). Also considering that, for these four com-
panies, there is a relatively low number of unique users (see
Table 1), we believe accounts tweeting about them are fewer
but more "popular." There are fewer outliers than the baseline,
which is not surprising since we do not expect many main-
stream accounts to tweet about genetic testing. Some outliers
appear for 23andMe and AncestryDNA, which, upon manual
examination, turn out to be Twitter accounts of newspapers or
known technology websites, reflecting how the two most pop-
ular companies also get more press coverage.
Following. The median number of 'following' (i.e., the ac-
counts followed by the users in our dataset) is usually higher
than baseline for DTC companies but similar for genomics ini-
tiatives (Figure 5(b)). This suggests that users interested in
DTC genetic testing might want to get more information off
Twitter and/or from more accounts.
Likes. We then measure the number of tweets each profile
has liked (Figure 5(c)). This measure, along with the number
of tweets, depicts, to a certain extent degree, a level of en-
gagement. We find that, for all keywords, profiles like fewer
tweets than baseline users. There is one interesting outlier
for 23andMe (@littlebytesnews), who liked more than 1M
tweets; this is likely to be a bot, as also confirmed by Botome-
ter [35]. Also, FamilyTreeDNA appears to have users liking
more tweets than others. However, these accounts appear not
to be bots, as we discuss later.
Tweets. We also quantify the number of tweets each account
posts (Figure 5(d)). As with the number of likes, users in our
datasets are less "active" than baseline users. There are inter-
esting outliers above 1M tweets, which are due to social bots.
We also find more tweets from Counsyl's users, seemingly
mostly due to a large number of profiles describing themselves
as "promoters" of science/digital life, technology enthusiasts,
and/or influencers. Finally, users tweeting about genomics ini-
tiatives appear to be even less active, with a lower median value
of tweets than the rest. Also considering that these users tweet
more about the same keyword (as discussed in Sec. 4) but fol-
low more accounts, we believe that they are somewhat more
passive than the average Twitter user, possibly using Twitter to
get information but actively engaging less than others.
Geographic Distribution. We then estimate the geographic
distribution of the users via the location field in their profile.
This is self-reported, and users use it in different ways, adding
their city (e.g., Miami), state (e.g., Florida), and/or country
(e.g., USA). In some cases, entries might be empty (7.5% of
the tweets in our dataset), ambiguous (e.g., Paris, France vs
Paris, Texas), or fictitious (e.g., "Hell"). Nevertheless, as done
Figure 4: Sentiment scores of the keyword dataset.
unable to reject the null hypothesis when comparing the base-
line dataset to the PathwayGenomics dataset (p = 0.77) and
when comparing DNAfit to Ubiome (p = 0.34). In general, the
genomics initiatives, and in particular Personalized Medicine
and Precision Medicine, have many outliers compared to most
DTC genetic companies, suggesting more users who reveal
strong feelings for or against these concepts. Genomics Eng-
land, however, has a median above zero, indicating generally
positive sentiment. Tweets about Counsyl are very neutral,
while Ubiome tweets seem to be the most positive.
5.4 Discussion
Our content analysis yields a few interesting findings. A
large part of the genetic testing discourse appears to be gen-
erated from news and technology websites, and from tech-
savvy users who rely on services to schedule social media
posts. Also, sentiment around DTC companies is overall neu-
tral, but positive for the genomics initiatives, however, tweets
about DTC companies include a lot of strongly opinionated
users (both positive and negative); we further explore tweets
with high negative score in Sec. 7. Finally, tweets related to
genetic testing not only contain a significantly higher number
of hashtags than a random baseline, but they are also used for
promotion. In general, we find several social media marketing
strategies at play, with some companies employing traditional
giveaways, others promoting mainly third-party articles about
the company, and others focusing their efforts across multiple
social media platforms. For instance, AncestryDNA is quite
active in this context, with one particular hashtag (#sweep-
stakes) found in 1 out of 8 AncestryDNA tweets. This has a
significant impact on how "regular" users engage in tweeting
about genetic testing, which we further analyze next.
6 User Analysis
In this section, we shed light on the accounts tweeting about
genetic testing. After a general characterization of the profiles,
we look for the presence of social bots [35]. Then, we select
a random sample of users tweeting about the two most pop-
ular DTC companies and analyze their latest 1,000 tweets to
7
(a) followers
(b) following
(c) likes
(d) tweets
Figure 5: Boxplots with statistics per user profile (note the log-scale in y-axis).
Figure 6: Geolocation of Twitter profiles, normalized by Internet us-
ing population per country.
in previous work [25], we use this field to estimate where most
of the tweets are coming from. We use the Google Maps Ge-
olocation API, which allows to derive the country from a text
containing a location.5 The API returns an error for 6.6% of
the profiles, mostly due to fictitious locations.
We find that the top 5 countries in our dataset are mostly
English-speaking ones: 69.1% of all profiles with a valid lo-
cation are from the US, followed by the UK (8.6%), Canada
(4.5%), India (2.1%), and Australia (1.4%). We then normal-
ize using Internet-using population estimates [20], and plot the
resulting heatmap, with the top 50 countries, in Figure 6. The
maximum value is obtained by the US (i.e., 0.000254 users per
Internet user), with 72.8K unique users, out of an estimated In-
ternet population of 286M, posting tweets in our dataset. This
suggest that US users dominate the conversation on genetic
testing on Twitter.
We also perform a geolocation analysis broken down to spe-
cific keywords. Unsurprisingly, the top country of origin for
Genomics England is the UK, as it is for DNAfit, which is
based in London. Similarly, the top country for India-based
company MapMyGenome tweets is India. Overall, we find that
tweet numbers are in line with the countries where the DTC
companies are based or operate – e.g., 23andMe health reports
are available in US, Canada, and UK, while AncestryDNA also
operates in Australia – as well as where the genomics initia-
tives are taking place.
5https://developers.google.com/maps/documentation/geolocation
Figure 7: Botometer scores for the keyword dataset.
6.2 Social Bot Analysis
Next, we investigate the presence of social bots in our
datasets, using the Botometer (botometer.iuni.iu.edu), a tool
that, given a Twitter handle, returns the probability of it being
a "social bot," i.e., an account controlled by software, algorith-
mically generating content and establishing interactions [35].
In Figure 7, we plot the distribution of Botometer scores for
all keywords. We compare the distributions using pairwise 2
sample KS tests, and reject the null hypothesis at α = 0.05 for
all datasets except Counsyl and MapMyGenome (p = 0.29),
DNAfit and VeritasGenetics (p = 0.17) and PrecisionMedicine
and VeritasGenetics (p = 0.10). We also find that all me-
dian scores are higher than the baseline (between 0.35 and 0.5
vs 0.3). This is not entirely surprising since we expect many
blogs, magazines, and news services covering genetic testing,
and these are likely to get higher scores than individuals since
they likely automate their activities. However, about 80% of
the accounts in our dataset have scores lower than 0.5 and 90%
lower than 0.6 (i.e., it is unlikely they are bots). We also find the
two most popular keywords, 23andMe and AncestryDNA, as
well as FamilyTreeDNA, somewhat stand out: accounts tweet-
ing about them get the lowest Botometer scores. Although for
FamilyTreeDNA this might be an artifact of the relatively low
number of tweets (2K users), the scores suggest there might
be more interaction/engagement from "real" individuals and/or
fewer tweets by automated accounts about 23andMe and An-
cestryDNA.
We then look at accounts with Botometer scores above 0.7,
8
1.34e−070.000254 RTs
9,077,066
1,399,804
10,476,870
Tweets Users
9,534,302 12,227
3,320
12,000,745 15,547
4,208,967
Likes Hashtags URLs Top 1M
23andMe
24.40% 63.62% 81.43%
AncestryDNA 2,466,443
34.21% 63.64% 78.86%
Total
26.41% 63.62% 80.89%
Baseline
17.47% 41.24% 88.41%
Table 4: Summary of the users' tweets dataset, with last 1K tweets of
a 20% sample of 23andMe and AncestryDNA users.
3,501,053
22,001,065
25,502,118
5,035 139,551,104 342,052,546
finding that, for most DTC keywords, they account for 3–5% of
the users; not too far from the baseline (2%) and the genomics
initiatives (1.5–2%). Counsyl and MapMyGenome have more
than 10% of users with scores above 0.7. We also quantify how
many tweets are posted by (likely) social bots: almost 15% of
all PathwayGenomics tweets come from users with score 0.7
or above (4.5% of all users), while for all other keywords so-
cial bots are not responsible for a substantially high number of
tweets in our datasets.
6.3 Last 1K Tweets of a Sample of Users
We then focus on the users tweeting about the two most pop-
ular companies – i.e., 23andMe and AncestryDNA – and study
their last 1K tweets aiming to understand the characteristics of
the accounts who show interest in genetic testing. We only do
so for 23andMe and AncestryDNA as these companies have
the highest numbers of tweets and users, and thus, are more
likely to lead to a representative and interesting sample.
Data Crawl. We select a random 20% sample of
the
users who have posted at least one tweet with keywords
23andMe/AncestryDNA (resp., 12.2K/64K and 3.3K/16.9K
users) and crawl their latest 1K tweets if their account is still
active.6 This yields a dataset of 12M tweets, outlined in Ta-
ble 4. For comparison, we also get the last 1K tweets of a ran-
dom sample of 5K users from the keyword dataset's baseline
users. Note that statistics in Table 4 refer to the latest 1K tweets
of the user sample, while those in Table 1 to tweets with a given
keyword.
The numbers of retweets and likes per tweet are, once again,
lower than the baseline. However, users tweeting about Ances-
tryDNA receive, for their last 1K tweets, one order of magni-
tude more likes than those tweeting about 23andMe. Moreover,
we observe relatively high percentages of tweets with hash-
tags (63%) and URLs (around 80%). How far back in time
the 1,000th tweet appears varies across users, depending on
how often they tweet. We measure the time between the most
recent and the 1,000th tweet, and find that baseline users are
more "active" than the users who have tweeted about 23andMe
and AncestryDNA, in line with what discussed previously. In
particular, AncestryDNA users appear to post less: for half of
them, it takes at least 359 days to tweet 1K tweets compared to
260 for the baseline and 287 for 23andMe.
Hashtag analysis. We conduct a hashtag analysis on tweets
in Table 4. In Table 5, we report the top 10 hashtags of the
users' last 1K tweets. For 23andMe, we find several hashtags
related to health in the top 10; also considering that the top 30
include #pharma, #cancer, and #biotech, it is likely that users
who have shown interest in 23andMe are also very much inter-
6We find 575 and 61 inactive accounts, resp., for 23andMe and AncestryDNA.
Baseline
gameinsight (0.55%)
AncestryDNA
giveaway (3.31%)
sweepstakes (2.01%) trecru (0.34%)
win (2.01%)
genealogy (1.01%)
tech (0.63%)
23andMe
tech (1.07%)
news (1.06%)
health (0.58%)
business (0.48%)
healthcare (0.43%)
digitalhealth (0.40%) ad (0.51%)
startup (0.39%)
socialmedia (0.34%) promotion (0.48%)
viral (0.34%)
technology (0.34%)
Table 5: The top 10 hashtags of the users' tweets dataset.
btsbbmas (0.33%)
nowplaying (0.30%)
android (0.28%)
androidgames (0.27%)
ipad (0.26%)
trump (0.24%)
perduecrew (0.47%) music (0.21%)
contest (0.44%)
ipadgames (0.20%)
entry (0.51%)
AncestryDNA
instagram.com (6.78%)
fb.me (5.48%)
techcrunch.com (4.42%)
youtu.be (4.04%)
23andMe
fb.me (4.00%)
instagram.com (3.06%)
youtu.be (2.18%)
buff.ly (2.17%)
techcrunch.com (1.53%) wn.nr (1.79%)
lnkd.in (1.02%)
mashable.com (0.65%)
entrepreneur.com (0.63%) buff.ly (1.08%)
swee.ps (0.80%)
nyti.ms (0.62%)
reddit.com (0.55%)
twittascope.com (0.41%)
woobox.com (1.51%)
giveaway.amazon.com (1.17%) fllwrs.com (0.40%)
Baseline
fb.me (5.85%)
instagram.com (4.42%)
youtu.be (2.94%)
twittascope.com (0.58%)
tmblr.co (0.56%)
buff.ly (0.54%)
gigam.es (0.33%)
soundcloud.com (0.32%)
vine.co (0.30%)
Table 6: The top 10 domains of the users' tweets dataset.
ested in (digital) health, which is one of the primary aspects of
23andMe's business. This happens to a lesser extent for Ances-
tryDNA results: while top hashtags include #genealogy (4th),
they also include #giveaway, #sweepstakes, #win, #ad, #pro-
motion, #perduecrew, and #contest, suggesting that these users
are rather interested in promotional products. This is line with
our earlier observation that AncestryDNA extensively uses ad-
vertising and marketing campaigns on Twitter.
URL analysis. In Table 6, we report the top 5 domains of
the three sets. Over the last 1K tweets, users tweeting about
23andMe and AncestryDNA share a substantial number of
links to techcrunch.com, a popular technology website; i.e.,
users who have tweeted at least once about these companies
have an interest about subjects related to new technologies.
In fact, the top 10 list of 23andMe's set of tweets also in-
clude lnkd.in, mashable.com, and entrepreneur.com. For An-
cestryDNA, we find wn.nr, another website related to con-
tests and sweeps. There are thousands of tweets like "En-
ter for a chance to win a $500 Gift Card! wn.nr/DRRrZq
#MemorialDaySweeps #Entry". We also note the presence of
woobox.com, a marketing campaign website, responsible for
organizing giveaways, as well as giveaway.amazon.com, an
Amazon site organizing promotional sweepstakes. We believe
this might be due to a large presence of bots, however, Botome-
ter scores actually indicate these accounts are not. Therefore,
this behavior might be related to the fact that AncestryDNA,
through their marketing campaigns, attract Twitter users who
are generally active in looking for deals and sweeps.
6.4 Discussion
Overall, our user-level analysis shows that Twitter discourse
related to genetic testing is dominated by US users and in gen-
eral by those in English-speaking countries, but not necessar-
ily by "mainstream" popular accounts. However, 23andMe and
AncestryDNA do regularly attract the attention of major news
9
sites. Also, the majority of users involved in genetic testing dis-
cussion are not bots, so Twitter conversation is, to some extent,
"genuine." However, promotion and marketing campaigns end
up attracting different kinds of users, and yield different levels
of engagement. Overall, we find that users interested in genetic
testing appear less active than a random baseline, however,
they are more likely to use Twitter to get information about
their topics of interest, and in particular they are interested in
technology and digital health subjects.
7 Case Studies
We take a closer look at "negative" tweets, following the
sentiment analysis presented previously. We also investigate
the presence of users who post their genetic test results.
7.1 Racism
We select all tweets with genetic testing keywords from
users who yield a total sentiment score below -3, obtaining
3,605 tweets from 3,209 unique users. We then manually ex-
amine those with keywords 23andMe or AncestryDNA (1,725
and 167, respectively), and find several of them containing
themes related to racism, hate, and privacy fears.
In particular, the "ethnic" breakdown provided by ances-
try reports7 seems to spur several
instances of negative-
sentiment tweets associated with racism and disapproval of
multi-cultural/multi-ethnic values. For instance, a user with
more than 3K followers self-describing as a "Yuge fan for Don-
ald Trump", tweets: "Get this race mixing shit off my time
line!!" (Mar 23, 2017) in response to a 23andMe video about
ancestry. Another posts: "I wanna do that 23andMe so bad!
I'm kinda scared what my results will be tho lmao I'm prob
like half black tbh"(Jan 13, 2017), and gets a response: "I was
too just do it and never tell anyone if you're a halfbreed haha".
Also, a user identifying as 'American Fascist' tweets: "I'd like
to get the @23andMe kit but, I'm worried about the results.
Just my luck, I'd have non-white/kike ancestors. #Ultimate-
Blackpill" (May 30, 2017).
Although we leave to future work an in-depth analysis of
genetic testing related racism on the Web, we assess whether
it may be systematic on Twitter, e.g., appearing also in tweets
not scored as negative by SentiStrength. To this end, we search
for the presence of hateful words using the hatebase.org dictio-
nary, a crowdsourced list of 1K terms that indicate hate when
referring to a third person, removing words that are ambiguous
or context-sensitive, as done by previous work [19]. Naturally,
this is a best-effort approach since hateful terms might be used
in non-hateful contexts (e.g., to refer to oneself), or, conversely,
racist behavior can occur without hate words. Also, Twitter
might be removing tweets with hate words as claimed in their
hateful conduct policy.8 Nonetheless, we do find instances of
hate speech, e.g., anti-semitic tweets such as: "as long as there
are khazar milkers to cause people to demand my 23andMe
results, i will always be here to shitpost" (Nov 19, 2016), or
"@*** i would be pleased if you posted your 23andMe so i
7E.g. https://permalinks.23andMe.com/pdf/samplereport ancestrycomp.pdf
8https://support.twitter.com/articles/20175050
can confirm your khazar milkers are indeed genuine" (Dec 23,
2016).
Note that "Khazar milkers" refers to an anti-semitic the-
ory on the origin of Jewish people from the 1900s [15] In
a nutshell, it posits that Ashkenazi Jews are not descendant
from Israelites, but from a tribe of Turkic origin that converted
to Judaism. 23andMe issued ancestry reports that suggested
Ashkenazi Jews in a given haplogroup were descendant from a
single Khazarian ancestor. Understanding the ancestry of Jew-
ish people has been of interest to the genetics community for
years, and the Khazar theory has been refuted repeatedly [4].
Nonetheless, the alt-right has exploited it to corroborate their
anti-semitic beliefs [32], and incorporate it into their collec-
tion of misleading/factually incorrect talking points. In partic-
ular, "khazar milkers" was allegedly coined by the "@***"
user mentioned above, and is used to imply a sort of succubus
quality of Jewish women.
7.2 Privacy
We also identify, among the most negative tweets, themes
related to fears of privacy violation and data misuse. Examples
include "Is it me? Does the idea of #23andMe seem a bit sin-
ister? Do they keep the results? Who owns the results? Who
owns 23andMe?"(Jan 1, 2016), "Same thing with 23andMe
and similar companies. Indefinitely stored data with possible
sinister future uses? #blackmirror"(Nov 13, 2016), and "Why
does this scare the hell out of me? How can our privacy ever
be assured?" (Feb 27, 2016). Searching for 'privacy' and 'pri-
vate' in our keyword dataset returns 1,991 tweets, mostly from
23andMe and Precision Medicine (1.1K and 625, resp.), which
we proceed to examine both manually and from a temporal
point of view (i.e., measuring daily volumes). Overall, we find
that privacy in the context of genetic testing appears to be a
theme discussed recurrently on social media and a concern far
from being addressed. This is not entirely unexpected, consid-
ering that both the DTC market and the genomics landscape
are evolving relatively fast, with regulation and understanding
of data protection as well as informed consent often lagging
behind, as also highlighted in prior work [26, 14, 30].
Interestingly, one of the peaks in tweets related to 23andMe
and privacy occurs on Oct 19, 2015 (with 152 tweets). As dis-
cussed in the Content Analysis section, this a relevant date
w.r.t. the FDA revoking their approval for 23andMe's health
reports, which yields a peak in 23andMe tweets overall. How-
ever, the FDA ruling had nothing to do with privacy, yet, it put
23andMe in the spotlight, possibly causing privacy concerns
to resurface. In fact, privacy and 23andMe discussions period-
ically appear in our dataset, even beyond tweets with negative
sentiment, e.g., "I want to do #23andMe but don't want a pri-
vate company owning my genetic data. Anyone heard of any
hacks to do it anonymously?" (Jul 13, 2017), "@23andMe ur
privacy policy describes how there is no privacy. How about u
not share any data at all. I pay u and u send the results. Period"
(Dec 8, 2015), "Should we be concerned about data collection
and privacy with direct to consumer DNA testing companies
like 23andMe?" (Apr 19, 2017).
10
7.3 Users Sharing Test Results
Finally, we investigate the presence of users who post their
genetic test results, aiming to estimate their number and shed
light on their profiles. Given their popularity, we only do so for
23andMe and AncestryDNA. Among other things, we believe
this is important because health/ancestry reports may contain
sensitive information about the individuals taking these tests,
including their predisposition to diseases and their ethnic her-
itage [3].
Methodology. Finding all tweets that may include test re-
sults is quite hard, and arguably out of scope, thus, we fo-
cus on screenshots of genetic test results as we anecdotally
find a non-negligible number in our dataset. These are actu-
ally almost exclusively ancestry results, even though 23andMe
also provides health reports. We start from the 4.5K/1.5K
23andMe/AncestryDNA tweets with images, but, since they
are too many to be all manually examined, we use the follow-
ing methodology. First, we build two sets with 100 "ground
truth" images with screenshots of ancestry test results, one
each for 23andMe and AncestryDNA, then, use Perceptual
Hashing [21] to find similar images that are likely screenshots
of test results too, and manually check them to exclude false
positives.9
Overall, our approach likely yields a conservative estimate,
nonetheless, it constitutes a best-effort approach to identify and
analyze tweets including test results. We obtain 366 and 204
images for, respectively, 23andMe and AncestryDNA. Upon
manual examination, we find and remove 58 (16%) and 26
(13%) false positives. Thus, we estimate a lower bound of
0.23% and 0.60% of 23andMe and AncestryDNA tweets con-
taining ancestry test screenshots (and 3.40% and 5.15% of
tweets with pictures).
Tweets content. We manually examine the 486 tweets with
screenshots of ancestry test results, finding that users often ap-
pear to be somewhat enthusiastic about their experience. In
some cases, we note a feeling of "relief", sometimes expressed
in a humorous way, when the results show they are predomi-
nantly "white": about 10% of tweets with screenshots include
the word white. Examples include: "23andMe confirms: I'm
super white.", "Got my @23andMe results back today. I'm su-
per white. Like, rice on a paper plate with a glass of milk in a
snowstorm, white."
User Analysis. We also crawl the last 1,000 tweets of the 308
users who have posted screenshots with test results, and ana-
lyze them as done for the random sample discussed in Sec. 6.
We find that their most commonly used hashtags are indeed
related to genetic testing, confirming that the users who do ge-
netic tests are actually generally interested in the subject. It
is also interesting to find #maga to be the second most com-
mon hashtag for users who post 23andMe results (appearing
in 431 tweets). Then, looking at the top domains shared by
these users, we do not observe surprising difference compared
9Perceptual hash functions extract features from multimedia content and cal-
culate hash values based on them. They can be used to compare two objects
by calculating a distance/similarity score between two hash values; the ob-
jects are labeled as (perceptually) equal if the distance is below a chosen
threshold [36]. We set the pHash distant threshold to 17 since it produces the
best results in our setting.
to those from a random sample users (see Table 6), although
we find more social media services like Instagram and Face-
book.
8 Conclusion
This paper presented a large-scale analysis of Twitter dis-
course related to genetic testing. We examined more than 300K
tweets containing 13 relevant keywords as well as 12M tweets
posted by more than 100K accounts that have shown interest in
genetic testing. We found that the discourse related to genetic
testing is often influenced by news and technology websites,
and by a group of tech-savvy users who are overall interested
in tech and digital health. Overall, users tweeting about ge-
netic testing are mostly in the US and other English-speaking
countries, while we do not find evidence of extensive influ-
ence of social bots. However, the broad conversation seems
to be dominated by users that might have a vested interest in
its success, e.g., specialist journalists, medical professionals,
entrepreneurs, etc. This is particularly evident in the tweets
related to genomics initiatives, which are mainly discussed
by highly engaged users and which are influenced, at least in
terms of volume, by important announcements. Moreover, we
noticed that the two most popular DTC companies, 23andMe
and AncestryDNA, also generate the most tweets, however, al-
though 23andMe has half the customers, it produces almost 5
times more tweets, also due to controversy around their failure
to get FDA approval in 2015. We also observed a clear distinc-
tion in the marketing efforts undertaken by different compa-
nies, which ends up influencing users' engagement on Twitter.
Our work is particularly timely as genetic testing and ge-
nomics initiatives are increasingly often associated to ethical,
legal, and societal concerns [16]. In this context, our analysis
sheds light not only on who tweets about genetic testing, what
they talk about, and how they use Twitter, but also on groups
utilizing genetic testing to push racist agendas and users ex-
pressing privacy concerns. We also found a number of enthu-
siastic users who broadcast their test results through screen-
shots notwithstanding possible privacy implications. Overall,
our findings motivate future work studying other social media
platforms and health forums/websites, as well as for in-depth
qualitative studies of the relation between genetic testing and
racism and privacy fears on social media.
References
[1] S. Abbar, Y. Mejova, and I. Weber. You Tweet What You Eat:
Studying Food Consumption Through Twitter. In CHI, 2015.
[2] E. A. Ashley. Towards Precision Medicine. Nature Reviews
Genetics, 17(9), 2016.
[3] E. Ayday, E. De Cristofaro, J.-P. Hubaux, and G. Tsudik. Whole
genome sequencing: Revolutionary medicine or privacy night-
mare? IEEE Computer, 2015.
[4] D. M. Behar, M. Metspalu, Y. Baran, et al. No Evidence from
Genome-Wide Data of a Khazar Origin for the Ashkenazi Jews.
Human Biology, 85(6), 2013.
[5] D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet alloca-
tion. JMLR, 2003.
11
[6] P. Burnap, M. L. Williams, L. Sloan, O. Rana, W. Housley,
A. Edwards, V. Knight, R. Procter, and A. Voss. Tweeting the
Terror: Modelling the Social Media Reaction to the Woolwich
Terrorist Attack. Social Network Analysis and Mining, 4(1),
2014.
[7] T. Caulfield, T. Bubela, and C. Murdoch. Myriad and the mass
media: the covering of a gene patent controversy. Genetics in
Medicine, 9(12), 2007.
[8] P. A. Cavazos-Rehg, M. J. Krauss, S. J. Sowles, and L. J. Bierut.
Hey Everyone, I'm Drunk. An Evaluation Of Drinking-Related
Twitter Chatter. JSAD, 76(4), 2015.
[9] D. Chatzakou, N. Kourtellis, J. Blackburn, E. De Cristofaro,
G. Stringhini, and A. Vakali. Measuring #GamerGate: A Tale
of Hate, Sexism, and Bullying. In WWW Companion, 2017.
[10] P. Chow-White, S. Struve, A. Lusoli, F. Lesage, N. Saraf, and
'Warren Buffet Is My Cousin': Shaping Public
A. Oldring.
Understanding of Big Data Biotechnology, Direct-To-Consumer
Genomics, and 23andMe on Twitter. Information, Communica-
tion & Society, 2017.
[11] G. Coppersmith, M. Dredze, and C. Harman. Quantifying Men-
tal Health Signals In Twitter. In CLPsych, 2014.
[12] L. Covolo, S. Rubinelli, E. Ceretti, and U. Gelatti.
Internet-
Based Direct-to-Consumer Genetic Testing: A Systematic Re-
view. JMIR, 17(12), 2015.
[13] M. De Choudhury, M. Gamon, S. Counts, and E. Horvitz. Pre-
dicting Depression via Social Media. In ICWSM, 2013.
[14] E. De Cristofaro. Genomic privacy and the rise of a new research
community. IEEE Security & Privacy, 12(2), 2014.
[15] A. Feldman.
23andMe Backpedals On Khazar The-
ory But The 'Alt-Right' Eats It Up, Anyway.
http:
//forward.com/news/national/381500/23andme-backpedals-on-
khazar-theory-but-the-alt-right-eats-it-up-anyway/, 2017.
[16] FTC.
tions.
kits-consider-privacy-implications, 2017.
the privacy implica-
DNA test kits: Consider
https://www.consumer.ftc.gov/blog/2017/12/dna-test-
[17] L. Goldsmith, L. Jackson, A. O'connor, and H. Skirton. Direct-
to-Consumer Genomic Testing: Systematic Review of the Liter-
ature on User Perspectives. European Journal of Human Genet-
ics, 20(8), 2012.
[18] L. Harley. White House hosts Precision Medicine Initia-
tive Summit. http://www.frontlinegenomics.com/white-house-
hosts-precision-medicine-initiative-summit/, 2016.
[19] G. E. Hine, J. Onaolapo, E. De Cristofaro, N. Kourtellis, I. Leon-
tiadis, R. Samaras, G. Stringhini, and J. Blackburn. Kek, Cucks,
and God Emperor Trump: A Measurement Study of 4chan's Po-
litically Incorrect Forum and Its Effects on the Web. In ICWSM,
2017.
[20] Internet Live Stats.
Internet Users by Country (2016). http:
//www.internetlivestats.com/internet-users-by-country/, 2017.
[21] E. Klinger and D. Starkweather. http://www.phash.org/, 2018.
[22] H. Kwak, C. Lee, H. Park, and S. Moon. What Is Twitter, A
Social Network Or A News Media? In WWW, 2010.
[23] E. S. Lander, L. M. Linton, B. Birren, C. Nusbaum, M. C. Zody,
J. Baldwin, K. Devon, K. Dewar, M. Doyle, W. FitzHugh, et al.
Initial Sequencing and Analysis of the Human Genome. Nature,
409(6822), 2001.
[24] K. Lerman, M. Arora, L. Gallegos, P. Kumaraguru, and D. Gar-
cia. Emotions, Demographics and Sociability in Twitter Inter-
actions. In ICWSM, 2016.
[25] A. Marcus, M. S. Bernstein, O. Badar, D. R. Karger, et al. Twit-
info: Aggregating and Visualizing Microblogs for Event Explo-
ration. In CHI, 2011.
[26] D. Mascalzoni, A. Hicks, P. Pramstaller, and M. Wjst. Informed
consent in the genomics era. PLoS Medicine, 5(9), 2008.
[27] NHGRI. The Cost of Sequencing a Human Genome. https:
//www.genome.gov/sequencingcosts/, 2017.
[28] B. Pardy. Tweet: FamilyTreeDNA Privacy. http://archive.is/
AUj6L, 2017.
[29] M. J. Paul and M. Dredze. You Are What You Tweet: Analyzing
Twitter for Public Health. In ICWSM, 2011.
[30] A. M. Phillips. DTC Genetics for Ancestry, Health, Love and
More: A View of the Business and Regulatory Landscape. Ap-
plied & Translational Genomics, 8, 2016.
[31] N. D. Prasetyo, C. Hauff, D. Nguyen, T. van den Broek, and
D. Hiemstra. On the Impact of Twitter-Based Health Cam-
paigns: A Cross-Country Analysis of Movember. In EMNPL,
2015.
[32] E. Reeve. Vice News – White Nonsense: Alt-right trolls are
arguing over genetic tests they think prove their whiteness. http:
//bit.ly/2DhP90h, 2016.
[33] G. Salton and M. J. McGill. Introduction to modern information
retrieval. McGraw-Hill, 1986.
[34] M. Thelwall, K. Buckley, G. Paltoglou, D. Cai, and A. Kappas.
Sentiment Strength Detection In Short Informal Text. JASIST,
61(12), 2010.
[35] O. Varol, E. Ferrara, C. A. Davis, F. Menczer, and A. Flam-
mini. Online Human-Bot Interactions: Detection, Estimation,
and Characterization. In ICWSM, 2017.
[36] C. Zauner. http://www.phash.org/docs/pubs/thesis zauner.pdf,
2010.
12
|
1704.02489 | 1 | 1704 | 2017-04-08T12:40:51 | Analyzing Social Interaction Networks from Twitter for Planned Special Events | [
"cs.SI"
] | The complex topology of real networks allows its actors to change their functional behavior. Network models provide better understanding of the evolutionary mechanisms being accountable for the growth of such networks by capturing the dynamics in the ways network agents interact and change their behavior. Considerable amount of research efforts is required for developing novel network modeling techniques to understand the structural properties such networks, reproducing similar properties based on empirical evidence, and designing such networks efficiently. First, we demonstrate how to construct social interaction networks using social media data and then present the key findings obtained from the network analytics. We analyze the characteristics and growth of such interaction networks, examine the network properties and derive important insights based on the theories of network science literature. We also discuss the application of such networks as a useful tool to effectively disseminate targeted information during planned special events. We observed that the degree-distributions of such networks follow power-law that is indicative of the existence of fewer nodes in the network with higher levels of interactions, and many other nodes with less interactions. While the network elements and average user degree grow linearly each day, densities of such networks tend to become zero. Largest connected components exhibit higher connectivity (density) when compared with the whole graph. Network radius and diameter become stable over time evidencing the small-world property. We also observe increased transitivity and higher stability of the power-law exponents as the networks grow. Data is specific to the Purdue University community and two large events, namely Purdue Day of Giving and Senator Bernie Sanders' visit to Purdue University as part of Indiana Primary Election 2016. | cs.SI | cs | Sadri, Hasan, Ukkusuri, Suarez Lopez
1
Analyzing Social Interaction Networks from Twitter for Planned Special Events
Arif Mohaimin Sadri, Ph.D.
Lyles School of Civil Engineering
Purdue University, West Lafayette, IN 47907, USA
Email: [email protected]
(Corresponding Author)
Samiul Hasan, Ph.D.
Assistant Professor
Department of Civil, Environmental, and Construction Engineering
University of Central Florida
12800 Pegasus Drive, Orlando, FL 32816
Phone: 407-823-2841
Email: [email protected]
Satish V. Ukkusuri, Ph.D.
Professor
Lyles School of Civil Engineering
Purdue University, West Lafayette, IN 47907, USA
Email: [email protected]
Juan Esteban Suarez Lopez
School of Civil Engineering
National University of Colombia
Medellin, Colombia
Email: [email protected]
Sadri, Hasan, Ukkusuri, Suarez Lopez
2
ABSTRACT
The complex topology of real networks allows its actors to change their functional behavior.
Network models provide better understanding of the evolutionary mechanisms being accountable
for the growth of such networks by capturing the dynamics in the ways network agents interact
and change their behavior. Considerable amount of research efforts is required for developing
novel network modeling techniques to understand the structural properties such networks,
reproducing similar properties based on empirical evidence, and designing such networks
efficiently. In this study, we first demonstrate how to construct social interaction networks using
social media data and then present the key findings obtained from the network analytics. We
analyze the characteristics and growth of online social interaction networks, examine the network
properties and derive important insights based on the theories of network science literature. We
also discuss the application of such networks as a useful tool to effectively disseminate targeted
information during planned special events. We observed that the degree-distributions of such
networks follow power-law that is indicative of the existence of fewer nodes in the network with
higher levels of interactions, and many other nodes with less interactions. While the network
elements and average user degree grow linearly each day, densities of such networks tend to
become zero. Largest connected components exhibit higher connectivity (density) when compared
with the whole graph. Network radius and diameter become stable over time evidencing the small-
world property. We also observe increased transitivity and higher stability of the power-law
exponents as the networks grow. Since the data is specific to the Purdue University community,
we also observe two very big events, namely Purdue Day of Giving and Senator Bernie Sanders'
visit to Purdue University as part of Indiana Primary Election 2016.
Keywords: planned special events, social media, Twitter, user mention, social interaction, network
science
Sadri, Hasan, Ukkusuri, Suarez Lopez
3
BACKGROUND AND MOTIVATION
Planned Special Events (PSE) include sporting events, concerts, conventions and similar large
events at specific venues such as stadiums and convention centers among others. PSEs possess
many operational needs that can be anticipated and managed in advance because of specific
locations and times of occurrence [1]. Organizing PSEs have several challenges including parking
management, crowd management, pedestrian facility design, and special facility for senior citizens
and handicapped individuals, providing transit facility for captive riders among others. In addition,
police enforcements often need to close several streets for security reasons, manage crowds who
walk together to the location and guide motorists to specific routes who are unfamiliar with the
area. Individuals attending these events travel by various travel modes, i.e. walk, private car and
public transit. Since the traffic patterns of PSEs vary significantly as compared to any given
weekday traffic patterns, accidents, or any other incidents, it is of great inconvenience for traffic
managers, drivers or freight movers to deal with PSEs [1]. Thus, PSEs are a major concern for
traffic planners and local transportation agencies because of increased traffic demand and
restricted roadway capacity causing disruptions to the regular traffic conditions [1, 2]. However,
this disruption and the associated operational needs can be anticipated and managed in advance
[3]. Participation from key stakeholders, development and implementation of effective traffic
management plan, and the flexibility to change plans to manage real-time traffic are among the
key strategies to efficiently handle PSEs [4].
Network Science is an emerging research field having multifaceted outlook in the study of large-
scale real networked systems which considers both the network topology and the behavior of
network actors. Complex networks with dynamic and irregular structure along with their statistical
properties are the primary focus of such research efforts i.e. the combined knowledge of network
structure and behavior which is significantly distinct from the straight-forward analysis of single
small graphs [5, 6]. The prevalence of networked systems has resulted in a number of studies with
applications in various domains over the last decade. Such domains include social organizations,
internet, systems biology, supply chains and logistics, information and communication systems,
financial markets, infrastructure systems among others [6-8]. A number of novel structural
properties and concepts have been observed and some unifying principles along with relevant
statistical distributions have been derived to characterize interdependence between the structure
and function of complex real networks. For example, the existence of small-world property in
many real networks suggest that most vertices in the network can be reached from every other
node by using relatively short paths despite large size of these networks [9-11]. On the other hand,
the scale-free property suggests the existence of large hubs i.e. a few nodes which are highly
connected to many other nodes in the network and such hubs result in a network degree
distributions (power-law) with highly right-skewed long tail referring to nodes with a much higher
degree than most other nodes [12]. Other properties include transitivity, network resilience, mixing
patterns, network homophily or similarity, degree correlations, preferential attachment,
community structure, network navigation, size of largest components among other which provide
valuable insights [6].
Information dissemination is the systematic way of distributing information and spreading
awareness to every individual and systematic planning, collection, organization, and delivery
technique are needed before circulating relevant information to any target audience by using
various media and communication means. Information dissemination thus constitutes an important
Sadri, Hasan, Ukkusuri, Suarez Lopez
4
and critical factor for the success of organizing PSEs. Despite many technical challenges to
manage PSEs, the empirical literature does not provide any specific guidelines to local traffic
managers and emergency response personnel to disseminate travel specific information as part of
traffic management plans for PSEs. Social media platforms (Twitter, Facebook and others) can be
considered as appropriate means of disseminating information dynamically. Studies have found
that an individual's real world actions can be inferred based on the connections and activities in
social media [13]. Twitter shows both the characteristics of a social network and an informational
network [14] and users can share short messages up to 140 characters along with the ability to
follow other. While the information network properties of Twitter instigate information contagion
globally, the social network properties allows access to geographically and personally relevant
information and [15]. Because of specific features, Twitter can be particularly useful for effective
information dissemination during PSEs.
The overwhelming usage and activity of Twitter users can be expressed by a 2013 statistic that
suggests over 143K tweets per second being generated on Twitter [16]. User activity in social
media has shown its prevalence in recent years and the world is more connected now than ever
before. The ease in information sharing and the ability to instigate others have primarily
contributed to such connectivity all over the world. However, the purpose and context of the online
activity on social media platforms, such as Facebook and Twitter, may vary from user to user. For
example, users' check-in activity can be referred as distinct from what users' post or share to
disseminate any specific information. One specific feature of such information sharing activity is
the ability of users' in mentioning (direct mentions, retweets and replies) others that they follow.
In this study, we used Twitter REST API for the keyword 'purdue' and obtained 56,159 tweets
over a month. Each tweet consisted of several words and user mentions which co-appeared with
the keyword 'purdue' and resembled higher likelihood of a Twitter subscriber belonging to the
Purdue University community. Next, we construct the 'purdue' specific online social interaction
networks by considering the user mentions. Finally, we analyze such network characteristics,
examine the properties and network growth, and derive relevant insights based on the theories of
network science literature.
RELATED WORK
The coupling between network structure and behavior of network agents has initiated a number of
studies related to the dynamical behavior of network actors communicating through complex
network topologies. This interdependence has significant outcomes when the robustness and
resilience of a real network is considered and the way networks respond to targeted failure due to
external disturbances [17]. Fundamental research questions, as explored in the empirical literature,
can be listed as (i) how to model emergence of new innovations or ideas based on agent
interactions, (iii) who are the key influential players in the network and how to find them, (iii) how
to maximize network influence based on certain mechanism, (iv) when and how networks execute
contagious behavior such as disease transmission or information propagation, (v) which agents are
more likely to connect and interact with agents of the same profile? A few examples of such studies
may include epidemic models of disease transmission [18, 19], email networks and computer virus
transmission [20, 21], collapse in power grid networks [22, 23], disruptions of trade markets [24],
information propagation in social networks [25], and many others.
Sadri, Hasan, Ukkusuri, Suarez Lopez
5
The successful spreading of awareness to every individual in a community solely depends on an
effectual information dissemination process [26-28]. The prevalence of user activities in social
media (Twitter and Facebook for example) has shown its prevalence over the last decade, allowing
people to be more connected than ever before worldwide. This has been possible because of the
flexibility in information sharing and the way people instigate their online neighbors. Unlike the
usual communication media, social media can be used as useful means to disseminate information
during any specific event such as PSEs. The benefits to convincingly collect, analyze and use such
large-scale and rich information from online information sources have also been addressed [29].
Many empirical studies have explored social media data for emergency response and disaster
management [30-35], crisis informatics [36-43] and many others. In fact, one could efficiently
analyze and predict real world human actions based on user activity and connectivity on social
media platforms [13, 15]. However, the purpose and context of the user activity over social media
platforms may vary from one user to another user.
Transportation researchers have started exploiting these large-scale data sources more extensively
in recent years. Such examples include problems related to travel survey techniques [44, 45],
activity-pattern modeling [46-48], origin-destination demand estimation [49-53] and transit
planning [54]. However, a few studies explored people's ego-centric offline social networks and
the effects of ego-centric social ties on joint activity participation [55, 56]. To the best of the
authors' knowledge, the empirical literature does not provide any specific evidence on how to
extract networks of user interactions from social media and the applicability of such networks to
solve transportation related problems, which is the primary scope of this study.
NETWORK DATA DESCRIPTION
We used twitter REST API to collect tweets for using a specific keyword 'purdue' which is
frequently used in Twitter within the Purdue community. The dataset selection was due to its
relevance in the study of information flow patterns of specific topics related to Purdue University.
It is equally important to understand how special events affect this behavior over time. By running
the twitter REST API between April 16, 2016 and May 16, 2016 for four consecutive weeks, we
were able to obtain 56,159 tweets for the query 'purdue' after initial data cleaning for non-English
tweets and common stop words. 19,532 of these tweets did not include any user mentions,
however, the rest of the tweets included at least one user mention in each tweet. The frequency
distribution of the top 50 words in these tweets, co-appeared with 'purdue', suggests that these top
50 most frequent words contribute up to ~25% of all the words that appeared in the collected data.
A word cloud of the most frequent words in the dataset is presented in Figure 1(a). Different
combinations of these words constitute specific topics based on which users influence one another
by using the user mention feature on Twitter. These frequently appeared words also suggest the
emergence of event specific topics such as Purdue Day of Giving, Senator Bernie Sanders's visit
during Indiana Primary among others. Celebrity players of Purdue such as Anthony Brown
(football), Spike Albrecht (basketball) and others also contributed to many topics. The differences
in the amount of user mentions in the tweets over days are plotted in Figure 1(b). It can be clearly
seen that the number of tweets having user mentions is almost twice as the number of tweets
without mentions. These tweets primarily contribute to the formation of networks of direct social
influence. In Table 1, we present the amount of user mentions as was observed in the tweets.
Sadri, Hasan, Ukkusuri, Suarez Lopez
6
In Twitter, users can post tweets up to 140 characters and each data point can be stored as a tuple
Tweet once collected with the following information:
Tweet (tweet_id) = {tweet, tweet_created_at, user_id, user_screen_name, user_location,
user_name, user_followers_count, user_friends_count, user_statuses_count,
user_favourites_count,, user_listed_count, user_mention, tweet_retweeted, tweet_lat, tweet_lon}
For this study, we are interested in using a sub-tuple tweet to infer the links of direct influence that
finally evolves into a highly connected network of a given context:
tweet (tweet_id) = { user_id, tweet, user_mention}
Let us consider the following three tuples from the tweets generated on Twitter on 04/28/2016
(02:45:40 +0000), 05/02/2016 (14:45:33 +0000) and 05/03/2016 (13:50:21 +0000), respectively.
tweet(725516302819938305) = {709920419529281537, 'at purdue university, we are in this
campaign to win and become the democratic nominee. - bernie sanders htt…', [null]}
tweet(727147016233558016) = {3239853627, 'rt @saracohennyc at purdue university, we are
in this campaign to win and become the democratic nominee. - bernie sanders htt…',
[709920419529281537]}
tweet(727495513277382656) = {325069363, 'rt @bernielovesall: rt @saracohennyc at purdue
university, we are in this campaign to win and become the democratic nominee. - bernie
sanders htt…', [3239853627, 709920419529281537]}
Based on the above tweets, there is a directed link from user 709920419529281537 to user
3239853627 and from user 709920419529281537 to 325069363. Please refer to Figure 2 (a) for
the details of this network construction. The preliminary analysis of these data waves suggest the
existence of 34,363 unique users and 38,442 unique undirected links (39,709 links if direction is
considered) of direct influence. The network elements of the graph (constructed based on the data)
are presented in Table 1.
Different network visualizations are presented from Figure 2(c)-2(f) to depict the network
configurations and how the network structure appears after four weeks. In Figure 2(c), we present
the undirected graph having 34,363 users from Twitter and 38,442 links. The network isolates
without any connectivity are also shown in the periphery. This graph includes 8,348 connected
components and the largest connected component is presented in Figure 2(e)-2(f). While Figure
2(f) better represents the hierarchical structure of the network with the most central node in the
center, Figure 2(e) presents weighted edges based on the number of appearance of these links. This
weighted graphs help to explain the existence of links having higher strength which also serves as
an evidence of higher influence. In Figure 2(d), we present the largest hub i.e. the most central
node having the largest degree. It is highly intuitive that the network will be under huge disruption
if such node disappears or remain active in cases.
Sadri, Hasan, Ukkusuri, Suarez Lopez
7
IMPLICATIONS FROM NETWORK ANALYSES
The user mentions, observed in the data for four consecutive weeks, construct a social interaction
network that includes 34,363 nodes and 39,709 links for the directed graph and 34,363 nodes and
38,442 links for the undirected graph (Table 1, Figure 2c). 6,096 of these nodes appeared as
network isolates (nodes without connectivity) in the periphery along with 8,348 connected
components (Figure 2c). 33,020 connections (links) among 21,045 users (nodes) were observed in
the largest connected component of this network (Figure 2e-2f). The radius and diameter of the
largest connected component were observed as 9 and 17, respectively. These are relevant to the
small world property of complex real networks that refers to the existence of relatively short paths
between any pair of nodes in most networks despite their large size. The existence of this property
has been observed in many real networks as studied in the empirical literature [9-11]. This property
has significant implications in the modeling of dynamic processes occurring on real networks. For
example, when effective information dissemination is considered, contagion will be faster through
the network because of short average path lengths [6]. Three important measures to explain this
property are eccentricity, radius and diameter. While the eccentricity of a node in a graph is the
maximum distance (number of steps or hops) from that node to all other nodes; radius and diameter
are the minimum and maximum eccentricity observed among all nodes, respectively.
Network Density, frequently used in the sociological literature [57], equals to 0 for a graph without
any link between nodes and 1 for a completely connected graph. The density of real graphs refer
to the proportion of links that exist in the graph and the maximum number of possible links in the
graph. For n users, the number of maximum links are n(n-1) for a directed graph and n(n-1)/2 for
a undirected graph. The densities that we observe in the social interaction network of 21,045 users
are 0.00003, 0.00007 and 0.00015 for the directed, undirected and the largest connected
component, respectively. This implies higher connectivity in the largest connected component,
more than twice as much as in the whole network. The node Degree is the number of edges
adjacent to that node, In-degree is the number of edges pointing in to the node and Out-degree is
the number of edges pointing out of the node. The degree of a node (𝑘) is the number of direct
edges to other nodes in a graph from that node and the degree distribution 𝑃(𝑘) in real networks,
(probability that a randomly chosen node has degree 𝑘), is significantly different from the Poisson
distribution, typically assumed in the modeling of random graphs. Real networks, in fact, exhibit
a power law (or scale-free) degree distribution characterized by higher densities of triangles such
as cliques in a social network [12]. Such networks also demonstrate significant correlations in
terms of node degrees or node-level attributes. Existence of hubs i.e. a few nodes that are highly
connected to other nodes, in the network can also be validated by the scale-free phenomenon. The
largest hub (or ego), as was observed in our dataset, is visualized in Figure 2(d). The presence of
large hubs results in a degree distribution with long tail (highly right-skewed), indicating the
presence of nodes with a much higher degree than most other nodes. For an undirected network,
the degree distribution 𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑘) can be written as follows:
𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑘) ∝ 𝑘−𝛾 …………..………….……….. (1)
where, 𝛾 is some exponent and 𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑘) decays slowly as the degree 𝑘 increases, increasing the
probability of obtaining a node with a very high degree. Networks with power-law distributions
are called scale-free networks that holds the same functional form (power laws) at all scales. The
Sadri, Hasan, Ukkusuri, Suarez Lopez
8
power law 𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑘) remains unchanged (other than a multiplicative factor) when rescaling the
independent variable 𝑘 by satisfying:
𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑥𝑘) = 𝑥−𝛾 𝑃𝑑𝑒𝑔𝑟𝑒𝑒(𝑘) …………………..….. (2)
The presence of hubs that are orders of magnitude larger in degree than most other nodes is a
characteristic of power law networks. The average degree of all the 34,363 users in the social
interaction network is 1.156 (Table 1) and the overall degree distributions are plotted in Figure
3(a). Figure 3(b) presents the degree distributions that were observed each day starting from the
data period of data collection. By using Alstott et al.'s python code, we obtained the best fitting to
the degree distributions [58] and the empirical data of this study fits close to being a power-law or
truncated power-law distributions. The code also returns a value of 𝑥𝑚𝑖𝑛 which refers to the
minimal value of 𝑥 at which the power law begins to become valid. For power-law, we obtain 𝛾 =
2.294 ± 0.046; 𝑥𝑚𝑖𝑛 = 11 and for truncated power-law 𝛾 = 2.278; 𝑥𝑚𝑖𝑛 = 11. Here, 𝛾 is the
slope of the distribution. When 𝛾 is high, the number of nodes with high degree is smaller than the
number of nodes with low degree. A low value of 𝛾 may refer to a more equal distribution, whereas
higher values of 𝛾 may denote more and more unfair degree distributions. It is important to note
here that the best fit power law may only cover a portion of the distribution's tail [58]. From Figure
3(d), it appears that the data also fits close to being a log-normal distribution. However, difficulties
in distinguishing the power law from the lognormal are common and well-described, and similar
issues apply to the stretched exponential and other heavy-tailed distributions [59, 60]. Our analysis
on the distributions fitting are based on pairwise comparison between power-law, truncated power-
law, log-normal, and exponential distributions. See Figure 3(c)-3(d) for details.
Another network property is Transitivity that implies the higher likelihood of any two given nodes
in a network to be connected, given each of these two nodes are connected to some other node.
This property refers to the fact that the friend of one's friend is likely also to be the friend of that
person in case of social networks and this is a distinctive deviation from the properties of random
graphs. In fact, this is indicative of heightened number of triangles (sets of three nodes each of
which is connected to each of the others) that exist in real networks [6] (Newman, 2003a). The
existence of triangles can be quantified by Clustering Coefficient. C:
𝐶 =
3∗ 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑟𝑖𝑎𝑛𝑔𝑙𝑒𝑠 𝑖𝑛 𝑡ℎ𝑒 𝑛𝑒𝑡𝑤𝑜𝑟𝑘
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑛𝑛𝑒𝑐𝑡𝑒𝑑 𝑡𝑟𝑖𝑝𝑙𝑒𝑠 𝑜𝑓 𝑛𝑜𝑑𝑒𝑠
……….……… (3)
A connected triple refers to a single node with links running to an unordered pair of others. In case
of an unweighted graph, the clustering coefficient (𝑐𝑐𝑖) of a node 𝑖 refers to the fraction of possible
triangles that exist through that node:
𝑐𝑐𝑖 =
2 𝑇𝑖
𝑑𝑒𝑔𝑖∗[𝑑𝑒𝑔𝑖−1]
…………………………………. (4)
Here, 𝑇𝑖 is the number of triangles that exist through node 𝑖 and 𝑑𝑒𝑔𝑖 is the degree of node 𝑖. The
average clustering coefficient in the undirected social interaction network was observed to be
0.149 (Table 1)
Turning to the results obtained from the network growth analysis, we present these results in Figure
4 and Figure 5. The unit of time for the analysis of network growth was set to be 24-hours. We
Sadri, Hasan, Ukkusuri, Suarez Lopez
9
observe that the growth of network elements [nodes and links in Figure 4(a), isolates and connected
components in Figure 4(b)] is almost linear over days except for the date 04/26/2016 for which we
could not obtain any data. One key insight here is that the growth rate is higher for days that
followed special events such as Purdue Day of Giving and Senator Bernie Sanders' visit to Purdue
University during Indiana Primary 2016. While the network elements grow linearly, the densities
tend to go down to zero because of less overall connectivity in the network (Figure 4c). However,
the density of the largest connected components remain slight higher over time. In addition, as the
social interaction network keeps growing based on difference in user interaction for various topics,
the diameter and radius keeps fluctuating initially, however becomes constant later. This is
indicative of network stability when the reachability from one node to another node is considered
(Figure 5a-5b). The average degree of the nodes shows similar pattern to that of the growth of
network elements initially, however becomes flat later (Figure 5c). The network transitivity, based
on average clustering coefficient, suggests that the network becomes more transitive over time,
however slight fluctuation still remains (Figure 5d). The power-law exponents each day are
presented in Figure 5e. After reducing slightly in the initial days, they turn to becoming flat and
take a value close to 2.3. This implies that the power-law property holds when social interaction
network is observed over a long period of time.
Finally, we present the existence of repetitions in terms of how elements of such networks appear
in the network data (Figure 6). This is of great significance within the context of finding highly
active nodes (users) in the social interaction networks along with the strength of relationships
between node pairs. The relevance of considering the dynamic strength of social ties in information
spreading has been duly addressed [61, 62]. The weighted graph, based on the number of times a
link has appeared, is presented in Figure 2(e). In order to assess the commonalities of network
elements (nodes and links) over time, we compute the fraction of nodes and links every day that
appeared at least once in any of the previous days. From Figure 6(a), it can be seen that 65.2% of
the total users (or nodes) on May 16, 2016 appeared in the data at least once. Similarly, we observe
that 28.3% of the total links of interaction (undirected graph) on May 16, 2016 appeared, in any
of the previous days, at least once (Figure 6b).
CONCLUSIONS AND KEY FINDINGS
Real networks having complex topologies demonstrate a unique interdependence between the
structure and functional behavior. In this study, we demonstrate such interdependence by
exploiting online social interaction networks based on network data obtained from Twitter. The
social interaction network was formed by following the user mentions appeared in the tweets
during four consecutive weeks which are specific to a university community. The network
characteristics and properties have been analyzed and the network growth has been monitored over
time. Key insights obtained from the network analyses are listed below:
• The network degree distributions exhibit a power-law which is indicative of the scale-free
property of most real networks. This property holds for any given day as evident from the
empirical data. This is indicative of the existence of fewer nodes in the network with higher
levels of interactions, and many other nodes with less interactions.
• Network visualization is indicative of some nodes (users) being highly active, some links
(relations) having higher strength, existence of network isolates, connected components,
Sadri, Hasan, Ukkusuri, Suarez Lopez
10
and hubs i.e. nodes having reachability to many other nodes. This is also evident when the
appearance of the network elements each day is compared to all previous days.
• Network elements and average user degree grow linearly each day, however, network
densities tend to become zero. Largest connected components exhibit higher connectivity
(density) when compared with the whole graph.
• Network radius and diameter become stable over time which suggests less variations when
the reachability from one node to another node is considered. These variables are related
to the small-world property.
Increased transitivity in the growth of such networks is observed following the pattern of
mean clustering coefficient. Initial fluctuations of the power-law exponents reduce as the
network grows.
•
The properties of social interaction networks, as observed in this study, have fundamental
implications towards effective information dissemination. For example, power-law degree
distributions is related to the resiliency of a communication network. The level of resilience, when
a random nodes in the network are removed, depends solely on the way the network is formed i.e.
network topology. In case of networks having many low-degree nodes would have less disruption
and higher resilience since this nodes lie on few paths between others. However, removal of hubs
(high degree nodes) would cause major disruption and network agents would fail to communicate
since the regular length of path will increase as a result of many disconnected pairs of nodes. For
any Planned Special Event (PSE), the assembling of vehicles and pedestrians in a short amount of
time cause transportation and transit authorities to often encounter significant challenges in
controlling the induced traffic coming from different origins before the event and departing from
the event location after the event. There is hardly any specific method in the empirical literature
that would allow local emergency managers or agencies to properly disseminate targeted
information to any specific audience as part of traffic management procedures for PSEs. A better
knowledge of social interaction network growth and properties would be worthwhile to be
considered for such events.
Acknowledgements
The authors are grateful to National Science Foundation for the grant CMMI-1131503 and CMMI-
1520338 to support the research presented in this paper. However, the authors are solely
responsible for the findings presented in this study.
Author Contributions Statement
All the authors have contributed to the design of the study, conduct of the research, and writing
the manuscript.
Additional Information
Competing financial interests: Authors declare no competing financial interests.
Sadri, Hasan, Ukkusuri, Suarez Lopez
11
REFERENCES
1. Skolnik J, Chami R, Walker M. Planned Special Events–Economic Role and Congestion
Effects. 2008.
2. Carson JL, Bylsma RG. Transportation planning and management for special events2003.
3. Latoski SP, Dunn WM, Wagenblast B, Randall J, Walker MD. Managing travel for planned
special events: final report. 2003.
4. Dunn Jr W. Traffic Management of Special Events: The 1986 US Open Golf Tournament.
Transportation Research Circular. 1989;(344).
5. Albert R, Barabási A-L. Statistical mechanics of complex networks. Reviews of modern
physics. 2002;74(1):47.
6. Newman ME. The structure and function of complex networks. SIAM review.
2003;45(2):167-256.
7. Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang D-U. Complex networks: Structure
and dynamics. Physics reports. 2006;424(4):175-308.
8. Ukkusuri SV, Mesa-Arango R, Narayanan B, Sadri AM, Qian X. Evolution of the
Commonwealth Trade Network. International Trade Working Paper 2016/07,
Commonwealth Secretariat, London. 2016.
9. Milgram S. The small world problem. Psychology today. 1967;2(1):60-7.
10. Travers J, Milgram S. An experimental study of the small world problem. Sociometry.
1969:425-43.
11. Watts DJ, Strogatz SH. Collective dynamics of 'small-world'networks. nature.
1998;393(6684):440-2.
12. Barabási A-L, Albert R. Emergence of scaling in random networks. science.
1999;286(5439):509-12.
13. Korolov R, Peabody J, Lavoie A, Das S, Magdon-Ismail M, Wallace W, editors. Actions are
louder than words in social media. Proceedings of the 2015 IEEE/ACM International
Conference on Advances in Social Networks Analysis and Mining 2015; 2015: ACM.
14. Myers SA, Sharma A, Gupta P, Lin J, editors. Information network or social network?: the
structure of the twitter follow graph. Proceedings of the 23rd International Conference on
World Wide Web; 2014: ACM.
15. Kryvasheyeu Y, Chen H, Obradovich N, Moro E, Van Hentenryck P, Fowler J, et al. Rapid
assessment of disaster damage using social media activity. Science advances.
2016;2(3):e1500779.
16. Krikorian R. New tweets per second record, and how. Twitter Engineering Blog. 2013;16.
17. Albert R, Jeong H, Barabási A-L. Error and attack tolerance of complex networks. nature.
2000;406(6794):378-82.
18. Anderson RM, May RM, Anderson B. Infectious diseases of humans: dynamics and control:
Wiley Online Library; 1992.
19. Murray JD. Mathematical biology I: an introduction, Vol. 17 of interdisciplinary applied
mathematics. Springer, New York, NY, USA; 2002.
20. Balthrop J, Forrest S, Newman ME, Williamson MM. Technological networks and the spread
of computer viruses. Science. 2004;304(5670):527-9.
21. Newman ME, Forrest S, Balthrop J. Email networks and the spread of computer viruses.
Physical Review E. 2002;66(3):035101.
Sadri, Hasan, Ukkusuri, Suarez Lopez
12
22. Kinney R, Crucitti P, Albert R, Latora V. Modeling cascading failures in the North American
power grid. The European Physical Journal B-Condensed Matter and Complex Systems.
2005;46(1):101-7.
23. Sachtjen M, Carreras B, Lynch V. Disturbances in a power transmission system. Physical
Review E. 2000;61(5):4877.
24. Sornette D. Why stock markets crash: critical events in complex financial systems: Princeton
University Press; 2009.
25. Coleman JS, Katz E, Menzel H. Medical innovation: A diffusion study: Bobbs-Merrill Co;
1966.
26. Cutter SL, Finch C. Temporal and spatial changes in social vulnerability to natural hazards.
Proceedings of the National Academy of Sciences. 2008;105(7):2301-6.
27. Helbing D. Globally networked risks and how to respond. Nature. 2013;497(7447):51-9.
28. Vespignani A. Predicting the behavior of techno-social systems. Science.
2009;325(5939):425-8.
29. Lazer D, Pentland AS, Adamic L, Aral S, Barabasi AL, Brewer D, et al. Life in the network:
the coming age of computational social science. Science (New York, NY).
2009;323(5915):721.
30. Bagrow JP, Wang D, Barabasi A-L. Collective response of human populations to large-scale
emergencies. PloS one. 2011;6(3):e17680.
31. Hughes AL, Palen L. Twitter adoption and use in mass convergence and emergency events.
International Journal of Emergency Management. 2009;6(3-4):248-60.
32. Li J, Rao HR. Twitter as a rapid response news service: An exploration in the context of the
2008 China earthquake. The Electronic Journal of Information Systems in Developing
Countries. 2010;42.
33. Van Hentenryck P, editor Computational Disaster Management. IJCAI; 2013.
34. Wang D, Lin Y-R, Bagrow JP. Social networks in emergency response. Encyclopedia of
Social Network Analysis and Mining: Springer; 2014. p. 1904-14.
35. Watts D, Cebrian M, Elliot M. Dynamics of social media. Public Response to Alerts and
Warnings Using Social Media: Report of a Workshop on Current Knowledge and Research
Gaps. The National Academies Press, Washington, DC; 2013.
36. Caragea C, McNeese N, Jaiswal A, Traylor G, Kim H-W, Mitra P, et al., editors. Classifying
text messages for the haiti earthquake. Proceedings of the 8th international conference on
information systems for crisis response and management (ISCRAM2011); 2011: Citeseer.
37. Earle PS, Bowden DC, Guy M. Twitter earthquake detection: earthquake monitoring in a
social world. Annals of Geophysics. 2012;54(6).
38. Freeman M. Fire, wind and water: Social networks in natural disasters. Journal of Cases on
Information Technology (JCIT). 2011;13(2):69-79.
39. Guy M, Earle P, Ostrum C, Gruchalla K, Horvath S, editors. Integration and dissemination of
citizen reported and seismically derived earthquake information via social network
technologies. International Symposium on Intelligent Data Analysis; 2010: Springer.
40. Pickard G, Pan W, Rahwan I, Cebrian M, Crane R, Madan A, et al. Time-critical social
mobilization. Science. 2011;334(6055):509-12.
41. Sakaki T, Okazaki M, Matsuo Y, editors. Earthquake shakes Twitter users: real-time event
detection by social sensors. Proceedings of the 19th international conference on World wide
web; 2010: ACM.
Sadri, Hasan, Ukkusuri, Suarez Lopez
13
42. Skinner J. Natural disasters and Twitter: Thinking from both sides of the tweet. First
Monday. 2013;18(9).
43. Ukkusuri S, Zhan X, Sadri A, Ye Q. Use of social media data to explore crisis informatics:
Study of 2013 Oklahoma tornado. Transportation Research Record: Journal of the
Transportation Research Board. 2014;(2459):110-8.
44. Abbasi A, Rashidi TH, Maghrebi M, Waller ST, editors. Utilising Location Based Social
Media in Travel Survey Methods: bringing Twitter data into the play. Proceedings of the 8th
ACM SIGSPATIAL International Workshop on Location-Based Social Networks; 2015:
ACM.
45. Maghrebi M, Abbasi A, Rashidi TH, Waller ST, editors. Complementing Travel Diary
Surveys with Twitter Data: Application of Text Mining Techniques on Activity Location,
Type and Time. 2015 IEEE 18th International Conference on Intelligent Transportation
Systems; 2015: IEEE.
46. Hasan S, Ukkusuri SV. Urban activity pattern classification using topic models from online
geo-location data. Transportation Research Part C: Emerging Technologies. 2014;44:363-81.
47. Hasan S, Ukkusuri SV. Location contexts of user check-ins to model urban geo life-style
patterns. PloS one. 2015;10(5):e0124819.
48. Zhao S, Zhang K, editors. Observing Individual Dynamic Choices of Activity Chains From
Location-Based Crowdsourced Data. Transportation Research Board 95th Annual Meeting;
2016.
49. Cebelak MK. Location-based social networking data: doubly-constrained gravity model
origin-destination estimation of the urban travel demand for Austin, TX. 2013.
50. Chen Y, Mahmassani HS, editors. Exploring Activity and Destination Choice Behavior in
Two Metropolitan Areas Using Social Networking Data. Transportation Research Board 95th
Annual Meeting; 2016.
51. Jin P, Cebelak M, Yang F, Zhang J, Walton C, Ran B. Location-Based Social Networking
Data: Exploration into Use of Doubly Constrained Gravity Model for Origin-Destination
Estimation. Transportation Research Record: Journal of the Transportation Research Board.
2014;(2430):72-82.
52. Lee JH, Gao S, Goulias KG, editors. Comparing the Origin-Destination Matrices from Travel
Demand Model and Social Media Data. Transportation Research Board 95th Annual
Meeting; 2016.
53. Yang F, Jin PJ, Wan X, Li R, Ran B, editors. Dynamic origin-destination travel demand
estimation using location based social networking data. Transportation Research Board 93rd
Annual Meeting; 2014.
54. Collins C, Hasan S, Ukkusuri SV. A novel transit rider satisfaction metric: Rider sentiments
measured from online social media data. Journal of Public Transportation. 2013;16(2):2.
55. Carrasco J-A, Miller EJ. The social dimension in action: A multilevel, personal networks
model of social activity frequency between individuals. Transportation Research Part A:
Policy and Practice. 2009;43(1):90-104.
56. Sadri AM, Lee S, Ukkusuri SV. Modeling Social Network Influence on Joint Trip Frequency
for Regular Activity Travel Decisions. Transportation Research Record: Journal of the
Transportation Research Board. 2015;(2495):83-93.
57. Scott J. Social network analysis: Sage; 2012.
58. Alstott J, Bullmore E, Plenz D. powerlaw: a Python package for analysis of heavy-tailed
distributions. PloS one. 2014;9(1):e85777.
Sadri, Hasan, Ukkusuri, Suarez Lopez
14
59. Malevergne Y, Pisarenko V, Sornette D. Gibrat's law for cities: uniformly most powerful
unbiased test of the Pareto against the lognormal. Swiss Finance Institute Research Paper.
2009;(09-40).
60. Malevergne* Y, Pisarenko V, Sornette D. Empirical distributions of stock returns: between
the stretched exponential and the power law? Quantitative Finance. 2005;5(4):379-401.
61. Granovetter MS. The strength of weak ties. American journal of sociology. 1973:1360-80.
62. Miritello G, Moro E, Lara R. Dynamical strength of social ties in information spreading.
Physical Review E. 2011;83(4):045102.
TABLE 1: Description of the Tweets and Network Elements
Description of the Tweets
Number of Total Tweets
Number of Tweets without any User Mentions
Number of Tweets with at least one User Mentions
Number of Tweets only including Self Mentions
Number of Words
Description of Network Elements
Number of Nodes (directed)
Number of Links (directed)
Network Density (directed)
Number of Nodes (undirected)
Number of Links (undirected)
Network Density (undirected)
Number of Nodes (lagest connected component)
Number of Links (lagest connected component)
Network Density (lagest connected component)
Radius (lagest connected component)
Diameter (lagest connected component)
Number of Connected Components
Number of Isolates
Average Degree (directed)
Average Clustering Coefficient (undirected)
56,159
19,532
36,627
20,645
3,589,732
34,363
39,709
0.00003
34,363
38,442
0.00007
21,045
33,020
0.00015
9
17
8,348
6,096
1.156
0.149
Sadri, Hasan, Ukkusuri, Suarez Lopez
15
FIGURE 1: Description of the Tweet Database. (a) Snapshot of 100 most frequent words in the
dataset, (b) Tweets collected over time (Linked tweets versus Non-linked tweets)
Sadri, Hasan, Ukkusuri, Suarez Lopez
16
FIGURE 2: Construction and Visualization of Social Interaction Network. (a) Directed graph, (b) Undirected graph, (c) Undirected
graph visualization, (d) Largest hub, (e) Weighted edges of the largest connected component, (f) Circular tree visualization of the largest
connected component
Sadri, Hasan, Ukkusuri, Suarez Lopez
17
FIGURE 3: Degree Distributions of Purdue Twitter Mention Network. (a) In-degree, Out-degree and Degree Distributions (b)
Degree distributions each day, (c) Comparison of data fitting with different distributions (d) Closer snapshot to power-law, truncated
power-law and log-normal fitting comparisons
Sadri, Hasan, Ukkusuri, Suarez Lopez
18
FIGURE 4: Growth of Social Interaction Network over Time. (a) Cumulative amount of
network elements (nodes and edges), (b) Cumulative network isolates and connected generated
each day, (c) Change in network densities each day due to additional elements.
Sadri, Hasan, Ukkusuri, Suarez Lopez
19
FIGURE 5: Change in Social Interaction Network Properties over Time. (a) Diameter, (b)
Radius, (c) Average Degree, (d) Average Clustering Coefficient, (e) Power-law and Truncated
Power-law Exponents.
Sadri, Hasan, Ukkusuri, Suarez Lopez
20
FIGURE 6: Existence of Common Nodes and Links Each Day as compared to All Previous
Days. (a) Common users, (b) Common links (undirected)
|
1308.6149 | 1 | 1308 | 2013-08-28T13:19:23 | The Extreme Right Filter Bubble | [
"cs.SI",
"cs.CY",
"physics.soc-ph"
] | Due to its status as the most popular video sharing platform, YouTube plays an important role in the online strategy of extreme right groups, where it is often used to host associated content such as music and other propaganda. In this paper, we develop a categorization suitable for the analysis of extreme right channels found on YouTube. By combining this with an NMF-based topic modelling method, we categorize channels originating from links propagated by extreme right Twitter accounts. This method is also used to categorize related channels, which are determined using results returned by YouTube's related video service. We identify the existence of a "filter bubble", whereby users who access an extreme right YouTube video are highly likely to be recommended further extreme right content. | cs.SI | cs | The Extreme Right Filter Bubble
Derek O’Callaghan
University College Dublin
Dublin 4, Ireland
[email protected]
Derek Greene
University College Dublin
[email protected]
Dublin 4, Ireland
Maura Conway
Dublin City University
Dublin 9, Ireland
[email protected]
Joe Carthy
University College Dublin
Dublin 4, Ireland
[email protected]
Pádraig Cunningham
University College Dublin
Dublin 4, Ireland
[email protected]
3
1
0
2
g
u
A
8
2
]
I
S
.
s
c
[
1
v
9
4
1
6
.
8
0
3
1
:
v
i
X
r
a
ABSTRACT
Due to its status as the most popular video sharing platform,
YouTube plays an important role in the online strategy of
extreme right groups, where it is often used to host associ-
ated content such as music and other propaganda. In this
paper, we develop a categorization suitable for the analysis
of extreme right channels found on YouTube. By combining
this with an NMF-based topic modelling method, we catego-
rize channels originating from links propagated by extreme
right Twitter accounts. This method is also used to cate-
gorize related channels, which are determined using results
returned by YouTube’s related video service. We identify
the existence of a “filter bubble”, whereby users who access
an extreme right YouTube video are highly likely to be rec-
ommended further extreme right content.
Keywords
extreme right; YouTube; filter bubble; matrix factorization.
Categories and Subject Descriptors
H.3.3 [Information Search and Retrieval]: Clustering;
H.3.5 [Online Information Services]: Web-based services
INTRODUCTION
1.
Social media platforms have become a key component in the
online strategy of extremist political groups, as they provide
direct access to audiences, particularly youth, for radicaliza-
tion and recruitment purposes [1, 12]. YouTube’s status as
the most popular video sharing platform means that it is es-
pecially useful to these groups. Extreme right content such
as music and other associated propaganda, some of which
is deemed illegal in certain countries, is made freely avail-
able there, often for long periods of time [4]. In addition to
hosting user-generated video content, YouTube also provides
recommendation services, where sets of related and recom-
mended videos are presented to users, based on factors such
as co-visitation count and prior viewing history [13].
In 2011, Pariser proposed the filter bubble concept, which
stated that recommendation services may limit the types of
content to which a user is exposed [27]. Recommendation in
social media can therefore have the undesirable consequence
of a user being excluded from information that is not aligned
with their existing perspective, potentially leading to im-
mersion within an ideological bubble. In this paper we show
that this can happen when a user accesses an extreme right
(ER) video or channel on YouTube. In studies on English
and German ER content, we show that users are likely to be
recommended further ER content within the same category,
or related ER content from a different category. They are
unlikely to be presented with non-ER content (Figure 7).
Our previous work found that Twitter is used by extreme
right groups to propagate links to content hosted on third-
party sites, including YouTube [26]. These links enabled
us to retrieve video and uploader channel information from
YouTube, where subsequent analysis is focused upon chan-
nels (synonymous with uploaders or accounts [29]), rather
than individual videos. We use a topic modelling strategy
to characterize these channels (see Section 4), where content
linked to these channels is then characterized in the same
fashion. Our proposed scheme includes categories identified
in the academic literature on the extreme right, which are
suited to the videos and channels found on YouTube. We
find evidence that an extreme right filter bubble does indeed
exist, as shown by the results discussed in Section 5.
In Section 2, we provide a description of prior work on You-
Tube categorization and recommendation, along with re-
search that involved the categorization of online extreme
right content. The retrieval of YouTube data based on links
originating from extreme right Twitter accounts is then dis-
cussed in Section 3. Next, in Section 4, we describe the
methodology used for related channel ranking, topic iden-
tification, and channel categorization. Our investigation
into the extreme right filter bubble can be found in Sec-
tion 5, where we focus on two data sets consisting primarily
of English and German language channels respectively. An
overview of the entire process can be found in Figure 1.
1
2. RELATED WORK
2.1 Video Recommendation
Video recommendation on YouTube has been the focus of a
number of studies. For example, Baluja et al. suggested that
standard approaches used in text domains were not easily
applicable due to the difficulty of reliable video labelling
[2]. They proposed a graph-based approach that utilised
the viewing patterns of YouTube users, which did not rely
on the analysis of the underlying videos. The recommen-
dation system in use at YouTube at the time was discussed
by Davidson et al. , where sets of personalized videos were
generated with a combination of prior user activity (videos
watched, favorited, liked) and the traversal of a co-visitation
graph [13]. In this process, recommendation diversity was
obtained by means of a limited transitive closure over the
generated related video graph. Zhou et al. performed a mea-
surement study on YouTube videos to determine the sources
responsible for video views, and found that related video
recommendation was the main source outside of the search
function for the majority of videos [33]. They also found
that the click through rate to related videos was high, where
the position of a video in a related video list played a critical
role. A similar finding was made by Figueiredo et al. , where
they demonstrated the importance of key mechanisms such
as related videos in the attraction of users to videos [14].
Turning to the task of YouTube video categorization, Filip-
pova et al. presented a text-based method that relied upon
metadata such as video title, description, tags and com-
ments, in conjunction with a pre-defined set of 75 categories
[15]. Using a bag-of-words model, they found that all of the
text sources contributed to successful category prediction.
More recently, a framework for the categorization of video
channels was proposed by Simonet, involving the use of se-
mantic entities identified within the corresponding video and
channel profile text metadata [29]. Following the judgement
that existing taxonomies were not well-suited to this partic-
ular problem, a new category taxonomy was developed for
YouTube content. Roy et al. investigated both video recom-
mendation and categorization in tandem, where videos were
categorized according to topics built from Twitter activity,
leading to the enrichment of related video recommendation
[28]. Video text metadata was used for this process, and the
topics were based on the categories proposed by Filippova
et al. , in addition to the standard YouTube categories at the
time. Separately, they also analyzed diversity among related
videos, where they found that there was a 25% probability
on average of a related video being from a different category.
2.2 Extreme Right Categorization
In the various studies that have analyzed online extreme
right activity, certain differences can be observed among re-
searchers in relation to their categorization of this activity
and associated organizations [5]. Burris et al. proposed a
set of eight primarily US-centric categories in their analy-
sis of a white supremacist website network; Holocaust Re-
visionists, Christian Identity Theology, Neo-Nazis, White
Supremacists, Foreign (non-US) Nationalists, Racist Skin-
heads, Music, Books/Merchandise [9]. A similar schema
was used by Gerstenfeld et al. , which also included Ku Klux
Klan and Militia categories [16]. They also discussed the dif-
ficulty involved in the categorization of certain sub-groups,
where a general category (Other) was applied in such cases.
These categories were adapted in separate studies of Italian
and German extreme right groups, where new additions in-
cluded Political Parties and Conspiracy Theorists, while oth-
ers such as Music and Skinheads were merged into a Young
category (Tateo [32], Caiani et al. [10]). Rather than focus-
ing on ideological factors, Goodwin proposed four organiza-
tional types found within the European extreme right milieu;
political parties, grassroots social movements, independent
smaller groups, and individual ‘lone wolves’. Other notable
categories include the Autonomous Nationalists identified
within Germany in recent years. These groups focus specif-
ically on attracting a younger audience, where social media
is often a critical component in this process [1].
The popularity of YouTube has led to its usage by extreme
right groups for the purpose of content dissemination. Its
related video recommendation service provides a motivation
for the current work to analyze the extent to which a viewer
may be exposed to such content. Separately, disagreements
over the categorization of online extreme right activity sug-
gests that a specific set of categories may be required for the
analysis of this domain.
3. DATA
In our previous work, we investigated the potential for Twit-
ter to act as one possible gateway to communities within
the wider online network of the extreme right [26]. Two
data sets associated with extreme right English language
and German language Twitter accounts were generated, by
retrieving profile data over an extended period of time. We
gathered all tweeted links to external websites, and used
these to construct an extended network representation. In
the current work, we are solely interested in tweeted You-
Tube URIs. Data for the Twitter accounts were retrieved
between June 2012 and May 2013, as limited by the Twit-
ter API restrictions effective at the time. YouTube URIs
found in tweets were analyzed to determine a set of chan-
nel (account) identifiers that were directly (channel profile
page URI) or indirectly (URI of video uploaded by channel)
tweeted. All identified channels were included, regardless of
the number of tweets in which they featured. Throughout
this work, we refer to these as seed channels; 26,460 and
3,046 were identified for the English and German data sets
respectively.
In order to explore the filter bubble hypothesis we must
first categorize YouTube channels. To do this, we use text
metadata associated with the videos uploaded by a particu-
lar channel, namely their titles, descriptions and associated
keywords. Although user comments have been employed
in other work [15], they were excluded here. This decision
followed an initial manual analysis of a sample of tweeted
videos, which found that comments were often not present,
or had been explicitly disabled by the uploader. We also
excluded the YouTube “category” field as it was considered
too broad to be useful in the extreme right domain. Using
the YouTube API, we initially retrieved the available text
metadata for up to 1,000 of the videos uploaded by each seed
channel, where the API returns videos in reverse chronologi-
cal order according to their upload time. In cases where seed
channels and their videos were no longer available (e.g. the
channel had been suspended or deleted since appearing in a
tweet), these channels were simply ignored.
2
Figure 1: Overview of the process used to investigate the extreme right filter bubble on YouTube.
In order to address the variance in the number of uploaded
videos per channel, and to reduce the volume of subsequent
data retrieval, we randomly sampled up to 50 videos for each
seed. For each video in this sample, metadata values were
retrieved for the top ten related videos returned by the API.
Using the default parameter settings, these videos appear to
be returned in order of “relevance”, as defined internally by
YouTube, similar to the default behaviour of video search re-
sults feeds described in the API documentation (June 2013).
We refer to the corresponding uploaders as related channels;
1,451,189 and 195,146 were identified for the English and
German data sets respectively. As before, we then retrieved
the available text metadata for up to 1,000 videos uploaded
by each unseen related channel, from which a random sam-
ple of up to 50 videos was generated.
For both seed and related channels, the corresponding up-
loaded video sample is used for categorization; this is de-
scribed in Section 4.2. Separately, we also note that the
videos in question may have been uploaded at any time prior
to retrieval, where these times are not necessarily restricted
to the period of either Twitter or YouTube data retrieval.
4. METHODOLOGY
Having retrieved the channel and video data from YouTube,
the next steps involved ranking the related channels for each
seed channel, identifying latent topics associated with the
uploaded videos, and using these to categorize both seed
and related channels. This section corresponds to part (b)
of the process overview diagram found in Figure 1.
4.1 Related Channel Ranking
For our analysis, it was first necessary to generate a set of
related channel rankings for each seed channel, according
to the related rankings returned by the YouTube API for
the sample of videos uploaded by the seed. We applied the
SVD rank aggregation method proposed by Greene et al. to
combine the rankings for each uploaded video into a single
ranked set across all videos for the seed in question, from
which we select the top ranked related channels [19]. We
restrict our focus to the top 10, given the impact of related
video position on click through rate [33]. The aggregation
process for each seed channel si and its ni uploaded videos
is as follows:
1. For each uploaded video vij (j ∈ [1, ni]), generate
a rank vector rvij between the seed si and its re-
lated channels using the retrieved YouTube API re-
lated video ranking. For channels related to seed videos
other than vij, a rank of (mi + 1) is applied, where mi
is the total number of channels related to si.
2. Stack all rvi rank vectors as columns, to form the mi×
ni rank matrix Ri, and normalise the columns of this
matrix to unit length.
3. Compute the SVD of Ri, and extract the first left sin-
gular vector. Arrange the entries in this vector in de-
scending order to produce a ranking of all channels
related to this seed.
From the ranking generated above, we then select the top 10
ranked related channels for each seed. These channels are
used for our filter bubble investigation in Section 5.
4.2 Topic Identification
For the purpose of channel categorization, we initially iden-
tified latent topics associated with the channels in both data
sets, based on their uploaded videos. Following the ap-
proach of Hannon et al. [20], we generated a “profile doc-
ument” for each seed and related channel, consisting of an
aggregation of the text metadata from their corresponding
uploaded video sample, from which a tokenized representa-
tion was produced. All available terms in the channel doc-
uments were used, where the tokenization process involved
the exclusion of URIs and normalization of diacritics. We
excluded a custom set of stopwords, such as official YouTube
category terms, additional YouTube-specific terms such as
“video” and “view”, and terms from multiple language stop-
word lists. Due to the frequent presence of names in the
data, we also excluded all identified first names. As sug-
gested by Fillipova et al. [15], terms were not stemmed due
to the expectation of both noisy terms and those of multiple
languages. Low-frequency terms appearing in < 20 channel
documents were excluded at this point.
These channel documents were transformed to log-based
TF-IDF vectors, and subsequently normalized to unit length.
However, in an attempt to reduce the term dimensionality
found in the combination of seed and related channel docu-
ments, we generated the TF-IDF vectors in two stages. The
first stage constructed vectors for the seed documents, from
which a reduced seed term vocabulary was derived. In the
second stage, vectors for the related documents were then
3
Related Channel RankingTopic IdentificationChannel Categorization(b) Methodology(c) Experiments Filter BubbleInvestigation(a) DataYouTubeData Retrieval Table 1: Categories of extreme right YouTube content, based on common categorizations found in academic literature on
extreme right ideology.
Category
Anti-Islam
Anti-Semitic
Description
Can include political parties (e.g. Dutch PVV ) or groups such as the English
Defence League (EDL), which often describe themselves as “counter-Jihad”.
All types of anti-Semitism, regardless of association (existing literature tends
to discuss this in relation to other categories).
Source
[1, 17]
[9, 32]
Conspiracy Theory Themes include New World Order (NWO), Illuminati etc. Not exclusively
[1, 30, 32]
Music
Neo-Nazi
Patriot
Political Party
Populist
Revisionist
Street Movement
White Nationalist
ER, but often related to Patriot in this context.
Includes any ER music such as Oi!, Rock Against Communism (RAC) etc.
Nazi references, such as to Hitler, WWII, SS etc.
US-centric, including groups such as “Birthers”, militia, anti-government,
anti-immigration, opposition to financial system. Some of these themes are
not exclusive to ER.
Primarily European parties such as the BNP, FP O, Jobbik, NPD, PVV,
Swedish Democrats, UKIP etc. Many of these parties are also categorized
as Populist.
Broader category that includes various themes such as anti-EU, anti-
establishment, anti-state/government, anti-immigration (as with Patriot,
some of these are not exclusive to ER). Although some disagreement about
this category exists [23], we have used it as it has proved convenient for
categorizing certain groups that span multiple themes.
References to Holocaust/WWII denial. Closely associated with Neo-Nazi.
Groups such as the EDL, Autonome Nationalisten, Spreelichter, Anti-Antifa
etc.
References to white nationalism and supremacism, also used to characterize
political parties such as the BNP or Jobbik.
[1, 9, 32]
[1, 9, 16]
[30]
[1, 10, 18]
[3, 24]
[9, 16, 32]
[1, 18]
[9, 16, 32]
constructed using the seed vocabulary. In both stages, short
documents containing < 10 terms were excluded. We use ms
and mr to refer to the number of seed and related channel
document vectors in a particular data set.
2. NMF is applied to Vs to produce two factors; Ws, an
n × T matrix containing topic basis vectors and Hs,
a T × ms matrix containing the topic assignments or
weights for each seed channel document.
Topic modelling is concerned with the discovery of latent
semantic structure or topics within a text corpus, which
can be derived from co-occurrences of words and documents
[31]. Popular methods include probabilistic models such as
latent Dirichlet allocation (LDA) [6], or matrix factoriza-
tion techniques such as Non-negative matrix factorization
(NMF) [21]. We initially evaluated both LDA and NMF-
based methods with the seed channel document representa-
tions described above (the former was applied to TF vec-
tors). However, as in our previous work, NMF was found
to produce the most readily-interpretable results, which ap-
peared to be due to the tendency of LDA to discover topics
that over-generalized [11, 25]. We were aware of the pres-
ence of smaller groups of channels associated with multiple
languages in both data sets, and opted for specificity rather
than generality by applying NMF to the TF-IDF channel
vectors. As we found previously, the IDF component en-
sured a lower ranking for less discriminating terms, thus
leading to the discovery of more specific topics.
To allow the seed channel vectors to determine the topic ba-
sis vectors, we undertook the process in two stages. Firstly,
we generate topics for the seeds:
1. Construct an n× ms term-document matrix Vs, where
each column contains a seed channel TF-IDF vector.
For both data sets, this results in a set of basis vectors con-
sisting of both ER and non-ER topics. The second stage
involves producing topic assignments for the related chan-
nel vectors:
1. Construct an n × mr term-document matrix Vr con-
taining the related channel TF-IDF vectors.
2. Generate a corresponding T × mr topic weights matrix
Hr by transforming Vr according to the Ws model.
The latter step was achieved with a single outer iteration
of the H non-negative least squares sub-problem from the
approach by Lin et al. [22]. As the basis vectors matrix Ws
was fixed, only one outer iteration was necessary in order to
generate an approximation of a converged Hr, thus permit-
ting large related channel matrices to be factorized given the
initial NMF operation on a far smaller seed matrix. In both
cases, to address the instability introduced by random ini-
tialization in standard NMF, we employed the deterministic
NNDSVD initialization method [7].
4.3 Topic and Channel Categorization
As discussed earlier in Section 2.1, some prior work has pro-
posed generic categories for use with YouTube videos and
channels. However, as these studies have focused on the
4
categorization of mainstream videos, they are not sufficient
for the present analysis where categories specifically associ-
ated with the extreme right are required. In Section 2.2, we
discuss prior work that characterized online extreme right
activity using a number of proposed categories, but as indi-
cated earlier, no definitive set of categories is agreed upon
in this domain [5]. Therefore, we propose a categorization
based on various schema found in a selection of academic lit-
erature on the extreme right, where this category selection
is particularly suited to the ER videos and channels we have
found on YouTube. Some categories are clearly delineated
while others are less distinct, reflecting the complicated ide-
ological make-up and thus fragmented nature of groups and
sub-groups within the extreme right [16].
In such cases,
we have proposed categories that are as specific as possible
while also accommodating a number of disparate themes and
groups. Details of the categories employed can be found in
Table 1.
As both data sets contained various non-ER channels, we
also created a corresponding set of non-ER categories con-
sisting of a selection of the general YouTube categories as of
June 2013, in addition to other categories that we deemed
appropriate following an inspection of these channels and
associated topics. These non-ER categories were:
• Entertainment
• Gaming
• Military
• Music
• News & Current Af-
fairs
• Politics
• Religion
• Science & Education
• Sport
• Television
Having produced a set of T topics for a data set, we then
proceeded to categorize them. For each topic, we manually
inspected the high-ranking topic terms from the correspond-
ing vector in Ws, in addition to profiles and uploaded videos
for a selection of seed channels most closely assigned to the
topic, according to their weights in Hs. Multiple categories
were assigned to topics where necessary, as using a single
category per topic would have been too restrictive while also
not reflecting the often multi-faceted nature of most topics
that were identified.
In many cases, categories for topics
were clearly identifiable, with a separation between ER and
non-ER categories. For example, an English Defence League
(EDL) topic was categorized as Anti-Islam and Street Move-
ment, while a topic having high-ranking terms such as “gui-
tar” and “band” was categorized as Music. For certain topics,
this separation was more ambiguous, where the channels as-
signed to a particular topic consisted of a mixture of both
ER and non-ER channels. A combination of both ER and
non-ER categories were assigned in such cases.
The set of categorized topics was then used to label both
seed and related channels, using the channel topic assign-
ment weights found in the H matrices. This supports the
potential assignment of multiple categories to a single chan-
nel. To achieve this, it was first necessary to determine a
weight threshold w for selecting discriminating topics to be
used in the categorization of a particular channel. Using a
(a) Categories assigned to topics.
(b) Channel H weights for topics.
(c) Channels categorized using topics where H ≥ w (in this
example, w = 0.05).
Figure 2: Topic and Channel categorization process.
range of values for w, we calculated µ, the mean number
of topics per channel with H weight ≥ w, and selected the
value for w where µ =∼ 1; the minimum number of topics
required for categorization. As we performed NMF in two
stages to produce Hs and Hr for seed and related channels
respectively, two corresponding ws and wr thresholds were
generated due to the dependency of the H weights’ upper
bound on the input matrices V . Using these thresholds, each
channel was categorized as follows:
1. Select all topics where the channel’s corresponding row
weights in H are ≥ w.
2. Add each selected topic weight to the totals maintained
for the topic’s corresponding categories.
3. Rank these categories in descending order based on
5
hitler, adolf, reichNeo-Nazicover, guitar, acousticRevisionistMusicCategoriesTopicsTopicsChannelshitler, adolf, reichcover, guitar, acousticChannel 1Channel 20.090.030.080.01Neo-NaziRevisionistMusicCategoriesChannelsChannel 1Channel 2 their totals across all selected topics, and assign the
highest ranked category to the channel. Multiple cat-
egories are assigned in the event of a tie.
A simple example illustrating the complete process of map-
ping of channels to categories is shown in Figure 2. We
found that a certain number of channels, both seed and re-
lated, had a flat profile with relatively low weights (< w) for
all topics; these could be considered as grey sheep. Further
analysis found this to be due to factors such as the original
documents being short or containing few unique terms. As
we were unable to reliably categorize such channels, they
were excluded from all subsequent analysis. Separately, al-
though the NMF process often identified topics with high-
ranking discriminating terms in languages other than that
of the data set, a small number of topics with less discrimi-
nating general language terms were also found. As it would
have been difficult to distinguish between ER and non-ER
channels closely associated with these topics, these were also
excluded. Further details on the w thresholds and numbers
of excluded channels are provided in Section 5.
5. THE ER FILTER BUBBLE
In this section, we discuss the filter bubble analysis of the
English and German language YouTube data sets. The ex-
perimental methodology is as follows:
1. Generate an aggregated ranking of related channels for
each seed channel.
2. Generate TF-IDF channel document vectors, and iden-
tify topics using NMF.
3. Categorize the identified topics according to the set
defined in Table 1.
4. Categorize the channels based on their topic weights
in H.
5. Investigate whether an extreme right filter bubble is
present.
For Step 5 above, we define an extreme right filter bubble in
terms of the extent to which the related channels for a par-
ticular extreme right seed channel also feature extreme right
content. It has been shown that the position of a video in a
related video list plays a critical role in the click through rate
[33]. Therefore, we investigate the presence of a filter bub-
ble using the top k ranked related channels, with increasing
values of k ∈ [1, 3, 5, 7, 10], as follows.
For each ER seed channel:
1. Select the top k ranked related channels, filtering any
excluded channels as defined in Section 4.3. Seed chan-
nels with no remaining related channels following fil-
tering are not considered for rank k.
2. Calculate the total proportion of each category as-
signed to the ≤ k related channels.
Then, for each ER category:
1. Select all seed channels to which the category has been
assigned.
2. Calculate the mean proportion of each related category
associated with these seed channels.
We consider a filter bubble to exist for a particular ER cat-
egory when its highest ranking related categories, in terms
of their mean proportions, are also ER categories.
5.1 English language categories
From the total number of channels in the English language
data set, we generated 24,611 seed and 1,376,924 related
channel documents, using a corresponding seed-based vo-
cabulary of 39,492 terms. The TF-IDF matrices Vs and
Vr were then generated for seed and related documents re-
spectively, and topics were identified by applying NMF to
Vs. To determine the number of topics T , we experimented
with values of T in [10, 100] to produce topics that were
as specific as possible, given prior knowledge of the pres-
ence of smaller groups of channels associated with multiple
languages within both data sets. This led to the selection
of T = 80, as larger values resulted in topic splits rather
than the emergence of unseen topics. Of these, 27 ER top-
ics (33.75%), 39 non-ER topics (48.75%), 8 topics that were
a combination of ER and non-ER categories (10%), and 6
topics based on general terms of a separate language (7.5%)
were found.
Figure 3: English mean topics per channel at ws, wr in
[0.01, 0.09]. ws and wr were set to 0.03 and 0.08 respectively.
These 80 topics were then categorized according to the set
defined in Table 1, which permitted the subsequent catego-
rization of the seed and related channels using their corre-
sponding Hs and Hr topic weights. As described in Section
4.3, we first generated the weights thresholds ws and wr.
Figure 3 contains the mean number of topics per channel
µ with H weight ≥ w, for values of w in [0.01, 0.09]. Us-
ing µ =∼ 1 as the selection criterion resulted in values of
0.03 and 0.08 for ws and wr respectively. Using these, we
excluded 8,225 (33.42%) seed and 482,226 (35.66%) related
grey sheep channels that could not be categorized. Related
channels ranked at k > 10 were also excluded, and all non-
ER seed channels were removed from the candidate seed
set. The remaining 6,573 ER seed and their 22,980 related
6
●●●●●●●●●●●●●●●●●●02460.010.020.030.040.050.060.070.080.09wMean topics●●SeedRelatedTopics=⇠1 ings with the introduction of certain non-ER categories, ER
categories consistently have the strongest presence. In the
case of Anti-Islam seed channels, it appears that the top
ranked related channels (k = 1) are mostly affiliated with
various groups from the UK and USA. Street Movement re-
lated channels at this rank are associated with the English
Defence League (EDL), a movement opposed to the alleged
spread of radical Islamism within the UK [18]. Channels
from various international individuals and groups that often
describe themselves as “counter-Jihad” can also be observed
[17]. The Conspiracy Theory and non-ER Religion cate-
gories appear to be associated with channels based in the
USA, where the dominance of these categories at lower rank-
ings (excluding the seed category) suggests that the channels
become progressively more US-centric.
The ER Music seed channels usually upload video and audio
recordings of high-profile groups associated with the extreme
right. For example, content from bands such as Skrewdriver
(UK) or Landser (Germany) can be found, along with other
groups from genres such as Oi!, RAC (Rock Against Com-
munism) and NSBM (National Socialist Black Metal) [1, 8].
Given this, the consistent presence of the White Nationalist
related category would appear logical. At the same time, we
also observe that non-ER Music becomes more evident as k
increases, perhaps reflecting the overlap between music gen-
res. For example, someone who is a fan of NSBM is often
a fan of other metal music that would not be categorized
as ER. For Populist seed categories, the related categories
generally appear more diverse. Channels affiliated with po-
litical parties can be observed, including the Eurosceptic
United Kingdom Independence Party (UKIP) or the British
National Party (BNP), where the latter is also considered
as White Nationalist [3, 24]. Opposition to establishment
organizations such as the EU may be a link to similar op-
position within the Patriot and Conspiracy Theory related
categories that are also present [30], while also explaining the
presence of the non-ER News & Current Affairs category. It
should also be mentioned that our definition of Populist is
broad and spans multiple themes (Table 1), where a certain
amount of disagreement about this category exists [23].
5.2 German language categories
A total of 2,766 seed and 177,868 related channel documents
were generated from the German language data set. Topics
were then identified by applying NMF to the corresponding
TF-IDF matrices. As before, we experimented with values
of T in [10, 100], and selected T = 60 given a similar ob-
servation of redundant topic splits for larger values of T .
Of these, 33 ER topics (55%), 20 non-ER topics (33.33%), 2
topics that were a combination of ER and non-ER categories
(3.33%), and 5 topics based on general terms of a separate
language (8.33%) were found. Having categorized these top-
ics, we calculated the mean number of topics per channel µ
with H weight ≥ w, for values of w in [0.01, 0.09], as seen in
Figure 5. Values of 0.06 and 0.087 were found for ws and wr
respectively at µ =∼ 1. We excluded 785 (28.38%) seed and
56,565 (31.8%) related grey sheep channels that could not be
categorized, in addition to related channels ranked at k > 10
and non-ER seed channels. The remaining 1,123 ER seed
and 4,973 related (ER and non-ER, k ≤ 10) channels were
used to calculate the mean related category proportions for
each ER category.
(a) Anti-Islam
(b) Music
(c) Populist
Figure 4: English mean category proportions of top k ranked
related channels (k ∈ [1, 3, 5, 7, 10]), for selected seed ER
categories.
channels were used to calculate the mean related category
proportions for each ER category.
Figure 4 contains plots for three ER categories that we have
selected for more detailed analysis, where ER and non-ER
categories have been prefixed with ER- and NER- respec-
tively. To assist visual interpretation, we have omitted any
weakly related categories whose mean proportion < 0.06 for
a particular k ranking. From inspecting these plots, two
initial observations can be made; (1) the seed category is
the dominant related category for all values of k, and (2) al-
though related category diversity increases at lower k rank-
7
0.00.20.40.6135710kMean proportionER−Anti−IslamER−Conspiracy TheoryER−Street MovementNER−EntertainmentNER−Religion0.00.20.40.6135710kMean proportionER−MusicER−White NationalistNER−EntertainmentNER−Music0.00.20.40.6135710kMean proportionER−Conspiracy TheoryER−PatriotER−Political PartyER−PopulistER−White NationalistNER−EntertainmentNER−News & Current Affairs Figure 6 contains plots for three ER categories that we have
selected for in-depth analysis. As seen in Figure 4, the seed
category is the dominant related category for all values of k,
and the ER related category presence is consistently stronger
than that of the non-ER categories, notwithstanding the in-
crease in diversity. The Populist and Political Party related
categories are prominent for Anti-Islam, given the inclu-
sion of channels affiliated with parties such as the National
Democratic Party of Germany (NPD), the Pro-Bewegung
collective, and the Freedom Party of Austria (FP O); all
strong opponents of immigration, particularly by Muslims
[1, 3, 17].
Figure 5: German mean topics per channel at ws, wr in
[0.01, 0.09]. ws and wr were set to 0.06 and 0.087 respec-
tively.
This data set also features many ER Music seed channels
that upload recordings of high-profile groups, with the main
difference being the prominence of the Neo-Nazi related cat-
egory for all rankings. These recordings and videos, along
with other non-music videos uploaded by these channels,
often feature recognizable Nazi imagery. Separately, chan-
nels that upload videos associated with groups that have
alleged ER ties, for example, Bohse Onkelz or Frei.Wild [1],
may explain in part the presence of the non-ER Music cat-
egory, given the mainstream success of these groups. This
may also be explained by material associated with hip-hop
groups such as “n’Socialist Soundsystem”, which provide an
alternative to traditional ER music based on rock and folk.
The close relationship with Music is also present for the
Neo-Nazi seed category, although further related diversity
can be observed. Seed channels featuring footage of German
participation in WWII, including speeches by high-ranking
members of the Nazi party, are likely to be the source of the
White Nationalist and Revisionist related categories. We
can safely assume that the prominence of Music is responsi-
ble for the appearance of its non-ER counterpart here.
5.3 Discussion
We conclude our filter bubble analysis by measuring the
mean proportions for the seed ER categories as a whole,
where the possible aggregated related categories were (1)
the same ER category as that of the seed, (2) a different
ER category, or (3) a non-ER category. The results for both
data sets can be found in Figure 7. As with the individual
seed categories, an ER filter bubble is also clearly identi-
fiable at the aggregate level. Although the increase in di-
versity for lower k rankings introduces a certain proportion
(a) Anti-Islam
(b) Music
(c) Neo-Nazi
Figure 6: German mean category proportions of top k
ranked related channels (k ∈ [1, 3, 5, 7, 10]), for selected seed
ER categories.
of non-ER categories, this is always outweighed by ER cate-
gories, where the seed ER category remains dominant for all
values of k. These findings would appear to contrast those
of certain prior work where greater related video diversity
was observed [28, 33]. Although we have analyzed related
channels rather than individual videos, we might have ex-
pected to also find this behaviour at both levels. However,
it would appear this is not always true, at the very least in
the case of ER channels.
8
●●●●●●●●●●●●●●●●●●2460.010.020.030.040.050.060.070.080.09wMean topics●●SeedRelatedTopics=⇠10.00.20.40.6135710kMean proportionER−Anti−IslamER−Political PartyER−PopulistNER−EntertainmentNER−News & Current Affairs0.00.20.40.6135710kMean proportionER−MusicER−Neo−NaziER−White NationalistNER−Music0.00.20.40.6135710kMean proportionER−MusicER−Neo−NaziER−PopulistER−RevisionistER−White NationalistNER−Music (a) English
(b) German
Figure 7: Aggregated mean category proportions of top k ranked related channels (k ∈ [1, 3, 5, 7, 10]), for seed ER categories.
The data retrieval process described in Section 3 involved
following related video links for only one step removed from
the corresponding seed video; no further related retrieval
was performed for related videos themselves. Of the ER
seed channels used in the filter bubble analysis in Sections
5.1 and 5.2, we identified those that appeared in top k re-
lated rankings of other ER seed channels, for k ≤ 10; 6,186
(94%) and 1,056 (94%) such channels respectively for the
English and German language data sets. These high per-
centages allude to the presence of cycles within the related
channel graph, where retrieving additional data by following
related videos for multiple steps may have been somewhat
redundant. They also further emphasize the existence of the
filter bubble.
Separately, it might be argued that our findings merely con-
firm that YouTube’s related video recommendation process
is working correctly. However, this is precisely what we were
trying to quantify, in terms of the categories of extreme right
content to which a user may be exposed following a short
series of clicks. As a consequence, users viewing this content
may rarely be presented with an alternative perspective.
6. CONCLUSIONS AND FUTURE WORK
YouTube’s position as the most popular video sharing plat-
form has resulted in it playing an important role in the online
strategy of the extreme right, where it is used to host as-
sociated content such as music and other propaganda. We
have proposed a set of categories that may be applied to
this YouTube content, based on a review of those found in
existing academic studies of the extreme right’s ideological
make-up. Using an NMF-based topic modelling approach,
we have categorized channels according to this proposed set,
permitting the assignment of multiple categories per chan-
nel where necessary. This categorization has helped us to
identify the existence of an extreme right filter bubble, in
terms of the extent to which related channels, determined
by the videos recommended by YouTube, also belong to ex-
treme right categories. Despite the increased diversity ob-
served for lower related rankings, this filter bubble maintains
a constant presence. The influence of related rankings on
click through rate [33], coupled with the fact that the You-
Tube channels in this analysis originated from links posted
by extreme right Twitter accounts, would suggest that it is
possible for a user to be immersed in this content following
a short series of clicks.
In future work, we would like to experiment with alternative
channel representations that address the existence of noise
found with certain channels and associated topics at lower
related rankings. We would also like to perform a longer
temporal analysis in an attempt to investigate any correla-
tion between changes in this filter bubble and extreme right
activity across multiple social media platforms. Separately,
although we have used the methodology presented in this
paper to study extreme right channels, it may be interest-
ing to investigate its application to other types of extremist
political content found on YouTube.
7. ACKNOWLEDGMENTS
This research was supported by 2CENTRE, the EU funded
Cybercrime Centres of Excellence Network and Science Foun-
dation Ireland (SFI) under Grant Number SFI/12/RC/2289.
8. REFERENCES
[1] Baldauf, J., Gross, A., Rafael, S., and Wolf, J.
Zwischen Propaganda und Mimikry.
Neonazi-Strategien in Sozialen Netzwerken. Amadeu
Antonio Stiftung (2011).
[2] Baluja, S., Seth, R., Sivakumar, D., Jing, Y., Yagnik,
J., Kumar, S., Ravichandran, D., and Aly, M. Video
suggestion and discovery for youtube: taking random
walks through the view graph. In Proc. 17th
international conference on World Wide Web, WWW
’08, ACM (2008), 895–904.
[3] Bartlett, J., Birdwell, J., and Littler, M. The New
Face of Digital Populism. Demos (2011).
[4] Bell, M. Frei verfugbarer Nazi-Rock: YouTube und die
braunen Musikanten. Der Spiegel (03 2013).
[5] Blee, K. M., and Creasap, K. A. Conservative and
Right-Wing Movements. Annual Review of Sociology
36 (2010), 269–286.
[6] Blei, D. M., Ng, A. Y., and Jordan, M. I. Latent
dirichlet allocation. J. Mach. Learn. Res. 3 (Mar.
2003), 993–1022.
9
0.000.250.500.751.00135710kMean proportionSame ER CategoryOther ER CategoryNon−ER Category0.000.250.500.751.00135710kMean proportionSame ER CategoryOther ER CategoryNon−ER Category [7] Boutsidis, C., and Gallopoulos, E. SVD based
Conference, WebSci ’13, ACM (2013), 118–121.
initialization: A head start for nonnegative matrix
factorization. Pattern Recogn. 41, 4 (Apr. 2008),
1350–1362.
[8] Brown, T. S. Subcultures, Pop Music and Politics:
Skinheads and ”Nazi Rock” in England and Germany.
Journal of Social History 38, 1 (2004), 157–178.
[9] Burris, V., Smith, E., and Strahm, A. White
Supremacist Networks on the Internet. Sociological
Focus 33, 2 (2000), 215–235.
[10] Caiani, M., and Wagemann, C. Online Networks of
the Italian and German Extreme Right. Information,
Communication & Society 12, 1 (2009), 66–109.
[11] Chemudugunta, C., Smyth, P., and Steyvers, M.
Modeling General and Specific Aspects of Documents
with a Probabilistic Topic Model. In Advances in
Neural Information Processing Systems (2006),
241–248.
[12] Conway, M., and McInerney, L. Jihadi Video and
Auto-radicalisation: Evidence from an Exploratory
YouTube Study. In Proc. 1st European Conference on
Intelligence and Security Informatics, EuroISI ’08,
Springer-Verlag (2008), 108–118.
[13] Davidson, J., Liebald, B., Liu, J., Nandy, P.,
Van Vleet, T., Gargi, U., Gupta, S., He, Y., Lambert,
M., Livingston, B., and Sampath, D. The YouTube
video recommendation system. In Proc. 4th ACM
conference on Recommender Systems, RecSys ’10
(2010), 293–296.
[14] Figueiredo, F., Benevenuto, F., and Almeida, J. M.
The Tube over Time: Characterizing Popularity
Growth of YouTube Videos. In Proc. 4th ACM
international conference on Web search and data
mining, WSDM ’11 (2011), 745–754.
[15] Filippova, K., and Hall, K. B. Improved video
categorization from text metadata and user comments.
In Proc. 34th Int. ACM SIGIR Conference on
Research and development in Information Retrieval,
SIGIR ’11 (2011), 835–842.
[16] Gerstenfeld, P. B., Grant, D. R., and Chiang, C.-P.
Hate Online: A Content Analysis of Extremist
Internet Sites. Analyses of Social Issues and Public
Policy 3, 1 (2003), 29–44.
[17] Goodwin, M. The Roots of Extremism: The English
Defence League and the Counter-Jihad Challenge.
Chatham House (2013).
[18] Goodwin, M., and Ramalingam, V. The New Radical
Right: Violent and Non-Violent Movements in
Europe. Institute for Strategic Dialogue (2012).
[19] Greene, D., and Cunningham, P. Producing a Unified
Graph Representation from Multiple Social Network
Views. In Proc. 5th Annual ACM Web Science
[20] Hannon, J., Bennett, M., and Smyth, B.
Recommending Twitter users to follow using content
and collaborative filtering approaches. In Proc. 4th
ACM Conference on Recommender Systems
(RecSys’10) (2010), 199–206.
[21] Lee, D. D., and Seung, H. S. Learning the parts of
objects by non-negative matrix factorization. Nature
401 (1999), 788–791.
[22] Lin, C.-J. Projected Gradient Methods for
Nonnegative Matrix Factorization. Neural Comput.
19, 10 (Oct. 2007), 2756–2779.
[23] Marli`ere, P. Populism and the enchanted world of
’moderate politics’. openDemocracy (2013).
[24] Mudde, C. Populist Radical Right Parties in Europe.
Cambridge University Press, 2007.
[25] O’Callaghan, D., Greene, D., Conway, M., Carthy, J.,
and Cunningham, P. An Analysis of Interactions
Within and Between Extreme Right Communities in
Social Media. In To Appear: MSM 2012 and MUSE
2012 Postproceedings (2013).
[26] O’Callaghan, D., Greene, D., Conway, M., Carthy, J.,
and Cunningham, P. Uncovering the Wider Structure
of Extreme Right Communities Spanning Popular
Online Networks. In Proc. 5th Annual ACM Web
Science Conference, WebSci ’13 (2013), 276–285.
[27] Pariser, E. The Filter Bubble: What the Internet Is
Hiding from You. Penguin Group USA, 2011.
[28] Roy, S. D., Mei, T., Zeng, W., and Li, S.
SocialTransfer: cross-domain transfer learning from
social streams for media applications. In Proc. 20th
ACM international conference on Multimedia, MM
’12, ACM (2012), 649–658.
[29] Simonet, V. Classifying YouTube Channels: a
Practical System. In Proc. 2nd International
Workshop on Web of Linked Entities (WoLE 2013),
ACM (2013), 1295–1304.
[30] SPLC. Ideology. Southern Poverty Law Center (2013).
[31] Steyvers, M., and Griffiths, T. Probabilistic Topic
Models. In Latent Semantic Analysis: A Road to
Meaning., T. Landauer, D. Mcnamara, S. Dennis, and
W. Kintsch, Eds. Laurence Erlbaum, 2006.
[32] Tateo, L. The Italian Extreme Right On-line Network:
An Exploratory Study Using an Integrated Social
Network Analysis and Content Analysis Approach.
Journal of Computer-Mediated Communication 10, 2
(2005).
[33] Zhou, R., Khemmarat, S., and Gao, L. The impact of
YouTube recommendation system on video views. In
Proc. 10th ACM SIGCOMM conference on Internet
measurement (2010), 404–410.
10
|
1312.6552 | 1 | 1312 | 2013-12-23T14:10:54 | Socially-Aware Networking: A Survey | [
"cs.SI",
"cs.NI",
"physics.soc-ph"
] | The widespread proliferation of handheld devices enables mobile carriers to be connected at anytime and anywhere. Meanwhile, the mobility patterns of mobile devices strongly depend on the users' movements, which are closely related to their social relationships and behaviors. Consequently, today's mobile networks are becoming increasingly human centric. This leads to the emergence of a new field which we call socially-aware networking (SAN). One of the major features of SAN is that social awareness becomes indispensable information for the design of networking solutions. This emerging paradigm is applicable to various types of networks (e.g. opportunistic networks, mobile social networks, delay tolerant networks, ad hoc networks, etc) where the users have social relationships and interactions. By exploiting social properties of nodes, SAN can provide better networking support to innovative applications and services. In addition, it facilitates the convergence of human society and cyber physical systems. In this paper, for the first time, to the best of our knowledge, we present a survey of this emerging field. Basic concepts of SAN are introduced. We intend to generalize the widely-used social properties in this regard. The state-of-the-art research on SAN is reviewed with focus on three aspects: routing and forwarding, incentive mechanisms and data dissemination. Some important open issues with respect to mobile social sensing and learning, privacy, node selfishness and scalability are discussed. | cs.SI | cs | Socially-Aware Networking: A Survey
Feng Xia, Senior Member, IEEE, Li Liu, Jie Li, Jianhua Ma, and Athanasios V. Vasilakos, Senior
Member, IEEE
1
Abstract—The widespread proliferation of handheld devices
enables mobile carriers to be connected at anytime and
anywhere. Meanwhile, the mobility patterns of mobile devices
strongly depend on the users’ movements, which are closely
related
to
their
social relationships and behaviors.
Consequently,
today’s mobile networks are becoming
increasingly human centric. This leads to the emergence of a
new field which we call socially-aware networking (SAN). One
of the major features of SAN is that social awareness becomes
indispensable
information for the design of networking
solutions. This emerging paradigm is applicable to various
types of networks (e.g. opportunistic networks, mobile social
networks, delay tolerant networks, ad hoc networks, etc) where
the users have social relationships and interactions. By
exploiting social properties of nodes, SAN can provide better
networking support to innovative applications and services. In
addition, it facilitates the convergence of human society and
cyber physical systems. In this paper, for the first time, to the
best of our knowledge, we present a survey of this emerging
field. Basic concepts of SAN are introduced. We intend to
generalize the widely-used social properties in this regard. The
state-of-the-art research on SAN is reviewed with focus on
three aspects: routing and forwarding, incentive mechanisms
and data dissemination. Some important open issues with
respect to mobile social sensing and learning, privacy, node
selfishness and scalability are discussed.
Index Terms—Social Awareness; Mobile Social Networks;
Data Dissemination; Routing; Forwarding
I. INTRODUCTION
T
striking
few decades have witnessed
HE
last
communications
developments
in wireless
and
networking technologies which yield essential network
architectures to constitute ubiquitous networks. The rich
diversity of wireless mobile devices, especially smart mobile
devices, has joined in the networks with unprecedented speed.
As predicted by Gartner, mobile phones will overtake PCs as
the most common Web access devices in the world by 2013
[1]. Clearly, the applications and services are being migrated
into mobile devices. Thus, mobility, or more generally the
This work was partially supported by the Natural Science Foundation of
China under Grant No. 60903153, Liaoning Provincial Natural Science
Foundation of China under Grant No. 201202032, and the Fundamental
Research Funds for the Central Universities (DUT12JR10).
Feng Xia, Li Liu, and Jie Li are with School of Software, Dalian University
of Technology, Dalian 116620, China (e-mail: [email protected]).
Jianhua Ma is with the Faculty of Computer and Information Sciences,
Hosei University, Japan (e-mail: [email protected]).
Athanasios V. Vasilakos
is with
the Dept of Computer and
Telecommunications Engineering, University of Western Macedonia, Greece
(e-mail: [email protected]).
dynamic nature stands out as the main factor affecting the
service. Many applications benefit from the mobile devices,
such as mobile social network services, environment
monitoring, urban sensing, etc. Currently, a great number of
mobile devices access the Internet through e.g. 3G cellular
networks, which leads to huge communication overload and
thus has negative influence on the quality of service (QoS) of
the applications.
With the increasing workload, it has become crucial to
make full use of the limited resources of mobile devices and
communication networks so that the resource efficiency can
be improved and hence more and more mobile applications
can be supported. Recently, researchers begin to address this
issue from a social perspective. As a consequence, the field of
socially-aware networking (SAN) is emerging as a new
paradigm to exploit the social properties of network nodes
(especially mobile devices) for designing networking
solutions. In addition , recent rapid advances in social
networking applications are also a major drive force for the
emergence of SAN.
The main starting point of SAN relies on the dramatic
development on fields of wireless communications and
sociology (or social network theory). Firstly, the mobile
devices are now equipped with more storage, higher
computing speed and various wireless communication
interfaces. By using short-range wireless technologies such
as Wi-Fi and Bluetooth, mobile devices could form
temporary ad hoc networks to communicate with each other.
In recent years, many research efforts related to mobile
devices have been made in order to explore the potential of
mobile devices. For instance, opportunistic networks [2] and
mobile ad hoc networks (MANET) [3] aim at exploring
communications among mobile devices by using the
aforementioned short-range telecommunication protocols.
On the other hand, mobile sensing [4], mobile cloud
computing [5] and opportunistic computing [6] utilize the
huge
information resources and computing capability
provided by mobile devices. However, ad hoc networks are
not stable enough which usually utilize the opportunistic
contact to communicate. The mobility of mobile devices will
directly affect the topology of networks.
Secondly, it has been found that mobile devices have close
relationship with objects in society, because they are usually
carried by e.g. people, animals or vehicles. The mobility of
devices’ carriers (i.e. users) represents the mobility of mobile
devices. Human beings, as the main actors of mobility, take
especially important role and attract increasing attention
from information technology both in academy and industry.
There are a lot of social properties or relationships hidden
among human beings. For example, people are often highly
sociable and people with similar properties usually spend
long time together, being more willing to share information
and resources. The gregarious feature is called community.
People from the same community may contact or share
information with higher probability. By exploring social
relationships, catching the mobility regularities of mobile
devices as well as predicting the contact opportunities of
them will be effective. Furthermore, the contact information
could be a basic evidence for designing the routing protocols
in mobile environments.
the
features driven by
its own
SAN has built
aforementioned spheres. First of all, SAN focuses on wireless
communications among ad hoc networks which generally
consist of mobile devices connected via e.g. Wi-Fi, Bluetooth ,
etc. Similar to opportunistic networks, SAN is characterized
by the intermittent disconnection and the absence of the
infrastructure. Furthermore, social properties are valuable to
find
the users’ mobility pattern and predict contact
opportunities more accurately. Therefore SAN is able to
discover more reliable contact opportunities (with higher
prediction accuracy) by taking advantage of the social
awareness of network nodes. Social awareness (also known
as social consciousness) originates from sociology, and it is
often used to describe the sociability and social behaviors of
human beings. For instance, social awareness could be
consciously shared within a society, which means that a user
should know what is socially acceptable and what to be
performed in that manner .
The SAN paradigm can be applied into many areas such as
pocket switched networks (PSN) [7], vehicular ad-hoc
networks (VANET) [8] and cyber-physical systems (CPS) [9],
etc. In the meantime, it requires support from a number of
technologies
such as mobile sensing, opportunistic
computing, and social network analysis. Recently, several
research fields have emerged as socially aware paradigms
which share some common attributes with SAN, such as
mobile social networks [10], social opportunistic networks
[11], vehicle social networks [12], opportunistic Internet of
Things (IoT) [13], etc. They all address mobile nodes and are
driven by social network theory. However, they are different
in terms of focus. Mobile social networks originate from
online social networks which emphasize the relationship of
similar interests or commonalities. Social opportunistic
networks are based mainly on the prediction of contact
opportunities and concentrate on the social relationships
between mobile nodes and their regular movements. Vehicle
social networks consider the social properties of vehicles and
passengers. Opportunistic IoT usually focuses on the
interaction between human and IoT.
Compared to these related areas, SAN is human-centric
and studies individuals’ social properties comprehensively.
They involve a great deal of information, such as personal
property (e.g. preference, habit, and
life regularity),
human-to-human relationship (e.g. friendship, colleague,
and
family), human-to-community
relationship
(e.g.
interests and popularity), and human-to-environment
relationship. Based on these social properties, the objective of
2
SAN is to design networking solutions to support mobile
applications. In this context, routing and forwarding
protocols and dissemination strategies are
the most
important components by which data is delivered or
disseminated. In addition, mobile devices usually involve
personal (or private) information which leads to selfish
behaviors that benefit the users. Thus incentive mechanisms
are also an essential component.
In this paper, we present a survey of the emerging field of
socially-aware networking. To the best of our knowledge,
this is the first effort of its kind. We present the architecture
of SAN. Following a discussion on social properties, we
summarize the state-of-the-art research efforts on routing
and forwarding protocols, incentive mechanisms and data
dissemination algorithms in SAN. Our aim is to provide a
better understanding of
research opportunities and
challenges in the field of SAN and to find appealing hints for
future explorative activities on this timely and exciting topic.
The remainder of this paper is structured as follows.
Section II presents the architecture of SAN and Section III
describes several important social properties. In Section IV,
we review socially-aware routing and forwarding protocols
from perspectives of unicast and multicast, respectively,
along with some discussion on congestion problem. Incentive
mechanisms for dealing with node selfishness are explored in
Section V. The paper examines socially-aware data
dissemination algorithms categorized into solicitation and
cache based approaches and forwarding based approaches in
Section VI. We discuss some important open issues in
Section VII and conclude the paper in Section VIII.
II. ARCHITECTURE OF SOCIALLY-AWARE NETWORKING
SAN studies the context information of network nodes
(especially mobile devices), captures and extracts the social
properties and formulates reasonable protocols to support
upper-layer applications. However, SAN is an emerging field
Applications
Routing and
Forwarding
Protocols
Incentive
Mechanism
Data
Dissemination
Privacy
and Security
Leaning and Analysis
Personal Property
Preference,
Willingness,
Habit, etc.
Social Property and Relationship
Human-to-
Community,
Human-to-
Human,
Human-to-
Environment
Community,
Similarity, Tie
Strength,
Mobility
Pattern, etc.
Sensing Data
Personal Operation
Environment Sensing
Contact Sensing
Fig. 1. Architecture of SAN
still in its infancy. Till now there is no formal (uniform)
definition for this terminology or its architecture. In th is
paper, we propose a simple architecture of SAN, as shown in
Fig. 1 where the above mentioned workflow of SAN is
depicted. The detailed description about the structure is
presented below.
The major task of the first two steps is to achieve social
awareness by sensing and analyzing data using some
intelligent technologies, such as data mining and machine
learning. To obtain the information about personal behaviors,
social contact and the environment, mobile sensing is an
effective way to sense the surroundings. Specifically, various
sensors deployed in real world could be utilized, as well as
sensors embedded in mobile devices. The available sensors
include e.g. accelerometer, digital compass, gyroscope, GPS,
microphone and camera, which enable personal, group, and
community scale sensing applications [4]. In this sense, SAN
relies on CPS, i.e. to obtain the information of social
awareness. On the other hand, SAN enhances CPS by
exploiting social properties, which enrich human -to-human,
human-to-object, and object-to-object interactions in the
physical world as well as in the virtual world.
The various data, which reflect either personal behaviors
or environment information, are the main evidences to
explore both the personal and social properties. The personal
properties involve e.g. one’s preference, willingness, habit,
life regularity, etc. They could be obtained through analyzing
the operations and behaviors of the user. SAN is based on
social properties and relationships of users, which are more
reliable and more natural. The learning and deduction
process requires more resources and computing capability,
which might need the support of e.g. mobile cloud computing
and data mining.
Through analyzing and learning from these data, SAN can
deduce important social properties. The most commonly used
social properties include community, centrality, similarity,
tie strength and human mobility pattern. These properties are
connected to the social relationships. To be exact, the
human-to-human relationship refers
to direct contact
relationship and it involves personal information such as
friendship, which indicates a kind of tie strength. The
human-to-community relationship indicates an individual’s
gregarious property and status in community. They can be
gathered by use of contact graph (physical contact) and social
graph (virtual graph, such as online social networks). The
human-to-environment
relationship
implies
that
the
environmental information is related to human beings.
The available personal and social properties are valuable
evidence
to devise
the routing protocols and data
dissemination algorithms. Thus, socially-aware routing and
forwarding protocols and socially-aware data dissemination
algorithms are essential issues in SAN. In addition, the
protocols in SAN rely heavily on the nodes’ cooperation due
to the exploration of social relationship and interactions
among the users. As a consequence, incentive mechanisms
are necessary to deal with selfishness of nodes. In addition ,
privacy and security protecting strategies are also an
important component of SAN. However,
it
is very
3
Fig. 2. Taxonomy of Protocols in SAN
challenging and we list it in open issues. For these reasons,
routing and forwarding protocols, incentive mechanisms and
data dissemination algorithms in the context of SAN are the
main topics of this paper. Fig. 2 illustrates a generalized
taxonomy for SAN protocols based on the above description,
which is used for organizing this paper.
SAN has some similarities with the SCI (social and
community intelligence) concept presented by Zhang et al.
[14]. However, SCI focuses on the collection and analysis of
social information and aims to reveal individual and group
behaviors, social interactions and community dynamics,
while SAN pays more attention to designing networking
protocols by use of social properties. It is worth noting that
the proposed architecture is just a starting point for extensive
study of SAN. We hope it will eventually lead to a converged
viewpoint and taxonomy of SAN.
III. SOCIAL PROPERTIES
The research on social networks was initiated by Milgram
in 1960’s [15]. Milgram introduced the idea of small world
phenomenon which indicates that any pair of people in the
world can connect with each other through small sequences
of relationships (typically five or six), and thereafter many
works reaffirmed it [16]-[18].
Contemporary researches mainly focus on understanding
the dynamics and structure of social networks with respect to
relationships which can be classified in terms of strength of
mutual familiarity and intensity. Social network analysis
(SNA) [19] studies relationships between human beings, as
well as patterns and implications of these relationships. SNA
is a useful and powerful tool for analyzing complex social
relationships among people in social sciences [20]. The
notion of social network and the methods for SNA attract
significant interest initially from the social and behavioral
communities, later data mining, and only recently from the
networking community. In this section, we concentrate on
social properties that are most popular in design of protocols
and algorithms in SAN.
A. Social Graph
Social networks exhibit the small world phenomenon that
node encounters are sufficient
to build a connected
relationship graph. Graph is a convenient tool to represent
the relational structure of social networks in a natural
manner , which is generally called social graph. In a social
graph, vertices (nodes) indicate human individuals and edges
(links) indicate social relationships between individuals. In
some degree, social network is equal to the social graph, so
they can be used alternatively. One significant challenge in
social networks is how to represent a link between two nodes
[21]. According to different link meanings, several social
graphs are proposed in recent literature. Contact graph is a
popular way to analyze and estimate relationships among
people by observing their inter-contact time in the history
[22]-[25]. Besides, neighbor graph [21], regularity graph and
interest graph [27] are proposed recently as well.
B. Community
A community is a structural subunit (which can be
represented as a set of individuals) of a social network with
high density of internal links [21]. Individuals have more
social connections with other individuals inside their own
community than with individuals outside. The social
connections may be family, friends, common location or
common interest, which are decided by the social graph. In
general, individuals in the same community may meet each
other more frequently. Therefore, community structure has
significant impact on people’s mobility patterns and thus is
beneficial for choosing appropriate forwarding path.
C. Centrality
Besides community, the centrality is another basic concept
in social networks, which considerably affects
the
performance of socially-aware forwarding algorithms. The
experiments in [28] demonstrate that it is important to find
appropriate centrality and community in the design of
socially-aware
dissemination
data
and
forwarding
algorithms.
Centrality is used to describe importan t and prominen t
nodes in a social graph. People have various roles and
popularities in society. A central node has stronger capability
of connection with other nodes. Experiments in [24] show
that there is a small number of nodes which have extremely
high relaying ability, and a large number of nodes which
have moderate or low centrality values, with 30 and 70
percentiles, respectively. This phenomenon is called human
heterogeneity (or node heterogeneity).
The most recognized centrality measures are introduced by
Freeman [29][30]: degree centrality, betweenness centrality
and closeness centrality. Here we give a brief introduction to
each of them.
1) Degree Centrality
Degree centrality is defined as the number of one-hop
neighbors of a node [19]. For a network consisting of n nodes,
the degree centrality DegCi of a node i is:
de g i
( )
(1)
D e gC
i
n
1
where deg(i) is the number of directly connected neighbors of
node i. DegCi indicates the connection ratio of node i
between the real connection number of node i and the
4
i
i
maximum number of possible connections, i.e. n-1. Degree
centrality identifies the most active nodes in the network. A
node with high degree centrality maintains large number of
links to others. As such, a central node occupies a structural
position (network location) that may act as a conduit for
information exchange [23].
2) Betweenness Centrality
Betweenness centrality is the percentage of the number of
shortest-paths including node i over all the shortest-paths
[31]. The betweenness centrality BetCi of a node i is:
sp
i
( )
(2)
B e tC
j k
,
sp
j
i
j k
,
where spj,k is the number of shortest paths linking nodes j and
k, and spj,k(i) is the number of shortest paths linking nodes j
and k which pass through node i. Betweenness centrality is a
measure of the extent to which a node has con trol over
information flowing between others [32]. Nodes with high
betweenness centrality can bridge two nonadjacent nodes and
may impact on data flow between communities. Therefore,
betweenness centrality is a key metric to determine the links
between communities.
3) Closeness Centrality
Closeness centrality is defined as the inverse of the sum of
the distances between a given actor and all other actors in the
network [14]. The closeness centrality CloCi of a node i is
1
(3)
C loC
j
d is i
( ,
)
j
i
where dis(i,j) is the distance between nodes i and j. A node
with the highest closeness means that the node has the
shortest path to other nodes in network. So closeness
centrality describes the efficiency of information propagation
from one node to all the others. In message forwarding and
data dissemination applications, closeness centrality can be
used to choose relay nodes to deliver the message with
success and/or faster to the other nodes in the community.
The analysis of data collected from ACM CoNEXT’07
concludes that centrality is the primary factor to decide
whether a node is a good next hop and the best performance
trade-off is obtained when several complementary rules are
combined [33].
D. Similarity
Sociologists have long realized that social network
displays a high degree of transitivity, that is to say, there is a
heightened probability of two people being acquainted if they
have one or more other acquaintances in common. Th is
phenomenon is also called clustering [23]. Similarity
indicates the group of nodes depending upon common
contacts or interests which can be measured by the ratio of
common links (e.g. contact, interest, neighbors) between
individuals. The higher similarity a node and the destination
share, the more opportunities they have to encounter. Nodes
with higher similarity can be good candidates for
information dissemination among clusters of nodes.
E. Tie Strength
The notion of tie strength was firstly introduced by
Granovetter in 1973, which is defined as “the amount of time,
the emotional intensity, the intimacy (mutual confiding), and
the reciprocal services, which characterize a tie” [34]. Tie
strength is a quantifiable property that characterizes the link
between two nodes. Strong ties are more likely to be activated
for information flow when compared to weak ties. The most
widely used
tie strength
indicators are:
frequency,
intimacy/closeness, longevity, reciprocity, recency, multiple
social context, and mutual confiding (trust) [22]. A
combination of the tie strength indicators can be used for
information flow to determine which contact has the
strongest social relationship to the destination. Meanwhile,
the effects of weak ties in social networks are also crucial to
data dissemination. Granovetter’s research modifies that the
weak ties may be beneficial to forming bridges between high
density clusters.
F. Human Mobility Pattern
Researchers observe that human mobility presents two key
properties from analysis of real mobility traces: spatial
regularity and
temporal regularity
[35][36]. Spatial
regularity is that nodes usually move around a set of locations
frequently and regularly in time schedule. For example, the
students usually move around among dormitory, classroom
and canteen at different periods of the day. Temporal
regularity is that human mobility pattern is repetitive at long
time. For instance, human repeats working mobility pattern
from Monday to Friday. The two properties of human
mobility can be used to predict the users’ future mobility
which plays important role in choosing the forwarders.
IV. SOCIALLY-AWARE ROUTING AND FORWARDING
PROTOCOLS
Due to the mobility of mobile devices, it is usually difficult
to find an end-to-end path between source and destination(s)
at the beginning of communications. All of the current
routing methods share a store-carry-and-forward paradigm
and utilize contact opportunity
to communicate
in
intermittently (dis)connected networks such as delay tolerant
networks (DTN) and opportunistic networks. Thus relaying
selection and forwarding decision are critical to be made by
the current node based on certain routing strategy. The
prediction of contact opportunity is one key issue for design
of efficient routing and forwarding protocols. SAN considers
the social properties during the course of routing and
forwarding protocol design to make better forwarding
decisions.
Routing can be classified into unicast and multicast
according to the number of destinations the data need to be
delivered. Unicast routing focuses on forwarding data to a
single
specific destination. Multicast
involves
the
distribution of the data to a group of users. In this section, we
will review unicast routing protocols and multicast routing
protocols based on social properties, respectively. In addition ,
congestion is a very challenging problem as well as an
important factor that influences the design of routing
protocols. Consequently, here we will also examine
state-of-the-art work addressing the congestion problem in
the context of SAN.
5
A. Unicast Routing
The key faction of routing protocols is to select the most
optimal relay nodes which have highest probability to meet
the destination(s) in order to maximize message delivery
ratio while minimizing message overhead and delay. The
relative stable social relationships (especially the community
relationship) of the users are reasonable information for
predicting future contact opportunities. The social network
structure can be viewed as
three
levels:
individual,
community, and whole network [37]. The community is very
widely used in routing protocols. Nodes in the same
community have more chances to meet than those in different
communities. A large number of community detection
algorithms have been proposed to divide a social network
into separated communities. Based on whether the support of
community is needed or not, we classify routing and
forwarding protocols as community-based routing and
community-independent routing.
1) Community-Based Routing
It is believed that nodes have more opportunities to contact
in community, which is beneficial to forwarding messages
for other members of their community. One of the first
research works on community in routing protocols is carried
out by Hui and Crowcroft [38]. They conducted an
experiment in INFOCOM’06 conference, which results in
community relationship through analyzing inter -contact
time distribution . The results proved that community
structure can improve forwarding efficiency. Inter-contact
time is the time elapsed between two successive contact
periods for a given pair of devices [39]. In experimental
analysis
the
distribution,
time
inter-contact
of
intra-community pair has higher power law coefficient than
the inter-community pair, which indicates that nodes in the
same community tend to meet more often. In the experiment,
they proposed a forwarding scheme named LABEL which
assumes each node has a label on behalf of its affiliation. It
directly forwards messages to destination, or chooses
next-hop nodes with the same label as the destination node.
They also presented the concept of friendship community
which can help improve delivery.
The general operation principle in community-based
routing is shown in Fig. 3. Community is the basis of
forwarding data. Firstly, mobile nodes are grouped into
communities by certain community detection algorithm.
According to the forwarding strategy, data are forwarded
among nodes. If the relay nodes are out of destination
community (the community where the destination belongs
to), the inter-community forwarding strategy is used to
forward data close to the destination community as quickly as
possible. If the relay nodes are in the destination community,
the intra-community forwarding strategy is used to forward
data to the central node and the central node can forward the
data to the destination .
Community detection and formation play an important
role in community-based routing. In LABEL, the affiliation
(or community) is labeled through configuration. To make
the community practically useful, many community detection
algorithms have been proposed. Most of the proposed
Community Detection
Out of Destination
Community
In Destination
Community
Inter-community
Forwarding
Forward to
Centrality Nodes
Forward to
Destination Node
Intra-community
Forwarding
Fig. 3. Operating Principle of Community-based Routing
algorithms are centralized and they focus on analysis of
offline mobile trace. Contrary to the centralized methods (see
[39]), Hui and Crowcroft proposed three algorithms, named
SIMPLE, K-CLIQUE and MODULARITY, for online
distributed community detection based on the contact graph
in [24]. Then they presented BUBBLE RAP [24] based on
socially-aware overlay. The overlay is constructed by
combining the community with centrality. BUBBLE RAP
calculates global ranking (i.e. global centrality) for the whole
network and local ranking (i.e. local centrality) for the local
community for each node based on degree centrality. When a
node has a message destined for another node, the node first
bubbles the message up through hierarchical ranking tree
using the global ranking until reach ing a node which is in the
same community as the destination node. Then in the same
community, the message is bubbled up through local ranking
tree using the local ranking until the destination is reached or
message expires.
Bulut and Szymansk [41] introduced direct and indirect
friendship to indicate the tie strength of virtual link and to
form friendship community. They consider three behavior
features of close friendship: high frequency, longevity and
regularity, and also define two metrics called SPM (social
pressure metric) and CSPM (conditional SPM) for direct and
indirect friendship based on contact history. Friendship
community is a set of nodes having a link quality larger than
a threshold. To reflect temporal distinctions, different
friendship communities in different periods of the day are
established for each node. When forwarding, current node
with the message will choose the node that belongs to the
same community with the destination and with a stronger
friendship of destination than current node.
Several works in the literature consider the relationship
between communities and formulate
inter-community
forwarding decision [21][42][43].
LocalCom [21] only requires limited local information to
form communities and it also considers the forwarding
between different communities. Firstly, the authors presented
a metric named “similarity” to construct the neighboring
graph which considers the encounter frequency, encounter
length, separation period in the encounter history. Then
based on
the neighboring graph and
the associated
“similarity”, the authors proposed a distributed algorithm for
6
community detection, and represented the communities with
an extended k-hop clique. LocalCom adapts different
forwarding plans for in tra- and inter-community packet
forwarding. Intra-community packet forwarding adapts
single hop source routing. The “similarity” metrics indicate
the quality of virtual links (tie strength) and the node with
larger “similarity” metrics is chosen to forward data.
Controlled flooding routing is used for inter-community
communication based on gateways. Gateways are nodes that
have direct neighboring relationships with nodes in other
communities and bridges are selected from gateways using
two marking and pruning schemes: static pre-pruning and
dynamic pruning based on betweenness centrality [21].
Gently [42] is based on the Context-aware Adaptive
Routing (CAR) protocol [44] and LABEL. LABEL will be
the same as WAIT when a sender does not meet a member of
the destination community (where the destination belongs to).
In contrast to WAIT, Gently uses CAR-like routing to
improve the efficiency when not meeting the destination
community. Gently supports unicast as well as multicast and
defines labels to identify communities of hosts. A recipient of
message can be a single host or a group (community) of hosts.
When the message carrier meets a node of destination
community, Gently adopts a LABEL based strategy. Then in
the destination community, CAR-like routing is used to
deliver the message to the destination.
Zhou et al. [43] observed that while habitual mobility is
useful in reducing the average communication latency,
irregular deviation from habit can seriously affect worst-case
performance. They proposed Diverse Routing (DR), a
tunable protocol, to cope with the nodes deviating from their
habitual activities. The main idea of DR is to statistically
cluster the network into proximity-based social cluster and
scatter at least one copy of a packet in a cluster such that even
deviant nodes will be close to at least one of these packets.
DR uses limited communication overhead to provide a
stability and flexibility propagation scheme. Like Gently, DR
supports unicast and multicast at the same time.
Community-based routing
the forwarding
improves
efficiency depending on the higher meeting probabilities in
the same community. Centrality and tie strength further
provide evidence to choose the appropriate forwarder as well.
However, distributed community detection and formation are
still a challenge due to the dynamic topology of the network
and difficulties in the information exchange and calculation.
On the other hand, special nodes such as gateways have to be
employed
inter-community
accommodate
to
communications. Hui and Crowcroft [38] introduced the
concept of friendship community where two or more
communities are very close or overlap and validated that the
power law coefficient of friendship community is between
intra-community
and
inter-community.
Exploiting
friendship among communities will be a potential way to
improve the forwarding efficiency between communities.
2) Community-Independent Routing
Considering the difficulty of community detection and
formation, many routing protocols have been presented
without community support. The concept of ego network is
exploited where only locally available information is
considered [22][23]. Ego networks can be defined as
networks consisting of a single actor (ego) together with the
actors they are connected to (alters) and all the links among
those alters [22]. These routing protocols are mainly based
on utility function to calculate the satisfied nodes.
SimBet routing [23] is based on utility function which
exploits betweenness centrality metrics and similarity to the
destination node using ego network. In routing process, the
encounter nodes firstly exchange the contact list to update the
betweenness and similarity values. Then they exchange a
summary list of destination values, calculate and compare the
SimBet utility. If a node has a higher utility for a given
destination, the destination is added to a request vector.
Finally, they exchange the request list for further routing.
Further introducing the tie strength relationship with the
destination, SimBetTS [22] was presented as an evolution of
SimBet. Measuring
tie strength
in SimBetTS
is an
aggregation of a selection of indicators based on frequency,
intimacy/closeness and
simulation
the
In
recency.
experiments, betweenness utility, similarity utility, and tie
strength utility have been examined respectively, as well as
the SimBetTS protocol combining these three metrics. The
results show that SimBetTS improves overall delivery
performance while reducing the load on central nodes, which
is better distributed across the network.
Some works embed context information of nodes into
routing protocols. For instance, PROPHET [45] exploits the
frequency of history contacts between users. MobySpace [46]
and MV [47] exploit information about users’ mobility
patterns and places the users visit. The social relationships
taking into
information worth
are
important context
consideration . HiBOp
[48] automatically
learns and
represents context information, the users’ behavior and their
social relations, and exploits this knowledge to drive the
forwarding process [6][76]. Nodes running HiBOp infer
acquaintance between users through analyzing the similarity
of their context information and behavior which include both
present and historical information. Messages are forwarded
through users closer to the destination. To improve the
successful delivery rate, for each message more than one
copy is injected into the network following a different route.
To control the delivery cost, only the sender is allowed to
inject multi-copies. In addition, HiBOp supports both unicast
and group communications without requiring any particular
customization.
Mtibaa et al. [49] exploited the information about social
interactions of users from online social network platforms
and applications such as Facebook, Orkut, or MySpace, etc.
They proposed PeopleRank to rank the nodes in the social
graph by using social relationships, inspired by the famous
PageRank algorithm. Such social relationships can be based
on explicit friendships (as defined in online social networks
for example), on personal communications (for example
extracted through the communication patterns available in
cell phones), or even common interests. The node with a
higher PeopleRank value will be more central in the social
network and be a better forwarder. In addition, PeopleRank
7
can be implemented in a centralized way or in a completely
distributed fashion.
Aforementioned approaches take the social relationships
among nodes into consideration. Several works leverage the
regular human mobility pattern to predict the best forwarder.
Some examples of predict-based routing protocols include
PROPHET, MobySpace, MaxProp [50], etc. PROPHET and
MaxProp are based on history contact data while MobySpace
is based on the visit probability of locations in history.
However, they all ignore the regularity of human mobility in
spatial and temporal aspects which can take more accurate
prediction for the future mobility.
In addition , the regularity of human mobility is considered
in some works. PER [51] employs a time-homogeneous
semi-markov process model to describe node mobility as
transitions between possible locations. Location transition
and sojourn time probability distributions are determined
from nodes’ mobility history. Nazir et al. [52] assumed that
people follow similar mobility patterns daily (i.e. Monday to
Friday) and proposed algorithms for social encounter based
content delivery system with the time critical property.
the
The
community-independent
routings
avoid
community detection, and leverage the context information,
social properties (such as centrality, similarity, and tie
strength) and the regular mobility pattern to predict the best
forwarder. It is simple and easy to implement these routing
protocols. One drawback of community-independent routing
is that they are less sensitive to deviations. The prediction
relies on historical mobility. If nodes deviate from their
regular mobility patterns, the prediction will have large
errors. For this reason, taking the deviation of mobility into
consideration is a critical issue in the future.
Temporal and spatial factors are also of great significance.
The history temporal/location information represent the
people’s movement pattern. The current temporal/location
information indicate the instantaneous state which is related
to the future state. Gao et al. took temporal factor into
consideration in [53]. The transient node contact patterns are
exploited. They define transient connectivity which indicates
that some nodes may remain connected with each other
during specific time periods and develop appropriate data
forwarding metrics for more accurate prediction of node
contact capability within the given time constraint. Existing
works in the literature pay more attention to history
temporal/location
information,
rather
than
current
information. In addition, the combination of social properties
and temporal/location information is worth investigating.
B. Multicast Routing
Multicast routing is a considerably fresh and challenging
problem in frequently disrupted and partitioned networks.
Multicast routing allows information to be delivered to
multiple destinations. Usually, the multiple destinations are
identified as a group. Group membership of a particular
multicast group may change with time as nodes join and
leave the group [54]. Patra et al. [55] classified multicast
routing strategies in DTNs on the basis of their basic working
mechanisms. The categories include flooding multicast
routing, tree-based multicast routing, probability-based
multicast routing, and intelligent multicast routing. None of
them consider social properties.
into
relationships group people
social
Inherent
communities, which can be leveraged for the purpose of
multicast routing. Some works aforementioned in unicast
routing support multicast such as community-based routing.
In community-based routing protocols, if the recipients of the
multicast are in the same community, it is feasible to support
multicast. The process is that the message is forwarded to any
member of the community and then replicated in an epidemic
manner among the members of the community. Taking
HiBOp as an instance, it supports context-aware multicasting
using the destinations’ identity table.
Gao et al. [54] are among the first to study multicast from
the social network perspective. They proposed a set of
multicast routing methods with single and multiple data
items based on community and centrality. In single-data
multicast, they introduced a new metric called cumulative
contact probability (CCP) to indicate the average probability
that a node meets a random node within a time constraint.
Based on CCP of nodes, the minimum number of relays is
selected (by solving a unified knapsack problem) to satisfy
the requirement on delivery ratio within certain time
constraint. In multi-data multicast, a node maintains its
destination-awareness about other nodes to select relays
among its neighbors. If the destinations are in other
communities, data forwarding is conducted through gateway
nodes to connect multiple communities. The multi-data
multicast is also modeled as a knapsack problem ensuring the
required delivery ratios.
In multicast, the selection of relays is usually based on the
relays’ cumulative probabilities in order to forward single or
multiple data to multiple destinations [54]. However, unicast
generally relies on the future meeting probability to decide
the relay. Compared to unicast, multicast protocols face more
complicated context and are more challenging to design.
However, multicast has a wide spectrum of application
scenarios which require sharing data among people in e.g. a
meeting, conference or battlefield scenario.
C. Congestion
One critical issue of routing protocols is the trade-off
between improving delivery ratio and reducing delivery
delay and overhead. The socially-aware routing protocols
mentioned above contribute to this target by selecting
appropriate forwarders which have more opportunities to
encounter the destination from the perspective of social
interactions. However, the problems of overhead and
congestion remain to be addressed. In the context of networks,
congestion generally refers to a network state where the
relevant node is overloaded with too many messages. A result
of congestion is that network QoS may be deteriorated. The
cause of overhead and congestion may come from e.g. the
redundant message duplications and over-use of the central
nodes. The physical buffer of any mobile node is limited.
When there are excessive messages to be exchanged,
congestion may happen [56].
8
Redundant message duplications might cause high
overhead and congestion of the network. In the worst case,
94% of duplicate packets reach the destination , which will
induce huge overhead on bandwidth, energy and memory
consumption [57]. Many protocols are trying to achieve an
efficient trade-off by controlling duplication.
Kawarabayashi et al. [57] proposed a strategy to lower
message duplication under a given delay or delivery
probability. Based on predictable working day model, the
authors formulated the problem of message duplication into a
spanning tree problem which is further optimized in delivery
time, duplication and storage space of messages. In [58], the
delegation forwarding algorithm forwards a message only if
it encounters another node whose quality metric is greater
than any other nodes seen by the message so far. The cost of
delegation forwarding algorithm is proportional to the square
root of population size, which is more optimal in contrast to
naive forwarding with linear cost.
Most of the routing protocols lead the routing to direct
most of the traffic through a small subset of good nodes (e.g.
central node). For instance, in the SimBet algorithm, the top
10% of nodes carry out 54% of all the forwards and 85% of
all the handover [59]. This unfair load distribution causes
local storage congestion, further increasing the discarding
rate and decreasing the delivery rate.
To deal with congestion problem, current methods mainly
focus on two strategies. One is to mitigate the role of the
central nodes. Fair routing [59] classifies the nodes into
different levels to limit the exchange of messages. It exploits
the social process of perceived interaction strength based on
the social relation between nodes in different time scales.
Then it forwards the message by the stronger social relation
and uses assortative-based queue control to limit the
exchange of messages to those users with similar “social
status”. In other words, nodes will only accept forwarding
requests from those nodes of equal or higher status. In
addition, Yuan et al. [62] introduced strangers to participate
in routing and exploited the optimized number of strangers
in order to achieve a better performance of overhead/packet
delivery ratio in pure darkness scenario. The other strategy is
based on detection of the congestion. Nile [55] keeps the link
loads in check to push replicas only to those promising paths
that may sustain more loads. Radenkovic and Grundy [60]
introduced congestion driven part in forwarding protocol to
avoid nodes that have lower availability and higher
“congestion rates”. In reality, the detection of congestion and
the prediction of idle paths are difficult to address due to
changing network topology. Recently, some researchers tried
to make use of strange nodes and nodes with weak ties, which
might be a promising approach to cope with congestion.
Table I summarizes representative socially-aware routing
and
forwarding protocols
in
the
literature. The
socially-aware routing and forwarding protocols using
community can support both unicast and multicast, which
have more flexibility and applicability. A lot of experiments
and simulations have been carried out in the protocols listed
in the table which are usually based on real trace data. The
results of experiments and simulations show that the
Solution
LABEL [33]
BUBBLE RAP [19]
LocalCom [16]
Friendship-based [36]
Gently [37]
DR [38]
SimBet [18]
SimBetTs [17]
HiBOp [43]
PeopleRank [44]
Social-based [49]
Fair Routing [52]
unicast, multicast
unicast, multicast
TABLE I
SUMMARY OF SOCIALLY-AWARE ROUTING AND FORWARDING PROTOCOLS
Social Characteristics
Unicast or Multicast
Trace-Based Mobility
InfoCom06
unicast, multicast
community
HongKong, Cambridge,
community
unicast, multicast
InfoCom06,
degree centrality
MIT Reality Mining
Haggle
community
MIT Reality Mining
tie strength
betweenness centrality
MIT Reality Mining
community, friendship,
similarity, tie strength
Implemented in Haggle
community
Duk trace, WTD trace
community
MIT Reality Mining
betweenness centrality
similarity
Intel, Cambridge, InfoCom06
betweenness centrality,
similarity, tie strength
context information,
social relationship, similarity
friendship, similarity
community, centrality
unicast, multicast
unicast, multicast
unicast
unicast
unicast, multicast
unicast
multicast
Community based mobility
model
MobiClique, SecondLife,
InfoCom06, Hope
MIT Reality Mining
InfoCom06
MIT Reality Mining
community,
centrality
unicast
9
Protocols for Comparison
MCP, WAIT, Control
MCP, LABEL,
FLOOD,
WAIT , PROPHE
PROPHET,
BubbleRap
PROPHET, SimBet
SimBet, Epidemic
PROPHET,
Epidemic
PROPHET,
Epidemic
PROPHET,
Epidemic
Epidemic, BubbleRap,
PROPHET, SimBet
Epidemic, PROPHET,
SimBet
socially-aware routing and forwarding protocols outperform
socially-oblivious protocols, as a general rule.
V. ROUTING AND FORWARDING WITH SELFISHNESS
All the approaches mentioned above assume all nodes in
the network are cooperative and altruistic. Therefore, all
nodes are willing to help forward messages for other nodes.
But in reality, many nodes exhibit non-cooperative behaviors,
such as selfish nodes or even malicious nodes, in order to e.g.
conserve limited resources (like power and buffer) and
increase their own benefits.
A selfish node always acts for its own interest, meaning
selfish nodes may not be willing to provide services for others
in order to conserve their limited buffer or power resources.
Whereas, a malicious node acts maliciously with the
intention to disrupt the main functionality of the networks, so
it might possibly drop packets, jam the wireless channel, and
even forge false packets [63].
Many efforts have been made in the literature to evaluate
the effects of cooperation and/or selfishness in opportunistic
networks from different aspects [64]-[69]. It has been proved
that node collaboration (even limited collaboration) can
dramatically
improve
performance
compared
to
non-cooperation scenarios and different non-cooperation
behaviors may have opposite impacts on different routing
algorithms.
Social selfishness is first introduced in [70]. In social
perspective, a selfish user is usually willing to help others
with whom he/she has social relationships (e.g., friends,
coworkers, roommates) or in same community. He/she will
provide better service to those with stronger social ties than
those with weaker ties, especially when there are resource
constraints. The social related selfishness is defined as social
selfishness. In contrast, individual selfishness refers to the
nodes having the same selfishness level to other nodes. Social
selfishness conveys the social tie between the nodes and can
be used to select trusted relay nodes. When a node has no
social ties to the outside world, e.g. a node is out of its own
community, it becomes individual selfishness. It would be
better to treat nodes’ selfishness behaviors in different ways
with respect to different scenarios.
To deal with selfish nodes, incentive mechanisms are
necessary
to
stimulate nodes’
cooperation,
check
misbehaviors and punish selfish nodes. Due to frequent
network partitions, resource constraints and longer delay, the
incentive mechanisms developed in wireless networks like
MANETs, may not work effectively in intermittently
(dis)connected scenarios [68]. This is a challenging problem
that has attracted a lot of attention from many researchers.
Most recent works addressing node selfishness fall into
one of the following four categories: reputation-based,
credit-based, Tit-For-Tat (TFT), and trust-based. Generally
speaking, the first three categories are traditional incentive
mechanisms that focus on individual selfishness, while
trust-based
strategies
take
social
selfishness
into
consideration. One could argue that trust-based solutions do
not belong to incentive mechanisms because they do not take
any action to stimulate cooperation or check the selfish node.
In the context of socially-aware networks, however, social
selfishness becomes an important issue to be addressed.
Consequently, we cover trust-based strategies here. Different
categories of solutions have different principles to stimulate
cooperation. At the same time, they need to cope with various
attacks coming from malicious nodes.
10
A. Reputation-Based Incentive Mechanisms
In general,
reputation-based
incentive mechanisms
identify misbehaving nodes and exclude them from the
network. Nodes can build up their good reputation scores by
forwarding packets for others, and thus will be rewarded with
higher priorities when transferring their own packets. The
corresponding reputation score decreases when a node
misbehaves. The node with bad reputation is detected and
excluded when its reputation becomes below a threshold
[74].
In reputation systems, sybil attack and whitewashing
attack are common attacks to be handled. A sybil attack [71]
indicates that a malicious node attempts to create multiple
identities in order to help itself gain good reputation,
abandon bad reputation or evade responsibility of its actions.
In a whitewashing attack [72], a node repeatedly leaves and
rejoins the system using newly created identities to avoid
suffering from bad reputation. In classical networks, trust is
established by a certificate authority (CA) through a public
key infrastructure (PKI) [73]. But it is not easy to implement
in non-infrastructure and intermittent connection scenarios.
RADON [74] is a reputation-assisted data forwarding
protocol which comprehensively evaluates an encounter’s
forwarding ability by
integrating
the
reputation of
forwarding data with the possibility of meeting a destination.
In RADON, a special message called Positive Feedback
Message (PFM) is used to help the Watchdog monitor a
node’s forwarding behavior in the reputation management
system. RADON prevents a malicious node from deliberately
dropping
and
arbitrarily
forwarding data, which
dramatically improves the network performance in a malign
environment.
Give2Get Epidemic Forwarding and Give2Get Delegation
Forwarding are the first protocols for packet forwarding in a
social setting to tolerant selfish behaviors, which are based
on Epidemic
forwarding and Delegation forwarding
respectively [74]. Give2Get consists of two phases: the relay
phase and the test phase. In the relay phase, the sender (S),
which generates the message, tries relaying the message to
(at least) the first two nodes it meets by negotiating a
cryptographic session key. Assume node B relays the
message, node B continues the relay phase to find two other
nodes and relay the message to them. By doing this, it can
collect two proofs of relay for itself. When meeting node S
again, during the test phase, if node B is not able either to
show the two proofs or to prove to have still the message in its
memory, then node S can broadcast a proof of misbehavior
(PoM) to the whole network, which will remove node B if it
cannot prove that node S is wrong. In addition, both
protocols are Nash equilibriums, i.e. no individual has an
interest to deviate.
MobiID [76] is a “user-centric” and dynamic reputation
based incentive scheme to stimulate cooperation in bundle
forwarding. Conventional reputation based schemes always
rely on neighboring nodes to monitor the traffic and keep
tracks of each other’s reputation. In contrast, MobiID allows
each node to maintain, update and show its reputation tickets
by itself on demand. This is called self-check. It also defines
a social metric which considers the forwarding willingness
from forwarding history and identifies the social community
based on
this metric. Through sharing
reputation
information in the community and forming consensus views
towards targets, MobiID implies community check. MobiID
exploits Offline System Manager
(OSM) which
is
responsible for key distribution. In addition, MobiID
addresses attacks such as modifying the forwarding history to
over-claim a high reputation (so as to attract bundles) and
isolate the node from the target user.
IRONMAN [77] is another incentive mechanism using
social information. However it uses pre-existing social
information to detect and punish selfish nodes, thus
incentivizing them to participate in the network. The social
information can be obtained through interview, or from an
online social network (e.g., Facebook friend lists). With the
proliferation of online social network applications such as
Facebook, Orkut, or MySpace,
the social interaction
information of users becomes available. The social
information is increasingly reliable and can be utilized for
many purposes such as predicting future encounters in
opportunistic networks. The integration of online social
network and mobile networks is a prominent problem in
future research work.
SRed [78] is a localized, link-state-based and multi-path
routing protocol which also mitigates a number of known
routing layer attacks such as black hole, den ial of service
(DoS) and wormhole. SRed uses reputation-based routing
generation mode in benign environments, otherwise the
probabilistic routing generation mode will be activated. The
dynamic window mechanism is used to switch between the
two modes in order to achieve a good trade-off between
efficiency and security.
The above protocols avoid central management
mechanisms to control the reputation estimation in order to
adapt to mobile environments. RADON, Give2Get and SRed
utilize
the successful
forwarding process
to verify
misbehaving nodes. Li and Cao [79] presented a similar
scheme to migrate routing misbehavior through detecting
packet dropping.
All these protocols can achieve high delivery efficiency,
but they benefit from different aspects. RADON, Give2Get,
SRed and Li and Cao’s scheme are easy to operate, while
MobiID and IROMAN are of high reliability thanks to the
exploring of social community and group strength.
B. Credit-Based Incentive Mechanisms
Credit-based incentive schemes introduce some forms of
virtual currency or credit to regulate the packet-forwarding
relationships among different nodes [80]. Nodes earn credits
by forwarding packets. As a return, these credits can then be
used to obtain forwarding service from other nodes in the
system. That is, it will take some credits for the source node
to send each packet. If a packet is delivered successfully, the
corresponding credits will be distributed to the intermediate
nodes participating in packet relay.
There are two significant difficulties in credit-based
incentive schemes. One is the management of nodes and
credits distribution. Sometimes, the realization of payments
requires an out-of-band trusted third party, which is unlikely
to be available in DTN. The other is cheat attacks from
selfish nodes. Due to the selfish nature, mobile nodes will try
to cheat the system to maximize their welfare through
injecting or deleting some relay nodes to achieve more
credits.
In SMART [80] and Pi [81], the systems employ an
Offline Security Manager (OSM), which is responsible for
key distribution , and a Virtual Bank (VB), which is
responsible for credit management. SMART is a secure
multilayer credit-based incentive scheme for DTN which can
be compatible with diverse data forwarding algorithms. It is
based on the notion of a layered coin that consists of one base
layer and several endorsed layers. The base layer is generated
by the source while the endorsed layers are generated by
forwarding nodes. This layered coin mechanism makes
SMART withstand a wide range of cheating actions (or
attacks) such as layer injection attack, nodular tontine attack
and submission refusal attack. Pi is a variant of SMART
which combines reputation- and credit-based schemes. In Pi,
if and only if the messages arrive at the destination, the
forwarding nodes can get credits from the source node. To
achieve fairness, the forwarding node still can get good
reputation from a trusted authority (TA) for the failure
forwarding.
MobiCent [73] also makes use of a Trusted Third Party
(TTP) to store key information for nodes and provides
verification
and
payment
services.
It
uses
incentive-compatible payment mechanisms to cater to client
that wants to minimize either payment or data delivery delay
and handle the edge insertion and edge deletion attacks. In
this scheme, nodes are paid for forwarding packets and the
destination makes the payment decision. No node will get
incentive to tamper with the path it reports to the destination.
C. TFT Incentive Mechanisms
TFT is based on the basic principle that “I’ll do for you as
much as you did for me” [84]. BitTorrent [82] is one of the
most popular P2P (peer-to-peer) systems using direct TFT
reciprocity strategy. In BitTorrent, a user’s download rate is
proportional to its upload rate. The TFT incentive scheme
often suffers from bootstrapping problems and injecting of
fake messages.
is a pair-wise TFT
Incentive-Aware Routing [83]
incentive routing protocol which combines generosity and
contrition. The generosity addresses the bootstrapping and
absorbs transient asymmetries, while contrition prevents
mistakes from causing endless retaliation.
Buttyán et al. [84] proposed a mechanism to discourage
selfish behaviors based on the principle of barter: a user who
trades in messages can download a limited volume of
messages from another user if he/she can give the same
volume of messages in return. In this scheme, digital
signature and reputation mechanism can be used to prevent
the injection of fake messages.
RELICS [85] is a mechanism to combat selfishness in
energy-constrained DTNs. The incentive mechanism used in
11
RELICS is called reciprocity of service. Every node is given
an explicit rank based on its transit behavior (i.e., forwarding
messages originated from others). The rank of a node is
accumulated when
in relaying
the node participates
messages, whereas the rank is decreased when the node sends
message. Based on the rank of the source node, messages
originated from high ly-ranked nodes are given priority over
lowly-ranked ones. Furthermore, RELICS, for the first time,
takes energy into consideration for the incentive mechanism
in DTNs. In the scheme, it considers energy to be the core
rationale behind selfishness, and each node is allowed to set a
delivery threshold and to adapt its energy depletion rate
based on its rank such that the rate is merely enough to
achieve the desired delivery ratio.
Existing credit-based and TFT incentive mechanisms do
not take social properties into consideration directly.
However, as traditional incentive mechanisms, they can be
applied to social-based routing protocols. Furthermore,
credit-based and TFT incentive mechanisms with social
awareness will be interesting research topics.
D. Trust-Based Strategies
Social selfishness is closely related to not only the
willingness in tention for forwarding but also the trusted
relationship between nodes. Furthermore, more relationship
trust implies stronger social tie between nodes, which can be
used in effective relay node selections during the course of
forwarding. Trust-based strategies often establish trusted
relationship to complete trusted routing by coping with social
selfishness.
Considering social selfishness and
following
the
philosophy of design for users, SSAR [70] allows user
selfishness and improves performance by considering user
willingness, resource constraints, and contact opportunity
when selecting relays. User willingness values can be
configured via user interface in the mobile device.
Chen et al. [63] integrated social trust and QoS trust into a
composite trust metric for determining the best message
carrier among new encounters for message forwarding. They
consider healthiness and cooperativeness for social trust to
account for a node’s trustworthiness for message delivery,
and connectivity and energy for QoS trust to account for a
node’s QoS capability to quickly deliver the message to the
destination node. Trust-threshold based routing (TTBR) was
proposed which designs trust thresholds for determining the
trustworthiness of node acting as a recommender or as the
next message carrier. TTBR is distributed and does not
require a complicated credit management system.
Considering the social network structure and its dynamics,
Trifunovic et al. [86] proposed two approaches for social
trust establishment that are robust to sybil attacks: explicit
social trust and implicit social trust. Explicit social trust is
based on consciously established friend ties by building a
robust tree-like graph of paired users. Implicit social trust
leverages mobility properties using complex network tools
and builds another graph up to two-hops based on the
familiarity of surrounding peers and the similarity to
reinforce trust in a user.
12
Mtibaa and Harras [87] leveraged social information and
proposed six trust based filters to establish trustworthy
communications over mobile opportunistic networks. The six
filters couple three socially-aware estimators of trust
including common interests, common friends, and the
distance in the social graph, with two major techniques of
trust
establishment
including Relay-to-Relay
and
Source-to-Relay. It has been shown that the trust filters yield
a fair trade-off between trust and success rate.
Trust-based strategies provide more effective and secure
forwarding solutions. But the establishment of trust may face
the difficulty that centralized mechanisms cannot be easily
deployed
in
infrastructure-less networks
such
as
opportunistic networks. It is also challenging to efficiently
and effectively propagate the social information.
VI. SOCIALLY-AWARE DATA DISSEMINATION
Large volume of data is being generated every day. Many
content-based services or information are on the fly in the
mobile networks. Meanwhile, with
the
emerging
User-Generated Content (UGC) service, the users are not
only the consumers but also the producers of the content. The
content-based services wish to push the data to their
describers, while people wish to produce and share the
content with their friends. Consequently, there is an
increasingly crucial demand to data dissemination in
multi-point asynchronous manner
in practice. With
multi-point asynchronous communication , the destination of
communication
can be
a group of nodes,
and
communications between nodes use asynchronous way. From
the social perspective, people always group into communities
and their behaviors are regular. These social properties are
beneficial for improving the efficiency of data dissemination.
The Publish/Subscribe (Pub/Sub for short) paradigm
recently emerges as a promising solution
to data
dissemination thanks to the decoupling characteristic. It has
attracted a lot of attention from many researches. The
decoupling characteristics give the advantage of removal of
static dependencies in a distributed environment, which is
beneficial in supporting highly dynamic and decentralized
systems [25].
In Pub/Sub paradigm, three roles are usually deployed:
publisher, subscriber and broker. The publisher is the
information producer which issues newly-detected events
without having to specify the receiver. The subscriber is the
information consumer which expresses their interest in
certain events without knowledge of what the publisher
might be. The broker is the interface between the publisher
and subscriber which provides routing, event matching, and
filtering services, etc. In the context of data dissemination ,
the nodes providing the contents are publishers. The nodes
that are interested in the contents are the subscribers. The
relay nodes are the brokers. As a consequence, mobile nodes
can be publisher, subscriber or broker alternatively [25].
Recent works about socially-aware data dissemination fall
into one of two categories depending on different point of
view. The first category is based on solicitation and cache,
from the viewpoint of nodes. For a node in this context,
besides its interested content, it can cache the uninterested
content for others to improve the dissemination efficiency.
However, the buffer size of any node is limited. It is
impossible to store all encountering contents. Frequent buffer
replacement will induce considerable consumption of power
and decrease the efficiency of the network. Thus, soliciting
and caching appropriate set of contents for future distribution
according to local and/or global environments is an effective
way to improve the efficiency of data dissemination and to
reduce energy consumption. The second category is based on
forwarding, from a perspective of the content. People sharing
common interest and/or activities can build communities
which facilitate communications and information sharing
between them. In this regard, how to find the proper
forwarders to carry them to the destination community as
quickly as possible
the core of successful data
is
dissemination.
In this section we will give an overview of solicitation and
cache based approaches, and forwarding based approaches
respectively. And most algorithms in these two research
directions employ similar mechanism with PodNet [88].
Therefore, we will give a simple introduction to the PodNet
firstly. In PodNet, the improvement approach for content
distribution focuses on solicitation and five solicitation
strategies are proposed. Accordingly, we classify PodNet into
the first category.
A. Solicitation and Cache Based Approaches
The PodNet project is a Pub/Sub paradigm application for
data dissemination in wireless ad hoc networks. In PodNet,
contents are organized into feed channels. Users subscribe to
channels they are interested in. Users are associated in
pair-wise way when they come into the transmission range of
each other. The behavior of the scheme can be described as a
receiver-driven broadcasting paradigm. The intermediate
nodes are allowed to solicit and cache the unsubscribed
contents (i.e. uninterested contents) according to their
available buffers so that it can better serve its potential
encounters in the future [89].
With PodNet, the cache structure of a node is separated
into two parts: private cache and public cache. The private
cache is used for storing subscribed channels and the public
cache is used for public channels that are solely for
redistribution. It is assumed that the capacity of private cache
is large enough to store the interested contents while the
public cache is limited. Upon pair -wise contacts, nodes first
exchange the subscribed content which is matched using
Bloom Filter. Then nodes use the remaining connection time
to update and download contents for public cache by using
certain solicitation strategy. Five solicitation strategies are
presented,
including Most Solicited, Least Solicited,
Uniform, Inverse Proportional, and No Caching. These
strategies are based on popularity of content which can be
obtained and maintained
locally by exploiting
the
information from the requests of other peers except for the
No Caching strategy. For example, Most Solicited caches the
most popular content while Least Solicited caches the least
popular content.
Due to buffer size limits, there are two key issues with the
Pub/Sub paradigm, including soliciting proper unsubscribed
content to cache and replicating proper unsubscribed content
when the buffer is full.
Ma et al. [90] proposed to make a soliciting and caching
decision by jointly considering the history encounter
information, the content preferences of the subscribers and
the popularity of the contents. Using these information, each
node evaluates all contents waiting for caching and solicits
those unsubscribed contents it prefers to cache for future
distribution.
Chuah et al. [91] proposed a data model that categorizes
data into sets. Each category has a certain set of associated
keywords. Users’ interests are described by using the same
keyword space as the data model. When two nodes encounter,
each node exchanges the meta-data descriptions of stored
data and counts how many nodes obtain data from itself in an
observation window. If the data is of interest, the node will
request for it. Otherwise, a node will request for data of no
interest only with a probability which is dynamically adjusted
based on the average rewards achieved from storing data of
this category in the previous observation window.
In the above-mentioned approaches, every node makes
independent dissemination decisions. Ma et al. [92]
presented a cooperative cache-based data dissemination
framework (CCCDF) to carry out the cooperative soliciting
and caching strategies for encountering nodes. Considering
two encountering nodes as a group, CCCDF produces
cooperative dissemination decisions based on all relevant
information such as nodes’ mobility characteristics and
content subscriptions. Through updating and exchanging
average meeting rates, two nodes solicit the subscribed and
unsubscribed content download sequence as well as drop
sequence in case the cache is full. Based on the CCCDF, Ma
et al. proposed two cooperative content dissemination
strategies for different motivations [92]: CCCDF (Optimal)
is an optimal strategy to maximize the overall content
delivery performance while CCCDF (Max-Min) is a
cooperative strategy to share the limited network resources
among the contents in a Max-Min fairness manner no matter
how popular the content is.
Considering that mobile nodes often move to places where
they can establish communications, Jaho and Stavrakakis [87]
integrated interest social group with locality-induced social
group to enhance data dissemination. They compared two
storage strategies: selfish and cooperative. When soliciting
contents, the cooperative strategy takes into account the
interests of the nodes they will (most) likely encounter in the
future besides its own interests, whereas the selfish strategy
only caches contents of its own interest. It has been shown
that the cooperative strategy outperforms the other one.
ContentPlace [94]-[96] employs the same community
detection mechanisms as [25], but does not need any overlay
infrastructure. ContentPlace is not limited by the assumption
that the members in a community have common interest. In
contrast, it assumes that different communities have different
interests and one community have different interests either.
ContentPlace considers the relationship of one user with
13
several communities, that is to say, users belong to several
different social communities. It assumes
that users
automatically learn the time spent on the communities, the
interested data types and the data spread in communities.
These kinds of information are used to evaluate the utility of
each encountered data. The core of ContentPlace is to select a
set of data objects to cache in order to maximize the local
utility of its buffer by solving the multi-constrained 0-1
knapsack problem. In addition, ContentPlace also considers
five policies to evaluate the social weight, i.e. Most
Frequently Visited (MFV), Most Likely Next (MLN), Future
(F), Present (P), and Uniform Social (US). The simulations
provide best results in the Most Likely Next and Future
policies.
B. Forwarding Based Approaches
The key factor of forwarding based approaches is how to
select appropriate nodes in order to complete the data
dissemination as quickly as possible. Human beings’ social
properties are beneficial to predicting the information of
destination and meeting probability. Community and
centrality are two popular properties in socially-aware data
dissemination. In the first place, people with common
interests always come into a community and there are some
“central” people in a community. The communities are the
sole destination of relevant contents and can help the content
find their destination easily. Central nodes can help the
content travel quickly in community. Additionally, people
always follow regular mobility patterns. They always access
several locations frequently. The mobility regularity can help
to select appropriate
relay nodes
to minimize
the
intermediate nodes and improve the efficiency of data
dissemination.
In [25], a socio-aware overlay is built for Pub/Sub
communications by using the community and centrality
concepts. It detects community dynamically in which
community members are well connected, implying that
socially they share the same interests with high probability.
Nodes with high closeness centrality in communities are
selected as brokers to construct an overlay to facilitate
multi-point Pub/Sub event dissemination. Since nodes with
high closeness centrality have the best visibility to other
nodes in the community, it will be relatively reliable to
delivery contents to any (other) member of the community
through these nodes. However, it is hard to maintain the
overlay network because nodes with highest centrality values
may change over time (due to e.g. mobility).
SocialCast [97] is an interest-based routing framework for
Pub/Sub paradigm and it is based on utility function to select
best carriers. The approach assumes that socially-bound
hosts are likely to be co-located regularly. It exploits
prediction based on metrics of social interaction (probability
of a user to be co-located with another one sharing the same
interests and change degree of connectivity) to identify the
best information carriers, which implies the dissemination of
messages through matching the subscriber’s interests.
SocialCast asks a publisher to insert r copies. When a better
carrier is encountered, only one copy is removed from the
local buffer and sent to the new carrier. Therefore, at any
time, the network contains at most r copies of the message.
This approach to message distribution is also explored by
other solutions such as Spry and Wait [98]. SocialCast works
well when a community is interested in the same type of
contents, but it is not clear how it works in more general
settings.
Li and Wu [26] proposed a Mobile cOmmunity-based
Pub/Sub (MOPS) scheme which utilizes the long-term social
network properties to facilitate content-based services.
MOPS defines the closeness metric based on nodes’ temporal
and spatial encounter information to depict the neighboring
relationship between nodes. The closeness-based local
community is defined as a clique of nodes where any
neighboring relationship is stronger than an adjustable
threshold. It considers not only the direct but also the indirect
neighboring relationship. MOPS combines push and pull
strategies
and
determines
the boundary between
communities as
the push-pull boundary. Within
the
community, nodes broadcast interests and publishers send
contents directly to the nodes. Brokers are then deployed as
the interfaces to match the interests and bridge communities.
Habit [27] leverages information about nodes’ regularity
of movement and their social network (or network of interest)
to construct regularity graph and interest graph respectively.
The regularity graph is responsible for describing the
familiar stranger relationship between nodes with temporal
and spatial relevant. If a node encounters anther node
frequently enough, it becomes a familiar stranger in a regular
time period. The interest graph is responsible for describing
the relationship of interested content transmission between
nodes. Using these two graphs, Habit enables relevant
content to reach interested nodes while minimizing the
computation and communication load on uninterested
intermediate nodes, while still achieving a high delivery rate.
Fan et al. [99] addressed data dissemination among
several communities and proposed to make decision on the
routing trajectory based on semi-Markov analytical model.
They defined geo-community based on the geographic
regularity of human mobility and geo-centrality as the super
user among communities. They formulated the super user
route design as a combinational optimization problem of
Convex Optimization and Traveling Salesman Problem to
achieve the goal of minimizing the total duration and
guaranteeing the required data dissemination probability.
Forwarding based approaches exploit social information
in the working environment and take into consideration
social relationships between users to select the data object to
exchange. The networks are generally divided into several
communities that provide more global information. However,
due to the dynamic topology of the network, more resources
will be consumed to detect and maintain the communities. In
addition, there is no consideration of the multi-interest
scenario. One challenge is that we must find a way to achieve
a fairly distributed delivery rate.
The main difference between solicitation and cache based
approaches and forwarding based approaches lies in what
they concern about. Solicitation and cache based approaches
14
emphasize more on messages’ properties. They usually
consider messages’ popularity and dissemination rate. Thus
they are suitable to content-aware environments. For
instance, the community is detected according to individuals’
interest. While forwarding based approaches emphasize
more on individual’s properties. They pay more attention to
the individual’s status in order to disseminate faster.
Therefore, they have been more widely adopted.
In addition, the cache freshness problem is attracting
researchers’ attention. Some messages such as news are
time-sensitive. As a result, to achieve cache freshness, the
date should be the newest in cache. However, previous
researches usually maintain the freshness of cached data by
refreshing periodically. Recently, Cao et al. [100], for the
first time, proposed a scheme to efficiently maintain the
cache freshness by organizing the caching nodes as a tree
structure during data access. Each node in the tree is
responsible for refreshing the data cached at its children in a
distributed and hierarchical manner.
VII. OPEN ISSUES
In previous sections, we have reviewed the state of the art
in the emerging field of SAN. Accordingly, we try to cover
the majority of categories of related research efforts, though
the content is not projected to be exhaustive. From the
analysis and comparison of these methods, we can conclude
that the social properties are the potential improvement
strength to mobile network design. There are large spaces to
improve the efficiency of protocols and algorithms by
exploring comprehensive use of social properties. In addition,
although security and mobility model from social aspects are
not involved in this paper, they are also hot research topics in
SAN and many research works have been conducted, e.g.
[63], [101]-[112]. Besides, quite a number of research
projects and initiatives closely related to SAN have been
launched in recent years. A list of these research projects is
provided in Table II. Despite the considerable amount of
ongoing research, the SAN research community is very
young. Many challenges remain yet to be addressed. On the
other hand, there are a lot of opportunities to improve the
efficiency and effectiveness of SAN to accommodate new
applications and services. In this section, several open issues
will be outlined.
A. Mobile Social Sensing and Learning
In the context of SAN, protocol and algorithm design takes
into account relevant social properties as main basis. It is
necessary to sense and learn the context information in order
to obtain the social properties. In present works, the social
properties mainly come from the analysis of the experiment
data and most existing studies are based on analytical or
simulation models. Then comes the challenge that how to
bring SAN into real applications with mobile social sensing
and learning.
Basically, it might be easy to obtain raw data of the context
information such as time, location, action, etc. by using
mobile devices equipped with appropriate sensors. In this
aspect, most of the projects listed in Table II have provided
Project
Organization(s)
SocialNets
FET
TABLE II
SUMMARY OF SOCIALLY-AWARE NETWORKING RELATED PROJECTS
Research Area(s)
Social networks for the delivery and acquisition of content covering
online opportunistic wireless network
PeerSoN
MetroSense
A Million People
PodNet
GroupMedia
Haggle
P2P infrastructure supporting features of online social networks.
KTH, EPFL, NTU
Singapore, etc
Dartmouth College Applications, classification techniques, and sensing paradigms for
mobile phones capable of societal-scale sensing
Mobile network, epidemiology, urban planning, and social science.
Content distribution protocol and system on mobile devices.
Perceptual socially-aware applications for cell phones and PDAs.
Situated and autonomic communications
CSG ETH Zürich
MIT
FU FET
15
URL
http://www.social-nets.eu/
http://www.peerson.net/
http://metrosense.cs.dartmouth.edu/
http://www.amillionpeople.net/
http://podnet.ee.ethz.ch/
http://groupmedia.media.mit.edu/
http://www.haggleproject.org/
large amounts of datasets of inevitable value. While, since
majority of the existing data is about information like time,
location, which is not enough for further analysis on social
behavior, novel mobile social sensing technologies need to be
tested, especially for collection of users’ words, actions, etc.
After being collected, the useful information needs to be
extracted from the sensor data for real usage, as a result of the
fact
that, different applications need different social
properties and different social properties require different
data at different scales. For example, the human mobility
pattern is sensitive to the time and location at individual scale,
while the community and centrality are sensitive to the
interest or interaction at group scale. This makes the mobile
social learning a vital issue. In other words, an SAN
application should be able to analyze the collected data or
classify them according to the social features they reflect. In
some occasions, the application should also be able to predict
the trend or future pattern to afford users a smarter life, for
example, remind a meeting, help schedule, or make an
appointment. To this end, the mobile devices need to support
large storage capacity and computation capacity. In this
regard, mobile cloud computing is a suitable solution to
computing and mining big data for very large numbers of
users [101], with the ability of providing available computing
and storage resources.
Based on the above statements, both mobile social sensing
and learning need further research on improving their own
effectiveness and efficiency considering the points of
application and social properties.
B. Privacy
Mobile devices, which are used to monitor people’s
context information , record their preferences and behaviors,
and hence possess the users’ private information. Most of the
protocols and algorithms for SAN require users to share their
personal information such as physical location [114],
preference, and social relation. Therefore, the privacy issue
for mobile users in SAN becomes crucial. A generalized
privacy policy is hard to implement because the granularity
of privacy concern may differ from user to user, even the
same user has different sensitivities to the same information
in different applications [115].
The solutions to privacy presented by existing works are
usually encryption or access control to private information .
Encryption only exposes the sensitive information to the
identified users, which is not suitable to SAN due to the
general requirement of central server. Information access
control allows users to disclose their private information to
“close” users. For instance, most mobile social networking
services allow a user to share his/her location with his/her
friends, even friends of friends (FoF). In the context of SAN,
access control to private information becomes difficult due to
various relationships involved. There is a tradeoff between
obtaining accurate social properties for socially-aware
networking protocol design and protecting privacy.
C. Node Selfishness
Dealing with selfish nodes is an important but challenging
issue in SAN. As mentioned before, node selfishness in this
context can be classified into social selfishness and
individual selfishness. Many research efforts on incentive
mechanisms are exploited to stimulate the individual selfish
nodes to cooperate [116]. However, the social selfishness is
usually used to help select trusted nodes in routing protocols
and they are not exploited sufficiently, particularly in
incentive mechanisms. It is necessary to detect and deal with
the individually selfish nodes and socially selfish nodes
differently according to the context information, particularly
social information. Addressing the problem of selfish nodes
effectively will be beneficial to not only the design of routing
protocols and data dissemination algorithms, but also
handling of privacy and security problems. As a consequence,
node selfishness is worth investigation.
D. Scalability
Scalability is an open problem that is seldom considered in
previous works [117]. Indeed it should be taken into account
on each level of SAN’s architecture. Here we take the routing
and data dissemination, and application levels as examples.
On the level of routing and data dissemination, most of the
existing protocols and algorithms require nodes to store their
encounter information with others or context information to
estimate the appropriate relays. However, mobile devices are
limited by the wireless spectrum and onboard resources,
especially the low buffer storage and energy. This causes the
problem of scalability and makes the routing protocols or
data dissemination algorithms badly compatible with the
increasing node density of the network in many aspects: (1) If
new nodes join the network or some nodes are powered off, a
new route between two nodes needs to be detected, and this
costs energy and time; (2) When the network is large, finding
an effective route will be difficult, especially in situations
where the topology is dynamic; (3) Larger memory space is
needed to store the increased encounter information; (4)
Sometimes, there are multiple feasible routes to disseminate
the information ; in this case, optimal route selection and
route management become challenging problems. All these
points require mobile devices to have good capacity in terms
of buffer, computation and energy efficiency.
On the level of application (or system), scalability is also
an important issue. For a robust system, the ability of being
able to be extended with new features easily is very necessary
when facing with changing requirements. When a new
device joins the system, it is necessary to integrate the device
into the system quickly, which is also an essential aspect of
scalability. Besides these scenarios, it is possible to identify
many requirements in terms of scalability in the context of
SAN.
VIII. CONCLUSION
Socially-aware networking is going to be a new hotspot of
research on network science and engineering. The close
connection between ubiquitous mobile devices and the users’
social relationships attracts researchers to explore the
potential of introducing social properties into network design.
We believe that SAN will benefit the engineering of next
generation networks as a promising paradigm.
In this paper, we have discussed basic concepts behind this
new terminology and presented a first survey of state of the
art. Through examining existing research results, it can be
seen that social properties are indeed a powerful source not
only for the design of outstanding networks, covering the
areas of routing and forwarding protocols, but also for
tackling the problems related to selfishness behavior in
routing and
forwarding situations. Additionally, we
described the major approaches on socially aware data
dissemination to make full use of the social properties.
However, a multitude of challenges remain to be addressed
before the full potential of SAN can be realized in practice.
We have examined some of these open issues in this paper.
We hope this survey will provide a better understanding of
the literature of SAN and spark new research interests and
developments in this field.
REFERENCES
[1] M. Walsh. “Gartner: mobile to outpace desktop web by 2013,” Online
Media Daily, January 2010.
[2] L. Pelusi, A. Passarella, and M. Conti, “Opportunistic networking: data
forwarding in disconnected mobile ad hoc networks,” Communications
Magazine, IEEE, vol. 44, no. 11, pp. 134–141, Nov. 2006.
[3] Handbook of Wireless Networks and Mobile Computing, 1st ed, John
Wiley & Sons, Inc., Ivan Stojmenović, New York, USA, 2002, pp.
325-346.
[4] N. D. Lane, E. Muluzzo, H. Lu, D. Peebles, T. Choudhury, and A. T.
Campbell, “A survey of mobile phone sensing,” Communications
Magazine, IEEE, vol. 48, no. 9, pp. 140–150, Sept. 2010.
[5] D. Huang, “Mobile cloud computing,” IEEE COMSOC Multimedia
Communications Technical Committee (MMTC) ELetter, 2011.
[6] M. Conti, and M. Kumar, “Opportunities in opportunistic computing,”
IEEE Computer, vol. 43, no. 1, pp. 42–50, Jan. 2010.
16
[7] P. Hui, A. Chaintreau, R. Gass, J. Scott, J. Crowcroft, and C. Diot,
“Pocket-switched networking: challenges, feasibility and implementation
issues,” Lecture Notes in Computer Science, vol. 3854, pp. 1-12, 2006.
[8] W. Chen, R. K. Guha, T. J. Kwon, J. Lee, and I. Y. Hsu, “A survey and
challenges in routing and data dissemination in vehicular ad-hoc
networks,” in Proc. IEEE Int. Conf. on Vehicular Electronics and Safety,
Columbus, 2008, pp. 328–333.
[9] F.-J. Wu, Y.-F. Kao, and Y.-C. Tseng, “From wireless sensor networks
toward cyber physical system,” Pervasive and Mobile Computing, vol, 7,
no. 4, pp. 397–413, Aug. 2011.
[10] K. Yang, X. Cheng, L. Hu, and J. Zhang, “Mobile social networks:
International Journal of
state-of-the-art and a new vision”,
Communication Systems, vol. 25, pp. 1245-1259.
[11] M. R. Schurgot, C. Comaniciu, K. Jaffrès-Runser, “Beyond traditional
DTN routing: social networks for opportunistic communication,”
Communications Magazine, IEEE, vol.50, no.7, pp.155-162, July 2012.
[12] D. Popovici, M. Desertot, S. Lecomte, T. Delot, “A framework for mobile
and context-aware applications applied to vehicular social networks”,
Social Network Analysis and Mining, pp.1-12. May 2012
[13] B. Guo, Z. Yu, X. Zhou, and D. Zhang, “Opportunistic IoT: Exploring the
social side of the internet of things,” CSCWD 2012, pp.925-929, 23-25
May 2012.
[14] D. Zhang, B. Guo, and Z. Yu, “The Emergence of Social and Community
Intelligence,” Computer, vol. 44, no. 7, pp. 21-28, July 2011.
[15] S. Milgram. “The small world problem,” Psychology Today, vol. 1, no. 1,
pp. 60–67, May 1967.
[16] D. Saxbe. “Six degrees of separation: Two new studies test ’six degrees of
separation’ hypothesis,” Psychology Today, 2003.
[17] W. X. Zhou, D. Sornette, R. A. Hill, and R. I. M. Dunbar, “Discrete
hierarchical organization of social group sizes,” Proc. the Royal Society
B, vol. 272, no. 1561, pp. 439–444, Feb. 2005.
[18] D. J. Watts, and S. H. Strogatz. “Collective dynamics of “small-world”
networks,” Nature, vol. 393, pp. 440–442, Jun. 1998.
[19] S. Wasserman, and K. Faust, “Social network analysis: methods and
applications, structural analysis in the social sciences series,” Cambridge
Univ. Press, 1994.
[20] O. Serrat, “Social network analysis”, Knowledge Solutions, Feb 2009.
[21] F. Li, and J. Wu, “LocalCom: A Community-based epidemic forwarding
scheme in disruption-tolerant networks,” in Proc. IEEE SECON '09,
Florida, 2009, pp.1–9.
[22] E. M. Daly, and M. Haahr, “Social network analysis for information flow
in disconnected delay-tolerant MANETs,” IEEE Transactions on Mobile
Computing, vol.8, no.5, pp. 606–621, May. 2009.
[23] E. M. Daly, and M. Haahr, “Social network analysis for routing in
disconnected delay-tolerant MANETs,” in Proc. ACM MobiHoc’07,
Montreal, 2007, pp. 32–40.
[24] P. Hui, J. Crowcroft, and E. Yoneki, “Bubble Rap: social –based
forwarding in delay tolerant networks,” in Proc. ACM MobiHoc’08,
Hong Kong, 2008, pp. 241–250.
[25] E. Yoneki, P. Hui, S. Chan, and J. Crowcroft, “A socio-aware overlay for
publish/subscribe communication in delay tolerant networks,” in Proc.
ACM MSWiM’07, Chania, 2007, pp. 225–234.
[26] F. Li, and J. Wu, “MOPS: providing content-based service in
disruption-tolerant networks,” in Proc. ICDCS’09, Florida, 2009, pp.
526–533.
[27] A. J. Mashhadi, S. B. Mokhtar, and L. Capra, “Habit: leveraging human
mobility and social network for efficient content dissemination in delay
tolerant networks,” in Proc. IEEE WoWMoM’09, London, 2009, pp. 1–6.
[28] N. Wang, and E. Yoneki, “Impact of social structure on forwarding
algorithms in opportunistic networks,” 2011 Int. Conf. on Selected
Topics in Mobile and Wireless Networking (iCOST) , Shanghai, 2011,
pp. 82–88.
[29] L. C. Freeman, “A set of measures of centrality based on betweenness,”
Sociometry, vol. 40, no. 1, pp. 35–41, Mar. 1977.
[30] L. C. Freeman, “Centrality in social networks conceptual clarification,”
Social Networks, vol. 1, no. 3, pp. 215–239, 1978-1979.
[31] Y. Zhang, and J. Zhao, “Social network analysis on data diffusion in
delay tolerant networks,” in Proc. ACM MobiHoc’09, New Orleans,
2009, pp. 345–346.
[32] D. Katsaros, N. Dimokas, and L. Tassiulas, “Social network analysis
concepts in the design of wireless ad hoc network protocols,” Network,
IEEE, vol.24, no.6, pp.23–29, Nov.-Dec. 2010.
[33] A. Mtibaa, A. Chaintreau, J. LeBrun, and E. Oliver, A.-K. Pietilainen,
and C. Diot. “Are you moved by your social network application?” in
Proc. WOSN’08: Wksp. Online social networks, New York, 2008, pp.
67–72.
[34] M. S. Granovetter, “The strength of weak ties,” The Am. J. Sociology, vol.
78, no. 6, pp. 1360-1380, May 1973.
[35] M. González, C. Hidalgo, and A.-L. Barabási, “Understanding individual
human mobility patterns,” Nature, vol. 453, no. 7196, pp. 779–782, Mar.
2008.
[36] W, Hsu, T . Spyropoulos, K. Psounis, and A. Helmy, “Modeling
time-variant user mobility in wireless mobile networks,” in Proc. IEEE
INFOCOM’07, Anchorage, 2007, pp. 758–766.
[37] S. Yang, X. Yang, C. Zhang, and E. Spyrou, “Using social network theory
for modeling human mobility,” Network, IEEE, vol. 24, no. 5, pp.6–13,
September-October 2010.
[38] P. Hui, and J. Crowcroft, “How small labels create big improvements,”
IEEE PerCom Workshops '07, New York, 2007, pp.65–70.
[39] A. Chaintreau, P. Hui, J. Crowcroft, C. Diot, R. Gass, and J. Scott,
“Impact of human mobility on the design of opportunistic forwarding
algorithms,” in Proc. IEEE INFOCOM’06, Barcelona, 2006, pp.
606–620,
[40] L. Danon, J. Duch, A. Diaz-Guilera, and A. Arenas, “Comparing
community structure identification,” Journal of Statistical Mechanics:
Theory and Experiment, vol. 2005, no. 9, pp. 1–10, Sep. 2005.
[41] E. Bulut, and B. K. Szymanski, “Friendship based routing in delay
tolerant mobile social networks,” in Proc. IEEE GLOBECOM’10, Miami,
2010, pp.1–5.
[42] M. Musolesi, P. Hui, C. Mascolo, and J. Crowcroft, “Writing on the clean
slate: Implementing a socially-aware protocol in Haggle,” in Proc. IEEE
WoWMoM’08, Newport Beach, 2008, pp. 1–6.
[43] T. Zhou, R. R. Choudhury, and K. Chakrabarty, “Diverse routing:
exploiting social behavior for routing in delay-tolerant networks,” in Proc.
CSE '09, vol. 4, pp. 1115–1122, Aug. 2009.
[44] M. Musolesi, S. Hailes, and C. Mascolo. “Adaptive routing for
intermittently connected mobile ad hoc networks,” in Proc. IEEE
WoWMoM’05, Taormina - Giardini Naxos, 2005, pp. 183–189.
[45] A. Lindgren, A. Doria, and O. Schelén, “Probabilistic routing in
intermittently connected networks,” Service Assurance with Partial and
Intermittent Resources, vol. 3126, pp. 239-254, 2004.
[46] J. Leguay, T. Friedman, and V. Conan, “DTN routing in a mobility
pattern space,” in Proc. ACM SIGCOMM WDTN '05. Philadelphia, 2005,
pp. 276–283.
[47] B. Burns, O. Brock, and B. N. Levine, “MV routing and capacity building
in disruption tolerant networks,” in Proc. IEEE INFOCOM’05, Miami,
2005.
[48] C. Boldrini, M. Conti, and A. Passarella, “Exploiting users’ social
relations to forward data in opportunistic networks: the HiBOp solution,”
Pervasive and Mobile Computing, vol. 4, no. 5, pp. 633–657, Oct. 2008.
[49] A. Mtibaa, M. May, C. Diot, and M. Ammar, “PeopleRank: social
opportunistic forwarding,” in Proc. IEEE INFOCOM’10, San Diego,
2010, pp. 1–5.
[50] J. Burgess, B. Gallagher, D. Jensen, and B. N. Levine, “Maxprop: routing
in Proc. IEEE
for vehicle-based disruption-tolerant networks,”
INFOCOM’06, Barcelona, 2006, pp. 1–11.
[51] Q. Yuan, I. Cardei, and J. Wu, “Predict and relay: an efficient routing in
disruption-tolerant networks,” in Proc. ACM MobiHoc’09, New Orleans,
2009, pp. 95–104.
[52] F. Nazir, J. Ma, and A. Seneviratne, “Time critical content delivery using
predictable patterns in mobile social networks,” in Proc. CSE '09,
Vancouver, 2009, pp.1066–1073.
[53] W. Gao, G. Cao, T . La Porta, and J. Han, “On Exploiting Transient Social
Contact Patterns for Data Forwarding in Delay Tolerant Networks,”
IEEE Transactions on Mobile Computing, January, 2013.
[54] W. Gao, Q. Li, B. Zhao, and G. Cao, “Social-aware multicast in
disruption-tolerant networks,” IEEE/ACM Trans. Netw., vol. PP, no. 99,
p. 1, 2012.
[55] S. Patra, S. Saha, V. Shah, S. Sengupta, K. G. Singh, and S. Nandi, “A
qualitative survey on multicast routing in delay tolerant networks,”
WiMo/CoNeCo 2011, Ankara, 2011, pp. 197–206.
[56] M. Radenkovic, and A. Grundy, "Efficient and Adaptive Congestion
Control for Heterogenious Delay Tolerant Networks", Ad Hoc Networks,
vol. 10, no. 7, Sep. 2012, pp. 1322-1345.
[57] K.-i. Kawarabayashi, F. Nazir, and H. Prendinger, “Message duplication
reduction in dense mobile social networks,” in Proc. IEEE ICCCN’10,
Zurich, 2010, pp.1–6.
[58] V. Erramilli, M. Crovella, A. Chaintreau, and C. Diot. “Delegation
forwarding,” in Proc. ACM MobiHoc’08, Hong Kong, 2008, pp.
251–260.
[59] J. M. Pujol, A. L. Toledo, and P. Rodriguez, “Fair routing in delay
tolerant networks,” in Proc. IEEE INFOCOM’09, Rio de Janeiro, 2009,
pp. 837–845.
[60] M. Radenkovic, and A. Grundy, “Congestion aware forwarding in delay
tolerant and social opportunistic networks,” in Proc. WONS’11,
Bardonecchia, 2011, pp. 60–67.
17
[61] M. A. Islam, and M. Waldvogel, “Optimizing message delivery in
mobile-opportunistic networks,” 2011 Baltic Congress on Future
Internet Communications, Riga, 2011, pp. 134–141.
[62] P. Yuan, H. Ma, and P. Duan, “On exploiting few strangers for data
forwarding in delay tolerant networks,” in Proc. MSN’11, Beijing, 2011,
pp. 282–288.
[63] I.-R. Chen, F. Bao, M. Chang, and J.-H. Cho, “Trust management for
encounter-based routing in delay tolerant networks,” in Proc. IEEE
GLOBECOM’10, Miami, 2010, pp.1–6.
[64] A. Panagakis, A. Vaios, and I. Stavrakakis, “On the effects of cooperation
in DTNs,” in Proc. COMSWARE’07, Bangalore, 2007, pp.1–6.
[65] K. Xu, P. Hui, V. O. K. Li, J. Crowcroft, V. Latora, and P. Lio, “Impact of
altruism on opportunistic communications,” in Proc. ICUFN’09, Hong
Kong, 2009, pp.153–158.
[66] P. Hui, K. Xu, V. O. K. Li, J. Crowcroft, V. Latora, and P. Lio,
“Selfishness, altruism and message spreading in mobile social networks,”
in Proc. IEEE INFOCOM’09, Rio de Janeiro, 2009, pp. 1–6.
[67] O. R. Helgason, F. Legendre, V. Lenders, M. May, and G. Karlsson,
“Performance of opportunistic content distribution under different levels
of cooperation,” 2010 European Wireless Conference (EW), Lucca,
2010, pp. 903–910.
[68] Y. Li, G. Su, and Z. Wang, “Evaluating the effects of node cooperation on
DTN routing,” AEU - International Journal of Electronics and
Communications, vol. 66, no. 1, pp. 62–67, Jan. 2012.
[69] Y. Li, G. Su, D. O. Wu, D. Jin, L. Su, and L, Zeng, “The impact of node
tolerant networks,” IEEE
selfishness on multicasting
in delay
Transactions on Vehicular Technology, vol. 60, no. 5, pp. 2224–2238,
Jun. 2011.
[70] Q. Li, S. Zhu, and G, Cao, “Routing in socially selfish delay tolerant
networks,” in Proc. IEEE INFOCOM’10, San Diego, 2010, pp.1–9.
[71] J. R. Douceur, “The sybil attack,” Lecture Notes of Computer Science,
vol. 2429, pp. 251-260, 2002.
[72] E. Friedman, and P. Resnick, “The social cost of cheap pseudonyms,”
Journal of Economics and Management Strategy, vol. 10, no. 2, pp.
173–199, Summer 2001.
[73] B. B. Chen, and M. C. Chan, “MobiCent: a credit-based incentive system
for disruption tolerant network,” in Proc. IEEE INFOCOM’10, San
Diego, 2010, pp.1–9.
[74] N. Li, and S. K. Das, “RADON: reputation-assisted data forwarding in
opportunistic networks,” in Proc. of 2nd Int. Workshop on Mobile
Opportunistic Networking, Pisa, 2010, pp. 8–14.
[75] A. Mei, and J. Stefa, “Give2get: forwarding in social mobile wireless
networks of selfish individuals,” in Proc. ICDCS’10, Genova, 2010, pp.
488–497.
[76] L. Wei, H. Zhu, Z. Cao, and X. Shen, “MobiID: a user-centric and
social-aware reputation based incentive scheme for delay/disruption
tolerant networks,” Ad-hoc, Mobile, and Wireless Networks, Lecture
Notes in Computer Science, vol. 6811, pp. 177–190.
[77] G. Bigwood, and T . Henderson, “IRONMAN: using social networks to
add incentives and reputation to opportunistic networks,” in Proc. IEEE
SocialCom’11, Boston, 2011, pp. 65–72.
[78] Z. Xu, Y. Jin, W. Shu, X. Liu, and J. Luo, “SReD: a secure
reputation-based dynamic window scheme for disruption-tolerant
networks,” in Proc. IEEE MILCOM’09, Boston, 2009, pp.1–7.
[79] W. Gao, G. Cao, M. Srivatsa and A. Iyengar, “Distributed Maintenance
of Cache Freshness in Opportunistic Mobile Networks,” in Proc. IEEE
ICDCS’12, Macau, 2012, pp. 132-141.
[80] H. Zhu, X. Lin, R. Lu, Y. Fan, and X. Shen, “SMART: a secure
multilayer credit-based incentive scheme for delay-tolerant networks,”
IEEE Transactions on Vehicular Technology, vol. 58, no. 8, pp.
4628–4639, Oct. 2009.
[81] R. Lu, X. Lin, H. Zhu, X. Shen, and B. Preiss, “Pi: a practical incentive
protocol for delay tolerant networks,” Wireless Communications, IEEE
Transactions on, vol. 9, no. 4, pp. 1483–1493, Apr. 2010.
[82] BitTorrent
Specification.
Protocol
http://wiki.theory.org/BitTorrentSpecification, 2003.
[83] U. Shevade, H. H. Song, L. Qiu, and Y. Zhang, “Incentive-aware routing
in DTNs,” in Proc. IEEE ICNP’08, Orlando, 2008, pp. 238–247.
[84] L. Buttyán, L. Dóra, M. Félegyházi, and I. Vajda, “Barter trade improves
message delivery in opportunistic networks”, Ad Hoc Networks, vol. 8,
no. 1, pp. 1-14, Jan. 2010.
[85] M. Y. S. Uddin, B. Godfrey, and T. Abdelzaher, “RELICS: in-network
realization of incentives to combat selfishness in DTNs,” in Proc. IEEE
ICNP’10, Kyoto, 2010, pp. 203–212.
[86] S. Trifunovic, F. Legendre, and C. Anastasiades, “Social trust in
opportunistic networks,” in Proc. IEEE INFOCOM’10, San Diego, 2010,
pp. 1–6.
[87] A. Mtibaa, and K. A. Harras, “Social-based trust in mobile opportunistic
networks,” in Proc. IEEE ICCCN’11, Hawaii, 2011, pp. 1–6
[88] V. Lenders, G. Karlsson, and M. May, “Wireless ad hoc podcasting,” in
Proc. IEEE SECON'07, San Diego, 2007, pp. 273–283.
[89] Y. Ma, and A. Jamalipour, “A cooperative cache-based content delivery
framework for intermittently connected mobile ad hoc networks,” IEEE
Transactions on Wireless Communications, vol. 9, no. 1, pp. 366–373,
Jan. 2010.
[90] Y. Ma, M. R. Kibria, and A. Jamalipour, “Cache-based content delivery
in opportunistic mobile ad hoc networks,” in Proc. IEEE GLOBECOM’
08, New Orleans, 2008, pp. 1–5.
[91] M. Chuah, P. Yang, and P. Hui, “Cooperative user centric information
dissemination in human content-based networks,” in Proc. IEEE
ICPAD’10, Shanghai, 2010, pp. 794–799.
[92] Y. Ma, and A. Jamalipour, “Cooperative content dissemination in
intermittently connected networks,” in Proc. IEEE ICC '09, Dresden,
2009, pp.1–5.
[93] E. Jaho, and I. Stavrakakis, “Joint interest- and locality-aware content
dissemination in social networks,” in Proc. WONS 2009, Snowbird, 2009,
pp. 173–180.
[94] C. Boldrini, M. Conti, and A. Passarella, “Design and performance
evaluation of ContentPlace, a social-aware data dissemination system for
opportunistic networks,” Computer Networks, vol. 54, no. 4, pp.
589–604, Mar. 2010.
[95] C. Boldrini, M. Conti, and A. Passarella, “Contentplace: social-aware
data dissemination in opportunistic networks,” in Proc. MSWIM’08,
Vancouver, 2008, pp. 203–210.
[96] C. Boldrini, M. Conti, and A. Passarella, “Context and resource
awareness in opportunistic network data dissemination,” in Proc. IEEE
WoWMoM’08. Newport Beach, 2008, pp.1–6.
[97] P. Costa, C. Mascolo, M. Musolesi, and G.-P. Picco, “Socially-aware
routing for publish-subscribe in delay-tolerant mobile ad hoc networks,”
IEEE JSAC Special issue on Delay-Tolerant Networks, vol. 26, no. 5, pp.
748–760, Jun. 2008.
[98] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, “Spray and wait: An
efficient routing scheme for intermittently connected mobile networks,” in
Proc. ACM SIGCOMM WDTN ’05, Philadelphia, 2005, pp. 252–259.
[99] J. Fan, Y. Du, W. Gao, J. Chen, and Y. Sun, “Geography-aware active
data dissemination in mobile social networks,” in Proc. IEEE MASS’10,
San Francisco, 2010, pp. 109–118.
[100] Q. Li, and G. Cao, “Mitigating Routing Misbehavior in Disruption
Tolerant Networks,” IEEE Transactions on Information Forensics and
Security, Vol. 7, No. 2, April 2012.
[101] F. Li, Y. Yang, and J, Wu, “CPMC: an efficient proximity malware
coping scheme in smartphone-based mobile networks,” in Proc. IEEE
INFOCOM’10, San Diego, 2010, pp. 1–9.
[102] V. Lenders, E, Koukoumidis, P, Zhang, and M, Martonosi,
“Location-based trust for mobile user-generated content: applications,
challenges and implementations,” Proc. the 9th workshop on Mobile
computing systems and applications, Santa Cruz, 2008, pp. 60-64.
18
[103] G. Palla, I. Derenyi, I. Farkas, and T, Vicsek, “Uncovering the
overlapping community structure of complex networks in nature and
society,” Nature, vol. 435, pp. 814–818, Jun. 2005.
[104] S. Fortunato, “Community detection in graphs,” Physics Reports, vol.
486, no. 3-5, pp.75–174, Feb. 2010.
[105] M. Musolesi, and C. Mascolo, “A community based mobility model for ad
hoc network research,” in Proc. ACM REALMAN’06, Florence, 2006, pp.
31–38.
[106] C. Boldrini and A. Passarella, “HCMM: modeling spatial and temporal
properties of human mobility driven by users’ social relationships,”
Computer Commun., vol. 33, no. 9, pp. 1056–1074, Jun. 2010.
[107] S. Yang, X. Yang, C, Zhang, and E, Spyrou, “Using social network theory
for modeling human mobility,” IEEE Network, vol. 24, no. 5, pp. 6–13,
Sep.-Oct. 2010.
[108] V. Borrel, F. Legendre, M. Dias de Amorim, and S. Fdida, “SIMPS: using
sociology for personal mobility,” IEEE/ACM Trans. Net., vol. 17, no. 3,
pp. 831–842, Jun. 2009.
[109] D. Fischer, K. Herrmann, and K. Rothermel, “GeSoMo - a general social
mobility model for delay tolerant networks,” IEEE 7th Int’l. Conf.
Mobile Ad Hoc and Sensor Sys, Stuttgart, 2010, pp. 99–108.
[110] A. Mei, and J. Stefa, “SWIM: a simple model to generate small mobile
worlds,” in Proc. IEEE INFOCOM’09, Rio de Janeiro, 2009, pp.
2106–2113.
[111] K. Lee, S. Hong, S. J. Kim, I. Rhee, and S. Chong, “SLAW: a new
mobility model for human walks,” in Proc. IEEE INFOCOM’09, Rio de
Janeiro, 2009, pp. 855–863.
[112] F. Ekman, A. Keranen, J, Karvo, and J. Ott, “Working day movement
model,” in Proc. 1st ACM SIGMOBILE Wksp. Mobility Models, Hong
Kong, 2008, pp. 33–40.
[113] A. Madan,M. Cebrian, D. Lazer, and A. Pentland, “Social sensing for
epidimiological behavior change,” in Proc. ACM UBICOMP’10,
Copenhagen, 2010, pp. 291–300.
[114] K. Rechert, K. Meier, R. M. Zahoransky, D. Wehrle, D. von Suchodoletz,
B. Greschbach, S. Wohlgemuth, and I. Echizen, “Reclaiming Location
Privacy in Mobile Telephony Networks - Effects and Consequences for
Providers and Subscribers”, IEEE Systems Journal, Vol. 7, No. 2, June
2013, pp. 211-222.
[115] N. Kayastha, D. Niyato, P. Wang, and E. Hossain, “Applications,
architectures, and protocol design issues for mobile social networks: a
survey,” Proceedings of the IEEE, vol. 99, no. 12, pp. 2130–2158, Dec.
2011.
[116] G. Vakili, and S. Khorsandi, “Self-Organized Cooperation Policy Setting
in P2P Systems Based on Reinforcement Learning”, IEEE Systems
Journal, Vol. 7, No. 1, March 2013, pp.151-160.
[117] X. Wang, and S. Li, “Scalable Routing Modeling for Wireless Ad Hoc
Networks by Using Polychromatic Sets”, IEEE Systems Journal, Vol. 7,
No. 1, March 2013, pp. 50-58
|
1806.02377 | 1 | 1806 | 2018-06-06T18:41:32 | Predictive Analysis on Twitter: Techniques and Applications | [
"cs.SI"
] | Predictive analysis of social media data has attracted considerable attention from the research community as well as the business world because of the essential and actionable information it can provide. Over the years, extensive experimentation and analysis for insights have been carried out using Twitter data in various domains such as healthcare, public health, politics, social sciences, and demographics. In this chapter, we discuss techniques, approaches and state-of-the-art applications of predictive analysis of Twitter data. Specifically, we present fine-grained analysis involving aspects such as sentiment, emotion, and the use of domain knowledge in the coarse-grained analysis of Twitter data for making decisions and taking actions, and relate a few success stories. | cs.SI | cs | Predictive Analysis on Twitter: Techniques and
Applications
Ugur Kursuncu1,2, Manas Gaur1, Usha Lokala1, Krishnaprasad Thirunarayan1,
Amit Sheth1, and I. Budak Arpinar2
1Kno.e.sis Center, Wright State University, Dayton, OH, USA
2Department of Computer Science, The University of Georgia, Athens, GA, USA
{ugur,manas,usha,prasad,amit}@knoesis.org
{kursuncu,budak}@uga.edu
Abstract. Predictive analysis of social media data has attracted con-
siderable attention from the research community as well as the business
world because of the essential and actionable information it can pro-
vide. Over the years, extensive experimentation and analysis for insights
have been carried out using Twitter data in various domains such as
healthcare, public health, politics, social sciences, and demographics. In
this chapter, we discuss techniques, approaches and state-of-the-art ap-
plications of predictive analysis of Twitter data. Specifically, we present
fine-grained analysis involving aspects such as sentiment, emotion, and
the use of domain knowledge in the coarse-grained analysis of Twitter
data for making decisions and taking actions, and relate a few success
stories.
Keywords. Social media analysis, Citizen sensing, Community evolu-
tion, Event analysis, Sentiment-emotion-intent analysis, Spatio-temporal-
thematic analysis, Election prediction, Harassment detection, Mental
health, Demographic prediction, Drug trends, Stock Market prediction,
Machine Learning, Semantic Social Computing.
1
Introduction
With the growing popularity of social media and networking platforms as an im-
portant communication and sharing media, they have significantly contributed
to the decision making process in various domains. In the last decade, Twitter
has become a significant source of user-generated data. The number of monthly
active users was 330 million as of third quarter of 2017, and the number of daily
active users was 157 million as of second quarter of 2017. Moreover, nearly 500
million tweets per day are shared on Twitter. Accordingly, significant technical
advancements have been made to process and analyze social media data us-
ing techniques from different fields such as machine learning, natural language
processing, statistics, and semantic web. This amalgamation and interplay of
multiple techniques within a common framework have provided feature-rich an-
alytical tools [1, 2], leading to valid, reliable and robust solutions.
8
1
0
2
n
u
J
6
]
I
S
.
s
c
[
1
v
7
7
3
2
0
.
6
0
8
1
:
v
i
X
r
a
2
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
Twitter provides multimodal data containing text, images, and videos, along
with contextual and social metadata such as temporal and spatial informa-
tion, and information about user connectivity and interactions. This rich user-
generated data plays a significant role in gleaning aggregated signals from the
content and making sense of public opinions and reactions to contemporary is-
sues. Twitter data can be used for predictive analysis in many application areas,
ranging from personal and social to public health and politics. Predictive analyt-
ics on Twitter data comprises a collection of techniques to extract information
and patterns from data, and predict trends, future events, and actions based on
the historical data.
Gaining insights and improving situational awareness on issues that matter
to the public are challenging tasks, and social media can be harnessed for a
better understanding of the pulse of the populace. Accordingly, state-of-the-
art applications, such as Twitris [3] and OSoMe [2], have been developed to
process and analyze big social media data in real time. Regarding availability
and popularity, Twitter data is more common than data from web forums and
Reddit1. It is a rich source of user behavior and opinions. Although analytical
approaches have been developed to process Twitter data, a systematic framework
to efficiently monitor and predict the outcome of events has not been presented.
Such a framework should account for the granularity of the analysis over a variety
of domains such as public health, social science, and politics, and it has been
shown in Figure 1.
We discuss a predictive analysis paradigm for Twitter data considering pre-
diction as a process based on different levels of granularity. This paradigm con-
tains two levels of analysis: fine-grained and coarse-grained. We conduct fine-
grained analysis to make tweet-level predictions on domain independent aspects
such as sentiment, topics, and emotions. On the other hand, we perform coarse-
grained analysis to predict the outcome of a real-world event, by aggregating
and combining fine-grained predictions. In the case of fine-grained prediction, a
predictive model is built by analyzing social media data, and prediction is made
through the application of the model to previously unseen data. Aggregation
and combination of these predictions are made from individual tweets form sig-
nals that can be used for coarse-grained predictive analysis. In essence, low-level
signals from tweets, such as sentiment, emotions, volume, topics of interest, loca-
tion and timeframe, are used to make high-level predictions regarding real-world
events and issues.
In this chapter, we describe use of Twitter data for predictive analysis, with
applications to several different domains. In Section 2, we discuss both processing
and analytic techniques for handling Twitter data and provide details of feature
extraction as well as machine learning algorithms. In Section 3, we explain a
predictive analysis paradigm for Twitter that comprises two levels: fine-grained
and coarse-grained. We also provide use cases, based on real-world events, of
how coarse-grained predictions can be made by deriving more profound insights
about a situation from social media using signals extracted through fine-grained
1 https://goo.gl/Jo1h9U
Predictive Analysis on Twitter: Techniques and Applications
3
Fig. 1. Overview of Predictive Analysis on Twitter Data.
predictions. We also describe common domain-independent building blocks that
can serve as the foundation for domain-specific predictive applications. In Sec-
tion 4, we give further details on specific state-of-the-art applications of Twitter
analytics that have been developed for different domains, such as public health,
social and political issues. In Section 5, we conclude with a discussion of the
impact of social media on the evolvement of real-world events and actions, chal-
lenges to overcome, for broader coverage and more reliable prediction. We also
provide a comparative table relating techniques used with corresponding appli-
cations.
2 Language Understanding of Tweets
Novel processing and analysis techniques are required to understand and de-
rive reliable insights to predict trends and future events from Twitter data due
to their unique nature – it contains slangs, unconventional abbreviations and
grammatical errors as a matter of course. Moreover, due to the evolving nature
of many events, may it be political, sports, or disaster-related, collecting relevant
information as the event unfolds is crucial [4, 5]. Overcoming the challenges posed
by the volume, velocity, and variety of incoming social, big data is non-trivial
[6]. Sole keyword-based crawling suffers from low precision as well as low recall.
For instance, obtaining tweets related to marijuana legislation [7] using its street
name spice pulls irrelevant content about pumpkin spice latte and spice in food.
To improve recall without sacrificing precision, Sheth et al. [8] provided a solu-
tion for adapting and enhancing filtering strategies that (a) obtains customized
tweet streams containing topics of user interest [9] by constructing a hierarchi-
cal knowledge base by analyzing each users tweets and profile information [10],
(b) selects and employs a domain-specific knowledge graph (e.g., using the Drug
TwitterComputational Techniques for Twitter DataML/DL/NLP/Sem. WebTextual ProcessingFeature ExtractionMachine LearningHealthPublic HealthPolitical IssuesSocial IssuesTransportationDisaster ManagementCommunity on Social MediaDemographicsAnomaly & Popularity PredictionSales & Stock PredictionFine-grained AnalysisCoarse-grained AnalysisBackground KnowledgePredictive Analysis Paradigm of TwitterApplication Domains 4
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
Abuse Ontology for opioid related analysis [11]) for focus, and (c) reuses a broad
knowledge graph such as DBPedia for coverage and generality. In Twitter data
analysis, the processing phase includes natural language processing using tech-
niques such as TF-IDF, word2vec, stemming, lemmatization, eliminating words
with a rare occurrence, and tokenizing. On the other hand, some of the com-
monly used techniques, such as removal of stop-words, have proven ineffective.
Saif [12] has compared six different stop words identification methods over six
different Twitter datasets using two well-known supervised machine learning
methods and assessed the impact of removing stop words by observing fluctu-
ations in the level of data sparsity, the size of the classifiers feature space and
the classifier performance. Saif concludes that in most cases that removing stop
words from tweets has a negative impact on the classification performance.
2.1 Unique Nature of Tweets
Twitters limit on the number of characters in a message encourages the use of
unconventional abbreviations, misspellings, grammatical errors and slang terms.
For instance, since a tweet was limited to 140 characters (until recent doubling to
280 character in December 2017), different sets of techniques and metadata have
been considered to identify the best features to optimize the overall performance
of the model being built. Due to the heterogeneous nature of the Twitter content,
one can develop a variety of features [13] ranging from textual, linguistic, visual,
semantic, network-oriented, to those based on the tweet and user metadata. Fur-
ther, to handle tweets textual data, the extracted features, techniques and tools
[3, 14, 15] have been customized to exploit as well as being robust concerning
misspellings, abbreviations, and slangs. Gimpel et al. [14] addressed this prob-
lem in the context of part-of-speech (PoS) tagging, by developing a new tagset
along with features specific to tweets, and reported 89% accuracy as opposed to
Stanford tagger with 85% accuracy.
Tweets also include hashtags, URLs, emoticons, mentions, and emoji in their
content. As these components contribute to the meaning of a tweet, it is imper-
ative that we incorporate them in the analysis, on a par with textual content.
Hashtags are meant to help in categorizing tweet's topics. They are fre-
quently used to collect and filter data as well as for sentiment [16–18], emotion
[6], and topical analysis [19, 20]. Wang et al. [16] used hashtags in their topical
hashtag level sentiment analysis incorporating co-occurrence and literal meaning
of hashtags as features in a graph-based model and reported better results com-
pared to a sentiment analysis approach at the tweet level. In emotion analysis,
Wang et al. [6] collected about 2.5 million tweets that contain emotion-related
hashtags such as #excited, #happy, and #annoyed, and used them as the self-
labeled training set for developing a high accuracy, supervised emotion classifier.
URL presence in a tweet is usually indicative the content being an index for
a longer explanatory story pointed to by the URL. Researchers found URLs in
a tweet to be discriminative in various studies such as sentiment analysis [21,
22], popularity prediction [23, 24], spam detection [25]. They reported that the
feature for URL presence in a tweet appeared as a top feature or has a substantial
contribution to the accuracy of the model.
Predictive Analysis on Twitter: Techniques and Applications
5
Emoticons (e.g., :), < 3) have been exploited by Liu et al. [26] in their
Twitter sentiment analysis study, such as by interpreting :) as conveying posi-
tive sentiment and :( as conveying the negative sentiment. They used all tweets
containing those emoticons as self-labeled training set and integrated them with
the manually labeled training set [27]. They have achieved significant improve-
ment over the model trained with only manually labeled data. Go et al. [21],
and other researchers [28, 29] conducted sentiment analysis on Twitter in 2009,
and they found that they were able to achieve a better accuracy using models
trained with emoticon data.
Emoji is a pictorial representation of facial expressions, places, food and
many other objects, being used very often on social media to express opinions
and emotions on contemporary issues of contentions and discussions. The use
of emoji is similar to emoticon since they both provide a shorter means of ex-
pression of an idea and thought. The difference is that an emoji use a small
image for the representation as opposed to emoticon that uses a sequence of
characters. Kelly et al. [30] studied the use of emoji in different contexts by
conducting interviews and found that the use of emoji goes beyond the context
that the designer intended. Novak et al. [31] created an emoji sentiment lexicon
analyzing the sentiment properties of emojis, and they pointed that the emoji
sentiment lexicon can be used along with the lexicon of sentiment-bearing words
to train a sentiment classifier. On the other hand, Miller et al. [32] found that
the emoji provided by different platforms are not interpreted similarly. Wijer-
atne et al. [33] gathered possible meanings of 2,389 emojis in a dataset called
EmojiNet, providing a set of words (e.g., smile), its POS tag (e.g., verb), and
its definition, that is called its sense. It associates 12,904 sense labels with 2,389
emojis, addressing the problem of platform-specific meanings by identifying 40
most confused emoji to a dataset.
2.2 Metadata for Tweet and User
There are mainly two types of metadata in a tweet object, namely, tweet meta-
data2 and user metadata3. Tweet metadata contains temporal and spatial infor-
mation along with user interactions and other information such as replies and
language. On the other hand, user metadata contains information pertaining to
the user that authored the tweet, such as screen-name and description. Some of
the available metadata are described below.
2.2.1 Tweet Metadata
createdAt: This field contains the information on when the tweet was created,
which is especially important when a time series analysis is being done [34].
favoriteCount: The users on Twitter can like a tweet, and this is one way of
interacting with the platform. The number of likes for a tweet has been used as
a feature in various applications that includes trend detection [34], identification
2 TweetObject.https://developer.twitter.com/en/docs/tweets/data-dictionary/overview/
tweet-object
3 UserObject.https://developer.twitter.com/en/docs/tweets/data-dictionary/overview/
user-object
6
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
of influence and popularity.
inReplyToScreenName: If this field of the tweet object is not null, it is a
reply to another tweet, and this field will hold the username of the user that
authored the other tweet. This information is valuable, especially to predict
the engagement of the audience over an issue that tweets relate to, and to find
influential users.
geoLocation: the Twitter platform has a feature that can attach the users
geolocation to the tweet, but this is up to the users to make it publically available.
Most of the users prefer not to share their geolocation.
retweet count: Twitter allows users to repost a tweet by retweeting to their
audience, and the original tweet holds this field to keep how many times this
tweet has been retweeted. This information is useful to incorporate the prediction
of popularity and trending topics.
2.2.2 User Metadata
description: This field holds the description of the account. As this metadata
carries information on characteristics of the user, it is mostly used in user clas-
sification.
followers count: This field holds the number of followers the user has, and as it
is changeable information over time, the information located in a specific tweet
may not be up to date.
friends count: Twitter calls the accounts that a user follows as "friends," but
it is also known as "followees." The numbers of followers and followees are used
to determine the popularity of user and topics.
statuses count: Twitter also calls tweets as status, and in this case, status
count refers to the number of tweets that a user has posted.
2.3 Network and Statistical Features
The users interact on the social networking platform Twitter with each other
through follows, replies, retweets, likes, quotes, and mentions. Centrality metrics
have been developed to compute and reveal users position and their importance
based on their connections in their network. These centrality measures can help
identify influential users. These metrics include in-degree, out-degree, closeness,
betweenness, PageRank and eigenvector centrality. Closeness centrality is defined
by Freeman [35] as the sum of distances from all other nodes, where the distance
from a node to another is defined as the length (in links) of the shortest path
from one to the other. The smaller the closeness centrality value, the more central
the node. Betweenness [36] measures the connectivity of a node by computing
the number of shortest paths which pass through the node. This aspect makes
this node, a user in a Twitter social network, an essential part of the network
as it controls the flow of information in the network. Therefore, removing this
node would disconnect the network. EigenVector [37, 38] metric measures the
importance of a node based on the importance of its connections within the
network. Therefore, the more critical connections a node gets, the more critical
the node becomes. These metrics were used in a user classification application as
Predictive Analysis on Twitter: Techniques and Applications
7
features by Wagner et al. [15] because of the intuition that similar users would
have similar network connectivity characteristics.
Statistical features such as min, max, median, mean, average, standard devia-
tion, skewness, kurtosis, and entropy can be computed for several data attributes
[34]. Machine learning determines a subset of these features that have the dis-
criminative power necessary for particular applications and domains, especially
for predicting user behaviors and user types [39]. For instance, [34] extracted
statistical features of a user, tweet, network. The statistical analysis was done
over attributes such as senders follower count, originators followee count, the
time between two consecutive tweets, and the number of hashtags in a tweet.
They conducted a time series analysis to predict if a trending meme is organic or
promoted by a group. On the other hand, [39] utilized statistical features to pre-
dict the type of users on social media based on their political leanings, ethnicity,
and affinity for a particular business. As they classified users, they computed
statistical characteristics of tweeting behavior of users such as average number
of messages per day, average number of hashtags and URLs per tweet, average
number and standard deviation of tweets per day.
2.4 Machine Learning and Word Embeddings
Machine learning algorithms play a crucial role in the predictive analysis for
modeling relationships between features. It is well-known that there is no uni-
versal optimal algorithm for classification or regression task, and in fact requires
us to tailor the algorithm to the structure of the data and the domain of dis-
course. Recent survey papers [40–43] and our comparative analysis (see Table 1)
of related influential studies show what algorithms we found to perform well for
various applications. As can be seen, this covers a wide variety – Random Forest,
Naive Bayes, Support Vector Machine, Artificial Neural Networks, ARIMA and
Logistic Regression.
Furthermore, deep learning (a.k.a advanced machine learning) enhanced the
performance of learning applications. Deep learning is a strategy to minimize the
human effort without compromising performance. It is because of the ability of
deep neural networks to learn complex representations from data at each layer,
where it mimics learning in the brain by abstraction4. The presence of big data,
GPU, and sufficiently large labeled/unlabeled datasets improve its efficacy. We
discuss some of the applications that make use of deep learning for prediction
task on social media in section 4.
Textual data processing benefits from the lexico-semantic representation of
content. TF-IDF [44], Latent/Hierarchical Dirichlet Allocation(LDA/HDA) [45],
Latent Semantic Analysis (LSA) [46] and Latent Semantic Indexing have been
utilized in prior studies for deriving textual feature representations. In a re-
cent paper [47], they put forward a word embedding approach called Word2Vec
that generates a numerical vector representation of a word that captures its
contextual meaning incorporating its nearby words in a sentence. Training the
word embedding model on a problem-specific corpus is essential for high-quality
4 How do Neural networks mimic the human brain? https://www.marshall.usc.edu/blog/
how-do-neural-networks-mimic-human-brain
8
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
domain-specific applications, since the neighborhood set of words for an input
term impacts its word embedding. For instance, pre-trained models of word2vec
on news corpora generate poor word embeddings over a Twitter corpus. Wijer-
atne et al. [48] used word embeddings to further enhance the prediction of gang
members on Twitter by training their model on a problem-specific corpus.
2.5 Multi-modality on Twitter
Visual elements such as images and videos are often used on social media plat-
forms. While users can attach images and videos to their tweets, they can also
upload a profile image and a header image. Since the latter images are mostly
related to the users characteristics, personality, interest or a personal prefer-
ence, these images are mostly used for classification of account type (e.g., me-
dia, celebrity, company), detection of user groups [49, 48] and identification of
demographic characteristics (e.g., gender, age) [50]. Balasuriya et al. [49] used
the profile image of users in their feature set for finding street gang members
on Twitter since gang members usually set their profile image in a particular
way to intimidate other people and members of rival gangs. They retrieved a set
of 20 words and phrases for each image through the Clarifai5 web service to be
used as features. As image processing is costly regarding time and computational
resources required for training a model to retrieve information from images, it
is usually preferred to use off-the-shelf web services that provide cheaper, yet
effective alternative, for scalable social media analytics.
3 Prediction on Twitter Data
Gaining understanding about and predicting an events outcome and its evolve-
ment over time using social media, requires incorporation of analysis of data
that may differ in granularity and variety. As tools [14, 51] are developed and
customized for Twitter, its dynamic environment requires human involvement
in many aspects. For instance, verification of a classification process [52] and an-
notation of a training dataset [33, 53, 54] are essential in the predictive analysis
that can benefit from human expert guidance in creating ground truth dataset.
Social media analysis in the context of complex and dynamic domains [55–57]
is challenging. Our approach to overcoming this challenge and dealing with a
variety of domains is to customize domain independent building blocks to de-
rive low-level/fine-grained signals from individual tweets. Then, we aggregate
and combine these signals to predict high-level/coarse-grained domain-specific
outcomes and actions with a human in the loop.
3.1 A Predictive Analysis Paradigm for Twitter
We consider predictive analysis on Twitter data as a two-phase approach: The
first phase is fine-grained predictive analysis and the second phase is coarse-
grained predictive analysis. An illustration of this paradigm is depicted in figure
2. The fine-grained analysis is a tweet-level prediction for individual signals,
such as sentiment and emotions, about an event that is being monitored. This
low-level prediction is made by building a predictive model that employs feature
5 https://www.clarifai.com
Predictive Analysis on Twitter: Techniques and Applications
9
Fig. 2. Two level Predictive Analysis Paradigm for Twitter.
engineering and machine learning algorithms. Aggregating the tweet-level pre-
dictions for a specific time frame and location generates signals. For instance,
a predictive model for sentiment predicts the sentiment of each tweet about an
event in question as negative (-1) neutral (0) or positive (+1), and we produce a
signal between -1 and +1 for a particular location and time frame. A collection
of such signals (e.g., emotions, topics) helps domain experts form insights while
monitoring or predicting the outcome of an event, in their higher level analysis.
Extraction of these signals is discussed further in subsequent section.
Coarse-grained analysis is a higher level prediction involving outcomes and
trends of a real-world event, such as elections[58], social movements[59] and dis-
aster coordination[60–63]. In this case, we gather the signals which we generated
from the fine-grained predictions and make a judgment call for the outcome by
making sense of these signals in the context of the event and the related domain.
Sentiment, emotions, volume, topics, and interactions between Twitter users can
be considered as signals, while the importance and informativeness of each of
these parameters may vary depending on the event and its domain. For instance,
gauging the sentiment of a populace towards an electoral candidate would be
very significant to predict the outcome of an election [56], but the same kind
of information may not be as critical in the context of disaster management
because, in the latter case, the sentiment may be largely negative. Further, for
reliable decision making, the sentiment must be interpreted in a broader context.
Predominantly positive sentiment towards democratic candidates in California
is not as significant as that in Ohio. Similarly, the context provided by county
demographics may be crucial in generalizing, predicting, and determining the
outcome of an election. Moreover, temporal and spatial context plays an im-
portant role to understand the ongoing events better and obtain more profound
Feature Extraction•Textual•Linguistic•Metadata•Network•Statistical•VisualFine-grained Predictive AnalysisBuild Predictive ModelGenerate Signals for Coarse-grained PredictionPrediction Making(Coarse-grained)Domain ExpertsSentimentEmotionsTopicsVolumeNetworkIntentHealth Rel. SignalsLocationPreferenceCoarse-grained Predictive Analysis 10
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
insights. In US presidential elections, some states, called the swing states (as the
electorates choice has changed between Republican and Democratic candidates
through the previous elections in these states), typically determine the eventual
outcome of US elections. Therefore, narrowing down the analysis to the state
level and gathering signals from these particular states would meaningfully con-
tribute to the prediction of the outcome of the Presidential election and the
future direction of the country.
In general, prediction analytics requires domain-specific labeled datasets cre-
ated with the assistance of domain experts, and customization of feature space,
classification algorithm, and evaluation. Real world events have a dynamic na-
ture in which critical happenings may change the course of discussions on social
media. For example, breaking news about a candidate in an election may change
the vibe in echo chambers of Twitter; thus, affecting the public opinion in one
or another direction. For this reason, it is imperative to conduct the analysis
accounting for essential milestone events happening during the process. There-
fore, the analysis of such events would require an active learning paradigm that
incorporates evolving domain knowledge in real-time.
3.2 Use Cases for Coarse-grained Prediction
Coarse-grained prediction requires taking into account many signals, and eval-
uating them concerning both present and historical context that varies with
location and time frame. Importance of the signals in some domains and their
related events may vary, and sole use of these signals would not be sufficient to
make a reliable judgment call, although these signals are essential parameters in a
real-world event context. For instance, an election usually whips up discussions
on various sub-topics, such as unemployment, foreign policy; and necessitates
proper cultivation of a diverse variety of signals following contextual knowledge
of the domain [56]. We provide two use cases in this subsection to illustrate how
a coarse-grained or high-level predictive analysis can be conducted.
3.2.1 US 2016 Presidential Election
During the 2016 US Presidential elections where swing states played a key role in
determining the outcome, many polling agencies failed to predict it accurately67.
On the other hand, researchers8 conducted a real-time predictive analysis using
a social media analytics platform [3], making the prediction accurately before
the official outcome was announced, by analyzing the state-level signals, such as
from Florida and Ohio. Temporal aspect was also important in this use case to
explain the evolution of the public opinion based on milestone events over the
period of the election, as well as the election day because people tend to express,
who they voted in the same day. They analyzed 60 million tweets by looking at
the sentiment, emotions, volume, and topics narrowing down their analysis to
state-level. On the election day, they focused on specific states such as Florida,
6 http://www.pewresearch.org/fact-tank/2016/11/09/why-2016-election-polls-missed-their-mark/
7 https://goo.gl/mFtzvb
8 https://goo.gl/AJVpKf
Predictive Analysis on Twitter: Techniques and Applications
11
which, before the election day, they predicted would be a pathway for Donald
Trump to win the election9. In their analysis of Florida, volume and positive
emotion (joy) for Trump was higher, whereas positive sentiment for Clinton was
higher, eliciting report10 such as limited to professed votes from Florida until
1pm is not looking in her favor. Later in the day, the volume of tweets for Trump
increased to 75% of all tweets based on the hashtag "#ivoted". Particularly in
critical states of Florida, North Carolina, and Michigan, volume and positive
emotions for Trump were significantly higher than for Clinton, although the
sentiment was countering the overall signal. They made the call that the winner
of Presidency and Congress as Donald Trump and the GOP respectively. While
conducting this analysis [56], they noticed that the predictive model that they
have built for sentiment signal was not successful due to the dynamic nature of
the election with changing topics in conversations. A similar analysis was made
for UK Brexit polls in 2012 by the same researchers, correctly predicting the
outcome utilizing the volume and sentiment signals111213.
3.2.2 US Gun Reform Debate 2018
Researchers14 monitored gun reform discussions on Twitter to predict the public
support using the Twitris platform after the tragic shooting at a high school in
Parkland, Florida, in February 2018. The public started demanding a gun control
policy reform, and it has attracted the attention of legislative and executive
branches of both state and federal governments. As polls measured the public
opinion15, researchers reported that the public support for a gun reform on
social media was increasing over time since the Parkland shooting, confirming the
overall outcome of these polls. They observed that reactions from public on social
media in terms of the volume, sentiment, emotions and topics of interest, are
strongly aligned with the milestone events related to this issue such as (i) POTUS
(President of the United States) meeting with families of the victims on February
21, (ii) CPAC(Conservative Political Action Conference) between February 22
and 24, (iii) POTUS meeting with lawmakers on February 28 expressing strong
support for a gun control policy change. These events significantly affected the
public opinion on social media based on the aforementioned signals.
At the beginning of the gun reform discussions on social media, sentiment
for pro-gun reform tweets was strong whereas the sentiment for anti-gun reform
was relatively weak. However, the CPAC meeting changed the climate on social
media, and it significantly boosted the momentum of anti-gun reform tweets,
especially after the NRA (National Rifle Association) CEO Wayne LaPierres
speech in the morning of February 2216. Overall the volume of tweets for pro-gun
9 https://goo.gl/sh7WNr
10 https://goo.gl/iCqzk3
11 https://goo.gl/i2Ztm6
12 https://goo.gl/dFCGL9
13 https://goo.gl/2EhSma
14 http://blog.knoesis.org/2018/04/debate-on-social-media-for-gun-policy.html
15 http://time.com/5180006/gun-control-support-has-surged-to-its-highest-level-in-25-years/
16 https://goo.gl/kgbqWC
12
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
reform was mostly higher than the anti-gun overhaul, except between February
22 and February 2417, which covers the CPAC meeting where NRA CEO, VP
Pence, and POTUS gave speeches. It surged the volume, positive sentiment and
emotions in anti-gun reform posts radically, and those parameters for pro-gun
reform posts dropped in the same manner. Effect of the meeting lasted a few
days, and boycott calls for NRA and NRAs sponsors started to pick up in the
meantime. After the meeting, sentiment for pro-gun reform tweets increased
consistently, and the emotions expressed in pro-gun reform tweets became more
intensified.
Emotions in anti-gun reform tweets were intense especially during and af-
ter the CPAC meeting, but later emotions in pro-gun reform tweets took over.
Especially volume, positive sentiment, and emotions were overwhelmingly high
right after the POTUS meeting with lawmakers on Wednesday, February 28,
expressing his support for a gun policy reform.
Furthermore, some of the most popular topics that users were discussing
in their tweets included midterm elections, parkland students, boycott the nra,
stupid idea and individual freedoms, where pro-gun reform arguments were ex-
pressed more frequently. The topic of midterm elections being one of the most
popular topics on social media in gun reform discussions, also suggests that
politicians from both Democrats and Republicans sensed the likely effect of this
public opinion change on the midterm elections on November 2018. They have
concluded in their predictive analysis that the public support for gun reform was
significantly higher based on the signals they observed in the context of related
events.
3.3 Extraction of Signals
We make predictions for the outcome of real-world events based on the insights
we collect from big social data, and these insights are extracted as various signals
such as sentiment, emotions, volume, and topics. The sentiment is a qualitative
summary of opinions on a particular issue, and sentiment analysis techniques
are utilized to extract such information computationally. The emotional analy-
sis provides another stream of qualitative summary that is expressed by users
about a particular event. The volume of tweets is an important signal about the
engagement of the public in an event or an issue of consequence. Topical analysis
is a process that extracts topics that contain particular themes in the domain of
interest. We can produce and make use of more specific signals depending on the
domain such as preference, intent, and symptoms. The signals described below
are commonly used parameters in higher level prediction tasks, and we describe
related state-of-the-art applications and their technical details in the following.
3.3.1 Sentiment Analysis
Sentiment is one of the essential signals that can be used to measure the public
opinion about an issue. As users on Twitter express their opinions freely, senti-
ment analysis of tweets attracted the attention of many researchers. Their ap-
17 https://goo.gl/LMFu3B
Predictive Analysis on Twitter: Techniques and Applications
13
proaches differ regarding the feature set, machine learning algorithm, and text
processing techniques. Considering feature set, [18] used n-grams, POS-tags,
emoticons, hashtags and subjectivity lexicon for sentiment analysis. For ma-
chine learning, Naive Bayes, SVM, and Conditional Random Fields (CRF) have
been employed, and Naive Bayes has shown good performance [29]. Also, text
processing techniques like stopwords removal, word-pattern identification, and
punctuation removal have shown to improve sentiment analysis in [17]. Nguyen
et al. [64] used time series analysis to be able to predict the public opinion so
that the stakeholders on a stock market can react or pro-act against the public
opinion by posting balanced messages to revert the public opinion based on the
measurement that they performed using social media. Their objective was to
use the sentiment change over time by identifying key features that contribute
to this change. They measured the sentiment change regarding the fraction of
positive tweets. They employed SVM, logistic regression and decision tree, and
found that SVM and logistic regression provided similar results outperforming
the decision tree. They modeled the sentiment change overall twitter data and
achieved around 73% F-score on sentiment prediction using time series analy-
sis. [65] employs a deep learning approach combining convolutional and gated
recurrent neural network (CGRNN) for a diverse representation of tweets for sen-
timent analysis. Such a system was trained on GloVe word embedding created
on a crawled dataset. The system was ranked among the top 10, evaluated using
average F1 score, average recall, mean absolute error (MAE), Kullback-Leibler
divergence (KLD), and EMD score [66] for SemEval-2016 sub-tasks B, C, D, E.
Exclusion of hand-crafted features and improved performance on SemEval 2016
shows the potency of the approach.
3.3.2 Emotion Analysis
Identification of emotions in tweets can provide valuable information about the
public opinion on an issue. Wang et al. [6] predicted seven categorical emo-
tions from the content of tweets using 131 emotion hashtags and utilizing the
features such as n-grams, emotion lexicon words, part-of-speech tags, and n-
gram positions. They used two machine learning algorithms: LIBLINEAR and
Multinomial Naive Bayes. In a similar study, Lewenberg et al. [54] examined the
relationship between the emotions that users express and their perceived areas of
interest, based on a sample of users. They used Ekmans emotion categories and
crowdsourced the task of examining the users and their tweets content to deter-
mine the emotions as well as their interest areas. They created a tweet-emotion
dataset consisting of over 50,000 labeled tweet-emotion pairs, then trained a lo-
gistic regression model to classify the emotions in tweets according to emotion
categories, using textual, linguistic and tweet metadata features. The model pre-
dicted a users emotion score for each emotion category, and they determined the
user's interest in areas such as sports, movies, technical computing, politics,
news, economics, science, arts, health, and religion.
14
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
3.3.3 Topical Analysis
Topical analysis is one of the essential strategies under the umbrella of infor-
mation extraction techniques that capture semantically relevant topics from the
social media content [67]. Extraction of topics in the context of social media
analysis helps understand the subtopics associated with an event or issue and
what aspects of the issue have attracted the most attention from the public.
As discussed in use cases for elections and gun reform debate, it is imperative
to have the topics extracted from tweets for a better understanding of the un-
derlying dynamics of relevant discussions. Chen et al. [68] associated topics of
interest with their relative sentiment to monitor the change in sentiments on
the extracted topics. Furthermore, utilizing the extracted topics as features for
a supervised model improved the performance of the classification task in [69].
In [70], researchers assessed quality of topics using coherence analysis, context-
sensitive topical PageRank based ranking and probabilistic scoring function. This
approach was used in a crime prediction application [71].
3.3.4 Engagement Analysis
The volume is the size of the dataset that has been collected and indicates the
user engagement on an event being monitored. In general, the larger the dataset,
the better is the accuracy and consistency of a predictive model because it min-
imizes the possibility of bias. Engagement analysis enables human experts to
improve their confidence in the learned representations/patterns for an accurate
high-level prediction. However, while maintaining the sufficient size of the dataset
to make reliable predictions from representative data is critical, data collection
strategies need to be chosen strategically since relying solely on keyword-based
crawling can bring in noise and irrelevant[72] data from a different context into
the dataset. Therefore, a suitable filtering mechanism is essential for better qual-
ity data with high recall as well as precision. A semantic filtering mechanism [8,
73] as in the Twitris platform, can be implemented that selects and employs
a domain-specific knowledge graph (e.g., using the Drug Abuse Ontology for
related opioid analysis [11]) for precision, and reuses a broad knowledge graph
such as DBPedia for coverage and generality (see section 2). Thus, a significant
and relevant dataset can be collected with high recall and precision that will
allow one to obtain insights on the user engagement.
4 Applications
Twitter data has enabled researchers and analysts to deal with diverse domains
ranging from healthcare, finance, and economy to socio-political issues and crisis
management. Approaches to retrieve as much information as possible requires
the inclusion of domain-specific features as well as the use of domain knowledge
in the analysis. In this section, we provide a list of domains where predictive
analysis applications on Twitter were implemented, along with the technical de-
tails. A comprehensive table is also included at the end to give a comparative
overview of application domains, the features and machine learning algorithms
being used, and their performance. The included applications were selected be-
Predictive Analysis on Twitter: Techniques and Applications
15
cause they were state-of-the-art in their respective domains or had been influ-
ential. The applications that we describe in this section combine a variety of
signals that can be the basis for coarse-grained predictive analysis. Since some
of the applications in this section make use of the Twitris platform; therefore,
we first provide background information about the platform. Purohit et al. [1]
introduced the Twitris platform for citizen sensing that performs analysis of
tweets, complemented by shared information from contextually relevant Web
of Data and background knowledge. They describe it as a scalable and interac-
tive platform which continuously collects, integrates, and analyzes tweets to give
more profound insights. They demonstrate the capabilities of the platform with
an analysis in various dimensions including spatio-temporal-thematic, people-
content network, and sentiment-emotion-subjectivity, with examples from busi-
ness intelligence including brand tracking, advertising campaigns, social/political
unrests, and disaster events.
4.1 Healthcare
Twitter data can be employed to shed light on many healthcare and disease-
related aspects of contemporary interest, ranging from Alzheimer and dementia
progression [74] to eating disorders [75] and mental health problems [76, 77].
We focus on applications to glean depression in individuals or at a community
level using self-reports about these conditions, their consequences, and patient
experiences on Twitter.
Depression is a condition that a sizable population in all walks of life ex-
periences in their daily life. Social media platforms including Twitter has been
used to voluntarily express the mood changes and feelings as they arise. From
these tweets, it is possible to predict whether a user is depressed or not, what
symptoms they show as well as the reasons for their depressive mood. Some
examples indicative of depression as expressed in tweets18 include: "I live such
a pathetic life.", "Cross the line if you feel insecure about every aspect of your
life." ,"That's how depression hits. You wake up one morning afraid that you're
going to live.", and "Secretly having a mental breakdown because nothing is
going right and all motivation is lost.". These tweets epitomize the expression of
emotional tumult that may underlie subsequent conscious actions in the physical
world.
An interesting study by Yazdavar et al. [76] explored the detection of clinical
depression from tweets by mimicking the PHQ-9 questionnaire which clinicians
administer to detect depression in patients. This study is different from tradi-
tional clinical studies that use questionnaires and self-reported surveys. They
crawled 23M tweets over 45K twitter users to uncover nine significant depres-
sive symptoms; (1) Lack of Interest, (2) Feeling Down, (3) Sleep Disorder, (4)
Lack of Energy, (5) Eating Disorder, (6) Low Self-esteem, (7) Concentration
Problems, (8) Hyper/Lower Activity, and (9) Suicidal Thoughts. A probabilistic
topic model with a semi-supervised approach is developed to assess clinical de-
pression symptoms. This hybrid approach is semi-supervised in that it exploits a
18 These tweets were modified before we share them in this chapter.
16
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
lexicon of depression symptoms as background information (top-down) and com-
bines it with generative model gleaned from the social media data (bottom-up)
to achieve a precision of 72% on unstructured text.
De Choudhury et al. [53] predicted the depression in an individual by exploit-
ing their tweets. For ground truth dataset, they used, crowdsourcing to collect
and label data. They utilized tweet metadata, network, statistical, textual and
linguistic features, and time series analysis over a year of data to train an SVM
model, obtaining an accuracy of 0.72.
The extraction of the location of people who experience depression using tex-
tual and network features can further assist in locating depression help centers.
[78] utilizes a multiview19 and deep learning based model, to predict the user
location. The multi-entry neural network architecture (MENET) developed for
location prediction uses words, the semantics of the paragraph (using doc2vec
[79]), network features and topology (using node2vec [80]) and time-stamps to
deduce users location. They achieved an accuracy over 60% for GeoText20, UT-
Geo1121 and 55% for TwitterWorld[81]. Furthermore, MENET achieves an ac-
curacy of 76% in region classification and 64.4% in state classification using
GeoText dataset.
4.2 Public Health
Social media platforms including web forums, Reddit and Twitter, has become
a venue where people seek advice and provide feedback for problems concern-
ing public health. These conversations can be leveraged to predict trends in
health-related issues that may threaten the well-being of the society. Moreover,
caregivers have also seen these sources to be a game changer in its potential for
actionable insights because of the information circulation. Particularly, cannabis
legalization issue in the U.S. has been a trending topic22 in the country as well
as social media. Prior research on Twitter data analysis in this domain proved
that it is an essential tool for epidemiological prediction of emerging trends.
Existing studies have involved identifying syntactic and statistical features
for public health informatics, such as PREDOSE (PRescription Drug abuse On-
line Surveillance and Epidemiology) which is a semantic web platform that uses
the web of data, background domain knowledge and manually created drug abuse
ontology for extraction of contextual information from unstructured social me-
dia content. PREDOSE performs lexical, pattern recognition (e.g., slang term
identification), trend analysis, triple extraction (subject-predicate-object) and
content analysis. It is helpful in detecting substance abuse involving marijuana
and related products. Not only can it analyze generic marijuana but also its
concentrates like butane hash oil, dabs, and earwax that are used in the form
of vaporizers or inhalers. In a similar analysis of Twitter data, the marijuana
concentrate use and its trends were identified in states where cannabis was le-
19 http://www.wcci2016.org/document/tutorials/ijcnn8.pdf
20 https://www.cs.cmu.edu/~ark/GeoText/README.txt
21 http://www.cs.utexas.edu/~roller/research/kd/corpus/README.txt
22 http://www.pewresearch.org/fact-tank/2018/01/05/americans-support-marijuana-legalization/
ft_18-01-05_marijuana_line_update/
Predictive Analysis on Twitter: Techniques and Applications
17
galized as well as not legalized. In 2014, utilizing the eDrugTrends23 Twitris
platform, researchers collected a total of 125,255 tweets for a two-month period,
and 22% of these tweets have state-level location information[82]. They found
that the percentage of dabs-related tweets was highest in states that allowed
recreational or medicinal cannabis use and lowest in states that have not passed
medical cannabis laws, where the differences were statistically significant. A sim-
ilar study in 2015 [83] reported adverse effects of Cannabis edibles and estimated
the relationship between edibles-related tweeting activity and local cannabis leg-
islation. Another study [82] was to automatically classify drug-related tweets by
user type and the source of communication as to what type of user has authored
the tweet, where the user types are defined as user, retailer and media. They
employed supervised machine learning techniques incorporating the sentiment
of tweets (e.g., positive, negative, neutral).
4.3 Political Issues
Political discussions on Twitter, which capture dynamic evolvement of public
opinion, can directly impact the outcome of any political process. Arab Spring
demonstrations [84–86] in the middle eastern countries, Gezi protests [87, 85] in
Turkey, as well as US Presidential elections in 2016 involving influence peddling
on several social media platforms [88] provide impactful illustrative examples.
Researchers have explored user classification and profiling in the context of such
political events on Twitter to predict the issue trends and eventual outcome.
Researchers [39, 89, 90] used Twitter data to predict political opinions of
users based on linguistic characteristics (e.g., Tf-IDF) of their tweet content.
While classification of users based on their political stance on Twitter has been
well studied, Cohen et al. [91] have claimed that much of the studies and their
datasets to date have covered very narrow portion of the Twittersphere, and
their approaches were not transferable to other datasets. Pennacchiotti et al.
[39] focused on the user profiling task on Twitter, and used user-centric features
such as profile, linguistic, behavioral, social and statistical information, to detect
their political leanings, ethnicity, and affinity for a particular business.
Moreover, prediction of dynamic groups of users has been employed [58] to
monitor the polarity during a political event by analyzing tweeting behavior and
content through clustering. Usage of hashtags and URL, retweeting behaviors
and semantic associations between different events were key to clustering. 56%
of the Twitter users participated in 2012 US Republican Primaries by posting
at least one tweet, while 8% of the users tweeted more than 10 tweets. 35%
of all users mostly retweet, separating them from the remaining. In terms of
dynamic user groups, they formed the following bilateral groups: silent majority
and vocal minority, high and low engaged users, right and left-leaning users,
where users were from different political beliefs and ages. They analyzed these
dynamic groups of users to predict the election outcomes of Super Tuesday
primaries in 10 states. They also reported that the characterization of users by
tweet properties (frequency of engagement, tweet mode, and type of content) and
23 http://wiki.knoesis.org/index.php/EDrugTrends
18
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
political preference provided insights to make reliable predictions. 8 weeks of data
comprising 6,008,062 tweets from 933,343 users about 4 Republican candidates:
Newt Gingrich, Ron Paul, Mitt Romney and Rick Santorum, was analyzed to
assess the accuracy of predicting the winner. Prediction of user location using
a knowledge base such as LinkedGeoData24 in tweets also contributed to the
election prediction. Furthermore, an error of 0.1 between the prediction and
actual votes attest to the efficacy of the approach. Such a low error rate in
prediction is attributed to original tweets (not retweets) from users who are
highly engaged and right leaned.
4.4 Social Issues
Social issues and related events have been a part of discussions on Twitter, which
gives opportunities to the researchers to address problems concerning individuals
as well as the society at large. Solutions to such problems can be provided by
measuring public opinion and identification of cues for detrimental behavior on
Twitter by employing predictive analysis. We explain three problems and their
respective solutions in this subsection.
4.4.1 Harassment
Harassment25 is defined as an act of bullying an individual through aggressive of-
fensive word exchanges leading to emotional distress, withdrawal from social me-
dia and then life. According to a survey from Pew Research Center26, 73% of the
adult internet users have observed, and 40% have experienced harassment, where
66% percent of them are attributed to social media platforms. Also, according
to a report from Cyberbullying27 research center, 25% of teenagers claimed to
be humiliated online. While it is imperative to solve this problem, frequency
and severe repercussions of online harassment exhibit social and technological
challenges.
Prior work [92] has modeled harassment on social media to identify the ha-
rassing content which was a binary classification approach. However, in their
predictive analysis, the context, network of users and dynamically evolving com-
munities shed more light on the activity than pure content-based analysis. For in-
stance, sarcastic communication between two friends on social media may not be
conceived as harassment while the aggressive conversation between two strangers
can be considered as an example of bullying. For reliably identifying and predict-
ing harassment on Twitter, it is essential to detect language-oriented features
(e.g., negation, offensive words), emotions, and intent. [93] employs machine
learning algorithms along with word embedding, and DBpedia knowledge graph
to capture the context of the tweets and user profiles for harassment prediction.
Edupuganti et al. [94] focused on reliable detection of harassment on Twit-
ter by better understanding the context in which a pair of users is exchanging
messages, thereby improving precision. Specifically, it uses a comprehensive set
24 http://linkedgeodata.org/About
25 http://wiki.knoesis.org/index.php/Context-Aware_Harassment_Detection_on_Social_Media
26 http://www.pewinternet.org/2014/10/22/online-harassment/
27 http://cyberbullying.us/facts
Predictive Analysis on Twitter: Techniques and Applications
19
of features involving content, profiles of users exchanging messages, and the se-
quence of messages, we call conversation. By analyzing the conversation between
users and features such as change of behavior during their conversation, length
of conversation and frequency of curse words, the harassment prediction can be
significantly improved over merely using content features and user profile infor-
mation. Experimental results demonstrate that the comprehensive set of features
used in our supervised machine learning classifier achieves F-score of 88.2 and
Receiver Operating Characteristic (ROC) of 94.3. Kandakatla et al. [95] presents
a system that identifies offensive videos on YouTube by characterizing features
that can be used to predict offensive videos efficiently and reliably. It exploits
using content and metadata available for each YouTube video such as comments,
title, description, and the number of views to develop Nave Bayes and Support
Vector Machine classifiers.The training dataset of 300 videos and test dataset of
86 videos were collected, and the classifier obtained an F-Score of 0.86.
4.4.2 Gang Communities & Their Members and Gun Violence
Gang communities and their members have been using Twitter to subdue their
rivals, and identification of such users on Twitter facilitates the law enforcement
agencies to anticipate the crime before it can happen. Balasuriya et al. [49] in-
vestigated conversations for finding street gang members on Twitter. A review
of the profiles of gang members segregates them from rest of the Twitter popu-
lation by checking hashtags, YouTube links, and emojis in their content [96]. In
[49], nearly 400 gang member profiles were manually identified using seed terms,
including gang affiliated rappers, their retweeters, followers as well as followees.
They used textual features of the tweet, YouTube video descriptions and com-
ments, emojis and profile pictures to power various machine learning algorithms
including Naive Bayes, Logistic Regression, Random Forest and Support Vec-
tor Machines, to train the model. Random Forest performed well for Gang and
Non-Gang classification. It is interesting to notice that gang members usually
make use of their profile images in a specific way to intimidate other people and
members of rival gangs.
As gun control policies in big cities, such as Chicago, have changed over the
years, the volume of the taunting and threatening conversations on social media
has also relatively increased [97]. Such conversations can be leveraged to assist
law enforcement officers by providing insights on situational awareness as well
as predicting a conflict between gang groups for a possible gun violence inci-
dent. Blevins et al. [97] used a Twitter dataset that was manually labeled by
a team of researchers with expertise in cyber-bullying, urban-based youth vio-
lence and qualitative studies. Their strategy was to collect all tweets, mentions,
replies, and retweets from a specific user profile between 29 March and 17 April
2014. Three experts developed the key types of content and used the work by
Bushman and Huesmann [98] to identify and categorize types of aggression. To
overcome the challenge of recognizing special slang terms and local jargons in
tweets as mentioned in Desmond et al. work, Blevins et al.[97] developed a part-
of-speech (POS) tagger for the gang jargon and mapped the vocabulary they
20
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
use to Standard English using machine translation alignment. They developed
emotion classifier that uses the extracted POS tags, and Dictionary of Affect in
Language (DAL) quantitative scores (Whissell, 2009) as key features. Ternary
classification is applied to the whole dataset (TCF) and binary classification
on the aggression-loss subset (BCS). Then they use a cascading classifier (CC),
which uses two SVM models. Initially, one SVM model is used to filter the tweets
into aggression/loss tweets, and all other tweets fall into the other category. After
this filtration, only aggression/loss tweets is passed to second SVM model which
is again a binary classifier for loss or aggression. So this Aggression Supervised
classifier is able to categorize loss with 62.3% F-score and aggression with 63.6%
F-score which beats the baseline model (Unigrams) by 13.7 points (aggression)
and 5.8 points (loss) [97].
4.5 Transportation
Congestion due to traffic is one of the prevalent problems in the United States
(U.S.). Even after having structured rules that govern the flow of the traffic in
the U.S., congestion due to non-recurring activities still affects the schedules
of people. However, the stationing of police officers to smooth the traffic is a
probable solution, although it would not be long-term. Having an estimation of
the flow of traffic in the advent of an event can help people to re-route their
path to the destination. Leveraging social media and machine learning to esti-
mate traffic is one such long-term solution that can be drafted for active traffic
monitoring. Social media is flooded with posts from people about an event. Such
posts can provide the location of the event or the tweeter, and it can be used
along with other textual features to estimate the traffic flow. In [99], textual fea-
tures, tweet and user-metadata such as text, hashtags, URLs, number of users
and retweeted tweets were used by combining with live event data to predict
traffic dynamics. They utilized autoregressive model, neural network, support
vector regression, and K-nearest neighbor for traffic prediction. The evaluation
was performed using mean absolute percentage error (MAPE), and root means
square error (RMSE), with support vector regression (SVR) performing better
over other regression models. SVR reduced the error in traffic prediction by 24%
in terms of RMSE.
4.6 Location Estimation
Social media serves a vital role in times when people struggle to survive a disas-
trous event such as hurricane or earthquake, to provide solutions for assisting the
public in recovery efforts. These solutions include identification of the demand
and its location, and mapping the identified demands with suitable suppliers
analyzing Twitter data.
In particular, location extraction plays a significant role in identifying the
area that is impacted by a disaster as well as providing assistance [100]. Mah-
mud et al. [101] developed an approach to predict the location of users at the
city level on Twitter combining several classifiers. They removed stop words,
performed part-of-speech tagging, extracted hashtags, and extracted a feature
called local term, a term used by local people to refer to the city. For detecting
Predictive Analysis on Twitter: Techniques and Applications
21
the local terms, several classification algorithms and found Nave Bayes, SVM and
Decision trees (J48) as the best performing algorithms. Al-Olimat et al. [102]
developed a tool called LNEx (Location Name Extraction), that extracts the lo-
cation from the tweet content by utilizing the OpenStreetMap[103], GeoNames
[104] and DBpedia [105] for disambiguation. The information retrieval process
from the tweet is two-fold, which are toponym extraction and geoparsing. To-
ponym is a process to extract city and street names, points of interest, from
unstructured text, tweets in particular for this study. Location names are usu-
ally abbreviated on Twitter; hence, a text normalization procedure is used for
expansion of such brevity. For instance, tweets may contain Rd as an abbrevi-
ation, and it is normalized to road. Furthermore, ambiguous location problems
are resolved by employing the geoparsing procedure using the OpenStreetMap
API28. LNEx improved the average F-Score by 98-145%, outperforming all the
state of art taggers.
4.7 Community on Social Media
People with distinct feelings, expression, solutions, and intelligence, share their
opinions on Twitter. Such a diversified content can be related to elections, foot-
ball game or a domain that is influenced by public views. With the abundance
of textual data, one can envision the power of collective intelligence that can be
harnessed for a wise recommendation, judgment and strategy building. Also, it is
a known fact that a judgment call made by a crowd is superior to an individuals
decision [106]. Formation of a diverse group can improve the decision-making
process through what is known as Wisdom of Crowd (WoC). WoC is meant to
minimize regional biases that may cloud objectivity associated with individuals
judgment and bring together different perspectives and knowledge that can en-
hance coverage and comprehensiveness of the analysis. For example, WoC can be
used to design a portfolio of stocks that maximize the profit in the stock market
trading. However, no existing work illustrates the notion of WoC statistically
and analytically. A methodological way for measuring the diversity of the crowd
is crucial to the rise of human social engagement on social media. According to
a recent survey from Pew Research Center, 76% of the American population is
active on social media. It attributes success to a significant amount of online
data and can aid in creating WoC of the social system. In [107], fantasy premier
league (FPL) is considered to exercise the better judgment of the diverse crowd.
In their work, they predict the best performing team captain in the premier
league, an element dictating the success of a team, based on the scores retrieved
from the fantasy football and content of Twitter users. They utilized Word2vec
similarity measure to quantify the diversity of two groups of users during cap-
tain selection in FPL. Furthermore, They defined and validated their statistical
objective scoring criteria to measure the quality of crowd judgment.
4.8 Demographics
In many applications, demographic information is a key to analysis that depends
on different segments of the population concerning age groups, ethnicity, and
28 https://wiki.openstreetmap.org/wiki/API_v0.6
22
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
gender. For example, age is critical for understanding drug abuse, while gender
is critical to understand vulnerability to depression. Twitter in its current state
does not require users to provide any demographic information.
4.8.1 Age Estimation using social media
Researchers developed a machine learning system coupled with the DBpedia
knowledge graph utilizing the user follower-followee networks to predict the most
probable age of a Twitter user, in [108]. They gathered pre-identified famous
people from DBpedia, based on their occupations and areas of interest, which
also included their birth dates. Then they extracted a sample of 23,120 users who
are in one/two hops of follower-followee network of famous people. Some of the
user profiles were spam/bot and hence they were removed. Then they selected
16K users among the followers of the top 50 famous figures as their training set
and 8K as their testing set. They achieved 84% accuracy in predicting the age of
these users. They selected Support Vector Regression (SVR) with K-Fold Cross
Validation [109] as their best performing model after evaluating using Linear
Regression [110], Least Absolute Shrinkage and Selection Operator (LASSO)
[111], and ElasticNet [112].
Zhang et al. [113] studied the problem of age prediction on Twitter, using
SVM and least square optimization algorithm in building the model. They uti-
lized various features such as linguistic, textual, and network, to improve their
model, achieving an F1 score of 0.81. They discovered that the characteristics
of users in the same age groups have similar content and interactions between
each other. On the other hand, Nguyen et al. [114] investigated the relationship
between the language used in tweets of a user and his/her age. They annotated
the dataset that was collected following a guideline formed based on the tweet
content of users in different age groups such as explicit or implicit age or life
stage mentions. They found that the language use of people in same age groups
is similar regarding the word and phrase selections as well as the topics that
they are talking about. For instance, the following two sets of words, school,
son, daughter, wish, enjoy, thanks, take care and haha, xd, internship, school
have been used by users in two different age groups. In their analysis, they used
linear and logistic regression models with unigram feature only, achieving an F1
score of 0.76.
4.8.2 Gender Estimation using Twitter
Estimation of the gender of a twitter user is beneficial to the analysis of Twit-
ter data for health-related, drug abuse, and harassment activities. Existing ap-
proaches utilized statistical features [115] and seldom involved background knowl-
edge along with social information. In [116], a dataset from Sina Weibo, which
is a counterpart of the micro-blogging platform Twitter, in China, was used to
assess their methodology for gender prediction. [117] exploits online behavioral
and textual features and choice of vocabulary for each user. Online behavioral
features include the number of fans, attention, messages, comments, forwards
and a ratio of original/forwarded messages. Textual features include hashtags,
Predictive Analysis on Twitter: Techniques and Applications
23
URLs, emoticons, and sentence length. They also made use of username and pic-
tures in content. Lexical features were extracted from the content using TF-IDF.
They used four algorithms for predicting gender: Decision Tree, Naive Bayes, Lo-
gistic Regression and Support Vector Machines (SVMs), and found that SVM
outperformed other classifiers by attaining accuracy of 94.3%.
4.9 Anomaly & Popularity Prediction
Twitter has become a playground for spammers. While public conversations on
Twitter are diverse and challenging to analyze and summarize, spammers and
bots further complicate the reliability of the outcome. Bots are automated soft-
ware that is programmed to post a predefined content. They are being used
mostly to propagate or promote bias and skew votes in politics, views on so-
cial issues, or provide impetus to promotional campaigns. On the other hand,
prediction of the popularity of trending topics or issues requires robust analysis
that takes into account anomalous accounts.
Thomas et al. [25] collected 1.8 billion tweets sent by 32.9 million users and
manually identified 1.1 million suspended accounts as spammer accounts along
with 80 million anomalous tweets. They used user behavior regarding interac-
tions with other users, public Twitter handler service usage and textual features
of tweet content such as shortened URLs created using free web hosting services.
Volkova et al. [118] also studied this problem by applying a deep learning tech-
nique, Recurrent Neural Networks (RNN), using tweet metadata and network
features. They compared their approach with state-of-the-art machine learning
methods such as log-linear models. Their RNN model outperformed all the ma-
chine learning models built using various combinations of features with 0.95
F1 score. Sentiment has also been used in spam detection works [119, 120] as
a feature to detect bots on Twitter. Varol et al. [120] also studied the detec-
tion of online bots on Twitter, and utilized Random Forests, AdaBoost, Logistic
Regression and Decision Tree algorithms. They found Random Forest classifier
achieved the best performance with 0.95 AUC score. They made use of senti-
ment features that they extracted from the text beside tweet and user metadata,
textual, linguistic and network features.
Poblete et al. [121] have investigated the tweet credibility issue in the news
disseminated on the platform. They crowdsourced the task of evaluating the
credibility of each tweet to determine if it has newsworthy topics, labeling each
tweet using automated credibility analysis. Labels given by crowdsourcing pro-
cess were used in the training phase. They used SVM, decision trees, decision
rules and Bayesian networks, and best results were given by J48 decision tree,
achieving an 86% F1 score. Ross et al. [122] created a robust and general feature
set for learning to rank tweets based on credibility and newsworthiness. In pre-
vious works by Gupta et al. [123–125], they have demonstrated that when the
training and testing data are from two distinct time periods, the ranker performs
poorly. Ross et al. [122] improved upon this by creating a feature set that does
not overfit a particular year or a set of topics, which is critical for robust analysis
of social media over time and across different domains.
Varol et al. [34] conducted a time series analysis to predict if a trending meme
24
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
is organic or promoted by a group. They aimed to predict memes that have
potential to trend before it becomes trending; therefore, the task of predicting
trends is naturally forced to utilize a sparse dataset. For this reason, they had
to reliably extract textual, linguistic, tweet and user metadata, network and
statistical features, from a small dataset. They used three learning algorithms
namely, K-Nearest Neighbor (KNN) with Dynamic Time Warping (KNN-DTW),
Symbolic Aggregate approXimation with Vector Space Model (SAX-VSM) and
KNN. KNN is a machine learning algorithm for classification and DTW for
multi-dimensional time series. They found KNN-DTW and KNN showed the
best performance in prediction. They used AUC as evaluation metric to measure
accuracy because it is not biased by the imbalance in classes (e.g., 75 promoted
trends versus 852 organic ones). Weng et al. [126] studied the prediction of the
popularity of meme on Twitter. They relied mostly on network features besides
tweet and user metadata, using random forest and linear regression. They ex-
tracted 13 features such as some early adopters, average shortest network path
length between users, the diameter between users, and the number of infected
communities. They built their model using random forest and tested against five
different baselines that used linear regression along with different combinations
of the 13 features. Their model achieved 0.85 F1 score, outperforming the base-
lines. Kobayashi et al. [127] predicted the popularity of a tweet in terms of the
number of retweets in a time window in the future. They used time series anal-
ysis using a method called time-dependent Hawkes process (TiDeH) calculating
infectious rate and using tweet and user metadata such as temporal information
from a tweet and number of followers of a user. They evaluated their system
against other existing methods that incorporated linear regression and Poisson
process and reported that it outperformed other approaches achieving around
5% mean error rate. Tsur et al. [128] also studied the popularity of hashtags
on Twitter, through linguistic features of the tweet text, specifically hashtags.
They obtained promising results using a modified version of Gradient Boosted
Trees called Gradient Boosted rank. They compared their approach with SVM
and Least-effort algorithms, obtaining 0.11 mean error rate. Ruan et al. [129]
predicted the volume of tweets, analyzing the user behavior on individual as
well as collective level. Besides tweeting activity and content analysis of users,
they utilized the underlying follower-followee network, user network structure,
neighboring friends influence and user past activity as features. They used lin-
ear regression model with multiple features that include network structure, user
interaction, content characteristics and past activity, and found that combining
features yields the best performance.
4.10 Sales & Stock Price Prediction
As social media, particularly Twitter, users share their satisfaction or frustration
with products on the platform, these user reviews can be exploited by companies
to generate actionable insights to meet customer expectations and eventually
provide better quality products and services. Industrial applications of predictive
analysis of social media have been gradually adopted, to gain the understanding
Predictive Analysis on Twitter: Techniques and Applications
25
Table 1. Comparative Analysis of Applications and their Evaluation. Acronyms for
Algorithms and Features are described in Table 2
Ref. & Evaluation
Application
Algorithms
Features
User Profiling
User Classification
SVM
LinR
CNN
RF
NB
LogR, LASSO
UsM, TwM
Ling, Nw,
Stat, Txt
Vis
[39]
F1=0.88
[113] F1=0.81
[130] F1=0.59
[131] F1=0.83
[49]
F1=0.77
[132] Acc=0.82
[89] Acc=0.92
F1=∼0.75
[90]
[114] F1=0.76
[54] AUC=∼0.7
[15] AUC=0.8
[108] Acc=0.84
[53] Acc=0.72
F1=0.62
[133]
User Attitude, Personality,
and Mood Prediction
SVM, NB, RF
TwM, Txt, Ling,
Nw, Stat
[134] AUC=0.8
Sales & Stock price prediction NB, RF, SVM
TwM
[59] Med error=0.32 Social and Political events,
[135]] F1=0.58
Elections, Collective action
[136] Acc=0.85
[137] AUC=0.91
PosR
NBR, SVM,
LogR, CNN, RF
TwM, UsM
Txt, Ling, Nw
[34] AUC=0.95
[126] F1=∼0.85
[127] Mean err= ∼0.179
[128] Mean err=∼0.11
Popularity prediction
KNN-DTW, SAX-VSM, Txt
KNN, RF,
TiDeH, LinR, LogR
GrB, SVM
Nw and Stat,
TwM and UsM
Ling
Spam bot detection
Troll detection
Credibility prediction
KNN-DTW, SAX-VSM, Text, Nw,
KNN, LinR, DT
NB, GrB
RF, AdB,
BNet, RNN,
SVM, LogR
Ling, TwM, UsM
Stat, Vis
Sentiment analysis
Emotion Detection
NB, SVM, CRF,
Ling,
LibLin, RBF-NN, AdB, Txt,
TwM
RT, REPTree,
BNet, LogR
and UsM
[25] Acc=0.89
[34] AUC=0.95
[119] AUC=0.73
[118] F1= 0.95
[138] F1=0.99
[120] AUC=0.95
[121] F1=0.86
[17] Pre=∼0.80
[16] F1=0.77
[18] F1=0.67
[22] F1=∼0.60
[29] F0.5=0.62
[21] Acc=0.82
[6] Acc=0.61
[139] Mean error=0.071
[140] Acc=∼0.71
[141] F0.5= 0.76
F1=∼0.73
[64]
F1=0.79
[26]
[101] Acc=∼0.83
[142] Mean error=0.39
[143] Acc=0.88
[144] Acc=0.79
Location Estimation
Traffic Estimation
NB, SVM, DT,
LinR, MxEnt
Ling, TwM, UsM
Txt
[145] F1= 0.70
[146] Pre=0.86
Finding Important Users,
Community Detection
RNN, SVM
Nw, UsM, Txt
[147] LFK-NMI=0.13 Topic Extraction,
[148] F1=0.63
Meme Extraction
HiCl, KM, LDA, SVM Txt, Ling, TwM, UsM
Nw
[149] Acc=0.84
[150] AUC=0.83
[151] F= 0.8736
[83] K Alpha=0.84
Public Health
Health-care
NB,
SVM, RF
LDA, ssToT
Txt, Ling, TwM
Sentiment,
[102]
F1=0.81
[152] R2= 0.67
Disaster Management
LogR
Txt, Ling, UsM, TwM
Nw
26
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
of the market. Some of the use-cases that have adopted social media data for
decision making are for:
1. Improvement of Customer Service: Delta Airline exploited social media to
identify the reasons for customer frustration. For instance, lost luggage or
poor service.
2. New Products Research and Development: JD Power quality assessment has
determined that car company modify car seats based user sentiments on the
social sphere [153].
3. Key Influencers: A cosmetics company L'Oreal uses social media follower-
followee network to find Influencers for promotions29.
4. Recommendations through deep learning: YouTube utilizes the deep neural
network to enhance their recommendation system using implicit feedback by
analyzing users comments and videos of interest [154].
Georgiev et al. [134] investigated the question of how the Olympic Games
impacted the sales of businesses in London. They used Twitter posts along with
the check-ins through Foursquare platform to extract mostly location-specific
features from tweet text and tweet metadata, such as the distance of businesses
to stadiums and sponsor businesses, transitions to entertainment places and
social areas. They evaluated their work using AUC, for Nave Bayes, Random
Forest, and SVM algorithms and reported that SVM performed best with 0.8 of
AUC.
[155] employs feed-forward network (FF) for predicting the likelihood of a
customer to buy a product. They restricted their dataset to tweets about mobile
phones and cameras, expensive products that people often buy after doing some
research online. Before predicting the likelihood of purchasing a product, they
predicted whether a tweet represents the respective users purchasing behavior.
Then they predict whether the user will purchase the product after 60 days time
window of tweeting. They compared the performance of their approach with
Long Short Term Memory (LSTM), Recurrent Neural Networks (RNNs) (with
varying dropout rates) based implementation and observed that their approach
with FF surpasses others by small margins. FF learning cycle involved RM-
Sprop [156], sigmoid activations and negative log-likelihood function with batch
training.
5 Conclusion
Twitter has positioned itself as an essential part of the social media environment
becoming an emergent communication medium. This development has opened
up new opportunities for researchers to gauge the pulse of the populace reliably
and use that to study public opinion, form policies, understand the impact of
events, and find newer ways to address certain problems. Social media data has
already enabled researchers to predict the trends and outcomes of several crit-
ical real-world events, and its reliability and coverage can further be improved
29 https://www.socialmediatoday.com/special-columns/adhutchinson/2015-09-09/big-brand-
theory-loreal-stays-connected-their-audience
Predictive Analysis on Twitter: Techniques and Applications
27
by incorporating background knowledge [85, 20]. Specifically, monitoring the en-
gagement and public opinion about ongoing events from temporal and spatial
perspectives can foretell their evolution as well as the outcome. Moreover, this
information can complement traditional surveys or polls that are conducted by
non-government agencies to improve our confidence in the prediction, as tradi-
tional methods alone can be misleading or sluggish in reacting to rapidly chang-
ing events. In order to account for the complex decision making that requires
consideration of a number of factors that can impact a situation or an event,
incorporation of as many signals as possible in comprehending the big picture
is necessary. We have explored a predictive analysis paradigm that comprises
two levels of prediction, using coarse-grained analysis built upon fine-grained
analysis. Such analysis have been conducted with creditable success for events
such as elections, gun violence, drug misuse or illicit drug use [3].
In this chapter, we have discussed processes, algorithms, and applications
concerning predictive analysis in different domains. We illustrated fine-grained
analysis by customizing domain-independent approaches to extract signals such
as sentiment, emotions, and topics through the application of machine learning
models, and coarse-grained analysis by aggregating and cultivating the signals to
make predictions. We have also provided details of related prominent studies in
ten different domains such as healthcare, public health, political and social issues,
disaster management, sales and stock prediction, and demographics. The follow-
ing table summarizes related work describing various applications and methods
used.
Table 2. The Acronyms used in the comparative table.
Acronym Algo. Description
LinR
RF
NB
LogR
PosR
NBR
GB
AdB
DT
BNet
LibLin
HiCl
KM
RT
Acronym Feature Descr.
UsM
Linear Regression
TwM
Random Forest
Ling
Nave Bayes
Nw
Logistic Regression
Poisson Regression
Stat
Negative Binomial Reg. Txt
Gradient Boosting
Vis
AdaBoost
Decision Trees
Bayes Net
LIBLINEAR
Hierarchical Clustering
K-Means
Random Tree
User metadata
Tweet metadata
linguistic
Network
Statistical
Textual
Visual
6 Acknowledgement
We are grateful to Amelie Gyrard, Mustafa Nural, Sanjaya Wijeratne, Shreyansh
Bhatt, Ankita Saxena for their assistance with their reviews and comments that
greatly improved this book chapter.
We acknowledge partial support from the National Science Foundation (NSF)
award CNS-1513721: "Context-Aware Harassment Detection on Social Media",
28
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
National Institutes of Health (NIH) award: MH105384-01A1: "Modeling Social
Behavior for Healthcare Utilization in Depression", NSF award EAR- 1520870
Hazards SEES: Social and Physical Sensing Enabled Decision Support for Dis-
aster Management and Response, Community in Social Media : This work was
supported by Army Research Office Grant No. W911NF-16-1-0300, National In-
stitute on Drug Abuse (NIDA) Grant No. 5R01DA039454-02 Trending: Social
media analysis to monitor cannabis and synthetic cannabinoid use. Any opin-
ions, conclusions or recommendations expressed in this material are those of the
authors and do not necessarily reflect the views of the NSF, NIH, NIDA or Army
Research Office.
References
1. H. Purohit and A. Sheth, "Twitris v3: From Citizen Sensing to Analysis, Coor-
dination and Action," in ICWSM, 2013.
2. C. A. Davis, G. L. Ciampaglia, L. M. Aiello, K. Chung, M. D. Conover, E. Ferrara,
A. Flammini, G. C. Fox, X. Gao, B. Gon¸calves, P. A. Grabowicz, K. Hong, P.-
M. Hui, S. Mccaulay, K. Mckelvey, M. R. Meiss, S. Patil, C. P. Kankanamalage,
V. Pentchev, J. Qiu, J. Ratkiewicz, A. Rudnick, B. Serrette, P. Shiralkar, O. Varol,
L. Weng, T.-L. Wu, A. J. Younge, and F. Menczer, "OSoMe: the IUNI observatory
on social media," PeerJ Computer Science.
3. A. Sheth, H. Purohit, G. A. Smith, J. Brunn, A. Jadhav, P. Kapanipathi, C. Lu,
and W. Wang, "Twitris: A System for Collective Social Intelligence," Encyclopedia
of Social Network Analysis and Mining, 2018.
4. K. B. Penuel and M. Statler, Encyclopedia of disaster relief. Sage Publications,
2011.
5. J. Malilay, M. Heumann, D. Perrotta, A. F. Wolkin, A. H. Schnall, M. N. Pod-
gornik, M. A. Cruz, J. A. Horney, D. Zane, R. Roisman, J. R. Greenspan, D. Thor-
oughman, H. A. Anderson, E. V. Wells, and E. F. Simms, "The Role of Applied
Epidemiology Methods in the Disaster Management Cycle," American Journal of
Public Health, vol. 104, no. 10, pp. 2092–2102, 2014.
6. W. Wang, L. Chen, K. Thirunarayan, and A. P. Sheth, "Harnessing Twitter 'Big
Data' for Automatic Emotion Identification," in IEEE International Confernece
on Social Computing (SocialCom), 2012.
7. F. R. Lamy, R. Daniulaityte, R. W. Nahhas, M. J. Barratt, A. G. Smith,
A. Sheth, S. S. Martins, E. W. Boyer, and R. G. Carlson, "Increases in synthetic
cannabinoids-related harms: Results from a longitudinal web-based content anal-
ysis," International Journal of Drug Policy, 2017.
8. A. Sheth and P. Kapanipathi, "Semantic Filtering for Social Data," IEEE Internet
Computing, 2016.
9. P. Kapanipathi, F. Orlandi, A. Sheth, and A. Passant, "Personalized Filtering of
the Twitter Stream," in SPIM Workshop at ISWC 2011, 2011.
10. P. Kapanipathi, P. Jain, C. Venkataramani, and A. Sheth, "User Interests Identi-
fication on Twitter Using a Hierarchical Knowledge Base," in European Semantic
Web Conference, 2014.
11. D. Cameron, G. A. Smith, R. Daniulaityte, A. P. Sheth, D. Dave, L. Chen,
G. Anand, R. Carlson, K. Z. Watkins, and R. Falck, "PREDOSE: A semantic web
platform for drug abuse epidemiology using social media," Journal of Biomedical
Informatics, vol. 46, pp. 985–997, 2013.
12. H. Saif, Semantic Sentiment Analysis in Social Streams, 2017.
Predictive Analysis on Twitter: Techniques and Applications
29
13. S. Wijeratne, A. Sheth, S. Bhatt, L. Balasuriya, H. S. Al-Olimat, M. Gaur, A. H.
Yazdavar, and K. Thirunarayan, "Feature Engineering for Twitter-based Appli-
cations," in Feature Engineering for Machine Learning and Data Analytics, 2017,
p. 35.
14. K. Gimpel, N. Schneider, B. O 'connor, D. Das, D. Mills, J. Eisenstein, M. Heil-
man, D. Yogatama, J. Flanigan, and N. A. Smith, "Part-of-Speech Tagging for
Twitter: Annotation, Features, and Experiments," Proceedings of ACL, 2011.
15. C. Wagner, S. Asur, and J. Hailpern, "Religious Politicians and Creative Photog-
raphers: Automatic User Categorization in Twitter," in SocialCom, 2013.
16. X. Wang, F. Wei, X. Liu, M. Zhou, and M. Zhang, "Topic Sentiment Analysis
in Twitter: A Graph-based Hashtag Sentiment Classification Approach," in Pro-
ceedings of the 20th ACM international conference on Information and knowledge
management. ACM, 2011.
17. D. Davidov, O. Tsur, and A. Rappoport, "Enhanced Sentiment Learning Using
Twitter Hashtags and Smileys," in Proceedings of the 23rd international confer-
ence on computational linguistics. ACM, 2010, pp. 241–249.
18. E. Kouloumpis, T. Wilson, and J. Moore, "Twitter sentiment analysis: The good
the bad and the omg!" Proceedings of the Fifth International AAAI Conference
on Weblogs and Social Media (ICWSM 11), pp. 538–541, 2011.
19. 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 Proceedings of the 20th international conference on
World wide web., 2011.
20. F. Morstatter, J. Pfeffer, H. Liu, and K. M. Carley, "Is the Sample Good Enough?
Comparing Data from Twitter's Streaming API with Twitter's Firehose," in
ICWSM, 2013, pp. 400–408.
21. A. Go, R. Bhayani, and L. Huang, "Twitter Sentiment Classification using Distant
Supervision," Tech. Rep., 2009.
22. A. Agarwal, B. Xie, and I. Vovsha, "Sentiment analysis of twitter data," in Pro-
ceedings of the Workshop on Language in Social Media (LSM 2011), no. June,
2011, pp. 30–38.
23. B. Suh, L. Hong, P. Pirolli, and E. H. Chi, "Want to be Retweeted? Large Scale
Analytics on Factors Impacting Retweet in Twitter Network," in IEEE interna-
tional conference on Social Computing Social computing (SocialCom), 2010.
24. N. Naveed, T. Gottron, J. Kunegis, and A. C. Alhadi, "Bad News Travel Fast : A
Content-based Analysis of Interestingness on Twitter," in Proceedings of the 3rd
International Web Science Conference. ACM, 2011.
25. K. Thomas, C. Grier, and V. Paxson, "Suspended Accounts in Retrospect: An
Analysis of Twitter Spam," in Proceedings of the 2011 ACM SIGCOMM confer-
ence on Internet measurement, 2011.
26. K.-L. Liu, W.-J. Li, and M. Guo, "Emoticon Smoothed Language Models for
Twitter Sentiment Analysis," in Proceedings of the Twenty-Sixth AAAI Confer-
ence on Artificial Intelligence, 2012.
27. C. Zhai, J. Lafferty, J. Lafferty, and C. Zhai, "A Study of Smoothing Methods
for Language Models Applied to Information Retrieval," ACM Transactions on
Information Systems, vol. 22, no. 2, pp. 179–214, 2004.
28. M. Boia and B. Faltings, "A :) Is Worth a Thousand Words: How People Attach
Sentiment to Emoticons and Words in Tweets," in SocialCom, 2013.
29. A. Pak and P. Paroubek, "Twitter as a Corpus for Sentiment Analysis and Opin-
ion Mining," LREc, vol. 10, 2010.
30
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
30. R. Kelly and L. Watts, "Characterising the Inventive Appropriation of Emoji as
Relationally Meaningful in Mediated Close Personal Relationships," Experiences
of Technology Appropriation: Unanticipated Users, Usage, Circumstances, and
Design, 2015.
31. P. K. Novak, J. Smailovi´c, B. Sluban, and I. Mozetic, "Sentiment of Emojis,"
PLOS One, 2015.
32. H. Miller, J. Thebault-Spieker, S. Chang, I. Johnson, L. Terveen, and B. Hecht,
"Blissfully happy or ready to fight: Varying Interpretations of Emoji," Interna-
tional AAAI Conference on Web and Social Media, no. ICWSM, pp. 259–268,
2016.
33. S. Wijeratne, L. Balasuriya, A. Sheth, and D. Doran, "EmojiNet: An Open Service
and API for Emoji Sense Discovery," in ICWSM, 2017.
34. O. Varol, E. Ferrara, F. Menczer, and A. Flammini, "Early detection of promoted
campaigns on social media," EPJ Data Science, vol. 6, no. 1, 2017.
35. L. C. Freeman, "Centrality in Social Networks Conceptual Clarification," Social
Networks, vol. 179, pp. 215–239, 1978.
36. L. Freeman, "A set of measures of centrality based on betweenness," Sociometry,
vol. 40, no. 1, pp. 35–41, 1977.
37. P. Bonacich, "Power and centrality : A family of measures," American Journal of
Sociology, vol. 92, no. 5, pp. 1170–1182, 1987.
38. G. Lawyer, "Understanding the influence of all nodes in a network," Nature Sci-
entific Reports, 2015.
39. M. Pennacchiotti and A.-M. Popescu, "Democrats, Republicans and Starbucks
Afficionados: User Classification in Twitter," in Proceedings of the 17th ACM
SIGKDD international conference on Knowledge discovery and data mining, 2011.
40. R. Irfan, C. K. King, D. Grages, S. Ewen, S. U. Khan, S. A. Madani, J. Kolodziej,
L. Wang, D. Chen, A. Rayes, N. Tziritas, C.-Z. Xu, A. Y. Zomaya, A. S. Alzahrani,
and H. L, "A Survey on Text Mining in Social Networks," The Knowledge Engi-
neering Review, vol. 000, pp. 1–24, 2004.
41. A. K. Nassirtoussi, S. Aghabozorgi, T. Y. Wah, D. Chek, and L. Ngo, "Text
mining for market prediction: A systematic review," EXPERT SYSTEMS WITH
APPLICATIONS, vol. 41, pp. 7653–7670, 2014.
42. F. Franch, "(Wisdom of the Crowds) : 2010 UK Election Prediction with Social
Media," Journal of Information Technology & Politics, vol. 10, no. 1, pp. 57–71,
jan 2013.
43. F. Bravo-Marquez, D. Gayo-Avello, M. Mendoza, and B. Poblete, "Opinion Dy-
namics of Elections in Twitter," in Eighth Latin American Web Congress, 2012.
44. L. Hong, O. Dan, and B. Davison, "Predicting Popular Messages in Twitter," in
WWW, 2011.
45. M. Sokolova, K. Huang, S. Matwin, J. Ramisch, V. Sazonova, R. Black, C. Orwa,
S. Ochieng, and N. Sambuli, "Topic Modelling and Event Identification from
Twitter Textual Data," ArXiv preprint, 2016.
46. S. T. Dumais, "Latent semantic analysis," Annual Review of Information Science
and Technology, vol. 3, no. 11, p. 4356, 2008.
47. T. Mikolov, I. Sutskever, K. Chen, G. Corrado, and J. Dean, "Distributed Repre-
sentations of Words and Phrases and their Compositionality," Advances in neural
information processing systems, 2013.
48. S. Wijeratne, L. Balasuriya, D. Doran, A. Sheth, and A. Org, "Word Embeddings
to Enhance Twitter Gang Member Profile Identification," in IJCAI Workshop on
Semantic Machine Learning, 2016.
Predictive Analysis on Twitter: Techniques and Applications
31
49. L. Balasuriya, S. Wijeratne, D. Doran, and A. Sheth, "Finding Street Gang Mem-
bers on Twitter," in ASONAM, 2016.
50. S. Sakaki, Y. Miura, X. Ma, K. Hattori, and T. Ohkuma, "Twitter User Gender
Inference Using Combined Analysis of Text and Image Processing," in Proceedings
of the 25th International Conference on Computational Linguistics, 2014, pp. 54–
61.
51. K. Bontcheva, L. Derczynski, A. Funk, M. a. Greenwood, D. Maynard, and
N. Aswani, "TwitIE : An Open-Source Information Extraction Pipeline for Mi-
croblog Text," Proceedings of Recent Advances in Natural Language Processing,
no. September, pp. 83–90, 2013.
52. T. Mitra and E. Gilbert, "CREDBANK: A Large-Scale Social Media Corpus with
Associated Credibility Annotations," in ICWSM, 2016.
53. M. De Choudhury, M. Gamon, S. Counts, and E. Horvitz, "Predicting Depression
via Social Media," in ICWSM, 2013.
54. Y. Lewenberg, Y. Bachrach, and S. Volkova, "Using emotions to predict user
interest areas in online social networks," in Data Science and Advanced Analytics
(DSAA), 2015.
55. H. Wang, D. Can, A. Kazemzadeh, F. Bar, and S. Narayanan, "A System for
Real-time Twitter Sentiment Analysis of 2012 U.S. Presidential Election Cycle,"
in Proceedings of the 50th Annual Meeting of the Association for Computational
Linguistics, 2012, pp. 115–120.
56. M. Ebrahimi, A. H. Yazdavar, and A. Sheth, "On the Challenges of Sentiment
Analysis for Dynamic Events," IEEE Intelligent Systems, 2017.
57. S. Vieweg, A. L. Hughes, K. Starbird, and L. Palen, "Microblogging During Two
Natural Hazards Events: What Twitter May Contribute to Situational Aware-
ness," in CHI - Crisis Informatics, 2010.
58. L. Chen, W. Wang, and A. P. Sheth, "Are Twitter Users Equal in Predicting Elec-
tions? A Study of User Groups in Predicting 2012 U.S. Republican Presidential
Primaries," in Social Informatics, 2012.
59. M. De Choudhury, S. Jhaver, B. Sugar, and I. Weber, "Social Media Participation
in an Activist Movement for Racial Equality," in ICSWM, no. Icwsm, 2016, pp.
92–101.
60. H. Purohit, A. Hampton, V. L. Shalin, A. P. Sheth, J. Flach, and S. Bhatt, "What
kind of #conversation is Twitter? Mining #psycholinguistic cues for emergency
coordination," Computers in Human Behavior, vol. 29, pp. 2438–2447, 2013.
61. H. Purohit, A. Hampton, S. Bhatt, V. L. Shalin, A. P. Sheth, and J. M. Flach,
"Identifying Seekers and Suppliers in Social Media Communities to Support Crisis
Coordination," in Computer Supported Cooperative Work (CSCW), 2014.
62. H. Purohit, S. Bhatt, A. Hampton, V. L. Shalin, and A. P. Sheth, "With Whom
to Coordinate, Why and How in Ad- Hoc Social Media Communications during
Crisis Response," in Proceedings of the 11 th International ISCRAM Conference,
2014, pp. 787–791.
63. Shreyansh Bhatt, Hemant Purohit, and Andrew Hampton, "Assisting Coordina-
tion during Crisis: A Domain Ontology based Approach to Infer Resource Needs
from Tweets," in Web Science, 2014.
64. L. T. Nguyen, P. Wu, W. Chan, W. Peng, and Y. Zhang, "Predicting collective
sentiment dynamics from time-series social media," in Proceedings of the First
International Workshop on Issues of Sentiment Discovery and Opinion Mining
(WISDOM), 2012, pp. 6:1–6:8.
32
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
65. D. Stojanovski, G. Strezoski, G. Madjarov, and I. Dimitrovski, "Finki at SemEval-
2016 Task 4: Deep Learning Architecture for Twitter Sentiment Analysis," in
Proceedings of SemEval, 2016, pp. 149–154.
66. A. Esuli, F. Sebastiani, C. Nazionale, and D. Ricerche, "Optimizing Text Quanti-
fiers for Multivariate Loss Functions," ACM Transactions on Knowledge Discov-
ery from Data ACM Trans. Knowl. Discov. Data. VV, vol. 26, 2015.
67. T. L. Griffiths, M. Steyvers, J. B. Tenenbaum, and T. Griffiths, "Topics in se-
mantic representation Topics in semantic representation," Psychological review,
2007.
68. L. Chen, C. Org, W. Wang, W. Org, M. Nagarajan, S. Wang, A. P. Sheth, and
A. Org, "Extracting Diverse Sentiment Expressions with Target-Dependent Po-
larity from Twitter," in Proceedings of the Sixth International AAAI Conference
on Weblogs and Social Media, 2012.
69. L. Hong and B. D. Davison, "Empirical Study of Topic Modeling in Twitter," in
1st Workshop on Social Media Analytics (SOMA '10), 2010.
70. W. X. Zhao, J. Jiang, J. He, Y. Song, P. Achananuparp, E.-P. Lim, and X. Li,
"Topical Keyphrase Extraction from Twitter," in Proceedings of the 49th Annual
Meeting of the Association for Computational Linguistics, 2011, pp. 379–388.
71. X. Wang, M. S. Gerber, and D. E. Brown, "Automatic Crime Prediction us-
ing Events Extracted from Twitter Posts," in International conference on social
computing, behavioral-cultural modeling, and prediction. Springer, 2012.
72. N. Bhattacharya, I. Arpinar, and U. Kursuncu, "Real Time Evaluation of Quality
of Search Terms during Query Expansion for Streaming Text Data Using Velocity
and Relevance," in Proceedings - IEEE 11th International Conference on Seman-
tic Computing, ICSC 2017, 2017.
73. L. Phillips, C. Dowling, K. Shaffer, N. Hodas, and S. Volkova, "Using Social Media
To Predict the Future: A Systematic Literature Review," in arXiv preprint, 2017.
74. J. M. Robillard, T. W. Johnson, C. Hennessey, B. L. Beattie, and J. Illes, "Aging
2.0: Health Information about Dementia on Twitter," Plos One, vol. 20, no. 87,
2013.
75. V. M. Prieto, S. Rgio Matos, M. Lvarez, F. Cacheda, J. L. Oliveira, and J. A. An,
"Twitter: A Good Place to Detect Health Conditions," PLoS ONE, vol. 9, no. 1,
2014.
76. A. H. Yazdavar, H. S. Al-Olimat, M. Ebrahimi, G. Bajaj, T. Banerjee,
K. Thirunarayan, J. Pathak, and A. Sheth, "Semi-Supervised Approach to Moni-
toring Clinical Depressive Symptoms in Social Media," in Proceedings of the 2017
IEEE/ACM International Conference on Advances in Social Networks Analysis
and Mining, 2017.
77. G. Coppersmith, M. Dredze, C. Harman, and K. Hollingshead Ihmc, "From
ADHD to SAD: Analyzing the Language of Mental Health on Twitter through
Self-Reported Diagnoses," in Proceedings of the 2nd Workshop on Computational
Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality,
2015, pp. 1–10.
78. T. H. Do, D. M. Nguyen, E. Tsiligianni, B. Cornelis, and N. Deligiannis, "Mul-
tiview Deep Learning for Predicting Twitter Users' Location," arXiv preprint,
2017.
79. J. H. Lau and T. Baldwin, "An Empirical Evaluation of doc2vec with Practical
Insights into Document Embedding Generation," arXiv preprint, 2016.
80. 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.
Predictive Analysis on Twitter: Techniques and Applications
33
81. H. Bo, P. Cook, T. Imoth, and B. Dw, "Geolocation Prediction in Social Media
Data by Finding Location Indicative Words," in Proceedings of COLING 2012,
2012, pp. 1045–1062.
82. R. Daniulaityte, R. W. Nahhas, S. Wijeratne, R. G. Carlson, F. R. Lamy, S. S.
Martins, E. W. Boyer, G. A. Smith, and A. Sheth, ""Time for dabs": Analyzing
Twitter data on marijuana concentrates across the U.S. HHS Public Access,"
Drug Alcohol Depend, vol. 155, pp. 307–311, 2015.
83. F. R. Lamy, R. Daniulaityte, A. Sheth, R. W. Nahhas, S. S. Martins, E. W. Boyer,
and R. G. Carlson Francois R Lamy, ""Those edibles hit hard": Exploration of
Twitter data on cannabis edibles in the U.S HHS Public Access," Drug Alcohol
Depend, vol. 1, no. 164, pp. 64–70, 2016.
84. P. N. Howard, M. Hussain, and W. Mari, "Opening Closed Regimes What Was
the Role of Social Media During the Arab Spring?" 2011.
85. Z. Tufekci, "Big Questions for Social Media Big Data: Representativeness, Validity
and Other Methodological Pitfalls," in ICWSM, 2014.
86. I. Arpinar, U. Kursuncu, and D. Achilov, "Social media analytics to identify and
counter islamist extremism: Systematic detection, evaluation, and challenging of
extremist narratives online," in Proceedings - 2016 International Conference on
Collaboration Technologies and Systems, CTS 2016, 2016.
87. G. Haciyakupoglu and W. Zhang, "Social Media and Trust during the Gezi
Protests in Turkey," Journal of Computer-Mediated Communication, vol. 20,
no. 4, pp. 450–466, 2015.
88. H. Allcott and M. Gentzkow, "Social Media and Fake News in the 2016 Election,"
Journal of Economic PerspectivesVolume, vol. 31, no. 2Spring, pp. 211–236, 2017.
89. T.-A. Hoang, W. W. Cohen, E.-P. Lim, D. Pierce, and D. P. Redlawsk, "Politics,
Sharing and Emotion in Microblogs," in ASONAM, 2013.
90. A. Makazhanov and D. Rafiei, "Predicting Political Preference of Twitter Users,"
Social Network Analysis and Mining, 2014.
91. R. Cohen and D. Ruths, "Classifying Political Orientation on Twitter: It's Not
Easy!" in ICWSM, 2013.
92. J.-M. Xu, K.-S. Jun, X. Zhu, and A. Bellmore, "Learning from Bullying Traces in
Social Media," in 2012 Conference of the North American Chapter of the Asso-
ciation for Computational Linguistics: Human Language Technologies, 2012, pp.
656–666.
93. Y. Chen, S. Zhu, Y. Zhou, and H. Xu, "Detecting Offensive Language in Social
Media to Protect Adolescent Online Safety," in Privacy, Security, Risk and Trust
(PASSAT), 2012 International Conference on and 2012 International Confernece
on Social Computing (SocialCom), 2012.
94. V. Edupuganti, "Harassment Detection on Twitter using Conversations," Ph.D.
dissertation, 2017.
95. R. Kandakatla, "Identifying Offensive Videos on YouTube," Ph.D. dissertation,
2016.
96. S. Wijeratne, D. Doran, A. Sheth, and J. L. Dustin, "Analyzing the Social Media
Footprint of Street Gangs," in Intelligence and Security Informatics (ISI), 2015.
97. T. Blevins, R. Kwiatkowski, J. Macbeth, K. Mckeown, D. Patton, and O. Ram-
bow, "Automatically Processing Tweets from Gang-Involved Youth: Towards De-
tecting Loss and Aggression," in Proceedings of COLING 2016, the 26th Inter-
national Conference on Computational Linguistics: Technical Papers, 2016, pp.
2196–2206.
98. B. Bushman and L. Huesmann, "Short-term and long-term effects of violent media
on aggression in children and adults," Arch Pediatr Adolesc Med, 2006.
34
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
99. M. Ni, Q. He, and J. Gao, "Using Social Media to Predict Traffic Flow under Spe-
cial Event Conditions," in The 93rd Annual Meeting of Transportation Research
Board, 2014.
100. R. Krishnamurthy, P. Kapanipathi, A. P. Sheth, K. Thirunarayan, and A. Sheth,
"Location Prediction of Twitter Users using Wikipedia," 2014.
101. J. Mahmud, J. Nichols, and C. Drews, "Where Is This Tweet From? Inferring
Home Locations of Twitter Users," in ICWSM, 2012.
102. H. S. Al-Olimat, K. Thirunarayan, V. Shalin, and A. Sheth, "Location Name
Extraction from Targeted Text Streams using Gazeeer-based Statistical Language
Models," arxiv preprint, vol. 11, no. 17, 2017.
103. M. Haklay and P. Weber, "Openstreetmap: User-generated street maps," IEEE
Pervasive Computing, 2008.
104. D. Ahlers, "Assessment of the accuracy of geonames gazetteer data," in GIR,
2013.
105. J. Lehmann, R. Isele, M. Jakob, A. Jentzsch, D. Kontokostas, P. N. Mendes,
S. Hellmann, M. Morsey, P. van Kleef, S. Auer, and C. Bizer, "Dbpedia a large-
scale, multilingual knowledge base extracted from wikipedia," Semantic Web,
vol. 1, pp. 1–5, 2012.
106. M. D. Lee and M. N. Lee, "The relationship between crowd majority and accuracy
for binary decisions," Judgment and Decision Making, vol. 12, no. 4, pp. 328–343,
2017.
107. S. Bhatt, B. Minnery, S. Nadella, B. Bullemer, V. Shalin, and A. Sheth, "En-
hancing crowd wisdom using measures of diversity computed from social media
data," in Proceedings of the International Conference on Web Intelligence, 2017.
108. A. Smith and M. Gaur, "What's my age?: Predicting Twitter User's Age using
Influential Friend Network and DBpedia," arxiv preprint, 2018.
109. P. Refaeilzadeh, L. Tang, and H. Liu, "Cross-Validation," Encyclopedia of
Database Systems. Springer, 2009.
110. D. Nguyen, N. A. Smith, and C. P. Ros´e, "Author Age Prediction from Text using
Linear Regression," in Proceedings of the 5th ACL-HLT Workshop on Language
Technology for Cultural Heritage, Social Sciences, and Humanities. Association
for Computational Linguistics, 2011.
111. C. Chen, Y. Chang, K. Ricanek, and Y. Wang, "Face age estimation using model
selection," 2010 IEEE Computer Society Conference on Computer Vision and
Pattern Recognition - Workshops, pp. 93–99, 2010.
112. A. Culotta, N. Kumar Ravi, and J. Cutler, "Predicting Twitter User Demograph-
ics using Distant Supervision from Website Traffic Data," Journal of Artificial
Intelligence Research, vol. 55, pp. 389–408, 2016.
113. J. Zhang, X. Hu, Y. Zhang, and H. Liu, "Your age is no secret: Inferring mi-
crobloggers' ages via content and interaction analysis," Proceedings of the 10th
International Conference on Web and Social Media, ICWSM 2016, no. Icwsm,
pp. 476–485, 2016.
114. D. Nguyen, R. Gravel, D. Trieschnigg, and T. Meder, ""How Old Do You Think
I Am?": A Study of Language and Age in Twitter," in ICWSM, 2013.
115. D. Bamman, J. Eisenstein, and T. Schnoebelen, "GENDER IN TWITTER:
STYLES, STANCES, AND SOCIAL NETWORKS," CoRR, 2012.
116. W. Li and M. Dickinson, "Gender Prediction for Chinese Social Media Data," in
Proceedings of Recent Advances in Natural Language Processing, 2017, pp. 438–
445.
Predictive Analysis on Twitter: Techniques and Applications
35
117. L. Li, M. Sun, and Z. Liu, "Discriminating Gender on Chinese Microblog: A
Study of Online Behaviour, Writing Style and Preferred Vocabulary," in 10th
International Conference on Natural Computation (ICNC), 2014.
118. S. Volkova and E. Bell, "Identifying Effective Signals to Predict Deleted and
Suspended Accounts on Twitter across Languages," in ICWSM, Association for
the Advancement of Artificial Intelligence, no. Icwsm, 2017, pp. 290–298.
119. J. P. Dickerson, V. Kagan, and V. S. Subrahmanian, "Using Sentiment to Detect
Bots on Twitter: Are Humans more Opinionated than Bots?" in ASONAM, 2014.
120. O. Varol, E. Ferrara, C. A. Davis, F. Menczer, and A. Flammini, "Online Human-
Bot Interactions: Detection, Estimation, and Characterization," in ICWSM, 2017.
121. B. Poblete, C. Castillo, and M. Mendoza, "Information Credibility on Twitter,"
pp. 45–59, 2011.
122. J. Ross and K. Thirunarayan, "Features for Ranking Tweets Based on Credibility
and Newsworthiness," in International Conference on Collaboration Technologies
and Systems, 2016.
123. A. Gupta, P. Kumaraguru, C. Castillo, and P. Meier, "TweetCred: A Real-
timeWeb-based System for Assessing Credibility of Content on Twitter," Lec-
ture Notes in Computer Science (including subseries Lecture Notes in Artificial
Intelligence and Lecture Notes in Bioinformatics), vol. 8851, no. November, 2014.
124. A. Gupta and P. Kumaraguru, "Credibility Ranking of Tweets during High Im-
pact Events," in PSOSM, 2012.
125. A. Gupta, H. Lamba, P. Kumaraguru, and A. Joshi, "Faking Sandy: Characteriz-
ing and Identifying Fake Images on Twitter during Hurricane Sandy," in WWW,
2013.
126. L. Weng, F. Menczer, and Y.-Y. Ahn, "Predicting Successful Memes using Net-
work and Community Structure," in IC, 2014, pp. 535–544.
127. R. Kobayashi and R. Lambiotte, "TiDeH: Time-Dependent Hawkes Process for
Predicting Retweet Dynamics," no. ICWSM, 2016, pp. 191–200.
128. O. Tsur and A. Rappoport, "Don't Let Me Be #Misunderstood: Linguisti-
cally Motivated Algorithm for Predicting the Popularity of Textual Memes," in
ICWSM, Ninth International AAAI Conference on Web and Social Media, 2015,
pp. 426–435.
129. Y. Ruan, H. Purohit, D. Fuhry, S. Parthasarathy, A. P. Sheth, and A. Sheth,
"Prediction of Topic Volume on Twitter," in 4th International ACM Conference
on Web Science, 2012, pp. 397–402.
130. N. Pattisapu, M. Gupta, P. Kumaraguru, and V. Varma, "Medical Persona Clas-
sification in Social Media," in ASONAM, 2017.
131. Z. Gilani, E. Kochmar, and J. Crowcroft, "Classification of Twitter Accounts into
Automated Agents and Human Users," in ASONAM, 2017.
132. J. S. Alowibdi, U. A. Buy, P. S. Yu, and L. Stenneth, "Detecting Deception in
Online Social Networks," in ASONAM, 2014.
133. J. Mahmud, G. Fei, A. Xu, A. Pal, and M. Zhou, "Predicting Attitude and Ac-
tions of Twitter Users," in Proceedings of the 21st International Conference on
Intelligent User Interfaces - IUI '16. ACM Press, 2016, pp. 1–6.
134. P. Georgiev, A. Noulas, and C. Mascolo, "Where Businesses Thrive: Predicting
the Impact of the Olympic Games on Local Retailers through Location-based
Services Data," in ICWSM, 2014, pp. 151–160.
135. X. Yang, R. Mccreadie, C. Macdonald, and I. Ounis, "Transfer Learning for Multi-
language Twitter Election Classification," in ASONAM, 2017.
36
Kursuncu, Gaur, Lokala, Thirunarayan, Sheth and Arpinar
136. R. Korolov, D. Lu, J. Wang, G. Zhou, C. Bonial, C. Voss, L. Kaplan, W. Wal-
lace, J. Han, and H. Ji, "On Predicting Social Unrest Using Social Media," in
IEEE/ACM International Conference on Advances in Social Networks Analysis
and Mining (ASONAM), 2016.
137. N. Kallus, "Predicting Crowd Behavior with Big Public Data," arXiv preprint,
2014.
138. J. Echeverria and S. Zhou, "Discovery, Retrieval, and Analysis of the 'Star Wars'
Botnet in Twitter," in ASONAM, 2017.
139. W. Gao and F. Sebastiani, "Tweet Sentiment: From Classification to Quantifica-
tion," in ASONAM, 2015.
140. A. Hassan, A. Abbasi, and D. Zeng, "Twitter Sentiment Analysis: A Bootstrap
Ensemble Framework," in SocialCom, 2013.
141. A. Kothari, W. Magdy, K. Darwish, A. Mourad, and A. Taei, "Detecting Com-
ments on News Articles in Microblogs," in ICWSM, 2013.
142. T. Georgiou, A. E. Abbadi, X. Yan, and J. George, "Mining Complaints for
Traffic-Jam Estimation: A Social Sensor Application," in ASONAM, 2015.
143. A. J. Aiswal, W. Peng, and T. Sun, "Predicting Time-sensitive User Locations
from Social Media," in ASONAM, 2013.
144. D. Rout, D. Preoiuc-Pietro, K. Bontcheva, and T. Cohn, "Where's @wally? A
Classification Approach to Geolocating Users Based on their Social Ties," in 24th
ACM Conference on Hypertext and Social Media, Paris, France, 2013.
145. B. Rath, W. Gao, J. Ma, and J. Srivastava, "From Retweet to Believability:
Utilizing Trust to Identify Rumor Spreaders on Twitter," in ASONAM, 2017.
146. I. Bizid, N. Nayef, P. Boursier, S. Faiz, and J. Morcos, "Prominent Users Detection
during Specific Events by Learning On-and Off-topic Features of User Activities,"
in ASONAM, 2015.
147. E. Ferrara, M. Jafariasbagh, O. Varol, V. Qazvinian, F. Menczer, and A. Flam-
mini, "Clustering Memes in Social Media," in ASONAM, 2013.
148. S. Yamamoto and T. Satoh, "Hierarchical Estimation Framework of Multi-Label
Classifying: A Case of Tweets Classifying into Real Life Aspects," in ICWSM,
2015.
149. A. Beykikhoshk, O. Arandjelovi, D. Phung, and S. Venkatesh, "Data-Mining
Twitter and the Autism Spectrum Disorder: A Pilot Study," in ASONAM, 2014.
150. Z. Yin, Y. Chen, D. Fabbri, J. Sun, and B. Malin, "#PrayForDad: Learning
the Semantics Behind Why Social Media Users Disclose Health Information," in
ICWSM, 2016.
151. R. Daniulaityte, L. Chen, F. R. Lamy, R. G. Carlson, K. Thirunarayan, and
A. Sheth, "When Bad' is Good': Identifying Personal Comm unication and Sen-
timent in Drug-Related Tweets," JMIR PUBLIC HEAL TH AND SURVEIL-
LANCE, 2016.
152. Y. Hu, S. Farnham, and K. Talamadupula, "Predicting User Engagement on
Twitter with Real-World Events," in ICWSM, 2015.
153. J. S. Kessler, M. Eckert, L. Clark, and N. J. Nicolov Power, "The ICWSM 2010
JDPA Sentiment Corpus for the Automotive Domain," in 4th International AAAI
Conference on Weblogs and Social Media Data Workshop Challenge (ICWSM-
DWC), 2010.
154. P. Covington, J. Adams, and E. Sargin, "Deep Neural Networks for YouTube
Recommendations," in Proceedings of the 10th ACM Conference on Recommender
Systems, 2016.
Predictive Analysis on Twitter: Techniques and Applications
37
155. M. Korpusik, S. Sakaki, F. Chen, and Y.-Y. Chen, "Recurrent neural networks
for customer purchase prediction on twitter." in CBRecSys@ RecSys, 2016, pp.
47–50.
156. T. Tieleman and G. Hinton, "Divide the gradient by a running average of its recent
magnitude. coursera: Neural networks for machine learning. technical report."
|
1311.2023 | 1 | 1311 | 2013-10-25T12:54:05 | Information dissemination processes in directed social networks | [
"cs.SI",
"physics.soc-ph"
] | Social networks can have asymmetric relationships. In the online social network Twitter, a follower receives tweets from a followed person but the followed person is not obliged to subscribe to the channel of the follower. Thus, it is natural to consider the dissemination of information in directed networks. In this work we use the mean-field approach to derive differential equations that describe the dissemination of information in a social network with asymmetric relationships. In particular, our model reflects the impact of the degree distribution on the information propagation process. We further show that for an important subclass of our model, the differential equations can be solved analytically. | cs.SI | cs |
Information dissemination processes in directed
social networks
K. Avrachenkov∗ 1, K. De Turck†2, D. Fiems‡2, and B.J. Prabhu§3,4
1INRIA Sophia Antipolis, 2004 route des Lucioles, 06902 Sophia
Antipolis Cedex, France
2Department of Telecommunications and Information Processing,
Ghent University, St. Pietersnieuwstraat 41, 9000 Gent, Belgium
3CNRS, LAAS, 7 avenue du colonel Roche, F-31400 Toulouse,
4Univ de Toulouse, LAAS, F-31400 Toulouse, France
France
August 11, 2018
Abstract
Social networks can have asymmetric relationships. In the online so-
cial network Twitter, a follower receives tweets from a followed person but
the followed person is not obliged to subscribe to the channel of the fol-
lower. Thus, it is natural to consider the dissemination of information in
directed networks. In this work we use the mean-field approach to derive
differential equations that describe the dissemination of information in a
social network with asymmetric relationships.
In particular, our model
reflects the impact of the degree distribution on the information propa-
gation process. We further show that for an important subclass of our
model, the differential equations can be solved analytically.
1 Introduction
We develop mathematical models for the dissemination of information on di-
rected graphs and investigate the influence of parameters such as the degree
distribution on the dynamics of the dissemination process. The directed graph
model, as opposed to the undirected model, is better suited for networks like
the one of Twitter because of the asymmetric relationship that exists between
∗[email protected]
†[email protected]
‡[email protected]
§[email protected]
1
different users. Specifically, in the Twitter network, a user can choose to re-
ceive the tweets -- in other words, become a follower -- of one or more other users
by subscribing to their accounts. Certain users, celebrities for example, have
several millions of followers who follow their tweets. These users do not neces-
sarily follow the tweets of all of their followers which results in an asymmetric
relationship between users. This asymmetry is modelled by a directed graph in
which an outgoing edge is drawn from a user to each of its followers. An edge
in the opposite direction from the follower to the user need not always exist and
is drawn only if the user subscribes to the channel of this follower.
A hashtag is a word or a phrase prefixed by # and used in social networks
as a keyword. The prefix facilitates the search for conversations related to the
prefixed word or phrase. The typical life cycle of a hashtag closely resembles an
epidemic. In the first phase the interest in the hashtag grows as users generate
tweets containing this hashtag. These tweets are received by followers who then
either retweet them or generate new tweets with this hashtag. The number of
users tweeting this hashtag ("infected" users) grows as a function of time during
this phase. At a certain point in time, the interest reaches its zenith and starts
to wane as users move on and get interested in other events. The second phase
begins at this point in time as users stop tweeting this hashtag (or, "recover"),
and the number of infected users decreases.
In this work we use the mean-field approach to derive the differential equa-
tions which describe the process of information dissemination. We obtain a
couple of differential equations which describe the evolution of the fractions of
infected and recovered persons. As a model for the underlying network we take
the Configuration-type model for directed graphs [2]. While epidemics have
been widely studied on undirected graphs, there is a hardly any analysis of the
epidemic-type processes on directed networks. In [7, 6], the mean-field approach
has been applied to the analysis of epidemics on an undirected configuration-
type graph model. In [5], the effect of network topology has been analysed in
the case of undirected graphs.
In particular, the authors of [5] applied their
general results to analyse the Erdos-R´enyi and preferential attachment random
graph models. An interesting approach combining a decomposition approach
with two-state primitive Markov chain has been proposed in [8] for undirected
networks with general topology. For an overview of various results about epi-
demic processes on undirected networks we refer the interested reader to the
books [4, 1, 3].
2 The mean-field model
Consider a network of N nodes structured according to the Configuration-type
model for directed graphs [2]. The in-degree and out-degree of the nodes are
drawn from a distribution f (k, l) = P(K = k, L = l), defined on the bounded
set D = {(k, l) : 0 ≤ k ≤ K, 0 ≤ l ≤ L, (k, l) 6= (0, 0)}, where the first (resp.
second) index corresponds to the in-degree (resp. out-degree) and K (resp. L) is
the maximal in-degree (resp. out-degree). In the remainder, we always assume
2
that EK = EL; in a network every outgoing link is an incoming link of some
other node. A generic node with in-degree k and out-degree l shall be referred
to as a (k, l)-node.
Each node in the network can be in one of the three states : infected, recov-
ered, or susceptible. An infected node infects its susceptible neighbours after
an exponentially distributed time with intensity λ. Note that, the infection is
spread simultaneously along all the outgoing edges and not just one edge at
a time. The simultaneous dissemination along all outgoing edges models the
spread of tweets on Twitter. An infected node recovers after an exponentially
distributed time of rate ν, at which time it stops spreading information in the
network.
We shall be mainly interested in a large-population model, that is when N →
∞. This assumption simplifies considerably the analysis of the dissemination
process while being realistic1.
Let ik,l(t) (resp. rk,l(t)) denote the fraction of infected (resp. recovered)
(k, l) nodes at time t. The following result describes the dynamics of these two
quantities.
Theorem 1. Let ik,l(0) > 0 for some (k, l) ∈ D. Then, ∀(k, l) ∈ D,
dik,l(t)
dt
= λk(f (k, l) − ik,l(t) − rk,l(t)) Pk′ ,l′ l′ik′,l′ (t)
Pk′ ,l′ l′f (k′, l′)
− ik,l(t)ν,
(1)
and
drk,l(t)
dt
= ik,l(t)ν.
(2)
Sketch of proof. Let I N
k,l(t)) be the number of infected (resp. re-
covered) (k, l) nodes in a network of N nodes. Then in a small time interval
∆,
k,l(t) (resp. RN
I (N )
k,l (t + ∆) = I (N )
k,l (t) + number of (k, l) nodes infected in time ∆
− number of (k, l) infected (k, l) nodes that recover in time ∆.
Since each infected node recovers after an exponentially distributed time of rate
ν, the number of (k, l) infected nodes that recover in ∆ will be approximately
I (N )
k,l (t)ν∆. There will be additional terms containing ∆2 which we neglect.
k,l − (I (N )
Let us compute the number of (k, l) nodes that get infected in time ∆. There
are N (N )
k,l (t)) susceptible (k, l) nodes. Assume that each (k, l)
node has a probability pk,l to get infected in the interval ∆. Then, expected
number of infected (k, l) nodes in time ∆ will be (N
k,l (t) + R(N )
Each (k, l) node has k incoming edges. Assuming that the edges are con-
nected independently, pk,l = (1 − (1 − qk,l)k), where qk,l is the probability that
the infection is transmitted along one of the edges in ∆. The number of (k, l)
nodes infected in ∆ is thus
(N )
k,l − (I
(N )
k,l + R
(N )
k,l ))pk,l.
(N (N )
k,l − (I (N )
k,l + R(N )
k,; )) · (1 − (1 − qk,l)k),
1Twitter has approximately 200 million registered users (source: Wikipedia).
3
which, for ∆ sufficiently small, can be approximated as
(N (N )
k,l − (I (N )
k,l + R(N )
k,l ))kqk,l.
Finally, we compute qk,l. In an interval ∆, each infected node transmits the
(N )
infection with probability λ∆. Thus, there are Pk′,′l l′I
k′ ,l′ λ∆ edges that are
infected and that transmit the infection in ∆. There are a total of Pk′,l′ l′N (N )
k′,l′ .
Assuming that an incoming edge is connected uniformly at random to an out-
going edge, we obtain
qk,l = Pk′ ,l′ l′I (N )
k′,l′ λ∆
Pk′,l′ l′N (N )
k′,l′
.
Consequently,
I (N )
k,l (t + ∆) − I (N )
k,l (t) = (N (N )
k,l − (I (N )
k,l + R(N )
k,l ))kPk′ ,l′ l′I (N )
k′,l′ λ∆
Pk′,l′ l′N (N )
k′,l′
− I (N )
k,l ν∆.
If the initial number of nodes is large, then we can divide the two sides of
the above equation to obtain the following difference equation in terms of the
fraction of the infected and the recovered nodes:
ik,l(t + ∆) − ik,l(t) = (f (k, l) − (ik,l(t) + rk,l(t)))kPk′,l′ l′ik′,l′ (t)λ∆
Pk′,l′ l′f (k′, l′)
− ik,l(t)ν∆.
To complete the picture, we take the limit ∆ → 0, and obtain the differential
equations (1) and (2).
Remark 1. In the above equation ik,l is the fraction of the (k, l) nodes that are
infected. This fraction varies between 0 and f (k, l). If instead, we want to look
at the evolution of the fraction of infected nodes and recovered nodes conditioned
on them being (k, l) nodes, then the corresponding differential equations for these
fractions will be
dik,l(t)
dt
drk,l(t)
dt
= λk(1 − ik,l(t) − rk,l(t))Pk′,l′ l′f (k′.l′)ik′,l′ (t)
Pk′,l′ l′f (k′, l′)
− ik,l(t)ν,
(3)
= ik,l(t)ν.
(4)
3 Epidemics without recovery
The solution of (1) and (2) can be computed numerically. In some specific case
we can obtain explicit solutions to these equations. In particular, this is the
case when there is no recovery: ν = 0, or in the language of Twitter, they keep
generating new tweets with the same hashtag. That is, a hashtag never gets out
of mode. This can well represent the case for the topics or personalities that
can sustain popularity over a long period of time.
4
Since there are no recovered nodes, rk,l(t) = 0, ∀t, and (1) takes the form
dik,l(t)
dt
= λk(f (k, l) − ik,l(t)) Pk′,l′ l′ik′,l′ (t)
Pk′,l′ l′f (k′, l′)
.
(5)
The differential equation (5) can be solved in terms of a reference value of (k, l),
say (k, l) = (1, 1) by noting that
1
dik,l(t)
k(f (k, l) − ik,l(t))
dt
=
1
di1,1(t)
(f (1, 1) − i1,1(t))
dt
,
whence
f (k, l) − ik,l(t) =
f (k, l) − ik,l(0)
(f (1, 1) − i1,1(0))k (f (1, 1) − i1,1(t))k =: ck,l(f (1, 1) − i1,1(t))k.
(6)
The fraction of infected nodes of degree (1, 1) can be obtained by substuting
the value of ik,l(t) in (5) and solving it:
di1,1(t)
dt
= λ(f (1, 1) − i1,1(t))Pk′,l′ l′ (cid:16)f (k′, l′) − ck′ ,l′(f (1, 1) − i1,1(t))k′(cid:17)
Pk′,l′ l′f (k′, l′)
. (7)
Deterministic in-degree
Assume that the in-degree K is deterministic and is equal to d. Then, equation
(5) becomes
did,l(t)
dt
= λd(f (d, l) − id,l(t))X
l′
l′id,l′(t)
Pj jf (d, j)
.
Since the expected in-degree and the expected out-degree coincide, Pj jf (d, j) =
d. Denote Θ(t) = Pj
, and rewrite the above equation as:
jid,j (t)
d
did,l
dt
= λd(f (d, l) − id,l(t))Θ(t).
(8)
Multiplying the above equation by l
the following equation for Θ:
d and summing over all values l, we obtain
dΘ
dt
= λd(1 − Θ)Θ,
which upon integration yields:
id,l(t) = f (d, l) − c1e−λd R Θ(t)dt = f (d, l) −
f (d, l) − id,l(0)
1 − Θ(0) + Θ(0)e−λd·t
.
(9)
5
4 Numerical experiments
In this section, we validate the mean-field model developed in Section 2.
In
the numerical experiments, first the in-degree and out-degree sequences are
generated according to the given degree distributions. So as to have the same
number of incoming stubs as outgoing stubs, the difference between the two is
added to the smaller quantity. A configuration-type graph is then created by
matching an incoming stub with an outgoing stub chosen uniformly at random.
It was shown in [2] that this procedure does indeed approximates closely the
configuration model. The information diffusion process is then simulated on
this graph.
For computing the solution of the system of differential equations (1) and
(2) numerically, the empirical degree distributions from the graph generated
previously are given as input.
The results of two such experiments with 20000 nodes is shown in Figure
1. The in-degree and the out-degree sequences were taken to be independent
of each other. The out-degree sequence was drawn from a Uniform distribution
in the set {1, 20} in the two simulations. For the figure on the left, the in-
degree distribution was taken to be deterministic with parameter 10, and for
the figure on the right it was the Zipf law on {1, 71} and exponent 1.2. In both
experiments, λ = 1 and ν = 0.5, and 5 percent of all nodes were assumed to be
infected at time 0.
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
s
e
d
o
n
d
e
t
c
e
f
n
i
f
o
n
o
i
t
c
a
r
F
diff. eqn.
simulation
0.6
0.5
0.4
0.3
0.2
0.1
s
e
d
o
n
d
e
t
c
e
f
n
i
f
o
n
o
i
t
c
a
r
F
diff. eqn.
simulation
0
0
2
4
6
Time
8
10
12
0
0
2
4
6
Time
8
10
12
Figure 1: Fraction of all nodes infected as a function of time for Deterministic
in-degree distribution (left) and Zipf in-degree distribution (right).
Observations
It is observed that the dissemination process is faster when the variance of
the in-degree distribution is smaller. This observation was reinforced by other
experiments in which the in-degree was drawn from a Uniform distribution. In
several other experiments that we conducted, it was also observed that the out-
degree distribution does not have any noticeable effect of the dynamics of the
epidemics.
6
Our on-going work is oriented towards investigating the influence of the
variance of the in-degree distribution and giving a theoretical foundation to the
above observations.
5 Acknowledgments
This work was partially supported by the Parternariat Hubert Curien PHC
TOURNESOL FR 2013 29053SF between France and the Flemish community
of Belgium, by the Inria Alcatel-Lucent Joint Lab ARC "Network Science", and
by the European Commission within the framework of the CONGAS project
FP7-ICT-2011-8-317672.
References
[1] Alain Barrat, Marc Barthelemy, and Alessandro Vespignani. Dynamical
Processes on Complex Networks. Cambridge University Press, 2008.
[2] N. Chen and M. Olvera-Cravioto. Directed random graphs with given degree
distributions. To appear in Stochastic Systems.
[3] Moez Draief and Laurent Massouli. Epidemics and Rumours in Complex
Networks. Cambridge University Press, 2010.
[4] Richard Durrett. Random Graph Dynamics, volume 20. Cambridge univer-
sity press, 2007.
[5] Ayalvadi Ganesh, Laurent Massouli´e, and Don Towsley. The effect of net-
work topology on the spread of epidemics. In INFOCOM 2005. 24th Annual
Joint Conference of the IEEE Computer and Communications Societies.
Proceedings IEEE, volume 2, pages 1455 -- 1466. IEEE, 2005.
[6] Yamir Moreno, Romualdo Pastor-Satorras, and Alessandro Vespignani. Epi-
demic outbreaks in complex heterogeneous networks. The European Physical
Journal B-Condensed Matter and Complex Systems, 26(4):521 -- 529, 2002.
[7] Romualdo Pastor-Satorras and Alessandro Vespignani. Epidemic spreading
in scale-free networks. Physical review letters, 86(14):3200, 2001.
[8] P. Van Mieghem, J. Omic, and R. Kooij. Virus spread in networks. Net-
working, IEEE/ACM Transactions on, 17(1):1 -- 14, 2009.
7
|
1711.06845 | 1 | 1711 | 2017-11-18T10:57:21 | Evaluating Roles of Central Users in Online Communication Networks: A Case Study of #PanamaLeaks | [
"cs.SI",
"cs.CY",
"cs.NI",
"nlin.AO",
"stat.AP"
] | Social media has changed the ways of communication, where everyone is equipped with the power to express their opinions to others in online discussion platforms. Previously, a number of stud- ies have been presented to identify opinion leaders in online discussion networks. Feng ("Are you connected? Evaluating information cascade in online discussion about the #RaceTogether campaign", Computers in Human Behavior, 2016) identified five types of central users and their communication patterns in an online communication network of a limited time span. However, to trace the change in communication pattern, a long-term analysis is required. In this study, we critically analyzed framework presented by Feng based on five types of central users in online communication network and their communication pattern in a long-term manner. We take another case study presented by Udnor et al. ("Determining social media impact on the politics of developing countries using social network analytics", Program, 2016) to further understand the dynamics as well as to perform validation . Results indicate that there may not exist all of these central users in an online communication network in a long-term manner. Furthermore, we discuss the changing positions of opinion leaders and their power to keep isolates interested in an online discussion network. | cs.SI | cs | Evaluating Roles of Central Users in Online Communication
Network: A Case Study of #PanamaLeaks
Mohsin Adalata, Muaz A. Niazi∗a
aComputer Science Department,
COMSATS Institute of Information Technology,
Islamabad, Pakistan
∗Corresponding author
7
1
0
2
v
o
N
8
1
]
I
S
.
s
c
[
1
v
5
4
8
6
0
.
1
1
7
1
:
v
i
X
r
a
Abstract
Social media has changed the ways of communication, where everyone is equipped with the power
to express their opinions to others in online discussion platforms. Previously, a number of stud-
ies have been presented to identify opinion leaders in online discussion networks. Feng ("Are
you connected? Evaluating information cascade in online discussion about the #RaceTogether
campaign", Computers in Human Behavior, 2016) identified five types of central users and their
communication patterns in an online communication network of a limited time span. However,
to trace the change in communication pattern, a long-term analysis is required. In this study, we
critically analyzed framework presented by Feng based on five types of central users in online
communication network and their communication pattern in a long-term manner. We take an-
other case study presented by Udnor et al. ("Determining social media impact on the politics of
developing countries using social network analystics", Program, 2016) to further understand the
dynamics as well as to perform validation . Results indicate that there may not exist all of these
central users in an online communication network in a long-term manner. Furthermore, we discuss
the changing positions of opinion leaders and their power to keep isolates interested in an online
discussion network.
Keywords: Complex network, Social network analysis, opinion formation
1. Introduction
To exchange opinions and information, internet is being used all-around the world. With in-
creased use of internet, social media also became popular and a central medium for communica-
tion. In recent times, social networking has become de facto medium for communication among
the same social group or network [1]. People participate in online discussion by using different
platforms like Facebook, Twitter, YouTube and other social networking tools [2].Twitter is one
of the popular social networking tool (micro blogs) among others, where people broadcast brief
update messages [3].
In scientific literature, a number of studies have been carried out using Twitter data. For
example, Ch'ng in [4] used Twitter data to find generation of online communities and Baek in
[5] argued roles and positions of participants in an online community. Geo located tweets are used
Preprint submitted to Elsevier
November 21, 2017
by Blanford in [6] to capture regional connections and cross border movement. Twitter data is
even used for evaluation of inter organizational disaster coordination networks, created in result of
disaster [7]. Lee et al. in [8] demonstrated how scholars are using Twitter communication network
for informal scholarly communication. Authors used social network analysis methods in all of the
above-mentioned studies. Social network analysis is a strategy for analyzing social structures [9].
Aim of all the above studies in different fields is to find the influential nodes in network. In online
communication network, these nodes are called opinion leaders.
Rogers defined opinion leadership as "the degree to which an individual is able to influence
other individuals' attitudes or overt behavior informally in desired way with relative frequency"
[10]. The concept of opinion leadership originates from two step flow theory, the theory posits
that information first comes from mass media to opinion leaders and then this information is trans-
mitted to society via opinion leaders [11]. Opinion leaders are opinion broker in two sense, first
opinion leaders influence is between two social groups rather than in-between social group and
second they are transition between two network mechanism responsible for spreading the idea
[12].
For identification of opinion leaders in Twitter network, Borge in [13] used social network
analysis techniques on following-follower and mention network. Some researchers proposed dif-
ferent algorithms to find opinion leaders in social networks. For example, Aleahmad et al. pro-
posed OLFinder in [14] and Huang proposed a Dynamic Opinion Rank algorithm in [15]. Li et
al.
in [16] proposed framework for opinion leader identification, which consist of five stages.
First is keyword blog search; second is ontology extraction; ontology-assisted extraction and hot
blog identification is third and fourth respectively; fifth and last stage of framework is opinion
leader identification. The method of identifying opinion leaders in virtual communities proposed
by Zhang in [17] consist of five steps: first, find twenty participators who are active with large
followers; second, to investigate details of these active users, construct a relational matrix; third,
centrality analysis; fourth, identify opinion leaders by comparing standardized degree of network
users; last, if centrality and map analysis fails then correlation analysis is used. Researchers also
applied two step flow theory [11] for identifying opinion leaders in online social network.
By adopting two step flow theory, Feng in [18] proposed another approach to identify central
users in the online communication network. In this work, the author identified five types of central
users in Twitter network: conversation starter, influencer, active engager, network builder, and
information bridge. However, Feng study analyzed tweets in only four-days time span to trace
communication patterns that is not appropriate because opinion in a society may vary with time.
The proposed approach needs to be applied for trend analysis on longitudinal network.
In this context, this study explores communication patterns among participants of online dis-
cussion network in longitudinal manner by adopting previous approach proposed by Feng in [18].
In this study, we use "PanamaLeaks" as case study.
International Consortium of Investigative
Journalists' raised issue of Panama Leaks by leaking 11.1 million records from Panama-based law
firm, Mossack Fonseca [19]. We retrieve Twitter data using #PanamaLeaks keyword from April
2017 to August 2017. The research questions we address are: Weather central users identified
by Feng study exist in online communication network in longitudinal manner?. If yes, then what
are the changes in communication pattern of central users?. For validation, we apply the same
approach on another case study of #NeigieraDecides presented in [20].
2
Our main contributions can be listed as follows:
• To identify most central users in longitudinal network of #PanamaLeaks.
• To analyze the flow of information among the participants of online longitudinal discussion
network in #PanamaLeaks.
The reminder of the paper is structured as follows: Section 2 presents background, Section 3
presents our proposed method and dataset followed by results presented in section 4, Discussion
is formulated in section 5.
2. Background
In this section, we present characteristics and communication patterns of the most central users
identified by Feng in [18].
2.1. Conversation starter
A conversation starter is a user in network with numerus "in-degree" links and a few or none
"out-degree" links. Conversation starter is the one, who is responsible for starting the original
topic and flow of information in the network. However, the control on flow of information in
network is not under control of the conversation starter. Figure 1 shows connection pattern for
conversation starter in the network.
Figure 1: Connection Pattern for "Conversation Starter" in online discussion network.Circles are presenting users in
the network. Circle filled with black color is "Conversation Starter". Many arrows pointing towards "Conversation
Starter" presents many users in the network mentioning/retweeting "Conversation Starter" tweet.
2.2. Influencer
An influencer is opinion leader in the network.
Influencer in a network has plentiful "in-
degree" links and few "out-degree" links. Influencer does not initiate the topic of conversation in
the network, however influencer do influence opinion of other users in the network by creating
frequent tweets that are retweeted by other users. Figure 2 shows connection pattern for influencer
in the network.
3
Figure 2: Connection Pattern for "Influencer" in online discussion network. Circles are presenting users in the
network. Circle filled with black color is "influencer". Many arrows pointing towards "Influencer" presents many
users in the network mentioning/retweeting "Influencer" tweet.
2.3. Active engager
An active engager in a user in online discussion network with many "out-degree" and a few or
none "in-degree" links. Active engager is eager to distribute information and make connections in
online discussion network. Active engager is also opinion expresser in network. Figure 3 shows
connection pattern for active engager in the network.
Figure 3: Presenting Connection Pattern for "Active Engager" in online discussion network. Circles are presenting
users in the network. Circle filled with black color is "Active Engager". Many arrows pointing towards users from
"Active Engager" presents that "Active Engager" mention/retweet many other users tweet in the network.
2.4. Network builder
Despite having a few "out-degree" and a few or none "in-degree" links in online discussion
network, network builder plays an important role in the network. Main role of network builder is
connecting two or more influencers in the network. Figure 4 shows connection pattern for network
builder in the network.
4
Figure 4: Connection Pattern for "Network Builder" in online discussion network. Circles are presenting users in the
network. Circle filled with black color is "network builder". Two arrows pointing towards "Influencers" in the
network is describing basic role of "Network Builder" in the network.
2.5. Information bridge
An information bridge is a user in online discussion network with a few "in-degree" and "out-
degree" links. Information bridge role is to assist influencer and active engager in the network to
connect with other users (Feng 2016). Note that in both case studies, #Panamaleaks and #Nige-
riaDecies which we took for validation. We are unable to find such user in the online discussion
network (see section 5). Figure 5 shows connection pattern for information bridge in the network.
Figure 5: Connection Pattern for "Information Bridge" in online discussion network. Circles are presenting users in
the network. Circle filled with black color is "Information Bridge". Arrow pointing towards "Influencer" in the
network is presenting that "Information Bridge" mention/retweet "Influencer" tweet and arrow from "Active
Engager" towards "Information Bridge" is presenting that same tweet is retweeted by the "Active Engager".
3. Proposed Methodology
3.1. Tool
We choose NodeXl and Gephi as tools to analyze #PanamaLeaks network. NodeXl is an exten-
sible toolkit to download and analyze network data [21]. NodeXl is capable of calculating network
5
matrices like "in-degree", "out-degree", "betweeness", "density", "modularity" etc. NodeXl gen-
erates excel-based reports and it is capable of downloading and analyzing data of several social
network platforms like Twitter, Flickr and Facebook. In Twitter user network, a Twitter user is
vertex in NodeXl and edge represents the relationship between users. Relationship between two
vertices can be direct or un-directed. In Twitter user network, "in-degree" of each vertex is how
many users in network have mentioned or retweeted his/her tweet. Retweets by a user in the
network and other users mentioned by that specific user is out-degree of that vertex.
Along with finding the most central users of #PanamaLeaks network, we aim to visualize
#PanamaLeaks network dynamically to find structural changes in the network. We choose Gephi
social network analysis tool [22] for network visualization due to its capability for visualizing
large, complex and dynamic networks. Gephi is capable of visualizing network with many differ-
ent algorithm like Force-Atlas, Fruchterman-Reingold, Yifan Hu etc [23].
3.2. Data Source
We used NodeXl Twitter Search Network data collector to get tweets having #PanamaLeaks. It
took us six months to collect these tweets because twitter streaming API only allows downloading
previous seven days of tweets.We choose #PanamaLeaks as keyword to download tweets. Our
dataset contains 10612 nodes and 24623 edges exist in the #Panamaleaks network during March
2017 to August 2017.
For validation purpose, we took #NigeriaDecides case study data set is used in [20]. The author
used #NigeriaDecides as a query to download tweets and generated network of tweets, mentions
and replies, which is same as #PanamaLeaks case study. #NigeriaDecides case study data set
consist of 1752 nodes and 6343 edges from 15 April 2015 to 20 April 2015.
In both case studies networks, nodes are the twitter users and edges are the relation between
two twitter users. NodeXl add an edge, when one user tweets, retweets or mention other user/users
in his/her tweet. Edges coming towards one users is "in-degree" of that specific user and edges
going from that user to others in network are "out-degree" of that specific user. If a user creates
new tweet/tweets without mentioning, replying or retweeting another users tweet, NodeXl adds
self-loop around that user in the network.
3.3. Methods
We combine all the data (Tweets downloaded using NodeXl during six months) and then divide
it into six months of different networks according to the dates of tweets from March 2017 to August
2017. We requested NodeXl to generate following matrices "in-degree", "out-degree", "between-
ness centrality", "density", "clustering coefficient", and "modularity" for each month indepen-
dently. "In-degree" is simple count of unique incoming connections/links towards a user/entity in
a directed graph. Whereas "out-degree" is simple count of unique links/connections from a user.
"Betweenness centrality" of the node (A unique user in a network) is how many times a node
appears in the shortest path between other nodes (other users in the same network) [24].
"In-degree", "out-degree" and "betweenness centrality" are the three measures used by Feng
in [18] for identifying most central users in online discussion network. That is the reason we also
used these three measures on #PanamaLeaks case study. We identified top ten central users based
on "betweenness centrality" for each months in the #PanamaLeaks network. We also identified
6
top ten central users in the #NigeriaDecides case study used in [20] for validation purpose. For
visualization of connection patterns for most central users in #PanamaLeaks network, we use Yifan
Hu algorithm for both case studies. Figure 7 shows connection pattern for conversation starter for
#PanamaLeaks and #NigeriaDecies. Figure 8, 10 and 12 shows connection patterns for influencer,
active engager and network builder in the #PanamaLeaks network for each month from April 2017
to August 2017. Figure 9, 11 and 13 shows the connection patterns for #NigeriaDecides network.
We used social network analysis measures like "density", "clustering coefficient", and "mod-
ularity" to analyze overall communication pattern among the participants of the #PanamaLeaks
network. To check the impact of conversation starter and influencer on the online discussion
network, we use number of unique users coming into #PanamaLeaks network every month.
For visualization of overall communication pattern among #PanamaLeaks network participant.
We created dynamic network of #PanamaLeaks using Gephi. We implemented month wise time
stamp on the network based on tweet date. We used Yifan Hu algorithm, one of the many avail-
able algorithm for network visualization in Gephi. Figure 15a, 15b, 15c, 15d, 15e and 15f are
showing communication patterns between all users of #PanamaLeaks network for March, April,
May, June, July and August respectively. Figure 16 shows overall visualization of the complete six
months network from March 2017 to August 2017 using Yifan Hu algorithm. Our overall research
methodology has been shown in figure 6.
Figure 6: Research method.
4. Results
In this section, we present series of experiments carried out in our study.
7
4.1. Conversation starter
According to the two-step flow theory, information flows from mass media to opinion leaders
and then transmitted to the population via opinion leaders [11]. In light of the above, we took
"siasatpk" as our conversation starter as it is a mass media platform. As per the result of NodeXl,
conversation starter topped the list as expected. According to results of March, April, May, June,
and July "in-degree" of "siasatpk" is 211, 137, 203, 446 and 1463 respectively. Whereas, "Out-
degree" of "siasatpk" for March, April, May, June, and July is 1, 3, 2, 7 and 22 respectively.
"siasatpk" who is playing the role as conversation starters in #PanamaLeaks network remains
the first or second most influential participant of the network from March to July. "In-degree" of
"siasatpk" is zero in the month of August because "siasatpk" stop taking participation in discussion
from August (See Table 1).
Table 1: Presenting "in-degree", "out-degree", "betweeness centrality" and "rank" of "siasatpk" playing role as
"Conversation Starter" in the #Panamaleaks network. The number "1" in rank column presents that "conversation
starter" is the most central user in #PanamaLeaks network.
Month
March
April
May
June
July
August
In-degree
211
137
203
446
1463
0
Out-degree
1
3
2
7
22
0
Betweenness
492974.973
461067.857
975715.667
1219538.696
7539507.557
-
Rank
1
2
2
1
1
-
"saharareporters" with (in − degree = 345, out − degree = 13 and betweenness − centrality =
616332.423) is playing the role of conversation starter in the #NigeriaDecides case study, which
we used for validation. As far as the connection pattern for conversation starter of #PanamaLeaks
network, Figures 7a,7b,7c,7d and 7e are showing the five months connection pattern of conversa-
tion starter. Mostly isolates are pointing towards the conversation starter (mentioning or retweet-
ing tweets of conversation starter). Conversation starter is acting as a hub in the #PanamaLeaks
network. Results of the #NigeriaDecides case study used for validation revels the same pattern of
communication between conversation starter and other participants of the network. (See Figure
7f).
8
(a) "Conversation Starter" of March
(b) "Conversation Starter" of April
(c) "Conversation Starter" of May
(d) "Conversation Starter" of June
(e) "Conversation Starter" of July
(f) "Conversation Starter" of the #NigeriaDecides
Figure 7: Connection Pattern for "Conversation Starter" of every month. Middle node of every figure is
"Conversation Starter". Size and color of nodes is according to their degree. Thus, the nodes bigger in size are high
degree nodes and smaller nodes are nodes having low degree. Whereas, green nodes are nodes having lowest degree
and color tends to change towards white with increase in degree.
9
4.2. Influencer
There can be more than one influencer in the online discussion network as result shows. Here
we present top three influencer in #PanamaLeaks network of each month based on "betweenness
centrality". In March, "qaiserthethia" with (in−degree = 139, out−degree = 12) is the third most
central user of the network. Fourth most central user of the March network is "alirazatweets" with
(in − degree = 150, out − degree = 1). "xaifoo official" is the fifth most central user in March
network with (in − degree = 90, out − degree = 1).
In April, "ptiofficial" is the most central user of the network acting as an influencer in the
network with (in − degree = 266, out − degree = 1) which means 266 other users have mentioned
or retweeted "ptiofficial" tweets. Third most central user is "qaiserthethia" with (in − degree =
107, out − degree = 12) and "amranafahad" with (in − degree = 116, out − degree = 2) is fourth
most central user of April network.
In May, the most central user is "khaleejmag" with (in − degree = 347, out − degree = 1).
"diyarahman99" is the third most central user of May network with (in − degree = 176, out −
degree = 1) and "maryamnsharif" with (in − degree = 59, out − degree = 0) is the fourth most
central user of May network.
In June, second, fourth and sixth most central users are "dunyanews", "imrankhanpti" and
"ptiofficial" with (in − degree = 290, out − degree = 17), (in − degree = 105, out − degree = 0)
and (in − degree = 80, out − degree = 1) respectively playing role as influencers in the network.
In July, "imrankhanpti" with (in − degree = 181, out − degree = 0), "dunyanews" with (in −
degree = 122, out − degree = 15) and "defencepk" with (in − degree = 111, out − degree = 1)
are the second, fourth and fifth most central user in the network.
In August, "Shahzad ind" with (in − degree = 721, out − degree = 2) and "yadavtejashwi"
with (in − degree = 622, out − degree = 0) are the top two most central users in network.
The overall results of the top three influencers of each month is shown in table 2.
10
Table 2: Top Three "Influencers" of #PanamaLeaks network for every month based on "Betweenness Centrality"
Month
March
April
May
June
July
August
Name of User
qaiserthethia
alirazatweets
xaifoo official
ptiofficial
qaiserthethia
amranafahad
khaleejmag
diyarahman99
maryamnsharif
dunyanews
imrankhanpti
ptiofficial
imrankhanpti
dunyanews
defencepk
shehzad ind
yadavtejashwi
In-Degree
139
150
90
266
107
116
347
176
59
290
105
80
181
122
111
721
622
Out-Degree Betweenness
315421.518
12
251914.020
1
1
200590.403
596473.970
1
329445.157
12
2
288166.206
1313558.254
1
608312.772
1
0
551766.163
836335.327
17
288851.658
0
213124.288
1
0
800178.225
705731.185
15
601869.039
1
2
328635.667
194985.667
0
Rank
3
4
5
1
3
4
1
3
4
2
4
6
2
4
5
1
2
Figure 8a shows the connection pattern of "qaiserthethia, who is playing role as influencer in
the month of March. Figure 8b shows connection pattern "ptiofficial", acting as top influencer in
the month of March. Many arrows are pointing towards this node can be seen in figure 8b and most
of these nodes are small. Which means mostly isolates like to mention or retweet influencer in the
network. "khaleejmag" is the top influencer of May network. Connection pattern of "khaleejmag"
is shown in 8c. June result shows, that conversation starter and active engager of June both
mentioned top influencer of that month. However, most nodes connecting with the "imrankhanpti"
are small, which means they are also isolates having low connectivity in the network. Figure 8d 8e
8f are showing the connection patterns of top influencer of June, July and August respectively.All
of the above figures are showing same pattern, which shows connection pattern of influencer in
the network remain similar.
11
(a) Top "Influencer" of March
(b) Top "Influencer" of April
(c) Top "Influencer" of May
(d) Top "Influencer" of June
(e) Top "Influencer" of July
(f) Top "Influencer" of August
Figure 8: Connection Pattern for "Influencer" of every month. Middle node of every figure is "Influencer". Size and
color of nodes is according to their degree. Thus, the nodes bigger in size are high degree nodes and smaller nodes
are nodes having low degree. Whereas, green nodes are nodes having lowest degree and color tends to change
towards white with increase in degree.
12
In #NigeriaDecides network, "thisisbuhari", "mbuhari","inecnigeria" are the top three influ-
encers ranked by "betweenness centrality" playing the role of influencer with (in − degree =
262, out−degree = 0, betweenness−centrality = 215550.136), (in−degree = 169, out−degree =
0, betweenness − centrality = 199496.055) , (in − degree = 67, out − degree = 0, betweenness −
centrality = 142580.817). Figure 9 shows the top influencer "thisisbuhari" of #NigeriaDecides.
Results are consistant with #PanamaLeaks case study as well as Feng study [18].
Figure 9: Top "Influencer" of #NigeriaDecides case study. Node in the center named as "thisisbuhari" is the top
"Influencer" of #NigeriaDecides network. Size and color of nodes is according to their degree. Thus, the nodes
bigger in size are high degree nodes and smaller nodes are nodes having low degree. Whereas, blue nodes are nodes
having lowest degree and color tends to change towards white with increase in degree.
4.3. Active engager
Result shows that in the network of March second most central user with (in − degree =
0, out − degree = 65) is "sugar9940" an active engager. In network of April "sugar9940" again
act as an active engager with (in − degree = 0, out − degree = 41) and became fifth most central
user in the network. Seventh most central user "umj3rry" is active engager in network of May with
(in − degree = 0, out − degree = 56) and in June it's once again "sugar9940" with (in − degree =
0, out − degree = 96). In network of July and August "lavajatonews" with (in − degree = 0, out −
degree = 65), (in− degree = 0, out − degree = 14) respectively is active engager. "lavajatonews"
is third most central user of July network and seventh most central user of August network playing
role as active engager (See table 3).
13
Table 3: Presenting "in-degree", "out-degree", "betweeness centrality" and "Betweeness wise rank" of the top
"Active Engager" of each month in the #PanamaLeaks network.
Month User name
sugar9940
March
April
sugar9940
umj3rry
May
sugar9940
June
July
lavajatonews
lavajatonews
August
In-degree
0
0
0
0
0
0
Out-degree Betweenness
356338.123
65
41
216878.111
357661.301
56
530891.138
96
65
798249.408
1360.000
14
Rank
2
5
7
3
3
7
Figures 10 is showing connection pattern for active engagers from March to August.
In
March, "sugar9940" is playing the role of active engager by retweeting/mentioning other par-
ticipants of the network. "sugar9940" mentiond participant having high degree like "qaiser-
thethia","alirazatweets" and "siasatpk", connection pattern of March active engager can be seen
in 10a."sugar9940" is again playing the role of active engager in April, Figure 10b shows the con-
nection pattern of "sugar9940". Mentioning "qaiserthethia","maryamnsharif" and "imrankhanpti"
three influencers in April network is the reason of his central position in the network. "umj3rry"
takes the role of active enager in May network. Figure 10c shows the connection pattern of
"umj3rry". By mentioning conversation starter, two influecers and other participants in June
"sugar9940" again is the active engager as shown in figure 10d. Figure 10e 10f shows the con-
nection pattern for active engager of July and August month respectively. Interesting thing to note
is that, in-degree for active engagers of all the months is zero. Which means either no participant
has mentioned or retweeted active engager's tweet.
14
(a) Top "Active engager" of March
(b) Top "Active engager" of April
(c) Top "Active engager" of May
(d) Top "Active engager" of June
(e) Top "Active engager" of July
(f) Top "Active engager" of August
Figure 10: Connection Pattern for "Active engager" of every month. Middle node of every figure is "Active engager".
Size and color of nodes is according to their degree. Thus, the nodes bigger in size are high degree nodes and smaller
nodes are nodes having low degree. Whereas, green nodes are nodes having lowest degree and color tends to change
towards white with increase in degree.
15
"wakiligentleman" is playing role as active engager in the network of #NigeriaDecides with
in − degree = 1 and out − degree = 26. He is the seventh most central user of the network. "Be-
tweenness centrality" of "wakiligentleman" is 112865.285. By mentioning/retweeting influencers
as well as conversation starter in his tweets, he got high betweenness centrality. Figure 11 shows
the connection pattern of "wakiligentleman", results are consistant with #PanamaLeaks case study
with many out-going connections.
Figure 11: Connection Pattern for "Active Engager" of #NigeriaDecides network. Node in the middle of figure is
"Active Engager". Size and color of nodes is according to their degree. Thus, the nodes bigger in size are high degree
nodes and smaller nodes are nodes having low degree. Whereas, blue nodes are nodes having lowest degree and
color tends to change towards white with increase in degree.
4.4. Network builder
"shahwar125" with only (in− degree = 0, out − degree = 2) is interestingly sixth most central
user of the March network playing role as network builder. With only (in − degree = 0, out −
degree = 7) "ambitiousfree" is again the sixth most central user of April netowork because he
mentioned two influencers "maryamnsharif" and "amranafahad" in his tweets and build a strategic
location in the network like "shahwar125" in the network of March month. "ali axhar" with
(in−degree = 20, out−degree = 8), "deziggnerumer" with (in−degree = 0, out−degree = 18) are
the network builders for month of May and June respectively. In July we could find " waqas788"
as network builder on thirteenth position with (in−degree = 5, out−degree = 21). In August with
(in − degree = 0, out − degree = 3) "foroneindia" is the network builder. He is the fourth most
central user in the netwrok. No influencer in the network of all months has mentioned network
builders in his/her tweet but still by connecting two or more Influential, network builders got
central positions in the network. The overall results of top network builder in the network is
shown in table 4.
16
Table 4: Presenting "in-degree", "out-degree", "betweeness centrality" and "rank" of top "Network Builder" of each
month in the #Panamaleaks network based on "Betweenness Centrality".
In-degree
Month User name
0
shahwar125
March
0
April
ambitiousfree
ali axhar
May
20
deziggnerumer 0
June
July
waqas788
5
0
foroneindia
August
Out-degree Betweenness
186462.694
2
7
162629.964
295700.335
8
267362.714
18
21
276354.023
2880.000
3
Rank
6
6
10
5
13
4
Node in the middle of Figure 12a "shahwar125" is network builder of March. With only two
out-degree links, he is the sixth most central user in network. "shahwar125" built a strategic loca-
tion in network by mentioning/retweeting only two participants having high degree. Figures 12b
12c 12d 12e 12f shows the connection pattern for network builder of April, May, June, July and
August respectively. All the above figures show similar pattern of communication. All network
builders have few "in-degree" and "out-degree" links; however, they build strategic location in the
#PanamaLeaks network by mentioning influencers or conversation starter.
17
(a) Top "Network builder" of March
(b) Top "Network builder" of April
(c) Top "Network builder" of May
(d) Top "Network builder" of June
(e) Top "Network builder" of July
(f) Top "Network builder" of August
Figure 12: Connection Pattern for "Network Builder" of every month. Middle node of every figure is "Network
builder". Size and color of nodes is according to their degree. Thus, the nodes bigger in size are high degree nodes
and smaller nodes are nodes having low degree. Whereas, green nodes are nodes having lowest degree and color
tends to change towards white with increase in degree.
18
"teamagbaje2015" is playing the role as network builder in #NigeriaDecides network with
in − degree = 13 and out − degree = 3. He is the fifth most centeral user of #NigeriaDecides net-
work by only mentioning two influencer "thisisbuhari", "demolarewaju" and conversation starter
"saharareporters" of the network. This result is also consistent with #PanamaLeaks network. Fig-
ure 13 is showing the links of network builder of #NigeriaDecides network.
Figure 13: Connection Pattern for "Network Builder" of #NigeriaDecides network. Node in the middle of figure is
"Active Engager". Size and color of nodes is according to their degree. Thus, the nodes bigger in size are high degree
nodes and smaller nodes are nodes having low degree. Whereas, blue nodes are nodes having lowest degree and color
tends to change towards white with increase in degree.
4.5. Information bridge
Results reveal that there is no information bridge in #PanamaLeaks six months network. In-
formation bridge is the user, who mention/retweets influencer tweet and that same tweet is again
retweeted by active engager [18]. We could not find any user with few "in-degree" and "out-
degree" links, who has retweeted influencer tweet and then that same tweet is retweeted by active
engager. We applied this aproach to #NigeriaDecides case study [20] for validation purpose and
according to the results; there exist no user like information bridge. Although, we were able to
find conversation starter, influencer, active engager and network builder in #NigeriaDecides case
study also.
4.6. Structural analysis
Figure 14a shows the number of unique users participating in #PanamaLeaks online discus-
sion every month. Number of unique users coming into #PanamaLeaks discussion network was
1277 in March. In April and May 1412 and 1750 unique users came into the network respec-
tively. 1487 unique users joined discussion in June and 3626 joined discussion in July. However,
there is a sudden decrease in number of unique users participating in #PanamaLeaks discussion
in August with 880 new unique users coming into the network. In September, there are negligi-
ble number of users discussing the issue of Panama Leaks anymore. That is the reason we only
analyzed this network until August. Figure 14b shows the density of #PanamaLeaks network.
Gradual decrees in the density of the network can be seen. However, there is a sudden increase
19
in density after July. Average clustering coefficient of #PanamaLeaks from March to August is
shown in Figure 14c. Average clustering coefficient of the network starts from 0.049 on March
and with gradual increase ends at 0.312 on August. Modularity of the #PanamaLeaks network is
0.773821, 0.656387, 0.696806, 0.579446, 0.49243 and 0.381433 for March, April, May, June, July
and August respectively (See Figure 14d).
(a) Unique users
(b) Density
(c) Avg. Clustering coefficient
(d) Modularity
Figure 14: Number of Unique users are shown in 14a. Overall Density of #PanamaLeaks network is shown in 14b.
14c is showing Average Clustering Cofficient of the network. Figure 14d is showing Modularity of the netwrok.
Figure 15 shows interaction patterns of users, there exist many isolates pointing towards the
influencers and conversation starter in the network. Figure 15a shows the network of March and
communities forming around conversation starter and influencers can be seen. These communities
keep on growing until July, which can be seen in figure 15b, 15c, 15d, 15e and 15f. However, in
August when major influencers and conversation starter left the discussion, communities started
to dissolve which can be seen in figure 15f. A combine visualization of #PanamaLeaks network
20
from March to August is shown in figure 16. A circle of participants around the graph can be seen.
These participants are not connected to any influencer or conversation starter in the #PanamaLeaks
network.
21
(a) March
(b) April
(c) May
(d) June
(e) July
(f) August
Figure 15: Dynamic Graph Visualization of #PanamaLeaks network for each by Gephi using Yifan Hu algorithm.
Communities forming around "Conversation Starter" and "Influencers" are presented in different colors.
22
Figure 16: Overall network structure of complete six months of #PanamaLeaks using Gephi.
5. Discussion
5.1. Conversation starter
A conversation starter is one, who is center of the network and connects many "isolates" who
otherwise have few or no links in the network [18]. These are the participants with high between-
ness [25]. Our results reveled the same as many "isolates" retweeted or mentioned "siasakpk"
in their tweets. There were some influential mentioning conversation starter in their tweets but
they are very few, mostly "isolates" tends to connect with conversation starter to gain informa-
tion. Twitter users with high connectivity are better at influencing information flow in Twitter
network [26]. This makes the conversation starter center of the network by various incoming
connections. However, when conversation starter stops taking participation in the network (Stop
Tweeting about the Topic) "isolates" also lose interest in the online conversation. As results rev-
eled that in August when conversation starter is no more in the network, there is sudden decrease
in the people discussing about the topic online. However, until the conversation starter keeps on
participating in the discussion. Conversation starter remains on the top central positions in the
network.(See Figure 17).
23
Figure 17: Changes in Ranking of Central users from March 2017 to August 2017. Months are shown on x-axis and
betweenness wise rank are shown along with y-axis.
5.2. Influencer
There exist influencers in the online discussion network [18] with plentiful "in-degree" and
few "out-degree" links. They are called opinion leaders in the network and isolates like to mention
them in their tweets. Isolates also like to retweet opinion leader's tweet and these opinion leaders
are usually mass media organizations or celebrities in different fields of life [18]. Opinion leaders
create many tweets to influence other users in online discussion. In this way, they act to influence
opinion in an online discussion network. Information is going from opinion leaders to users with
less connectivity and this supports the argument by [27, 28] that opinion leaders by using their
network connectivity replicate asymmetric offline power in online discussion. The participant like
influencer who has high degree in network act as hub in the network [29]. However, Opinion
leader's influence is not consistent as the results show in #PanamaLeaks study their influence can
increase/decrease with the passage of time. There can be changes in the"betweenness centrality"
of network influencers as well. Figure 18 shows the changing in betweenness centrality of five
selected influencers in the #PanamaLeaks network.
24
Figure 18: Changes in "Betweenness Centrality" of selected "Influencers" in #PanamaLeaks network.
5.3. Active engager
A user with many "out-degree" and none or few "in-degree" links is active engager in the
network [18]. Results also reveal that there always exist an active engager in the network from
start of conversation until end. However, position of active engager is not consistent and other
active engager can take his/her place with the passage of time. Active engager is also among the
top ten most central users of the network. Results show that active engager is one who is active
and with intention to get information. However, there are usually few or no user in the network
to follow or mention him in their tweets. From the centrality point of view, active engager is an
important user. However, Active engager in a socity is a person with special knowledge, who has
potential to become influencer [30].
5.4. Network builder
Network builder in a network has few or none "in-degree" and few "out-degree" links [18]. His
basic role in the network is to associate two or more influential nodes in network. By this, network
builder builds a strategic location in network due to which it gets high centrality. However, in [18]
"wnkr" is a network builder and its in-degree is "0". Which means no one either has mentioned
or retweet "wnkr" in their tweets. In #PanamaLeaks network, network builders follow the same
tendency. In this context, we can conclude that network builder do help in creating connection
between two or more influencers in the online discussion network. However, influencers usually
do not mention or reply on their tweets. Therefore, their impact on the opinion in a network is
petite.
5.5. Information bridge
Information bridge is a user with few "in-degree" and "out-degree" links, who connects an
influencer and active engager as noted in [18]. However, results of six months of network shows
that there is no user such as information bridge exist in the network. We implemented this concept
on the dataset #NigeriaDecides used in [20] for validation purpose. Result shows that in the
#NigeriaDecides, there does not exist any user who has retweeted influencer tweet and that tweet
25
is retweeted by active engager. Active engager's basic role in network is to get engage with other
users [18], and for this purpose; active engager likes to retweet mostly influencer/conversation
starter tweets. Active engagers do not mostly retweet/mention "isolates" in their tweets. In this
context, we can conclude that there is very less chance that an active engager retweets same tweet
in which a person with few "in-degree" and "out-degree" has mentioned influencer.
5.6. Network structure
Unique users have increased until June and then there is a sudden decrease in the number
of unique users as shown in figure 14a. When conversation starter stop taking participation in
the online discussion there is a sudden decrease in number of unique users participation. Users
with high connectivity like conversation starter can influence the participation of users [31]. In
this context, opinion leaders have the power to keep isolates interested in the online discussion.
Network density can decrease with the number of increase users in the network and increase when
number of participants decrease [8], density of the network is showing same tendency. Participant
of the online discussion tends to cluster around the influential participants and with the passage of
time in presence of conversation starter and influencers, participant became less independent as
modularity decreased [25]. However, with many participants having low degree pointing towards
the conversation starter and influencer. We can conclude that they play role as hub in the online
discussion network.
26
Reference
References
[1] Sheela Gole and Bharat Tidke. Frequent itemset mining for big data in social media using clustbigfim algorithm.
In Pervasive Computing (ICPC), 2015 International Conference on, pages 1–6. IEEE, 2015.
[2] Itai Himelboim, Stephen McCreery, and Marc Smith. Birds of a feather tweet together: Integrating network and
content analyses to examine cross-ideology exposure on twitter. Journal of Computer-Mediated Communication,
18(2):40–60, 2013.
[3] Victor R Martinez, Miguel A Escalante, Mariano Beguerisse-D´ıaz, Elmer Garduno, and Victor M Gonzalez.
Understanding human behavior in urban spaces using social network data: A mobility graph approach. 2016.
[4] Eugene Ch'ng. The bottom-up formation and maintenance of a twitter community: analysis of the# freejahar
twitter community. Industrial Management & Data Systems, 115(4):612–624, 2015.
[5] Seung Ik Baek and Young Min Kim. Longitudinal analysis of online community dynamics. Industrial Manage-
ment & Data Systems, 115(4):661–677, 2015.
[6] Justine I Blanford, Zhuojie Huang, Alexander Savelyev, and Alan M MacEachren. Geo-located tweets. enhanc-
ing mobility maps and capturing cross-border movement. PloS one, 10(6):e0129202, 2015.
[7] Alireza Abbasi and Naim Kapucu. A longitudinal study of evolving networks in response to natural disaster.
Computational and mathematical organization theory, 22(1):47–70, 2016.
[8] Mi Kyung Lee, Ho Young Yoon, Marc Smith, Hye Jin Park, and Han Woo Park. Mapping a twitter scholarly
communication network: a case of the association of internet researchers conference. Scientometrics, pages
1–31, 2017.
[9] Evelien Otte and Ronald Rousseau. Social network analysis: a powerful strategy, also for the information
sciences. Journal of information Science, 28(6):441–453, 2002.
[10] Everett M Rogers. Diffusion of innovations. free press. New York, page 551, 2003.
[11] Elihu Katz and F Paul. Lazarsfeld (1955), personal influence. The Part Played by People in the Flow of Mass
[12] Ronald S Burt. The social capital of opinion leaders. The Annals of the American Academy of Political and
Communication. New York, 1955.
Social Science, 566(1):37–54, 1999.
[13] Rosa Borge Bravo and Marc Esteve Del Valle. Opinion leadership in parliamentary twitter networks: A matter
of layers of interaction? Journal of Information Technology & Politics, pages 1–14, 2017.
[14] Abolfazl Aleahmad, Payam Karisani, Maseud Rahgozar, and Farhad Oroumchian. Olfinder: Finding opinion
leaders in online social networks. Journal of Information Science, 42(5):659–674, 2016.
[15] Baocheng Huang, Guang Yu, and Hamid Reza Karimi. The finding and dynamic detection of opinion leaders in
social network. Mathematical Problems in Engineering, 2014, 2014.
[16] Feng Li and Timon C Du. Who is talking? an ontology-based opinion leader identification framework for
word-of-mouth marketing in online social blogs. Decision Support Systems, 51(1):190–197, 2011.
[17] Xiaofei Zhang and Dahai Dong. Ways of identifying the opinion leaders in virtual communities. International
Journal of Business and Management, 3(7):21, 2009.
[18] Yang Feng. Are you connected? evaluating information cascades in online discussion about the# racetogether
campaign. Computers in Human Behavior, 54:43–53, 2016.
[19] Pamela L Cross and Braek RJ Urquhart. The panama papers: Official statements from caribbean countries.
Trusts & Trustees, 23(1):8–13, 2017.
[20] Collins Udanor, Collins Udanor, Stephen Aneke, Stephen Aneke, Blessing Ogechi Ogbuokiri, and Bless-
ing Ogechi Ogbuokiri. Determining social media impact on the politics of developing countries using social
network analytics. Program, 50(4):481–507, 2016.
[21] Marc A Smith, Ben Shneiderman, Natasa Milic-Frayling, Eduarda Mendes Rodrigues, Vladimir Barash, Cody
Dunne, Tony Capone, Adam Perer, and Eric Gleave. Analyzing (social media) networks with nodexl.
In
Proceedings of the fourth international conference on Communities and technologies, pages 255–264. ACM,
2009.
27
[23] Ken Cherven. Mastering Gephi network visualization. Packt Publishing Ltd, 2015.
[24] Linton C Freeman. A set of measures of centrality based on betweenness. Sociometry, pages 35–41, 1977.
[25] Weiai Wayne Xu, Ji-young Park, and Han Woo Park. Longitudinal dynamics of the cultural diffusion of kpop
on youtube. Quality & Quantity, 51(4):1859–1875, 2017.
[26] Weiai Wayne Xu, Yoonmo Sang, Stacy Blasiola, and Han Woo Park. Predicting opinion leaders in twitter
activism networks: The case of the wisconsin recall election. American Behavioral Scientist, 58(10):1278–
1293, 2014.
[27] Lincoln Dahlberg. Rethinking the fragmentation of the cyberpublic: from consensus to contestation. New media
& society, 9(5):827–847, 2007.
[22] Mathieu Bastian, Sebastien Heymann, Mathieu Jacomy, et al. Gephi: an open source software for exploring and
manipulating networks. Icwsm, 8:361–362, 2009.
[28] Graham Murdock and Peter Golding. Dismantling the digital divide: Rethinking the dynamics of participation
and exclusion. Toward a political economy of culture: Capitalism and communication in the twenty-first century,
pages 244–260, 2004.
[29] Manju K Ahuja, Dennis F Galletta, and Kathleen M Carley. Individual centrality and performance in virtual
r&d groups: An empirical study. Management science, 49(1):21–38, 2003.
[30] Daniele Quercia, Michal Kosinski, David Stillwell, and Jon Crowcroft. Our twitter profiles, our selves: Predict-
ing personality with twitter. In Privacy, Security, Risk and Trust (PASSAT) and 2011 IEEE Third Inernational
Conference on Social Computing (SocialCom), 2011 IEEE Third International Conference on, pages 180–185.
IEEE, 2011.
[31] Jinlou Zhao, Hongyu Gao, Yongli Li, and Jiaguo Liu. Which factors affect the duration of hot topics on social
media platforms? Quality & Quantity, 51(5):2395–2407, 2017.
28
|
1704.03913 | 2 | 1704 | 2018-01-05T03:35:00 | Higher-order clustering in networks | [
"cs.SI",
"cond-mat.stat-mech",
"physics.soc-ph",
"stat.ML"
] | A fundamental property of complex networks is the tendency for edges to cluster. The extent of the clustering is typically quantified by the clustering coefficient, which is the probability that a length-2 path is closed, i.e., induces a triangle in the network. However, higher-order cliques beyond triangles are crucial to understanding complex networks, and the clustering behavior with respect to such higher-order network structures is not well understood. Here we introduce higher-order clustering coefficients that measure the closure probability of higher-order network cliques and provide a more comprehensive view of how the edges of complex networks cluster. Our higher-order clustering coefficients are a natural generalization of the traditional clustering coefficient. We derive several properties about higher-order clustering coefficients and analyze them under common random graph models. Finally, we use higher-order clustering coefficients to gain new insights into the structure of real-world networks from several domains. | cs.SI | cs |
Higher-order clustering in networks
Hao Yin∗
Institute for Computational and Mathematical Engineering, Stanford University, Stanford, CA, 94305, USA
Department of Computer Science, Cornell University, Ithaca, NY, 14850, USA
Austin R. Benson†
Computer Science Department, Stanford University, Stanford, CA, 94305, USA
(Dated: January 8, 2018)
Jure Leskovec‡
A fundamental property of complex networks is the tendency for edges to cluster. The extent
of the clustering is typically quantified by the clustering coefficient, which is the probability that
a length-2 path is closed, i.e., induces a triangle in the network. However, higher-order cliques
beyond triangles are crucial to understanding complex networks, and the clustering behavior with
respect to such higher-order network structures is not well understood. Here we introduce higher-
order clustering coefficients that measure the closure probability of higher-order network cliques
and provide a more comprehensive view of how the edges of complex networks cluster. Our higher-
order clustering coefficients are a natural generalization of the traditional clustering coefficient. We
derive several properties about higher-order clustering coefficients and analyze them under common
random graph models. Finally, we use higher-order clustering coefficients to gain new insights into
the structure of real-world networks from several domains.
I.
INTRODUCTION
Networks are a fundamental tool for understanding
and modeling complex physical, social,
informational,
and biological systems [1]. Although such networks are
typically sparse, a recurring trait of networks through-
out all of these domains is the tendency of edges to ap-
pear in small clusters or cliques [2, 3].
In many cases,
such clustering can be explained by local evolutionary
processes. For example, in social networks, clusters ap-
pear due to the formation of triangles where two indi-
viduals who share a common friend are more likely to
become friends themselves, a process known as triadic
closure [2, 4]. Similar triadic closures occur in other net-
works: in citation networks, two references appearing in
the same publication are more likely to be on the same
topic and hence more likely to cite each other [5] and in
co-authorship networks, scientists with a mutual collab-
orator are more likely to collaborate in the future [6]. In
other cases, local clustering arises from highly connected
functional units operating within a larger system, e.g.,
metabolic networks are organized by densely connected
modules [7].
The clustering coefficient quantifies the extent to
which edges of a network cluster in terms of triangles.
The clustering coefficient is defined as the fraction of
length-2 paths, or wedges, that are closed with a trian-
gle [3, 8] (Fig. 1, row C2). In other words, the clustering
coefficient measures the probability of triadic closure in
∗ [email protected]
† [email protected]
‡ [email protected]
1. Start with 2. Find an adjacent edge 3. Check for an
(ℓ + 1)-clique
to form an ℓ-wedge
an ℓ-clique
C2
C3
C4
FIG. 1. Overview of higher-order clustering coefficients as
clique expansion probabilities. The ℓth-order clustering co-
efficient Cℓ measures the probability that an ℓ-clique and an
adjacent edge, i.e., an ℓ-wedge, is closed, meaning that the
ℓ − 1 possible edges between the ℓ-clique and the outside node
in the adjacent edge exist to form an (ℓ + 1)-clique.
the network. However, the clustering coefficient is in-
herently restrictive as it measures the closure probabil-
ity of just one simple structure-the triangle. Moreover,
higher-order structures such as larger cliques are crucial
to the structure and function of complex networks [9–
11]. For example, 4-cliques reveal community structure
in word association and protein-protein interaction net-
works [12] and cliques of sizes 5–7 are more frequent than
triangles in many real-world networks with respect to cer-
tain null models [13]. However, the extent of clustering
of such higher-order structures has not been well under-
stood nor quantified.
Here, we provide a framework to quantify higher-order
clustering in networks by measuring the normalized fre-
quency at which higher-order cliques are closed, which
we call higher-order clustering coefficients. We derive our
higher-order clustering coefficients by extending a novel
interpretation of the classical clustering coefficient as a
form of clique expansion (Fig. 1). We then derive sev-
eral properties about higher-order clustering coefficients
and analyze them under the Gn,p and small-world null
models.
Using our theoretical analysis as a guide, we analyze
the higher-order clustering behavior of real-world net-
works from a variety of domains. We find that each
domain of networks has its own higher-order clustering
pattern, which the traditional clustering coefficient does
not show on its own. Conventional wisdom in network
science posits that practically all real-world networks ex-
hibit clustering; however, we find that not all networks
exhibit higher-order clustering. More specifically, once
we control for the clustering as measured by the classical
clustering coefficient, some networks do not show signifi-
cant clustering in terms of higher-order cliques. In addi-
tion to the theoretical properties and empirical findings
exhibited in this paper, our related work also demon-
strates a connection between higher-order clustering and
community detection [14].
II. DERIVATION OF HIGHER-ORDER
CLUSTERING COEFFICIENTS
In this section, we derive our higher-order clustering
coefficients and some of their basic properties. We first
present an alternative interpretation of the classical clus-
tering coefficient and then show how this novel interpre-
tation seamlessly generalizes to arrive at our definition
of higher-order clustering coefficients. We then provide
some probabilistic interpretations of higher-order clus-
tering coefficients that will be useful for our subsequent
analysis.
A. Alternative interpretation of the classical
clustering coefficient
Here we give an alternative interpretation of the clus-
tering coefficient that will later allow us to generalize
it and quantify clustering of higher-order network struc-
tures (this interpretation is summarized in Fig. 1). Our
interpretation is based on a notion of clique expansion.
First, we consider a 2-clique K in a graph G (that is,
a single edge K; see Fig. 1, row C2, column 1). Next,
we expand the clique K by considering any edge e adja-
cent to K, i.e., e and K share exactly one node (Fig. 1,
row C2, column 2). This expanded subgraph forms a
wedge, i.e., a length-2 path. The classical global cluster-
ing coefficient C of G (sometimes called the transitivity
of G [15]) is then defined as the fraction of wedges that
are closed, meaning that the 2-clique and adjacent edge
induce a (2 + 1)-clique, or a triangle (Fig. 1, row C2, col-
2
umn 3) [8, 16]. The novelty of our interpretation of the
clustering coefficient is considering it as a form of clique
expansion, rather than as the closure of a length-2 path,
which is key to our generalizations in the next section.
Formally, the classical global clustering coefficient is
C =
6K3
W
,
(1)
where K3 is the set of 3-cliques (triangles), W is the
set of wedges, and the coefficient 6 comes from the fact
that each 3-clique closes 6 wedges-the 6 ordered pairs
of edges in the triangle.
We can also reinterpret the local clustering coeffi-
cient [3] in this way.
In this case, each wedge again
consists of a 2-clique and adjacent edge (Fig. 1, row C2,
column 2), and we call the unique node in the intersection
of the 2-clique and adjacent edge the center of the wedge.
The local clustering clustering coefficient of a node u is
the fraction of wedges centered at u that are closed:
C(u) =
2K3(u)
W (u)
,
(2)
where K3(u) is the set of 3-cliques containing u and W (u)
is the set of wedges with center u (if W (u) = 0, we say
that C(u) is undefined). The average clustering coeffi-
cient ¯C is the mean of the local clustering coefficients,
¯C =
C(u),
(3)
1
eV Xu∈ eV
where eV is the set of nodes in the network where the
local clustering coefficient is defined.
B. Generalizing to higher-order clustering
coefficients
Our alternative interpretation of the clustering coef-
ficient, described above as a form of clique expansion,
leads to a natural generalization to higher-order cliques.
Instead of expanding 2-cliques to 3-cliques, we expand
ℓ-cliques to (ℓ + 1)-cliques (Fig. 1, rows C3 and C4). For-
mally, we define an ℓ-wedge to consist of an ℓ-clique and
an adjacent edge for ℓ ≥ 2. Then we define the global ℓth-
order clustering coefficient Cℓ as the fraction of ℓ-wedges
that are closed, meaning that they induce an (ℓ+1)-clique
in the network. We can write this as
(ℓ2 + ℓ)Kℓ+1
Cℓ =
(4)
,
Wℓ
where Kℓ+1 is the set of (ℓ + 1)-cliques, and Wℓ is the set
of ℓ-wedges. The coefficient ℓ2 + ℓ comes from the fact
that each (ℓ + 1)-clique closes that many wedges: each
(ℓ + 1)-clique contains ℓ + 1 ℓ-cliques, and each ℓ-clique
contains ℓ nodes which may serve as the center of an ℓ-
wedge. Note that the classical definition of the global
clustering coefficient given in Eq. 1 is equivalent to the
definition in Eq. 4 when ℓ = 2.
We also define higher-order local clustering coefficients:
Cℓ(u) =
ℓKℓ+1(u)
Wℓ(u)
,
(5)
where Kℓ+1(u) is the set of (ℓ + 1)-cliques containing
node u, Wℓ(u) is the set of ℓ-wedges with center u (where
the center is the unique node in the intersection of the
ℓ-clique and adjacent edge comprising the wedge; see
Fig. 1), and the coefficient ℓ comes from the fact that
each (ℓ+1)-clique containing u closes that many ℓ-wedges
in Wℓ(u). The ℓth-order clustering coefficient of a node
is defined for any node that is the center of at least one
ℓ-wedge, and the average ℓth-order clustering coefficient
is the mean of the local clustering coefficients:
¯Cℓ =
Cℓ(u),
(6)
1
eVℓ Xu∈ eVℓ
where eVℓ is the set of nodes that are the centers of at
To understand how to compute higher-order clustering
least one ℓ-wedge.
coefficients, we substitute the following useful identity
Wℓ(u) = Kℓ(u) · (du − ℓ + 1),
where du is the degree of node u, into Eq. 5 to get
Cℓ(u) =
ℓ · Kℓ+1(u)
(du − ℓ + 1) · Kℓ(u)
.
(7)
(8)
From Eq. 8, it is easy to see that we can compute all
local ℓth-order clustering coefficients by enumerating all
(ℓ + 1)-cliques and ℓ-cliques in the graph. The compu-
tational complexity of the algorithm is thus bounded by
the time to enumerate (ℓ+1)-cliques and ℓ-cliques. Using
the Chiba and Nishizeki algorithm [17], the complexity
is O(ℓaℓ−2m), where a is the arboricity of the graph, and
m is the number of edges. The arboricity a may be as
large as √m, so this algorithm is only guaranteed to take
polynomial time if ℓ is a constant. In general, determin-
ing if there exists a single clique with at least ℓ nodes is
NP-complete [18].
For the global clustering coefficient, note that
Wℓ = Xu∈V
Wℓ(u).
(9)
3
C. Probabilistic interpretations of higher-order
clustering coefficients
To facilitate understanding of higher-order clustering
coefficients and to aid our analysis in Section III, we
present a few probabilistic interpretations of the quanti-
ties. First, we can interpret Cℓ(u) as the probability that
a wedge w chosen uniformly at random from all wedges
centered at u is closed:
Cℓ(u) = P [w ∈ Kℓ+1(u)] .
(10)
The variant of this interpretation for the classical clus-
tering case of ℓ = 2 has been useful for graph algorithm
development [19].
For the next probabilistic interpretation, it is useful to
analyze the structure of the 1-hop neighborhood graph
N1(u) of a given node u (not containing node u). The
vertex set of N1(u) is the set of all nodes adjacent to u,
and the edge set consists of all edges between neighbors
of u, i.e., {(v, w) (u, v), (u, w), (v, w) ∈ E}, where E is
the edge set of the graph.
Any ℓ-clique in G containing node u corresponds to a
unique (ℓ − 1)-clique in N1(u), and specifically for ℓ = 2,
any edge (u, v) corresponds to a node v in N1(u). There-
fore, each ℓ-wedge centered at u corresponds to an (ℓ−1)-
clique K and one of the du − ℓ + 1 nodes outside K (i.e.,
in N1(u)\K). Thus, Eq. 8 can be re-written as
ℓ · Kℓ(N1(u))
(du − ℓ + 1) · Kℓ−1(N1(u))
,
(11)
where Kk(N1(u)) denotes the number of k-cliques in
N1(u).
If we uniformly at random select an (ℓ − 1)-clique K
from N1(u) and then also uniformly at random select a
node v from N1(u) outside of this clique, then Cℓ(u) is
the probability that these ℓ nodes form an ℓ-clique:
Cℓ(u) = P [K ∪ {v} ∈ Kℓ(N1(u))] .
(12)
Moreover, if we condition on observing an ℓ-clique from
this sampling procedure, then the ℓ-clique itself is se-
lected uniformly at random from all ℓ-cliques in N1(u).
Therefore, Cℓ−1(u) · Cℓ(u) is the probability that an
(ℓ − 1)-clique and two nodes selected uniformly at ran-
dom from N1(u) form an (ℓ + 1)-clique. Applying this
recursively gives
Thus, it suffices to enumerate ℓ-cliques (to compute Wℓ
using Eq. 7) and to count the total number of ℓ-cliques.
In practice, we use the Chiba and Nishizeki to enumerate
cliques and simultaneously compute Cℓ and Cℓ(u) for all
nodes u. This suffices for our clustering analysis with
ℓ = 2, 3, 4 on networks with over a hundred million edges
in Section IV.
ℓYj=2
Cj(u) = Kℓ(N1(u))
(cid:0)du
ℓ(cid:1)
.
(13)
In other words, the product of the higher-order local clus-
tering coefficients of node u up to order ℓ is the ℓ-clique
density amongst u's neighbors.
u
u
u
C2(u)
C3(u)
C4(u)
1
1
1
d
2(d − 1)
≈
1
2
0
0
d−2
4d−4 ≈
1
4
d − 4
2d − 4
d − 6
2d − 6
≈
≈
1
2
1
2
FIG. 2. Example 1-hop neighborhoods of a node u with degree
d with different higher-order clustering. Left: For cliques,
Cℓ(u) = 1 for any ℓ. Middle: If u's neighbors form a complete
bipartite graph, C2(u) is constant while Cℓ(u) = 0, ℓ ≥ 3.
Right: If half of u's neighbors form a star and half form a
clique with u, then Cℓ(u) ≈ pC2(u), which is the upper
bound in Proposition 1.
III. THEORETICAL ANALYSIS AND
HIGHER-ORDER CLUSTERING IN RANDOM
GRAPH MODELS
We now provide some theoretical analysis of our
higher-order clustering coefficients. We first give some
extremal bounds on the values that higher-order clus-
tering coefficients can take given the value of the tradi-
tional (second-order) clustering coefficient. After, we an-
alyze the values of higher-order clustering coefficients in
two common random graph models-the Gn,p and small-
world models. The theory from this section will be a
useful guide for interpreting the clustering behavior of
real-world networks in Section IV.
A. Extremal bounds
We first analyze the relationships between local higher-
order clustering coefficients of different orders. Our tech-
nical result is Proposition 1, which provides essentially
tight lower and upper bounds for higher-order local clus-
tering coefficients in terms of the traditional local cluster-
ing coefficient. The main ideas of the proof are illustrated
in Fig. 2.
Proposition 1. For any fixed ℓ ≥ 3,
0 ≤ Cℓ(u) ≤pC2(u).
Moreover,
(14)
1. There exists a finite graph G with a node u such
that the lower bound is tight and C2(u) is within ǫ
of any prescribed value in [0, ℓ−2
2. There exists a finite graph G with a node u such
that Cℓ(u) is within ǫ of the upper bound for any
prescribed value of C2(u) ∈ [0, 1].
ℓ−1 ].
4
Proof. Clearly, 0 ≤ Cℓ(u) if the local clustering coefficient
is well defined. This bound is tight when N1(u) is (ℓ− 1)-
partite, as in the middle column of Fig. 2. In the (ℓ− 1)-
partite case, C2(u) = ℓ−2
ℓ−1 . By removing edges from this
extremal case in a sufficiently large graph, we can make
C2(u) arbitrarily close to any value in [0, ℓ−2
ℓ−1 ].
To derive the upper bound, consider the 1-hop neigh-
borhood N1(u), and let
δℓ(N1(u)) = Kℓ(N1(u))
(15)
(cid:0)du
ℓ(cid:1)
denote the ℓ-clique density of N1(u). The Kruskal-
Katona theorem [20, 21] implies that
δℓ(N1(u)) ≤ [δℓ−1(N1(u))]ℓ/(ℓ−1)
δℓ−1(N1(u)) ≤ [δ2(N1(u))](ℓ−1)/2.
Combining this with Eq. 8 gives
Cℓ(u) ≤ [δℓ−1(N1(u))]
1
ℓ−1 ≤pδ2(N1(u)) =pC2(u),
where the last equality uses the fact that C2(u) is the
edge density of N1(u).
The upper bound becomes tight when N1(u) consists
of a clique and isolated nodes (Fig. 2, right) and the
neighborhood is sufficiently large. Specifically, let N1(u)
consist of a clique of size c and b isolated nodes. When
ℓ = 2,
(c + b − 1)(c + b) →(cid:18) c
(c − 1)c
c + b(cid:19)2
Cℓ(u) = (cid:0)c
2(cid:1)
2 (cid:1) =
(cid:0)c+b
ℓ ·(cid:0)c
ℓ(cid:1)
and by Eq. 11, when 3 ≤ ℓ ≤ c,
(c + b − ℓ + 1) ·(cid:0) c
ℓ−1(cid:1) =
Cℓ(u) =
c − ℓ + 1
c + b − ℓ + 1 →
c
c + b
.
By adjusting the ratio c/(b + c) in N1(u), we can con-
struct a family of graphs such that C2(u) takes any value
in the interval [0, 1] as du → ∞ and Cℓ(u) →pC2(u) as
du → ∞.
The second part of the result requires the neighbor-
hoods to be sufficiently large in order to reach the upper
bound. However, we will see later that in some real-world
data, there are nodes u for which C3(u) is close to the
upper bound pC2(u) for several values of C2(u).
Next, we analyze higher-order clustering coefficients
in two common random graph models: the Erdos-R´enyi
model with edge probability p (i.e., the Gn,p model [22])
and the small-world model [3].
B. Analysis for the Gn,p model
Now, we analyze higher-order clustering coefficients
in classical Erdos-R´enyi random graph model, where
each edge exists independently with probability p (i.e.,
the Gn,p model [22]). We implicitly assume that ℓ is
small in the following analysis so that there should be
at least one ℓ-wedge in the graph (with high probabil-
ity and n large, there is no clique of size greater than
(2 + ǫ) log n/ log(1/p) for any ǫ > 0 [23]). Therefore, the
global and local clustering coefficients are well-defined.
In the Gn,p model, we first observe that any ℓ-wedge
is closed if and only if the ℓ − 1 possible edges between
the ℓ-clique and the outside node in the adjacent edge
exist to form an (ℓ + 1)-clique. Each of the ℓ − 1 edges
exist independently with probability p in the Gn,p model,
which means that the higher-order clustering coefficients
should scale as pℓ−1. We formalize this in the following
proposition.
Proposition 2. Let G be a random graph drawn from
the Gn,p model. For constant ℓ,
1. EG [Cℓ] = pℓ−1
2. EG [Cℓ(u) Wℓ(u) > 0] = pℓ−1 for any node u
3. EG(cid:2) ¯Cℓ(cid:3) = pℓ−1
Proof. We prove the first part by conditioning on the set
of ℓ-wedges, Wℓ:
E[Cℓ] = EG [EWℓ [Cℓ Wℓ]]
= EGhEWℓh 1
= EGhEWℓh 1
= EG(cid:2)pℓ−1(cid:3)
= pℓ−1.
WℓPw∈Wℓ
WℓPw∈Wℓ
P [w is closed]ii
pℓ−1ii
As noted above, the second equality is well defined (with
high probability) for small ℓ. The third equality comes
from the fact that any ℓ-wedge is closed if and only if the
ℓ − 1 possible edges between the ℓ-clique and the outside
node in the adjacent edge exist to form an (ℓ + 1)-clique.
The proof of the second part is essentially the same,
except we condition over the set of possible cases where
Wℓ(u) > 0.
least one ℓ-wedge. To prove the third part, we take the
Recall that eV is the set of nodes at the center of at
conditional expectation over eV and use our result from
the second part.
The above results say that the global, local, and av-
erage ℓth order clustering coefficients decrease exponen-
tially in ℓ.
It turns out that if we also condition on
the second-order clustering coefficient having some fixed
value, then the higher-order clustering coefficients still
decay exponentially in ℓ for the Gn,p model. This will be
useful for interpreting the distribution of local clustering
coefficients on real-world networks.
Proposition 3. Let G be a random graph drawn from
the Gn,p model. Then for constant ℓ,
EG [Cℓ(u) C2(u), Wℓ(u) > 0]
=(cid:2)C2(u) − (1 − C2(u)) · O(1/d2
≈ (C2(u))ℓ−1.
u)(cid:3)ℓ−1
Proof. Similar to the proof of Proposition 3, we look at
the conditional expectation over Wℓ(u) > 0:
5
1
EG [Cℓ(u) C2(u), Wℓ(u) > 0]
= EG(cid:2)EWℓ(u)>0 [Cℓ(u) C2(u), Wℓ(u)](cid:3)
= EGhEWℓ(u)>0h
Wℓ(u)Pw∈Wℓ(u)
Now, note that N1(u) has m = C2(u)·(cid:0)du
ing that w ∈ Wℓ(u) accounts for(cid:0)ℓ−1
symmetry, the other q = m−(cid:0)ℓ−1
the remaining r =(cid:0)du
2(cid:1) −(cid:0)ℓ−1
at random. There are (cid:0)r
which (cid:0)r−ℓ+1
q−ℓ+1(cid:1) would close the wedge w. Thus,
P [w is closed C2(u)]
= (r−ℓ+1)!q!
=
P [w closed C2(u)]ii .
2(cid:1) edges. Know-
2 (cid:1) of these edges. By
2 (cid:1) edges appear in any of
2 (cid:1) pairs of nodes uniformly
q(cid:1) ways to place these edges, of
(q−ℓ+1)!r! = (q−ℓ+2)(q−ℓ+3)···q
(r−ℓ+2)(r−ℓ+3)···r .
(r−ℓ+1
q−ℓ+1)
(r
q)
Now, for any small nonnegative integer k,
q − k
r − k
=
C2(u)·(du
2 )−(ℓ−1
2 )−k
2 )−(ℓ−1
(du
2 )−k
= C2(u) − (1 − C2(u))(cid:20)
(ℓ−1
2 )+k
2 )−(ℓ−1
= C2(u) − (1 − C2(u)) · O(1/d2
u).
(du
2 )−k(cid:21)
(Recall that ℓ is constant by assumption, so the big-O no-
tation is appropriate). The above expression approaches
(C2(u))ℓ−1 when C2(u) → 1 as well as when du → ∞.
Proposition 3 says that even if the second-order local
clustering coefficient is large, the ℓth-order clustering co-
efficient will still decay exponentially in ℓ, at least in the
limit as du grows large. By examining higher-order clique
closures, this allows us to distinguish between nodes u
whose neighborhoods are "dense but random" (C2(u) is
large but Cℓ(u) ≈ (C2(u))ℓ−1) or "dense and structured"
(C2(u) is large and Cℓ(u) > (C2(u))ℓ−1). Only the latter
case exhibits higher-order clustering. We use this in our
analysis of real-world networks in Section IV.
C. Analysis for the small-world model
We also study higher-order clustering in the small-
world random graph model [3]. The model begins with a
ring network where each node connects to its 2k nearest
neighbors. Then, for each node u and each of the k edges
(u, v) with v following u clockwise in the ring, the edge
is rewired to (u, w) with probability p, where w is chosen
uniformly at random.
With no rewiring (p = 0) and k ≪ n, it is known that
¯C2 ≈ 3/4 [3]. As p increases, the average clustering coef-
ficient ¯C2 slightly decreases until a phase transition near
p = 0.1, where ¯C2 decays to 0 [3] (also see Fig. 3). Here,
we generalize these results for higher-order clustering co-
efficients.
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
-
C2
-
C3
-
C4
10-3
10-2
10-1
100
Rewiring probability (p)
FIG. 3. Average higher-order clustering coefficient ¯Cℓ as a
function of rewiring probability p in small-world networks for
ℓ = 2, 3, 4 (n = 20, 000, k = 5). Proposition 4 shows that the
ℓth-order clustering coefficient when p = 0 predicts that the
clustering should decrease modestly as ℓ increases.
Proposition 4.
rewiring (p = 0),
In the small-world model without
¯Cℓ → (ℓ + 1)/(2ℓ)
for any constant ℓ ≥ 2 as k → ∞ and n → ∞ while
2k < n.
Proof. Applying Eq. 8, it suffices to show that
Kℓ(u) =
ℓ
(ℓ − 1)! · kℓ−1 + O(kℓ−2)
(16)
as
Cℓ(u) =
ℓ · (ℓ+1)kℓ
ℓ!
(2k − ℓ + 1) · ℓkℓ−1
(ℓ−1)!
,
which approaches ℓ+1
2ℓ as k → ∞.
Now we give a derivation of Eq. 16. We first label the
2k neighbors of u as 1, 2, . . . , 2k by their clockwise order-
ing in the ring. Since 2k < n, these nodes are unique.
Next, define the span of any ℓ-clique containing u as the
difference between the largest and smallest label of the
ℓ − 1 nodes in the clique other than u. The span s of
any ℓ-clique satisfies s ≤ k − 1 since any node is directly
connected with a node of label difference no greater than
k − 1. Also, s ≥ ℓ − 2 since there are ℓ − 1 nodes in an ℓ-
clique other than u. For each span s, we can find 2k−1−s
pairs of (i, j) such that 1 ≤ i, j ≤ 2k and j − i = s. Fi-
nally, for every such pair (i, j), there are(cid:0)s−1
ℓ−3(cid:1) choices of
ℓ − 3 nodes between i and j which will form an ℓ-clique
together with nodes u, i, and j. Therefore,
Kℓ(u) =Pk−1
=Pk−1
=Pk−ℓ+2
s=ℓ−2(2k − 1 − s) ·(cid:0)s−1
ℓ−3(cid:1)
s=ℓ−2(2k − 1 − s) · (s−1)(s−2)···(s−ℓ+3)
(2k + 2 − t − ℓ) · t(t+1)···(t+ℓ−4)
.
(ℓ−3)!
(ℓ−3)!
t=1
If we ignore lower-order terms k and note that t = O(k),
we get
6
Kℓ(u) =Pk
= 1
(ℓ−3)! + O(kℓ−3)i
t=1(2ktℓ−3 − tℓ−2) + O(kℓ−2).
ℓ−1i + O(kℓ−2),
t=1h (2k−t)tℓ−3
(ℓ−3)!Pk
(ℓ−3)!h2k · kℓ−2
ℓ−2 − kℓ−1
(ℓ−1)! · kℓ−1 + O(kℓ−2).
= 1
= ℓ
Proposition 4 shows that, when p = 0, ¯Cℓ decreases
as ℓ increases. Furthermore, via simulation, we observe
the same behavior as for ¯C2 when adjusting the rewiring
probability p (Fig. 3). Regardless of ℓ, the phase tran-
sition happens near p = 0.1. Essentially, once there is
enough rewiring, all local clique structure is lost, and
clustering at all orders is lost. This is partly a conse-
quence of Proposition 1, which says that Cℓ(u) → 0 as
C2(u) → 0 for any ℓ.
IV. EXPERIMENTAL RESULTS ON
REAL-WORLD NETWORKS
We now analyze the higher-order clustering of real-
world networks. We first study how the higher-order
global and average clustering coefficients vary as we in-
crease the order ℓ of the clustering coefficient on a collec-
tion of 20 networks from several domains. After, we con-
centrate on a few representative networks and compare
the higher-order clustering of real-world networks to null
models. We find that only some networks exhibit higher-
order clustering once the traditional clustering coefficient
is controlled. Finally, we examine the local clustering of
real-world networks.
A. Higher-order global and average clustering
We compute and analyze the higher-order clustering
for networks from a variety of domains (Table I). We
briefly describe the collection of networks and their cat-
egorization below:
1. Two synthetic networks-a random instance of an
Erdos-R´enyi graph with n = 1, 000 nodes and edge
probability p = 0.2 and a small-world network with
n = 20, 000 nodes, k = 10, and rewiring probability
p = 0.1;
2. Four neural networks-the complete neural sys-
tems of the nematode worms P. pacificus and C.
elegans as well as the neural connections of the
Drosophila medulla and mouse retina;
3. Four online social networks-two Facebook friend-
ship networks between students at universities from
2005 (fb-Stanford, fb-Cornell) and two complete
online friendship networks (Pokec and Orkut);
Network
Erdos-R´enyi [22]
Small-world [3]
P. pacificus [24]
C. elegans [3]
Drosophila-medulla [25]
mouse-retina [26]
Nodes
1,000
20,000
50
297
1,781
1,076
99,831
100,000
576
2,148
32,311
577,350
fb-Stanford [27]
fb-Cornell [27]
Pokec [28]
Orkut [29]
11,621
18,660
568,330
790,777
1,632,803 22,301,964
3,072,441 117,185,083
arxiv-HepPh [30]
arxiv-AstroPh [30]
congress-committees [31]
DBLP [32]
12,008
18,772
871
317,080
email-Enron-core [33]
email-Eu-core [14, 30]
CollegeMsg [34]
wiki-Talk [35]
148
1005
1,899
2,394,385
118,505
198,050
248,848
1,049,866
1356
16064
41,579
4,659,565
oregon2-010526 [36]
as-caida-20071105 [36]
p2p-Gnutella31 [30, 37]
as-skitter [36]
11,461
26,475
62,586
32,730
53,381
147,892
1,696,415 11,095,298
Edges
C2
C3
C4
¯C2
¯C3
¯C4
0.200 0.040 0.008
0.480 0.359 0.229
0.200 0.040 0.008
0.489 0.350 0.205
0.015 0.051 0.035
0.181 0.080 0.056
0.000 0.002 0.001
0.008 0.038 0.029
0.157 0.107 0.116
0.136 0.106 0.121
0.047 0.044 0.046
0.041 0.022 0.019
0.659 0.749 0.788
0.318 0.326 0.359
0.037 0.080 0.063
0.306 0.634 0.821
0.383 0.245 0.192
0.267 0.170 0.135
0.004 0.005 0.003
0.002 0.011 0.010
0.037 0.085 0.097
0.007 0.012 0.015
0.004 0.003 0.000
0.005 0.007 0.011
0.073 0.052 0.034
0.308 0.137 0.062
0.116 0.061 0.024
0.033 0.100 0.085
0.253 0.181 0.157
0.225 0.169 0.148
0.122 0.084 0.061
0.170 0.131 0.110
0.698 0.586 0.519
0.677 0.609 0.561
0.082 0.142 0.126
0.732 0.613 0.517
0.496 0.363 0.277
0.450 0.329 0.264
0.053 0.017 0.006
0.201 0.081 0.051
0.494 0.294 0.300
0.333 0.159 0.134
0.010 0.001 0.000
0.296 0.126 0.109
7
eV2/V eV3/V eV4/V
1.000
1.000
0.999
1.000
1.000
1.000
0.880
0.949
0.803
0.998
0.955
0.973
0.900
0.978
0.876
0.932
1.000
0.864
0.966
0.887
0.829
0.262
0.711
0.625
0.542
0.871
0.580
0.926
0.616
0.996
0.922
0.951
0.675
0.949
0.723
0.839
1.000
0.675
0.946
0.847
0.591
0.077
0.269
0.171
0.067
0.633
0.440
0.808
0.425
0.994
0.877
0.923
0.508
0.878
0.567
0.740
1.000
0.489
0.946
0.784
0.332
0.027
0.121
0.060
0.001
0.335
TABLE I. Higher-order clustering coefficients on random graph models, neural connections, online social networks, collaboration
networks, human communication, and technological systems. Broadly, networks from the same domain have similar higher-
order clustering characteristics. Since eVℓ is the set of nodes at the center of at least one ℓ-wedge (see Eq. 6), eVℓ/V is the
fraction of nodes at the center of at least one ℓ-wedge (the higher-order average clustering coefficient ¯Cℓ is only measured over
those nodes participating in at least one ℓ-wedge).
4. Four collaboration networks-two co-authorship
networks constructed from arxiv submission cat-
egories (arxiv-AstroPh and arxiv-HepPh), a co-
authorship network constructed from DBLP, and
the co-committee membership network of United
States congresspersons (congress-committees);
(email-Enron-core,
5. Four human communication networks-two email
networks
email-Eu-core), a
Facebook-like messaging network from a college
(CollegeMsg), and the edits of user talk pages by
other users on Wikipedia (wiki-Talk); and
6. Four technological systems networks-three au-
tonomous
as-caida-
20071105, as-skitter) and a peer-to-peer connection
network (p2p-Genutella31).
(oregon2-010526,
systems
In all cases, we take the edges as undirected, even if the
original network data is directed.
Table I lists the ℓth-order global and average clustering
coefficients for ℓ = 2, 3, 4 as well as the fraction of nodes
that are the center of at least one ℓ-wedge (recall that
the average clustering coefficient is the mean only over
higher-order local clustering coefficients of nodes partici-
pating in at least one ℓ-wedge; see Kaiser [38] for a discus-
sion on how this can affect network analyses). We high-
light some important trends in the raw clustering coeffi-
cients, and in the next section, we focus on higher-order
clustering compared to what one gets in a null model.
Propositions 2 and 4 say that we should expect the
higher-order global and average clustering coefficients
to decrease as we increase the order ℓ for both the
Erdos-R´enyi and small-world models, and indeed ¯C2 >
¯C3 > ¯C4 for these networks. This trend also holds for
most of the real-world networks (mouse-retina, congress-
committees, and oregon2-010526 are the exceptions).
Thus, when averaging over nodes, higher-order cliques
are overall less likely to close in both the synthetic and
real-world networks.
The relationship between the higher-ordrer global clus-
tering coefficient Cℓ and the order ℓ is less uniform
over the datasets. For the three co-authorship net-
works (arxiv-HepPh, arxiv-AstroPh, and DBLP) and the
three autonomous systems networks (oregon2-010526, as-
caida-20071105, and as-skitter), Cℓ increases with ℓ, al-
though the base clustering levels are much higher for co-
authorship networks. This is not simply due to the pres-
ence of cliques-a clique has the same clustering for any
order (Fig. 2, left). Instead, these datasets have nodes
that serve as the center of a star and also participate
in a clique (Fig. 2, right; see also Proposition 1). On
the other hand, Cℓ decreases with ℓ for the two email
networks and the two nematode worm neural networks.
Finally, the change in Cℓ need not be monotonic in ℓ.
8
C. elegans
fb-Stanford
arxiv-AstroPh
email-Enron-core
oregon2-010526
original CM MRCN original CM MRCN original CM MRCN original CM MRCN original CM MRCN
¯C2
¯C3
0.31
0.14
0.15∗ 0.31
0.04∗ 0.17†
0.25
0.18
0.03∗ 0.25
0.00∗ 0.14∗
0.68
0.61
0.01∗ 0.68
0.00∗ 0.60
0.50
0.36
0.23∗ 0.50
0.08∗ 0.35
0.49
0.29
0.25∗ 0.49
0.10∗ 0.14∗
TABLE II. Average higher-order clustering coefficients for five networks as well as the clustering with respect to two null
models: a Configuration Model (CM) that samples random graphs with the same degree distribution [39, 40], and Maximally
Random Clustered Networks (MRCN) that preserve degree distribution as well as ¯C2 [41, 42]. For the random networks, we
report the mean over 100 samples. An asterisk (∗) denotes when the value in the original network is at least five standard
deviations above the mean and a dagger (†) denotes when the value in the original network is at least five standard deviations
below the mean. Although all networks exhibit clustering with respect to CM, only some of the networks exhibit higher-order
clustering when controlling for ¯C2 with MRCN.
In three of the four online social networks, C3 < C2 but
C4 > C3.
Overall, the trends in the higher-order clustering co-
efficients can be different within one of our dataset cat-
egories, but tend to be uniform within sub-categories:
the change of ¯Cℓ and Cℓ with ℓ is the same for the two
nematode worms within the neural networks, the two
email networks within the communication networks, and
the three co-authorship networks within the collabora-
tion networks. These trends hold even if the (classical)
second-order clustering coefficients differ substantially in
absolute value.
While the raw clustering values are informative, it is
also useful to compare the clustering to what one expects
from null models. We find in the next section that this
reveals additional insights into our data.
B. Comparison against null models
For one real-world network from each dataset category,
we also measure the higher-order clustering coefficients
with respect to two null models (Table II). First, we com-
pare against the Configuration Model (CM) that samples
uniformly from simple graphs with the same degree dis-
In real-world networks, ¯C2 is much
tribution [39, 40].
larger than expected with respect to the CM null model.
We find that the same holds for ¯C3.
Second, we use a null model that samples graphs pre-
serving both degree distribution and ¯C2. Specifically,
these are samples from an ensemble of exponential graphs
where the Hamiltonian measures the absolute value of
the difference between the original network and the sam-
pled network [41]. Such samples are referred to as as
Maximally Random Clustered Networks (MRCN) and
are sampled with a simulated annealing procedure [42].
Comparing ¯C3 between the real-world and the null net-
work, we observe different behavior in higher-order clus-
tering across our datasets. Compared to the MRCN null
model, C. elegans has significantly less than expected
higher-order clustering (in terms of ¯C3), the Facebook
friendship and autonomous system networks have signif-
icantly more than expected higher-order clustering, and
the co-authorship and email networks have slightly (but
not significantly) more than expected higher-order clus-
tering (Table II). Put another way, all real-world net-
works exhibit clustering in the classical sense of triadic
closure. However, the higher-order clustering coefficients
reveal that the friendship and autonomous systems net-
works exhibit significant clustering beyond what is given
by triadic closure. These results suggest the need for
models that directly account for closure in node neigh-
borhoods [43, 44].
Our finding about the lack of higher-order cluster-
ing in C. elegans agrees with previous results that 4-
cliques are under-expressed, while open 3-wedges re-
lated to cooperative information propagation are over-
expressed [9, 45, 46]. This also provides credence for the
"3-layer" model of C. elegans [46]. The observed clus-
tering in the friendship network is consistent with prior
work showing the relative infrequency of open ℓ-wedges
in many Facebook network subgraphs with respect to
a null model accounting for triadic closure [47]. Co-
authorship networks and email networks are both con-
structed from "events" that create multiple edges-a pa-
per with k authors induces a k-clique in the co-authorship
graph and an email sent from one address to k others in-
duces k edges. This event-driven graph construction cre-
ates enough closure structure so that the average third-
order clustering coefficient is not much larger than ran-
dom graphs where the classical second-order clustering
coefficient and degree sequence is kept the same.
We emphasize that simple clique counts are not suffi-
cient to obtain these results. For example, the discrep-
ancy in the third-order average clustering of C. elegans
and the MRCN null model is not simply due to the
presence of 4-cliques. The original neural network has
nearly twice as many 4-cliques (2,010) than the samples
from the MRCN model (mean 1006.2, standard deviation
73.6), but the third-order clustering coefficient is larger in
MRCN. The reason is that clustering coefficients normal-
ize clique counts with respect to opportunities for closure.
Thus far, we have analyzed global and average higher-
order clustering, which both summarize the clustering of
the entire network. In the next section, we look at more
localized properties, namely the distribution of higher-
order local clustering coefficients and the higher-order
average clustering coefficient as a function of node degree.
A
B
C
D
E
9
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
10-3
100
-
C2
-
C3
-
C4
C. elegans
101
Degree
102
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
100
fb-Stanford
101
102
Degree
103
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
100
arxiv-AstroPh
101
102
103
Degree
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
email-Enron-core
10-1
101
Degree
102
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
10-3
100
oregon2_010526
101
102
Degree
103
FIG. 4. Top row: Joint distributions of (C2(u), C3(u)) for (A) C. elegans (B) Facebook friendship, (C) arxiv co-authorship,
(D) email, and (E) autonomous systems networks. Each blue dot represents a node, and the red curve tracks the average over
logarithmic bins. The upper trend line is the bound in Eq. 14, and the lower trend line is expected Erdos-R´enyi behavior from
Proposition 3. Bottom row: Average higher-order clustering coefficients as a function of degree.
A
B
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
10-3
102
Erd s Rényi
103
Degree
.
f
f
e
o
c
.
t
s
u
c
.
l
g
v
A
100
10-1
10-2
Small-world
101
Degree
-
C2
-
C3
-
C4
102
FIG. 5.
Analogous plots of Fig. 4 for synthetic (A)
Erdos-R´enyi and (B) small-world networks. Top row: Joint
distributions of (C2(u), C3(u)). Bottom row: Average higher-
order clustering coefficients as a function of degree.
C. Higher-order local clustering coefficients and
degree dependencies
We now examine more localized clustering properties
of our networks. Figure 4 (top) plots the joint distribu-
tion of C2(u) and C3(u) for the five networks analyzed
in Table II, and Fig. 5 (top) provides the analogous plots
for the Erdos-R´enyi and small-world networks. In these
plots, the lower dashed trend line represents the expected
Erdos-R´enyi behavior, i.e., the expected clustering if the
edges in the neighborhood of a node were configured ran-
domly, as formalized in Proposition 3. The upper dashed
trend line is the maximum possible value of C3(u) given
C2(u), as given by Proposition 1.
For many nodes in C. elegans, local clustering is nearly
random (Fig. 4A, top), i.e., resembles the Erdos-R´enyi
joint distribution (Fig. 5A, top). In other words, there
are many nodes that lie on the lower trend line. This
provides further evidence that C. elegans lacks higher-
order clustering.
In the arxiv co-authorship network,
there are many nodes u with a large value of C2(u) that
have an even larger value of C3(u) near the upper bound
of Eq. 14 (see the inset of Fig. 4C, top). This implies that
some nodes appear in both cliques and also as the center
of star-like patterns, as in Fig. 2. On the other hand,
only a handful of nodes in the Facebook friendships, En-
ron email, and Oregon autonomous systems networks are
close to the upper bound (insets of Figs. 4B,4D, and 4E,
top).
Figures 4 and 5 (bottom) plot higher-order average
clustering as a function of node degree in the real-world
and synthetic networks. In the Erdos-R´enyi, small-world,
C. elegans, and Enron email networks, there is a distinct
gap between the average higher-order clustering coeffi-
cients for nodes of all degrees. Thus, our previous find-
ing that the average clustering coefficient ¯Cℓ decreases
with ℓ in these networks is independent of degree. In the
Facebook friendship network, C2(u) is larger than C3(u)
and C4(u) on average for nodes of all degrees, but C3(u)
and C4(u) are roughly the same for nodes of all degrees,
which means that 4-cliques and 5-cliques close at roughly
the same rate, independent of degree, albeit at a smaller
rate than traditional triadic closure (Fig. 4B, bottom).
In the co-authorship network, nodes u have roughly the
same Cℓ(u) for ℓ = 2, 3, 4, which means that ℓ-cliques
close at about the same rate, independent of ℓ (Fig. 4C,
bottom).
In the Oregon autonomous systems network,
we see that, on average, C4(u) > C3(u) > C2(u) for
nodes with large degree (Fig. 4E, bottom). This explains
how the global clustering coefficient increases with the
order, but the average clustering does not, as observed
in Table I.
V. DISCUSSION
We have proposed higher-order clustering coefficients
to study higher-order closure patterns in networks, which
generalizes the widely used clustering coefficient that
measures triadic closure. Our work compliments other
recent developments on the importance of higher-order
information in network navigation [11, 48] and on tem-
poral community structure [49]; in contrast, we examine
higher-order clique closure and only implicitly consider
time as a motivation for closure.
Prior efforts in generalizing clustering coefficients have
focused on shortest paths [50], cycle formation [51], and
triangle frequency in k-hop neighborhoods [52, 53]. Such
approaches fail to capture closure patterns of cliques, suf-
fer from challenging computational issues, and are dif-
ficult to theoretically analyze in random graph models
more sophisticated than the Erdos-R´enyi model. On
the other hand, our higher-order clustering coefficients
are simple but effective measurements that are analyz-
10
able and easily computable (we only rely clique enumer-
ation, a well-studied algorithmic task). Furthermore, our
methodology provides new insights into the clustering be-
havior of several real-world networks and random graph
models, and our theoretical analysis provides intuition
for the way in which higher-order clustering coefficients
describe local clustering in graphs.
Finally, we focused on higher-order clustering coeffi-
cients as a global network measurement and as a node-
level measurement, and in related work we also show
that large higher-order clustering implies the existence
of mesoscale clique-dense community structure [14].
ACKNOWLEDGMENTS
This research has been supported in part by NSF IIS-
1149837, ARO MURI, DARPA, ONR, Huawei, and Stan-
ford Data Science Initiative. We thank Will Hamil-
ton and Marinka Zitnik for insightful comments. We
thank Mason Porter and Peter Mucha for providing the
congress committee membership data.
[1] M. E. J. Newman, SIAM Review 45, 167 (2003).
[2] A. Rapoport, The Bulletin of Mathematical Biophysics
[18] R. M. Karp, in Complexity of computer computations
(Springer, 1972) pp. 85–103.
15, 523 (1953).
[3] D. J. Watts and S. H. Strogatz, Nature 393, 440 (1998).
[4] M. S. Granovetter, American Journal of Sociology , 1360
(1973).
[19] C. Seshadhri, A. Pinar, and T. G. Kolda, in Proceed-
ings of the 2013 SIAM International Conference on Data
Mining (SIAM, 2013) pp. 10–18.
[20] J. B. Kruskal, Mathematical Optimization Techniques
[5] Z.-X. Wu and P. Holme, Physical Review E 80, 037101
10, 251 (1963).
(2009).
[21] G. Katona, in Theory of Graphs: Proceedings of the Col-
[6] E. M. Jin, M. Girvan, and M. E. J. Newman, Physical
loquium held at Tihany, Hungary (1966) pp. 187–207.
Review E 64, 046132 (2001).
[22] P. Erdos and A. R´enyi, Publicationes Mathematicae (De-
[7] E. Ravasz and A.-L. Barab´asi, Physical Review E 67,
brecen) 6, 290 (1959).
026112 (2003).
[8] A. Barrat and M. Weigt, The European Physical Jour-
nal B: Condensed Matter and Complex Systems 13, 547
(2000).
[23] B. Bollob´as and P. Erdos, in Mathematical Proceedings of
the Cambridge Philosophical Society, Vol. 80 (Cambridge
University Press, 1976) pp. 419–427.
[24] D. J. Bumbarger, M. Riebesell, C. Rodelsperger, and
[9] A. R. Benson, D. F. Gleich, and J. Leskovec, Science
R. J. Sommer, Cell 152, 109 (2013).
353, 163 (2016).
[10] O. N. Yaveroglu, N. Malod-Dognin, D. Davis, Z. Lev-
najic, V. Janjic, R. Karapandza, A. Stojmirovic, and
N. Przulj, Scientific Reports 4 (2014).
[11] M. Rosvall, A. V. Esquivel, A. Lancichinetti, J. D. West,
and R. Lambiotte, Nature Communications 5 (2014).
[12] G. Palla, I. Der´enyi, I. Farkas, and T. Vicsek, Nature
435, 814 (2005).
[13] N. Slater, R. Itzchack, and Y. Louzoun, Network Science
2, 387 (2014).
[14] H. Yin, A. R. Benson, J. Leskovec, and D. F. Gleich, in
Proceedings of the 23rd ACM SIGKDD international con-
ference on Knowledge discovery and data mining (2017)
(To appear).
[15] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and
D.-U. Hwang, Physics reports 424, 175 (2006).
[16] R. D. Luce and A. D. Perry, Psychometrika 14, 95 (1949).
[17] N. Chiba and T. Nishizeki, SIAM Journal on Computing
14, 210 (1985).
[25] S.-y. Takemura, A. Bharioke, Z. Lu, A. Nern, S. Vita-
ladevuni, P. K. Rivlin, W. T. Katz, D. J. Olbris, S. M.
Plaza, P. Winston, et al., Nature 500, 175 (2013).
[26] M. Helmstaedter, K. L. Briggman, S. C. Turaga, V. Jain,
H. S. Seung, and W. Denk, Nature 500, 168 (2013).
[27] A. L. Traud, P. J. Mucha, and M. A. Porter, Physica
A: Statistical Mechanics and its Applications 391, 4165
(2012).
[28] L. Takac and M. Zabovsky, in International Scientific
Conference and International Workshop Present Day
Trends of Innovations, Vol. 1 (2012).
[29] A. Mislove, M. Marcon, K. P. Gummadi, P. Dr-
uschel,
in Proceedings of
the 5th ACM/Usenix Internet Measurement Conference
(IMC'07) (San Diego, CA, 2007).
and B. Bhattacharjee,
[30] J. Leskovec, J. Kleinberg, and C. Faloutsos, ACM Trans-
actions on Knowledge Discovery from Data (TKDD) 1,
2 (2007).
[31] M. A. Porter, P. J. Mucha, M. E. J. Newman,
and
C. M. Warmbrand, Proceedings of the National Academy
of Sciences 102, 7057 (2005).
[43] U. Bhat, P. Krapivsky, R. Lambiotte, and S. Redner,
Physical Review E 94, 062302 (2016).
[32] J. Yang and J. Leskovec, Knowledge and Information
[44] R. Lambiotte, P. Krapivsky, U. Bhat, and S. Redner,
11
Systems 42, 181 (2015).
[33] B. Klimt and Y. Yang, in CEAS (2004).
[34] P. Panzarasa, T. Opsahl, and K. M. Carley, Journal of
the Association for Information Science and Technology
60, 911 (2009).
[35] J. Leskovec, D. P. Huttenlocher, and J. M. Kleinberg, in
Proceedings of the Internatonal Conference on Web and
Social Media (2010).
[36] J. Leskovec, J. Kleinberg, and C. Faloutsos, in Proceed-
ings of the eleventh ACM SIGKDD international con-
ference on Knowledge discovery in data mining (ACM,
2005) pp. 177–187.
[37] M. Ripeanu, A. Iamnitchi, and I. Foster, IEEE Internet
Computing 6, 50 (2002).
[38] M. Kaiser, New Journal of Physics 10, 083042 (2008).
[39] B. Bollob´as, European Journal of Combinatorics 1, 311
(1980).
[40] R. Milo, N. Kashtan, S. Itzkovitz, M. E. J. Newman, and
Physical Review Letters 117, 218301 (2016).
[45] R. Milo, S. Shen-Orr, S.
Itzkovitz, N. Kashtan,
D. Chklovskii, and U. Alon, Science 298, 824 (2002).
[46] L. R. Varshney, B. L. Chen, E. Paniagua, D. H. Hall,
and D. B. Chklovskii, PLOS Computational Biology 7,
e1001066 (2011).
[47] J. Ugander, L. Backstrom, and J. Kleinberg, in Proceed-
ings of the 22nd international conference on World Wide
Web (ACM, 2013) pp. 1307–1318.
[48] I. Scholtes, arXiv:1702.05499 (2017).
[49] V. Sekara, A. Stopczynski, and S. Lehmann, Proceedings
of the National Academy of Sciences 113, 9977 (2016).
[50] A. Fronczak,
and
J. Sienkiewicz, Physica A: Statistical Mechanics and its
Applications 316, 688 (2002).
J. A. Ho lyst, M. Jedynak,
[51] G. Caldarelli, R. Pastor-Satorras,
and A. Vespignani,
The European Physical Journal B: Condensed Matter
and Complex Systems 38, 183 (2004).
U. Alon, arXiv preprint cond-mat/0312028 (2003).
[52] R. F. Andrade, J. G. Miranda, and T. P. Lobao, Physical
[41] J. Park and M. E. J. Newman, Physical Review E 70,
Review E 73, 046101 (2006).
066117 (2004).
[42] P. Colomer-de Sim´on, M. ´A. Serrano, M. G. Beir´o, J. I.
Alvarez-Hamelin, and M. Bogun´a, Scientific Reports 3,
2517 (2013).
[53] B. Jiang and C. Claramunt, Environment and Planning
B: Planning and Design 31, 151 (2004).
|
1902.02372 | 1 | 1902 | 2019-02-06T19:30:15 | Modeling and Analysis of Tagging Networks in Stack Exchange Communities | [
"cs.SI",
"physics.soc-ph",
"stat.ML"
] | Large Question-and-Answer (Q&A) platforms support diverse knowledge curation on the Web. While researchers have studied user behavior on the platforms in a variety of contexts, there is relatively little insight into important by-products of user behavior that also encode knowledge. Here, we analyze and model the macroscopic structure of tags applied by users to annotate and catalog questions, using a collection of 168 Stack Exchange websites. We find striking similarity in tagging structure across these Stack Exchange communities, even though each community evolves independently (albeit under similar guidelines). Using our empirical findings, we develop a simple generative model that creates random bipartite graphs of tags and questions. Our model accounts for the tag frequency distribution but does not explicitly account for co-tagging correlations. Even under these constraints, we demonstrate empirically and theoretically that our model can reproduce a number of statistical properties of the co-tagging graph that links tags appearing in the same post. | cs.SI | cs | Modeling and Analysis of Tagging Networks in Stack Exchange Communities
Xiang Fu∗, and Shangdi Yu∗, and Austin R. Benson
Computer Science Department
{xf74, sy543}@cornell.edu, [email protected]
Cornell University
9
1
0
2
b
e
F
6
]
I
S
.
s
c
[
1
v
2
7
3
2
0
.
2
0
9
1
:
v
i
X
r
a
Abstract
Large Question-and-Answer (Q&A) platforms support diverse
knowledge curation on the Web. While researchers have stud-
ied user behavior on the platforms in a variety of contexts,
there is relatively little insight into important by-products
of user behavior that also encode knowledge. Here, we ana-
lyze and model the macroscopic structure of tags applied by
users to annotate and catalog questions, using a collection of
168 Stack Exchange websites. We find striking similarity in
tagging structure across these Stack Exchange communities,
even though each community evolves independently (albeit
under similar guidelines). Using our empirical findings, we
develop a simple generative model that creates random bi-
partite graphs of tags and questions. Our model accounts for
the tag frequency distribution but does not explicitly account
for co-tagging correlations. Even under these constraints, we
demonstrate empirically and theoretically that our model can
reproduce a number of statistical properties of the co-tagging
graph that links tags appearing in the same post.
1
Introduction
Question-and-Answer (Q&A) platforms are now a standard
context for social interaction on the Web with platforms such
as Quora and Stack Exchange supporting large user bases. As
a result, the social networks that these platforms support have
undergone a great deal of study, including, for example, how
people find interesting and popular questions on Quora (Wang
et al. 2013), prediction of "best answer" selection on Yahoo
Answers (Adamic et al. 2008), market design for knowledge
base construction with Google Answers (Chen, Ho, and Kim
2010), and badge collection on Stack Overflow (Anderson et
al. 2013). These studies have largely focused on models and
analysis of the user behavior. However, the users also create
other types of richly structured data. In this paper, we model
and analyze the structure revealed by tags on Stack Exchange,
which are used to annotate and catalog questions. Thus, our
principal object of study is the tags (and their relationships
through co-tagging), rather than the users; however, tags are
still a by-product of user behavior since users apply the tags.
A Stack Exchange website is a Q&A forum for a partic-
ular community. The platform began with Stack Overflow,
which is a community for computer programming. Stack
Overflow is the the largest and arguably most well-known
∗The first two authors contributed equally to this work.
Stack Exchange community, but the Stack Exchange ecosys-
tem supports a diverse set of communities ranging from pet
ownership1 to coffee2 to philosophy.3 For the most part, these
Stack Exchange communities evolve independently under the
same Q&A format (Fig. 1). A linchpin of every Stack Ex-
change community is the tagging system. When posting a
question, users are encouraged to apply a small number of
tags (at least one and at most five) that provide a reasonable
abstraction of the question's topics. In addition to describing
the question's content, tags also serve users in information re-
trieval of similar questions as well as questions they might be
able to answer. Tags on Stack Exchange are not taken lightly --
users cannot immediately create new tags and are encouraged
to use existing and popular tags (Fig. 1, bottom); moreover,
there are also official tagging guidelines.4 Thus, tags on Stack
Exchange are fundamentally different from, e.g., hashtags
on social media platforms such as Twitter which are largely
free from regulation. The value placed on tags means that
they can contain rich information about the community. For
example, tag frequencies can show popular topics and the
change of tag frequency over time can reveal the change of a
community's interests over time.
Here, we provide the first large-scale study of the macro-
scopic structure of tagging behavior by analyzing a collection
of 168 Stack Exchange communities. We frame our study
through the lens of network analysis, focusing on two net-
works constructed from the tagging behavior of users. The
first is the bipartite network of tags and questions, where
there is an edge between a tag and all of the questions to
which the tag was applied. The second is the co-tagging net-
work, or the projection of the first network onto the tags; in
this case, two tags are connected by an edge if the two tags
jointly annotate at least one question. (We also consider a
weighted version of the second network, where the weight is
the number of questions containing the two tags.)
Oftentimes, network analyses suffer from the fact that there
is only "one sample" of a social system to study. For example,
there is only one Facebook friendship graph (Ugander et al.
2011) and one Twitter follower network (Kwak et al. 2010)
1https://pets.stackexchange.com/
2https://coffee.stackexchange.com/
3https://philosophy.stackexchange.com/
4https://stackoverflow.com/help/tagging
ing the "co-tagging network" induced by the bipartite tag-
question network. Specifically, we analyze the graph where
the nodes are tags and there is an edge connecting two tags if
they are "co-tagged" on at least one question (with possible
weighting on edges corresponding to the number of ques-
tions on which the two tags appear). Our analysis focuses on
three macroscopic properties of the data. First, the weighted
number of co-tags of a given tag is well-approximated by a
linear function of the number of questions in which the tag
appears. Second, the number of unique co-tags of a given tag
is well-approximated by a simple third-degree polynomial
of the number of the tag frequency. Qualitatively, as we in-
crease the number of questions that a tag has appeared in,
the number of unique co-tags will also increas; however, this
growth tapers for popular tags, when it is difficult to accumu-
late more unique co-tags. Third, we measure three versions
of the clustering coefficient for weighted and unweighted
networks networks and find various levels of clustering and
find that the unweighted clustering coefficient is only mildly
correlated with the size of the Stack Exchange community
(as measured by the number of questions), but two versions
of the weighted versions both negatively correlate with size.
All three macroscopic properties are replicated by our
model, which we validate with both empirical and theoretical
analysis across the collection of 168 Stack Exchange net-
works. Importantly, the model does not bake in any notions
of correlation or clustering in the co-tagging but can still
replicate important co-tagging network properties. Thus, we
can conclude that these network properties could actually
be simply explained by our simple generative model that
only makes a strong assumption on the frequency distribution
of the tags. These findings contrast sharply with traditional
social network analysis in measuring clustering. Standard
random graph models for social networks that do not bake in
clustering structure do not exhibit the same clustering levels
as the real-world social system (Newman 2003). However,
in our case, the co-tagging network constructed from our bi-
partite tag-question generative model matches the clustering
levels in the empirical data.
2 Related Work
We summarize below how our research relates to several areas
in social media, information retrieval, and network science.
Online Q&A platforms Question-and-Answer (Q&A)
platforms have been a staple of online discussion for sev-
eral years, involving major web companies such as Ya-
hoo!, Google, and Quora. Research on these platforms has
spanned a variety of topics, including reputation mecha-
nisms (Bosu et al. 2013; Paul, Hong, and Chi 2012), answer
quality measurement (Wang et al. 2013; Posnett et al. 2012;
Anderson et al. 2012), network structure (Adamic et al. 2008;
Paranjape, Benson, and Leskovec 2017); social behav-
ior (Yang et al. 2011); answer prediction (Adamic et al. 2008;
Tian, Zhang, and Li 2013); topic popularity (Maity, Sahni,
and Mukherjee 2015); and expertise evaluation (MacLeod
2014; Posnett et al. 2012; Pal, Chang, and Konstan 2012).
This research has largely focused on the questions, answers,
and user behavior. Our paper, in contrast, treats tags as the
Figure 1: Stack Exchange tagging. (Top) A question
on the COFFEE Stack Exchange community with two
tags: espresso and nespresso (https://coffee.
stackexchange.com/q/1572). We study the tag fre-
quency distributions across a large collection of Stack
Exchange communities, as well as networks constructed
(Bottom)
from tags applied to the same questions.
User
tagging guidelines on the COFFEE
Stack Exchange (https://coffee.stackexchange.
com/questions/ask). The last rule says that users can-
not immediately create new tags without due process; thus,
tagging is fundamentally different from hashtags on other
social media platforms such as Twitter or Instagram.
interface of
to analyze. While such studies provide valuable insights into
real-world social systems, it is also well-known that there
can be randomness in the evolution of social networks when
crafted in a controlled setting (Salganik 2006). The Stack Ex-
change communities provide a unique opportunity to study
a collection of similar networks of tags with highly similar
dynamics that have evolved largely independently and dif-
fer most in the community topic (and implications of the
community topic, such as the number of users).
We begin with an empirical analysis on the frequency
distribution of the tags across our collection of 168 Stack Ex-
change communities. We find that this distribution is heavy-
tailed and well-approximated by a lognormal distribution,
and the two parameters of this distribution are themselves
well-approximated by a normal distribution when estimated
over the large collection of Stack Exchange communities.
From our findings, we devise a simple generative model for
creating random bipartite graphs with links connecting tags
to questions. The model takes as input the desired number of
questions, number of tags, number of total tag occurrences,
and two parameters of a lognormal distribution, and produces
as output a bipartite graph linking tags to questions.
We futher explore the Stack Exchange data by analyz-
fundamental object of study. Furthermore, most prior work
has only examined at most a few Q&A web sites, whereas
we study a large collection of Stack Exchange networks.
Folksonomy The tag-question network that we study is
related to the idea of folksonomy, a term coined by Thomas
Vander Wal to describe the practice of users tagging in-
formation for personal retrieval in an open social environ-
ment (Vander Wal 2005). Folksonomy has been a lens for
analysis on social media platforms such as CiteULike,
del.icio.us, and BibSonomy (Cattuto et al. 2007;
Capocci and Caldarelli 2008; Cattuto et al. 2009). A ma-
jor difference of these folksonomy studies and the present
work is that folksonomies are much less restricted in the
annotations -- users can add many (possibly new) annotations
freely -- whereas the Stack Exchange system is restricted (be-
tween one and five tags with systematic vetting of new tags).
And again, we analyze a large collection of Stack Exchange
communities and not just a few folksonomies.
Bipartite network models and co-tagging networks Bi-
partite graph (network) models are employed across a broad
range of scientific disciplines, including ecology (Bascompte,
Jordano, and Olesen 2006), biomedicine (Goh et al. 2007),
and information science (Akoglu, Chandy, and Faloutsos
2013). The model that we develop in this paper is a genera-
tive (random) model for a bipartite graph (network) between
tags and questions. Other generative models for bipartite (or
multipartite) graphs include the bipartite stochastic block
model (Larremore, Clauset, and Jacobs 2014), evolution-
ary affiliation networks (Lattanzi and Sivakumar 2009), and
generative models for folksonomy (Chojnacki and Kłopotek
2010). In contrast to prior research, the goal with our model
is to develop a simple generative model that captures the
empirical properties that we observe to persist across Stack
Exchange communities. Our model is designed to capture
the tag frequency distribution amongst questions, but we find
that properties of the co-tagging network -- where tags are
connected if they have appeared in a question together -- are
still replicated with our model. Properties and statistics of
co-tagging networks, such as clustering coefficients, char-
acteristic path lengths, and number of co-tags have been
used to analyze online communities such as del.icio.us
and BibSonomy, have been studied (Cattuto et al. 2007;
Halpin, Robu, and Shepherd 2007). Co-tagging networks
have also been used for application on connecting users with
similar interests (Wang, Liu, and Fan 2011).
3 Data Description and Preliminary Analysis
A Stack Exchange is a self-moderating online Q&A forum,
and each Stack Exchange community centers on a different
topic. Questions are annotated with at least one and at most
five tags that serve as essential descriptors of the question
(Fig. 1). Importantly, these platforms also largely evolve inde-
pendently, allowing us to perform a better statistical analysis
compared to analyzing a single Stack Exchange community.
We now describe our dataset collection and provide prelim-
inary statistical analyses that will serve the development of
our generative model in the next section.
Figure 2: Number of unique tags (green squares), number
of questions (yellow circles), and tag-question pairs (blue
triangles) of the 168 Stack Exchange communities that we
analyze in this paper. Datasets are sorted in ascending order
by number of questions. The APPLE and COFFEE communi-
ties are annotated as examples. In this paper, we analyze and
model the relationships between tags and questions.
Data Collection
We collected data from https://archive.org/
details/stackexchange, which hosts the entire his-
tory of every Stack Exchange community, including the tags
used to annotate questions. In total, we collected the sets of
tags applied to each question from 168 Stack Exchange com-
munities. In order to ensure that we could analyze data by
inspection, we omitted communities where the predominant
language was not English (thus, we do not consider the ES,
JA, PT, RU, RUS, and UKRAINIAN communities). However,
we do include Stack Exchange communities such as RUS-
SIAN, where people discuss the Russian language in English.
We also omitted so-called "meta" communities that discuss a
particular Stack Exchange community since these meta com-
munities have a different set of goals as well as a dependence
on the community that they discuss. Finally, we also omitted
Stack Overflow, which is over an order of magnitude larger
than any other community, and has already been the subject
of much research (Anderson et al. 2012; Bosu et al. 2013;
Wang et al. 2013). We will release our processed dataset
collection with the publication of this paper.
Figure 2 presents an overview of the basic statistics of our
collection of tags. Among the 168 Stack Exchange commu-
nities that we study in this paper, the number of unique tags
ranges from 70 (ARABIC) to 5,318 (SUPERUSER), and the
number of questions ranges from 122 (again, ARABIC) to
994,983 (MATH). Although the Stack Exchange communities
vary in size and topic and also evolve largely independently,
we see in the next section (and later in the paper) that there
are broad similarities across the communities.
Lognormal Distribution of Tag Frequencies
In this section, we study the distribution of tag frequen-
cies, i.e., the number of times that a tag is applied to a
question or, when normalized, the fraction of questions that
contains a given tag. One consistent trait is that tags used
Datasets101102103104105106CountsAPPLE: 90213APPLE: 256698APPLE: 1065COFFEE: 937COFFEE: 1849COFFEE: 107Tag-Question PairsQuestionsTags Figure 3: Tag frequencies of four diverse Stack Exchange
communities (APPLE -- 90,213 questions and 1,065 tags; COF-
FEE -- 937 questions and 107 tags; MATH -- 994,983 questions
and 1,749 tags; FITNESS -- 7,626 questions and 393 tags). We
find that tag frequencies are well-modeled by a lognormal
distribution in Stack Exchange communities (Fig. 4).
only a few times are much more common than tags used
many times, and the distribution of tags is heavy-tailed.
Many communities have tags appearing at much higher fre-
quencies than most other tags; as an extreme example, the
magic-the-gathering tag appears in more than 3000
questions in the BOARDGAMES community, while all other
tags appear in fewer than 500 questions.
Such heavy-tailed distributions are common on the Web
and other domains (Mitzenmacher 2004; Clauset, Shalizi,
and Newman 2009). Here, we find that the tag frequencies
are well-modeled by a lognormal distribution. Figure 3 illus-
trates four representative cases and also provides a compar-
ison against other commonly-used heavy-tailed probability
distributions such as a power law, truncated power law, and
stretched exponential. (Fig. 3 shows four such cases). We find
that a lognormal tends to match both the head and tail of the
distribution, while other common heavy-tailed distributions
can only capture either the head or tail of the distribution
(e.g., in Fig. 3, the truncated power law captures the head of
the tag frequency distribution in APPLE but not the tail and
the tail of the COFFEE distribution but not the head). The lone
outlier is the PATENT community, which does not seem to
be well-approximated by any commonly-used heavy-tailed
distribution.
More formally, we fit the parameters of a lognormal, power
law, truncated power law, and stretched exponential distri-
butions to the tag frequency of each Stack Exchange com-
munity using the powerlaw Python package (Alstott, Bull-
more, and Plenz 2013). Figure 4 (top left) shows the fitted
parameters, which are themselves approximately normally
distributed. We use two standard procedures for evaluating
the fit of the lognormal: the Kolmogorov-Smirnov (KS) statis-
tic and the likelihood ratio test comparing the lognormal to
other heavy-tailed degree distributions (Clauset, Shalizi, and
Newman 2009). The distribution of the KS statistics is much
(Top Left) The distribution of fitted lognormal
Figure 4:
parameters for tag frequency across 167 Stack Exchange
communities (we exclude the outlier PATENT), which are ap-
proximately normally distributed. (Top Right) The CDF of
the KS Statistic (D) of fitted heavy-tailed distribution to the
empirical data. The lognormal distribution has the smallest
statistics, and D < 0.06 for 80% of datasets; the only com-
munity with D > 0.15 is PATENT. (Bottom) The p-values
for comparing an alternative heavy tailed degree distribution
hypothesis to the null hypothesis of the lognormal (left) and
the distribution of the p value with the log-likelihood ratio
R (right). The small p-values and positive log-likelihood ra-
tios indicate that the lognormal is a better fit for the data
compared to other common heavy-tailed distributions.
smaller for the lognormal compared to the other distributions
(Fig. 4, top right) and is less than 0.06 for 80% of the Stack
Exchange communities. Furthermore, the p-values from the
likelihood ratio test show that the power law, truncated power
law, and stretched exponentials are not likely alternatives to
the null of a lognormal (Fig. 4, bottom).
To summarize, a lognormal distribution is an appropriate
model for the distribution of tag frequencies. In the next
section, we describe a simple generative model for random
bipartite graphs of tags and questions based on this lognormal
distribution. We will then later see that this model matches
the real data in a number of characteristics related to the
co-tagging, i.e., how multiple tags are used on the same
question.
4 A Generative Model for Bipartite
Tag-Question Networks
In this section, we propose a simple generative model for the
bipartite tag-question network. Later, we will see that this
model is able to recover many properties of the co-tagging
network of Stack Exchange communities, i.e, the graph where
nodes correspond to tags, and edges connect tags that have
been applied to the same question. Formally, the bipartite
tag-question graph B consists of disjoint vertex sets T and
Q, each corresponding to the set of tags and questions, as
well as a set of undirected edges E; where (t, q) ∈ E with
t ∈ T and q ∈ Q signifies that tag t is applied to question q.
The frequency, or number of occurrences, of a tag t is then
100101102103104Tag Occurrences106104102100CDFAPPLEPower LawTrunc. Power LawStretched Exp.Empirical DataLognormal100101102Tag Occurrences102101100CDFCOFFEE101103105Tag Occurrences104102100CDFMATH100101102103Tag Occurrences103102101100CDFFITNESS42024Parameters of Lognormal Fits0.00.51.01.5FrequencyN(2.46, 0.89)N(1.59, 0.24)0.00.10.20.30.4KS Statistic0.000.250.500.751.00CDFLognormalPower LawTrunc. Power LawStretched Exp.0.00.20.40.60.81.0p-value for Distribution Comparisons0.000.250.500.751.00CDFPower LawTrunc. Power LawStretched Exp.p = 0.1010203040R values10171013109105101Max(p, 1e-16)Power LawTrunc. Power LawStretched Exp. corrections.
Algorithm 1: Simple generative model for creating ran-
dom bipartite graphs of tags and questions.
Input: number of tags NT ; number of questions NQ;
target number of tag occurrences m; µ, σ2
Output: tag-question bipartite graph B = (T ∪ Q, E)
/* Sample tag occurrences and compute
*/
t ∼ LogNormal(µ, σ2), t = 1, . . . , NT .
1 x(cid:48)
2 xt ← round(m · x(cid:48)
t), t = 1, . . . , NT .
3 Solve NQ − NQ exp(−m/ NQ) = NQ for NQ.
4 NQ ← round( NQ).
/* Construct bipartite graph
5 T ← {1, . . . , NT}, Q ← {1, . . . , NQ}.
6 for each tag t ∈ T do
7
8
9 end
10 Q ← {q ∈ Q ∃t ∈ T for which (t, q) ∈ E}
Qt ← uniform sample of xt questions from Q.
for q ∈ Qt do add edge (t, q) to edge set E.
t/(cid:80)NT
t=1 x(cid:48)
*/
are scaled by a constant so that(cid:80)
simply the degree of t in the graph B.
Our random network model has two basic steps. First,
given NT = T, NQ = Q, and the parameters µ and σ of
a lognormal distribution, we first generate a sequence of tag
occurrence counts xt ∼ Lognormal(µ, σ2). These samples
t xt = m (where m is
the total number of tag occurrences in the original dataset)
and then rounded to an integer. Since scaling a lognormal
random variable by a constant is still lognormally distributed,
we maintain this property of the tag distribution, and this
preserves the total number of tag-question pairs in the dataset.
Second, we assign tag t to xt questions chosen uniformly at
random without replacement. In this simplified version of the
model, the output deviates from the Stack Exchange networks
in two ways: (i) it is possible that a question has no tags and
(ii) it is possible that a question is assigned more than five
tags. We now show how to account for these deviations, and
Algorithm 1 describes the full algorithm.
Correction for question counts To fix the problem where
questions can have no tags, we make a "correction" in the
number of questions. More specifically, we increase the num-
ber of questions from NQ to NQ so that after the random
assignment, the expected number of questions with at least
one tag is close in expectation to NQ, the number of questions
in the empirical dataset. We then simply discard questions
with no tags (Algorithm 1).
We approximate the expected number of questions with no
tags under a simplification where tags can be duplicated in
questions (the approximation is not necessary, but it makes
the calculations simpler, has small variance theoretically, and
provides good results empirically). Here, the probability that
a question gets 0 tags is the same for each question -- it is just
the probability that all tags are assigned to the other NQ − 1
Figure 5: (Left) The CDF of the ratio of the (theoretical)
standard deviations to the corrected number of questions.
The small ratio (less than 0.008 for 80% of datasets) shows
that our correction for the number of questions is accurate.
(Right) The CDF of the fraction of questions with more than
five tags in one sample of the random graph for each dataset.
This fraction is small -- less than 0.045 for 80% of datasets,
which justifies the relaxation in our random graph model.
questions:
NT(cid:89)
xi−1(cid:89)
(cid:104)
i=1
j=0
(cid:105) ≈ (1 − 1/ NQ)m,
1 − 1/( NQ − j)
where m is total number of tag occurrences. Thus, since NQ
and m are generally large, when assigning tags uniformly at
random to NQ questions, the expected number of questions
with 0 tags is
NQ(1 − 1/ NQ)m ≈ NQ exp(−m/ NQ).
There are NQ questions if the following equation is satisfied:
NQ − NQ exp(−m/ NQ) = NQ.
(1)
We claim that Eq. (1) has a unique positive solution NQ >
NQ. Since m and NQ are positive constants, the left hand
side of Eq. (1) is a function f of NQ. Moreover, the function
f is continuous and monotonically increasing in NQ, and
f (NQ) = NQ(1 − exp(−m/NQ)) < NQ. Therefore, the
above equation has a unique positive solution for NQ that
is larger than NQ. We can find the solution efficiently with
binary search, and then round NQ to the nearest integer.
In our experiments, using the corrected number of ques-
tions with our model is accurate, even with our approxima-
tions. Generating one sample for each dataset, the relative
error between the number of questions with at least one tag
in the model deviates from the true number of questions by
0.32% on average and by at most 3.75% across all datasets.
While these statistics are for just one sample in each network,
the variance in the number of questions with 0 tags is ap-
proximately NQp(1 − p). The ratio between the theoretical
standard deviation and the corrected number of questions is
small -- -less than 0.008 for 80% of the datasets (Fig. 5, left).
Number of tags per question We next justify our second
model deviation, which is that questions can be assigned more
than five tags. Our argument is that only a small fraction of
questions are actually assigned more than five tags with our
generative model. We generated tag-question bipartite graphs
0.000.010.020.03Standard dev.\ NQ0.000.250.500.751.00CDF0.0000.0250.0500.0750.100Fraction of Questions with More Than 5 Tags0.000.250.500.751.00CDF Figure 6: Distributions of number of tags per post. Markers
indicate number of tags: 1 -- blue triangle; 2 -- yellow 'Y';
3 -- green diamond; 4 -- red square; and 5 -- purple '+'. (Left)
Fraction of questions with a given number of tags as a func-
tion of the number of questions in a datasets. The distribution
of tags per post is roughly independent of the size of the Stack
Exchange community. (Right) Comparison of the distribu-
tion of the number of tags per post in the data and a sample
from our generative model. The distributions are strongly
correlated (shown in legend).
with Algorithm 1 for each Stack Exchange community using
the fitted lognormal parameters (Fig. 4, bottom). The mean
fraction of questions with more than 5 tags in the generated
networks across 168 Stack Exchange platforms is only 2.5%
and more than 80% of datasets have less than a 4.5% of
questions with more than five tags (Fig. 5, right).
Summary Algorithm 1 is a simple generative model for bi-
partite tag-question networks that generates tag occurrences
with the lognormal distribution that we found to be common
across nearly all Stack Exchange communities. As a first
look at how our model matches the empirical data, we con-
sider the distribution of the number of tags per question. In
the empirical data, this distribution tends to be uncorrelated
with the size of the dataset (Fig. 6, left). We also find that
the distribution of the number of tags per question in the
model closely matches the empirical data (Fig. 6, right). In
the next section, we analyze co-tagging, i.e., how tags jointly
annotate questions. Our model has no built-in notion of corre-
lations in co-tagging, yet we find that the model still matches
macroscopic co-tagging properties in the data.
5 Co-tagging Analysis
In addition to the bipartite tag-question network, we also
build a "co-tagging network" for each Stack Exchange com-
munity. Recall that the tag-question network B = (T ∪ Q, E)
is given by vertex sets T and Q corresponding to tags and
questions and has edges (t, q) ∈ E connecting tags to ques-
tions. The co-tagging network G is the projection of this
graph onto the set of tags. Formally, G = (T, F ), where
(s, t) ∈ F if and only if there is some question q ∈ Q such
that (s, q), (t, q) ∈ E. In this case, we say that s and t co-tag
with each other. We also associate a weight with each edge in
G corresponding to the number of questions containing the
two tags (the number of times that two nodes are co-tagged):
(2)
In the rest of this section, we show that co-tagging net-
works constructed from samples of our generative model
ws,t = {q ∈ Q (s, q), (t, q) ∈ E}.
Figure 7: (Left) The weighted number of co-tags is approx-
imately a linear function of tag frequency. Here, we show
the distribution of slopes from the linear regression over
our collection of Stack Exchange communities. The regres-
sion has an r2 value greater than 0.95 in 95% of the empir-
ical datasets and greater than 0.97 in 97% of the generated
datasets. (Right) The relationship between the fitted slope
on the data and in the model across the Stack Exchange
communities, which are strongly correlated.
(Algorithm 1) match statistical properties of the co-tagging
networks of empirical data, even though our model does not
explicitly account for co-tagging behavior. Again, we use
the lognormal parameters µ and σ fitted for each dataset
(Fig. 4) to generate a random graph for each Stack Exchange
network. We focus our attention on three properties of the
co-tagging network: (i) the expected number of co-tags (i.e.,
the weighted degree in G) as a function of tag frequency; (ii)
the expected number of unique co-tags (i.e., the unweighted
degree in G), again as a function of tag frequency; and (iii)
weighted and unweighted versions of the clustering coeffi-
cient of the graph G.
co-tags of tag t is kt =(cid:80)
Weighted Co-tags and Tag Frequency
We first examine the relationship between the number of
co-tags of a given tag as a function of its frequency (the
number of questions in which it appears). Here, we consider
the number of co-tags to be weighted, i.e., the number of
s∈T ws,t, following Eq. (2). In the
empirical data, this relationship is essentially linear -- a linear
model of the number of co-tags in regressed on the number of
questions containing the tag has a coefficient of determination
(r2 value) greater than 0.95 in 95% of the Stack Exchange
communities. Figure 7 (left) shows the distribution of the
slopes, which concentrate around 1.82.
We now show why we would also expect this behavior
from our model. Recall that the generative model samples tag
frequencies xt ∼ LogNormal(µ, σ2) and then scaled so that
these variables to match the total number of tag occurrences.
The number of co-tags between s and t then follows ws,t ∼
Hypergeom( NQ, xs, xt), where NQ is the corrected number
of questions in Algorithm 1. Thus, the expected number of
co-tags kt of a given tag t is
(cid:88)
E[kt] =
xsxt
NQ
=
t(cid:54)=s,s=1,...,NT
(m − xt)xt
,
NQ
where m is the target number of tag occurrences (the first
equality comes from the independence in assignment of the
102103104105106Number of Questions0.00.20.40.60.8Fraction of Questions0.00.20.40.60.8Fraction (Model)0.00.20.40.60.8Fraction (Data)c1=0.96c2=0.77c3=0.92c4=0.96c5=0.9212Slope (Data)0.000.250.500.751.00Frequencymean=1.82median=1.87123Slope (Model)12Slope (Data)Corr = 0.932 Figure 8: (Left) The CDF of the mean-squared error in third-
degree and first-degree (linear) polynomial models of of the
log number of unique co-tags in terms of the log of tag fre-
quency in both the data and the samples from the generative
model. The third-degree polynomial is a good approximation
and matches the expected value of the model (Fig. 9). (Right)
The CDF of the mean error in the expected number of unique
co-tags in the model and the actual number of co-tags in the
dataset. The error is less than 0.5 in 80% of the datasets. The
model slightly over-estimates the number of unique co-tags
by not taking into account tag correlations (see also Fig. 9).
d
dxt
tags). Although there is a quadratic relationship between
kt and xt, we know that xt is typically small compared to
m. Thus, the gradient is well-approximated by the linear
kt ≈ m/ NQ, independent of xt.
function m/ NQ, i.e.,
Our analysis here is independent of the lognormal distribution
of the tag frequency -- we only relied on independence in the
way that tags are assigned to questions.
In actual random samples, the linear relationship holds.
We performed the same linear regression on random samples
from our generative model using the fitted parameters in
Fig. 4 as we did for the empirical datasets. In the model, 97%
of the 168 datasets have a correlation coefficient r2 > 0.97.
Furthermore, the slopes from the regression on the generated
data are highly correlated with the slopes on the empirical
data (the correlation is 0.932; Fig. 7, right), and the mean
squared error between the slope derived from a sample from
the generative model and the computed slope on the empirical
data across all Stack Exchange communities is just 0.10.
Unique Co-tags and Tag Frequency
In the above analysis, we saw that the number of co-tags of a
given tag is approximately linear in the number of questions
in which the tag appears -- in both the empirical data and our
model-generated data. In this section, we instead consider
the number of unique co-tags of a given tag t as a function
of the number of questions containing tag t. In this case, the
number of unique co-tags is equal to the unweighted degree
of tag t in the co-tagging network G defined above.
We find that the log of the number of unique co-tags is
well-approximated as a third-degree polynomial of the log
of the number of question that contain the tag. Formally, let
dt denote the unweighted degree of tag t in the co-tagging
network G and xt the number of questions containing tag t.
We then fit a the following polynomial model:
log(d(t) + 1) =
ai log(xt + 1)i.
(3)
3(cid:88)
i=0
Figure 9: Relationship between the number of unique co-
tags and tag frequency on four Stack Exchange communities,
which is well-approximated by a degree-three polynomial
(see also Fig. 8). The model has the same shape, albeit slightly
above the data.
Figure 8 (left) shows the CDF of the mean-squared error of
the polynomial fit. The third-degree polynomial is a good
fit for both the empirical data and the model across the col-
lection of Stack Exchange communities. Figure 9 shows the
distributions and fit of the third-degree polynomial for a few
representative networks. In these cases, the polynomial fit
is accurate and captures the fact that the number of unique
co-tags does not grow linearly with tag frequency. Instead,
the growth in unique co-tags tapers for some of the most
frequently used tags. This happens because there is a limited
total number of tags (Fig. 2), so tags that occur frequently
have fewer options to increase the number of unique co-tags.
Interestingly, the fitted third-degree polynomial coeffi-
cients {ai}, when taken as a collection across the Stack
Exchange communities, largely lie on a lower-dimensional
subspace. In the empirical datasets, the first principal compo-
nent explains 86% of the variability, and the second principal
component explains an additional 13% of the variability. Sim-
ilar results hold for the fitted coefficients in datasets generated
with our model -- 89% of the variability is explained with the
first principal component and an addition 10% is explained
by the second principal component.
We can easily compute the expected number of unique
co-tags with a simple summation. We argued in the previous
section that the weighted number of co-tags between tags s
and t is ws,t ∼ Hypergeom( NQ, xs, xt). Thus, the expected
number of unique co-tags dt of tag t is
(cid:88)
s(cid:54)=t
E[dt] =
P(ws,t > 0) =
=
(cid:88)
(cid:88)
s(cid:54)=t
s(cid:54)=t
1 −
1 − P(ws,t = 0)
,
(cid:1)
(cid:0) NQ−xs
(cid:0) NQ
(cid:1)
xt
xt
where xs is the sampled number of questions for tag s in
0.00.10.20.3MSE0.00.20.40.60.81.0CDFDegree 3 Fit (Model)Degree 1 Fit (Model)Degree 3 Fit (Data)Degree 1 Fit (Data)0.250.500.751.00Mean Error (ln(Expected+1) - ln(Data+1))0.000.250.500.751.00CDF100101102103104Number of Tag Occurrences, xi100101102103Number of Unique Cotags, kUAPPLEData fitModel fitDataModelExpected100101102Number of Tag Occurrences, xi100101102Number of Unique Cotags, kUCOFFEE100101102103104105Number of Tag Occurrences, xi100101102103Number of Unique Cotags, kUMATH.STACKEXCHANGE.COM100101102103Number of Tag Occurrences, xi100101102Number of Unique Cotags, kUFITNESS u,v = w(cid:48)
u,v/ maxx,y w(cid:48)
x,y and summations over
u,v = 0 (i.e., with no edge) are ignored.
where w(cid:48)
cases where w(cid:48)
Figure 10 (top row) shows that all three clustering coef-
ficients are approximately normally distributed across the
collection of Stack Exchange communities. Furthermore,
the unweighted coefficients are only weakly correlated with
the size of the community, measured by the log-number of
questions on the Stack Exchange (Fig. 10, middle row). We
conclude that the size of a Stack Exchange community is
likely not a driving factor in the unweighted clustering of
the network, which backs up conventional wisdom for the
analysis of real-world networks (Newman 2003), differs from
the behavior of random graph models that produce heavy-
tailed degree distributions, where clustering decreases with
size (Bollob´as and Riordan 2004). On the other hand, the
weighted clustering coefficients tend to decrease with the size
of the Stack Exchange community (Fig. 10, middle row).
The co-tagging networks derived from samples of our gen-
erative model reproduce these clustering coefficients remark-
ably closely and with strong positive correlations (Fig. 10,
bottom row). Again, we emphasize that our model does not
bake in any explicit notion of clustering. Instead, our model
only matches the lognormal distribution of the tag frequency
and the total number of tags applied to all questions. Thus,
clustering in the co-tagging in Stack Exchange communities
could be explained simply by these simpler statistics. This
finding contrasts sharply with typical (social) network analy-
sis, where clustering is exhibited at a much higher level than
is expected by random graph models (Newman 2003). The
key difference is that our model is based on a projection of a
bipartite tag-question graph rather than directly modeling the
co-tagging network. This type of modeling has a long history
in sociology (Breiger 1974) but has received relatively less
theoretical attention in social network analysis (Lattanzi and
Sivakumar 2009).
6 Discussion
In addition to providing answers to questions, the users of
Q&A platforms create knowledge through annotation of ques-
tions. With their tagging system, Stack Exchange provides a
unique opportunity to study these annotations for two main
reasons. First, tags cannot be created freely and there are
community guidelines for their application, which differs
substantially from tagging norms on other social media plat-
forms. Second, there is a collection of Stack Exchange com-
munities that have largely evolved independently, enabling us
to model and analyze tagging with more statistical evidence.
And we indeed found similarities in macroscopic tagging
structure -- in terms of tag frequency and co-tagging network
structure -- across 168 Stack Exchange communities span-
ning a diverse range of topics. This contrasts from typical
network analyses that study a single snapshot of a social
network. Previously, researchers have circumvented this is-
sue by looking at, for example, sets of disparate subgraphs
from a larger graph (Traud, Mucha, and Porter 2012; Patania,
Petri, and Vaccarino 2017); samples of ego networks (Ugan-
der, Backstrom, and Kleinberg 2013; Benson et al. 2018;
Mcauley and Leskovec 2014); and collections of snapshots
Figure 10: (Top) The clustering coefficients of the Stack Ex-
change communities are approximately normally distributed.
(Middle) The unweighted clustering coefficient C has a weak
correlation with respect to the size of the community as mea-
sured by the log-number of questions; the weighted versions
are negative correlated. (Bottom) The clustering coefficients
in the co-tagging networks generated by our model are similar
to the clustering coefficients of the empirical Stack Exchange
communities.
Algorithm 1 and NQ is the corrected number of questions.
Figure 9 shows that the generated model data matches this
expectation.
Clustering in the co-tagging networks
Finally, we analyze the clustering coefficient of the co-
tagging networks, which is one of the fundamental measure-
ments in networks (Watts and Strogatz 1998; Newman 2003).
Let ∆u, du, and wu,v be the number of triangles containing
node u, the unweighted degree of du, and the weight of edge
(u, v) in the co-tagging graph G. We consider three clustering
coefficients:
1. The unweighted clustering coefficient (Watts and Strogatz
1998):
(cid:88)
u∈T
C =
1
T
2∆u
du(du − 1)
(cid:88)
v,z
(cid:88)
v,z
2. The weighted clustering clustering coefficient:
Cw =
1
T
1
du(du − 1)
( wu,v wu,z wv,z)1/3,
where wu,v = wu,v/ maxx,y wx,y (Onnela et al. 2005).
We will analyze log(Cw).
3. The log-weighted clustering coefficient, which is the same
as the mean weighted clustering coefficient, except the
weight wu,v is replaced by w(cid:48)
u,v = log(wu,v + 1):
Clw =
1
T
1
du(du − 1)
( w(cid:48)
u,v w(cid:48)
u,z w(cid:48)
v,z)1/3,
(cid:88)
u∈T
n(cid:88)
u∈T
0.20.40.6C0246FrequencyN(0.56, 0.06)864ln(Cw)0.00.10.20.3FrequencyN(-5.09, 1.28)0.10.2Clw01020FrequencyN(0.12, 0.03)103105Number of Questions0.20.40.6CCorr = 0.160103105Number of Questions864ln(Cw)Corr = -0.934103105Number of Questions0.10.2ClwCorr = -0.7120.20.40.60.8C (model)0.20.40.6C (data)Corr = 0.786642ln(Cw) (model)864ln(Cw) (data)Corr = 0.8390.10.2Clw (model)0.00.10.2Clw (data)Corr = 0.798 of time-evolving networks (Yaveroglu et al. 2014).
One macroscopic property across communities is that the
distribution of tag frequencies is well-modeled by a lognor-
mal distribution. The fitted parameters of the lognormal dis-
tributions (Fig. 4) themselves are approximately normally
distributed across our collection of Stack Exchange communi-
ties. Thus, one could incorporate this information as a simple
prior in bayesian modeling of tag-question networks.
We used the tag frequency distribution to develop a simple
generative model for random tag-question bipartite graphs,
which was able to reproduce a number of the co-tagging
and clustering properties of the datasets, without explicitly
modeling correlations or clustering in the co-tagging process.
Further understanding of the process producing this distri-
bution is an avenue for future research. For example, multi-
plicative growth models are a well-known generative process
for lognormal distributions (Mitzenmacher 2004). Although
outside the scope of this paper, the availability of temporal
information from Stack Exchange provides a path towards
more robust understanding of the underlying processes of tag
use, similar to other methods for estimating growth on the
Web and in social networks (Huberman and Adamic 1999;
Overgoor, Benson, and Ugander 2018).
Code and data. Code to reproduce our results, along with
processed data, are available at https://github.com/
yushangdi/stack-exchange-cotagging.
Acknowledgments. This research was supported by NSF
award DMS-1830274 and ARO award 86798.
References
[Adamic et al. 2008] Adamic, L. A.; Zhang, J.; Bakshy, E.;
and Ackerman, M. S. 2008. Knowledge sharing and yahoo
answers: Everyone knows something. In Proceedings of the
17th International Conference on World Wide Web (WWW),
665 -- 674.
[Akoglu, Chandy, and Faloutsos 2013] Akoglu, L.; Chandy,
R.; and Faloutsos, C. 2013. Opinion fraud detection in
online reviews by network effects. In International AAAI
Conference on Web and Social Media (ICWSM), volume 13,
2 -- 11.
[Alstott, Bullmore, and Plenz 2013] Alstott, J.; Bullmore, E.;
and Plenz, D. 2013. Powerlaw: a python package for analysis
of heavy-tailed distributions.
[Anderson et al. 2012] Anderson, A.; Huttenlocher, D.; Klein-
berg, J.; and Leskovec, J. 2012. Discovering value from com-
munity activity on focused question answering sites: a case
study of Stack Overflow. In Proceedings of the 18th ACM
SIGKDD International Conference on Knowledge Discovery
and Data Mining (KDD).
[Anderson et al. 2013] Anderson, A.; Huttenlocher, D.; Klein-
berg, J.; and Leskovec, J. 2013. Steering user behavior with
badges. In Proceedings of the 22nd International Conference
on World Wide Web (WWW).
[Bascompte, Jordano, and Olesen 2006] Bascompte, J.; Jor-
dano, P.; and Olesen, J. M. 2006. Asymmetric coevolu-
tionary networks facilitate biodiversity maintenance. Science
312(5772):431 -- 433.
[Benson et al. 2018] Benson, A. R.; Abebe, R.; Schaub, M. T.;
Jadbabaie, A.; and Kleinberg, J. 2018. Simplicial closure
and higher-order link prediction. Proceedings of the National
Academy of Sciences 115(48):E11221 -- E11230.
[Bollob´as and Riordan 2004] Bollob´as, B., and Riordan,
O. M. 2004. Mathematical results on scale-free random
graphs. In Handbook of Graphs and Networks. Wiley. 1 -- 34.
[Bosu et al. 2013] Bosu, A.; Corley, C. S.; Heaton, D.; Chat-
terji, D.; Carver, J. C.; and Kraft, N. A. 2013. Building
reputation in StackOverflow: An empirical investigation. In
Working Conference on Mining Software Repositories, 89 --
92.
[Breiger 1974] Breiger, R. L. 1974. The duality of persons
and groups. Social Forces 53(2).
[Capocci and Caldarelli 2008] Capocci, A., and Caldarelli, G.
2008. Folksonomies and clustering in the collaborative sys-
tem citeulike. Journal of Physics A: Mathematical and Theo-
retical 41(22).
[Cattuto et al. 2007] Cattuto, C.; Schmitz, C.; Baldassarri, A.;
Servedio, V. D. P.; Loreto, V.; Hotho, A.; Grahl, M.; and
Stumme, G. 2007. Network properties of folksonomies. AI
Communications 20(4):245 -- 262.
[Cattuto et al. 2009] Cattuto, C.; Barrat, A.; Baldassarri, A.;
Schehr, G.; and Loreto, V. 2009. Collective dynamics of
social annotation. Proceedings of the National Academy of
Sciences 106(26):10511 -- 10515.
[Chen, Ho, and Kim 2010] Chen, Y.; Ho, T.-H.; and Kim, Y.-
M. 2010. Knowledge market design: A field experiment
at google answers. Journal of Public Economic Theory
12(4):641 -- 664.
[Chojnacki and Kłopotek 2010] Chojnacki, S., and Kłopotek,
M. 2010. Random graph generative model for folksonomy
network structure approximation. Procedia Computer Sci-
ence 1(1):1683 -- 1688.
[Clauset, Shalizi, and Newman 2009] Clauset, A.; Shalizi,
C.; and Newman, M. 2009. Power-law distributions in
empirical data. SIAM Review 51(4):661 -- 703.
[Goh et al. 2007] Goh, K.-I.; Cusick, M. E.; Valle, D.; Childs,
B.; Vidal, M.; and Barabasi, A.-L. 2007. The human disease
network. Proceedings of the National Academy of Sciences
104(21):8685 -- 8690.
[Halpin, Robu, and Shepherd 2007] Halpin, H.; Robu, V.;
and Shepherd, H. 2007. The complex dynamics of col-
laborative tagging. In Proceedings of the 16th International
Conference on World Wide Web (WWW).
[Huberman and Adamic 1999] Huberman, B. A.,
and
Adamic, L. A. 1999. Growth dynamics of the world-wide
web. Nature 401(6749):131 -- 131.
[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 19th International Conference on World
Wide Web (WWW).
[Larremore, Clauset, and Jacobs 2014] Larremore, D. B.;
Clauset, A.; and Jacobs, A. Z. 2014. Efficiently inferring
community structure in bipartite networks. Physical Review
E 90(1).
2005.
Folk-
http://www.vanderwal.net/random/
[Ugander, Backstrom, and Kleinberg 2013] Ugander,
J.;
Backstrom, L.; and Kleinberg, J. 2013. Subgraph frequen-
cies: mapping the empirical and extremal geography of large
graph collections. In Proceedings of the 22nd International
Conference on World Wide Web (WWW).
[Ugander et al. 2011] Ugander, J.; Karrer, B.; Backstrom, L.;
and Marlow, C. 2011. The anatomy of the facebook social
graph. arXiv:1111.4503.
[Vander Wal 2005] Vander Wal, T.
sonomy.
entrysel.php?blog=1622.
[Wang et al. 2013] Wang, G.; Gill, K.; Mohanlal, M.; Zheng,
H.; and Zhao, B. Y. 2013. Wisdom in the social crowd: An
analysis of quora. In Proceedings of the 22nd International
Conference on World Wide Web (WWW), 1341 -- 1352.
[Wang, Liu, and Fan 2011] Wang, X.; Liu, H.; and Fan, W.
2011. Connecting users with similar interests via tag network
inference. In Proceedings of the 20th ACM International
Conference on Information and Knowledge Management,
1019 -- 1024.
[Watts and Strogatz 1998] Watts, D. J., and Strogatz, S. H.
1998. Collective dynamics of 'small-world' networks. Nature
393(6684):440 -- 442.
[Yang et al. 2011] Yang, J.; Morris, M. R.; Teevan, J.;
Adamic, L. A.; and Ackerman, M. S. 2011. Culture matters:
A survey study of social Q&A behavior. In International
AAAI Conference on Web and Social Media (ICWSM).
[Yaveroglu et al. 2014] Yaveroglu, O. N.; Malod-Dognin, N.;
Davis, D.; Levnajic, Z.; Janjic, V.; Karapandza, R.; Sto-
jmirovic, A.; and Przulj, N. 2014. Revealing the hidden
language of complex networks. Scientific Reports 4(1).
[Lattanzi and Sivakumar 2009] Lattanzi, S., and Sivakumar,
D. 2009. Affiliation networks. In Proceedings of the 41st
annual ACM Symposium on Theory of Computing (STOC).
[MacLeod 2014] MacLeod, L. 2014. Reputation on Stack Ex-
change: Tag, you're it! In 28th International Conference on
Advanced Information Networking and Applications Work-
shops, 670 -- 674.
[Maity, Sahni, and Mukherjee 2015] Maity, S.; Sahni,
J.
S. S.; and Mukherjee, A. 2015. Analysis and Prediction
of Question Topic Popularity in Community Q&A Sites: A
Case Study of Quora. In International AAAI Conference on
Web and Social Media (ICWSM).
[Mcauley and Leskovec 2014] Mcauley, J., and Leskovec, J.
2014. Discovering social circles in ego networks. ACM
Transactions on Knowledge Discovery from Data 8(1):1 -- 28.
[Mitzenmacher 2004] Mitzenmacher, M. 2004. A brief his-
tory of generative models for power law and lognormal dis-
tributions. Internet Mathematics 1(2):226 -- 251.
[Newman 2003] Newman, M. E. J. 2003. The structure and
function of complex networks. SIAM Review 45(2):167 -- 256.
[Onnela et al. 2005] Onnela, J. P.; Saramki, J.; Kertsz, J.; and
Kaski, K. 2005.
Intensity and coherence of motifs in
weighted complex networks. Physical Review E.
[Overgoor, Benson, and Ugander 2018] Overgoor, J.; Ben-
son, A. R.; and Ugander, J. 2018. Choosing to grow
a graph: Modeling network formation as discrete choice.
arXiv:1811.05008.
[Pal, Chang, and Konstan 2012] Pal, A.; Chang, S.; and Kon-
stan, J. A. 2012. Evolution of experts in question answering
communities. In International AAAI Conference on Web and
Social Media (ICWSM).
[Paranjape, Benson, and Leskovec 2017] Paranjape, A.; Ben-
son, A. R.; and Leskovec, J. 2017. Motifs in temporal
networks. In Proceedings of the Tenth ACM International
Conference on Web Search and Data Mining.
[Patania, Petri, and Vaccarino 2017] Patania, A.; Petri, G.;
and Vaccarino, F. 2017. The shape of collaborations. EPJ
Data Science 6(1).
[Paul, Hong, and Chi 2012] Paul, S. A.; Hong, L.; and Chi,
E. H. 2012. Who is authoritative? understanding reputation
mechanisms in Quora. In Collective Intelligence.
[Posnett et al. 2012] Posnett, D.; Warburg, E.; Devanbu, P.;
and Filkov, V. 2012. Mining Stack Exchange: Expertise is ev-
ident from initial contributions. In International Conference
on Social Informatics, 199 -- 204.
[Salganik 2006] Salganik, M. J. 2006. Experimental study of
inequality and unpredictability in an artificial cultural market.
Science 311(5762):854 -- 856.
[Tian, Zhang, and Li 2013] Tian, Q.; Zhang, P.; and Li, B.
2013. Towards predicting the best answers in community-
based question-answering services. In International AAAI
Conference on Web and Social Media (ICWSM).
[Traud, Mucha, and Porter 2012] Traud, A. L.; Mucha, P. J.;
and Porter, M. A. 2012. Social structure of Facebook net-
works. Physica A: Statistical Mechanics and its Applications
391(16):4165 -- 4180.
|
1904.05329 | 3 | 1904 | 2019-08-14T22:37:02 | GraSPy: Graph Statistics in Python | [
"cs.SI",
"stat.ML"
] | We introduce GraSPy, a Python library devoted to statistical inference, machine learning, and visualization of random graphs and graph populations. This package provides flexible and easy-to-use algorithms for analyzing and understanding graphs with a scikit-learn compliant API. GraSPy can be downloaded from Python Package Index (PyPi), and is released under the Apache 2.0 open-source license. The documentation and all releases are available at https://neurodata.io/graspy. | cs.SI | cs |
GraSPy: Graph Statistics in Python
Jaewon Chung1, †, Benjamin D. Pedigo1, †, Eric W. Bridgeford2, Bijan K.Varjavand1, Hayden S. Helm3,
and Joshua T. Vogelstein1, 3, 4, (cid:63)
1Department of Biomedical Engineering, Johns Hopkins University, Baltimore, MD 21218
2Department of Biostatistics, Johns Hopkins University of Public Health, Baltimore, MD 21218
3Center for Imaging Science, Johns Hopkins University, Baltimore, MD 21218
4Kavli Neuroscience Discovery Institute, Institute for Computational Medicine, Johns Hopkins Univer-
sity, Baltimore, MD 21218
†Denotes equal contribution
(cid:63)Corresponding author
We introduce GraSPy, a Python library devoted to statistical inference, machine learning, and
visualization of random graphs and graph populations. This package provides flexible and easy-
to-use algorithms for analyzing and understanding graphs with a scikit-learn compliant
API. GraSPy can be downloaded from Python Package Index (PyPi), and is released under the
Apache 2.0 open-source license. The documentation and all releases are available at https:
//neurodata.io/graspy.
1 Introduction
Graphs, or networks, are a mathematical representation of data that consists of discrete objects (nodes
or vertices) and relationships between these objects (edges). For example, in a brain, regions of in-
terest can be vertices, the edges represent the presence of a structural connection between them [1].
Since graphs necessarily deal with relationships between nodes, classical statistical assumptions about
independence are violated. Thus, novel methodology is required for performing statistical inference on
graphs and populations of graphs [2]. While the theory for inference on graphs is highly developed, to
date, there has not existed a numerical package implementing these methods. GraSPy fills this gap
by providing implementations of algorithms with strong statistical guarantees, such as graph and multi-
graph embedding methods, two-graph hypothesis testing, and clustering of vertices of graphs. Many
of the algorithms implemented in GraSPy are flexible and can operate on graphs that are weighted
or unweighted, as well as directed or undirected. Table 1 provides a comparison of GraSPy to other
existing graph analysis packages [3 -- 5].
2 Library Overview
GraSPy includes functionality for fitting and sampling from random graph models, performing dimen-
sionality reduction on graphs or populations of graphs (embedding), testing hypotheses on graphs,
and plotting of graphs and embeddings. The following provides brief overview of different modules of
GraSPy. An example workflow using these modules is shown in Figure 1. More detailed overview
and code usage can be found in the tutorial section of GraSPy documentation at https://graspy.
neurodata.io/tutorial. All descriptions here correspond to GraSPy version 0.1.1.
Simulations Several random graph models are implemented in GraSPy, including the Erdos-Rényi
(ER) model, stochastic block model (SBM), degree-corrected Erdos-Rényi (DCER) model, degree-
corrected stochastic block model (DCSBM), and random dot product graph (RDPG) [6 -- 8]. The
1
Graph Theory
Statistical Modeling
Other
al
ers
v
tra
package
GraSPy 0.1.1
NetworkX 2.3
graph-tool 2.29
Snap.py 4.1
stats.
ork
etw
n
s
nitie
u
m
m
o
c
d
e
b
m
e
d
e
b
m
e
ult.
m
fit
el
d
o
m
s
n
ulatio
sim
st
te
p.
y
h
g
plottin
stall
in
pip
Table 1: Qualitative comparison of Python graph analysis packages. GraSPy is largely complementary
to existing graph analysis packages in Python. GraSPy does not implement many of the essential
algorithms for operating on graphs (rather, it leverages NetworkX for these implementations). The
focus of GraSPy is on statistical modeling of populations of networks, with features such as multiple
graph embeddings, model fitting, and hypothesis testing. A is given for packages that implement the
respective feature, a for packages that partially implement the respective feature, and a is given
for packages that do not implement the respective feature. Note that while a shows that the feature
exists in the corresponding package, it does not imply that the specific algorithms are the same for all
packages.
Figure 1: Illustration of modules and procedure for statistical inference on graphs, populations of graphs,
or simulated data. A detailed description of each module is given in Section 2.
2
SimulationsSample random graphsSingle graphInput dataMultiple graphsInput dataUtilsImport, preprocessingPlotVisualize graphs, graph embeddingsEmbedGraph, graph population dimensionality reductionModelsFit graph models to dataClusterCluster vertex or graph embeddingsInferenceTest graph hypotheses simulations module allows the user to sample random graphs given the parameters of one of
these models. Additionally, the user can specify a distribution on the weights of graph edges.
Utils GraSPy includes a variety of utility functions for graph and graph population importing and
preprocessing. GraSPy can import graphs represented as NetworkX objects or NumPy arrays.
Preprocessing examples include finding the largest connected component of a graph, finding the
intersection or union of connected components across multiple graphs, or checking whether a
graph is directed.
Embed Inferences on random graphs can leverage low-dimensional Euclidean representations of the
vertices, known as latent positions. Adjacency spectral embedding (ASE) and Laplacian spectral
embedding (LSE) are methods for embedding a single graph [2]. Omnibus embedding and mul-
tiple adjacency spectral embedding (MASE) allows for embedding multiple graphs into the same
subspace such that the embeddings can be meaningfully compared [9, 10]. GraSPy includes a
method for choosing the number of embedding dimensions automatically [11].
Models GraSPy includes classes for fitting random graph models to an input graph (Figure 2). Cur-
rently, ER, SBM, DCER, DCSBM, and RDPG are supported for model estimation. After fitting
a model to data, the model class can also output goodness-of-fit metrics (mean squared error,
likelihood) and the number of estimated model parameters, allowing for model selection. The
model classes can also be used to sample new simulated graphs based on the fit model.
Inference Given two graphs, a natural question to ask is whether these graphs are both random
samples from the same generative distribution. GraSPy provides two types of test for this null
hypothesis: a latent position test and a latent distribution test. Both tests are framed under the
RDPG model, where the generative distribution for the graph can be modeled as a set of latent
positions. The latent position test can only be performed on two graphs of the same size and with
known correspondence between the vertices of the two graphs [13]. The latent distribution test
can be performed on graphs without vertex alignment, or even with slightly different numbers of
vertices [14].
Cluster GraSPy extends Gaussian mixture models (GMM) and k-means from scikit-learn to
sweep over a specified range of parameters and choose the best clustering [15]. The number
of clusters and covariance structure for each GMM is chosen by Bayesian information criterion
(BIC), which is a penalized likelihood function to evaluate the quality of estimators [16]. Silhouette
score is used to choose the number of clusters for k-means [17]. Clustering is often useful for
computing the the community structure of vertices after embedding.
Plot GraSPy extends seaborn to visualize graphs as adjacency matrices and embedded graphs as
paired scatter plots [18]. Individual graphs can be visualized using heatmap function, and multiple
graphs can be overlaid on top of each other using gridplot. The nodes in both graph visualizations
can be sorted by various node metadata, such as node degree or assigned node labels. Pairplot
can visualize high dimensional data, such as embeddings, as a pairwise scatter plot.
3
Figure 2: Connectome model fitting using GraSPy. Heatmaps show the probability of potential edges
for models of graphs fit to the Drosophila larva right mushroom body connectome (unweighted, di-
rected) [12]. The node labels correspond to cell types: P) projection neurons, O) mushroom body
output neurons, I) mushroom body input neurons. The graph models are: inhomogeneous Erdos-Rényi
(IER) model in which all potential edges are specified, random dot product graph (RDPG), degree-
corrected stochastic block model (DCSBM), degree-corrected Erdos-Rényi (DCER), stochastic block
model (SBM), and Erdos-Rényi (ER). Blocks (cell types) are sorted by number of member vertices and
nodes are sorted by degree within blocks. The code used to generate the figure is tutorial section at
https://neurodata.io/graspy.
4
IERRDPGDCSBMDCERSBMERKPOIKPOIKPOIKPOIKPOIKPOIKPOIKPOIKPOIKPOIKPOIKPOI0.00.20.40.60.81.0 3 Code example
Given the connectomes of the Drosophila larva left and right mushroom bodies, one natural question
to ask is: how similar are these graphs [12]? We can frame this question as whether these graphs are
generated from the same distribution of latent positions [14]. We can use the latent distribution test to
test this hypothesis:
from graspy.datasets import load_drosophila_left, load_drosophila_right
from graspy.inference import LatentDistributionTest
# Load data
left_graph = load_drosophila_left()
right_graph = load_drosophila_right()
# Initialize hypothesis test object and compute p-value
ldt = LatentDistributionTest(n_components=3, n_bootstraps=500)
p_value = ldt.fit(left_graph, right_graph)
print("p-value: " + str(p_value))
p-value: 0.002
4 Conclusion
GraSPy is an open-source Python package to perform statistical analysis on graphs and graph popu-
lations. Its compliance with the scikit-learn API makes it an easy-to-use tool for anyone familiar
with machine learning in Python [19].
In addition, GraSPy is implemented with an extensible class
structure, making it easy to modify and add new algorithms to the package. As GraSPy continues to
grow and add functionality, we believe it will accelerate statistically principled discovery in any field of
study concerned with graphs or populations of graphs.
Acknowledgements
This work is graciously supported by the DARPA, under agreement numbers FA8650-18-2-7834 and
FA8750-17-2-0112. We thank all the contributors for assisting with writing GraSPy. We thank the
NeuroData Design class, the NeuroData lab, and Carey E. Priebe at Johns Hopkins University for
helpful feedback.
5
Bibliography
[1] J. T. Vogelstein, E. W. Bridgeford, B. D. Pedigo, J. Chung, K. Levin, B. Mensh, and C. E. Priebe,
"Connectal coding: discovering the structures linking cognitive phenotypes to individual histories,"
Current Opinion in Neurobiology, vol. 55, pp. 199 -- 212, 2019.
[2] A. Athreya, D. E. Fishkind, M. Tang, C. E. Priebe, Y. Park, J. T. Vogelstein, K. Levin, V. Lyzinski,
Y. Qin, and D. L. Sussman, "Statistical
inference on random dot product graphs: a survey,"
Journal of Machine Learning Research, vol. 18, no. 226, pp. 1 -- 92, 2018. [Online]. Available:
http://jmlr.org/papers/v18/17-448.html
[3] A. Hagberg, P. Swart, and D. S Chult, "Exploring network structure, dynamics, and function using
networkx," Los Alamos National Lab.(LANL), Los Alamos, NM (United States), Tech. Rep., 2008.
[4] T. P. Peixoto,
"The graph-tool python library," figshare, 2014.
[Online]. Available:
http:
//figshare.com/articles/graph_tool/1164194
[5] J. Leskovec and R. Sosic, "Snap: A general-purpose network analysis and graph-mining library,"
ACM Transactions on Intelligent Systems and Technology (TIST), vol. 8, no. 1, p. 1, 2016.
[6] P. W. Holland, K. B. Laskey, and S. Leinhardt, "Stochastic blockmodels: First steps," Social net-
works, vol. 5, no. 2, pp. 109 -- 137, 1983.
[7] B. Karrer and M. E. Newman, "Stochastic blockmodels and community structure in networks,"
Physical review E, vol. 83, no. 1, p. 016107, 2011.
[8] S. J. Young and E. R. Scheinerman, "Random dot product graph models for social networks,"
in International Workshop on Algorithms and Models for the Web-Graph. Springer, 2007, pp.
138 -- 149.
[9] K. Levin, A. Athreya, M. Tang, V. Lyzinski, and C. E. Priebe, "A central limit theorem for an omnibus
embedding of multiple random dot product graphs," pp. 964 -- 967, 2017.
[10] J. Arroyo, A. Athreya, J. Cape, G. Chen, C. E. Priebe, and J. T. Vogelstein, "Inference for multiple
heterogeneous networks with a common invariant subspace," arXiv preprint arXiv:1906.10026,
2019.
[11] M. Zhu and A. Ghodsi, "Automatic dimensionality selection from the scree plot via the use of profile
likelihood," Computational Statistics & Data Analysis, vol. 51, no. 2, pp. 918 -- 930, 2006.
[12] K. Eichler, F. Li, A. Litwin-Kumar, Y. Park, I. Andrade, C. M. Schneider-Mizell, T. Saumweber,
A. Huser, C. Eschbach, B. Gerber et al., "The complete connectome of a learning and memory
centre in an insect brain," Nature, vol. 548, no. 7666, p. 175, 2017.
[13] M. Tang, A. Athreya, D. L. Sussman, V. Lyzinski, Y. Park, and C. E. Priebe, "A semiparametric two-
sample hypothesis testing problem for random graphs," Journal of Computational and Graphical
Statistics, vol. 26, no. 2, pp. 344 -- 354, 2017.
[14] M. Tang, A. Athreya, D. L. Sussman, V. Lyzinski, and C. E. Priebe, "A nonparametric two-sample
hypothesis testing problem for random dot product graphs," Journal of Computational and Graphi-
cal Statistics, Sep. 2014.
6
[15] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Pretten-
hofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and
E. Duchesnay, "Scikit-learn: Machine learning in Python," Journal of Machine Learning Research,
vol. 12, pp. 2825 -- 2830, 2011.
[16] G. Schwarz et al., "Estimating the dimension of a model," The annals of statistics, vol. 6, no. 2, pp.
461 -- 464, 1978.
[17] P. J. Rousseeuw, "Silhouettes: a graphical aid to the interpretation and validation of cluster analy-
sis," Journal of computational and applied mathematics, vol. 20, pp. 53 -- 65, 1987.
[18] M. Waskom, O. Botvinnik, D. O'Kane, P. Hobson, J. Ostblom, S. Lukauskas, D. C.
Gemperline, T. Augspurger, Y. Halchenko, J. B. Cole, J. Warmenhoven, J. de Ruiter,
C. Pye, S. Hoyer, J. Vanderplas, S. Villalba, G. Kunter, E. Quintero, P. Bachant, M. Martin,
K. Meyer, A. Miles, Y. Ram, T. Brunner, T. Yarkoni, M. L. Williams, C. Evans, C. Fitzgerald,
Brian, and A. Qalieh, "mwaskom/seaborn: v0.9.0 (july 2018)," Jul. 2018. [Online]. Available:
https://doi.org/10.5281/zenodo.1313201
[19] L. Buitinck, G. Louppe, M. Blondel, F. Pedregosa, A. Mueller, O. Grisel, V. Niculae, P. Pretten-
hofer, A. Gramfort, J. Grobler, R. Layton, J. VanderPlas, A. Joly, B. Holt, and G. Varoquaux, "API
design for machine learning software: experiences from the scikit-learn project," in ECML PKDD
Workshop: Languages for Data Mining and Machine Learning, 2013, pp. 108 -- 122.
[20] D. G. Albertson and J. Thompson, "The pharynx of caenorhabditis elegans," Phil. Trans. R. Soc.
Lond. B, vol. 275, no. 938, pp. 299 -- 325, 1976.
[21] J. G. White, E. Southgate, J. N. Thomson, and S. Brenner, "The structure of the nervous system
of the nematode caenorhabditis elegans," Philos Trans R Soc Lond B Biol Sci, vol. 314, no. 1165,
pp. 1 -- 340, 1986.
[22] H. Akaike, "A new look at the statistical model identification," in Selected Papers of Hirotugu Akaike.
Springer, 1974, pp. 215 -- 222.
7
|
1912.00634 | 1 | 1912 | 2019-12-02T08:50:27 | Meta-Path Constrained Random Walk Inference for Large-Scale Heterogeneous Information Networks | [
"cs.SI",
"cs.LG"
] | Heterogeneous information network (HIN) has shown its power of modeling real world data as a multi-typed entity-relation graph. Meta-path is the key contributor to this power since it enables inference by capturing the proximities between entities via rich semantic links. Previous HIN studies ask users to provide either 1) the meta-path(s) directly or 2) biased examples to generate the meta-path(s). However, lots of HINs (e.g., YAGO2 and Freebase) have rich schema consisting of a sophisticated and large number of types of entities and relations. It is impractical for users to provide the meta-path(s) to support the large scale inference, and biased examples will result in incorrect meta-path based inference, thus limit the power of the meta-path. In this paper, we propose a meta-path constrained inference framework to further release the ability of the meta-path, by efficiently learning the HIN inference patterns via a carefully designed tree structure; and performing unbiased random walk inference with little user guidance. The experiment results on YAGO2 and DBLP datasets show the state-of-the-art performance of the meta-path constrained inference framework. | cs.SI | cs | Meta-Path Constrained Random Walk Inference
for Large-Scale Heterogeneous Information Networks
Chenguang Wang
Amazon Web Services
[email protected]
9
1
0
2
c
e
D
2
]
I
S
.
s
c
[
1
v
4
3
6
0
0
.
2
1
9
1
:
v
i
X
r
a
Abstract
Heterogeneous information network (HIN) has
shown its power of modeling real world data as a
multi-typed entity-relation graph. Meta-path is the
key contributor to this power since it enables infer-
ence by capturing the proximities between entities
via rich semantic links. Previous HIN studies ask
users to provide either 1) the meta-path(s) directly
or 2) biased examples to generate the meta-path(s).
However, lots of HINs (e.g., YAGO2 and Freebase)
have rich schema consisting of a sophisticated and
large number of types of entities and relations. It
is impractical for users to provide the meta-path(s)
to support the large scale inference, and biased ex-
amples will result in incorrect meta-path based in-
ference, thus limit the power of the meta-path. In
this paper, we propose a meta-path constrained in-
ference framework to further release the ability of
the meta-path, by efficiently learning the HIN infer-
ence patterns via a carefully designed tree structure;
and performing unbiased random walk inference
with little user guidance. The experiment results on
YAGO2 and DBLP datasets show the state-of-the-
art performance of the meta-path constrained infer-
ence framework.
1 Introduction
Heterogeneous information networks (HINs) consisting of
multi-typed entities and relations have been extensively stud-
ied recently, and shown the power against many other state-
of-the-art data models in lots of real world applications, such
as link prediction [Taskar et al., 2003; Lu and Zhou, 2011]
and classification [Ji et al., 2010; Wang et al., 2016a], etc.
The main reason behind the success of the HIN studies is the
concept of meta-path [Sun et al., 2011a]. Meta-path is a se-
quence of consecutive entity types and relation types captur-
ing semantic proximity between entities in HINs, thus pro-
vides the opportunities to deeply understand the data.
Most HIN tasks can be formalized as inference with meta-
paths problems in HINs. The goal of HIN inference is
to find the best assignment to the output variables accord-
ing to the given model and input instances. HIN infer-
ence particularly leverages the proximities between the in-
1
put and output variables provided by the meta-paths, i.e.,
y∗ = arg maxy f (θ, Φ(x, P, y)), where f is an HIN model,
θ is the model parameters, Φ is a set of representation (i.e.,
feature) functions based on input instance x, output variables
y and meta-paths P relevant to x and y. For example, in HIN
based link prediction task, given the linear regression based
prediction model f and the feature function set Φ built upon
source entity s and target entity t (e.g., x = [s, t]), meta-path
based link prediction can be seen as finding the best assign-
ment to output variable y (one-dimension vector, e.g., [1] rep-
resents link exists, [0] represents link doesn't exist), via the
meta-paths P between x and y. As we can see, the power of
the inference in HINs is brought by meta-path(s). The ability
of how well we can handle the meta-paths is the key to better
inference in HINs.
However, there are two possible limits in the current infer-
ence methods in large scale HINs.
User needs to provide the meta-path(s). Most of the
previous HIN studies ask users or experts to provide meta-
path(s) as explicit inference rules to perform relevant tasks in
HINs, such as similarity search [Sun et al., 2011a]. Since in
traditional HINs, such as DBLP, there are only four types of
entities, Paper, Venue, Author and Term and relevant relation
types between the entities. It will be possible for users to pro-
vide high-quality meta-paths. However, in large scale HINs,
such as YAGO, it consists of a sophisticated and large num-
ber of entity types (e.g., millions) and relation types (e.g.,
hundreds). Users will feel harder to provide meta-paths in
such cases. Moreover, multi-order (length larger than one)
meta-paths could carry more important information for infer-
ence than that of first-order meta-path (length equals to one).
Thus it becomes harder to rely on users to provide the relevant
meta-paths for inference in large scale HINs.
Biased examples lead to incorrect inference. To relieve
the efforts of users, random walk based inference [Lao et al.,
2011a; Lao et al., 2011b] is proposed for large scale HINs.
Existing random walk procedure aims to enumerate the meta-
paths within fixed length l, then perform jointly inference.
Since the time complexity of enumerating the meta-paths
grows exponentially with the length of the meta-paths, tradi-
tional random walk will be impractical in large scale network.
Besides, another potential problem is that the inference per-
formance is very sensitive to l. For example, if l is small (e.g.,
equals to 1), multi-order and meaningful meta-paths will be
ignored; if l is large, meaningless and duplicated meta-paths
will be generated. Meng et al. [Meng et al., 2015] recently
provides a more general inference framework by requiring
users to provide example entity pairs. Then the meta-paths
are generated if relevant meta-paths are of high proximity be-
tween the entity pairs. The algorithm generates meta-paths
by considering the local (near) randomly generated negative
examples. Since there could be some positive examples intro-
duced by randomly producing negative samples, the negative
examples could be biased and noisy towards the inference.
We thus consider an efficient HIN inference framework that
could further release the power of meta-path based inference
in large scale HINs by relieving the aforementioned issues.
In this paper, we propose a meta-path constrained inference
framework for large scale HIN. Since for a particular HIN in-
ference task, we may only care about a (or a set of) relevant
inference rules, then the parts of the network only distantly
connected to inference goals are likely to have a small in-
fluence. Intuitively, we consider the HIN inference as graph
random walk inference with constraints. We carefully design
an efficiency-optimized meta-path tree data structure to con-
strain the random walk to follow the tree structure. Compared
to the running time of the original random walk, the proposed
meta-path constrained random walk can improve efficiency
by mostly two orders of magnitude. The proposed inference
scheme is initialized with several sample pairs to capture the
inference target. Then based on the meta-path based prox-
imities in the samples, the scheme iterates random walk in-
ference in the meta-path tree until it becomes convergence.
Notice that the inference method is weakly supervised. The
supervision information complies with the intuition that the
inference is only relevant to parts of the HINs.
below:
which is important and has broad applications.
The main contributions of this paper can be highlighted as
• We study the problem of large scale HIN inference,
• We propose a meta-path constrained inference frame-
work, where many of the HIN tasks can be unified under
this framework. In particular, we propose a meta-path
tree constrained random walk inference method, which
is weakly supervised and efficient to model the inference
targets.
• We conduct experiments on two large scale HIN. Our
proposed inference method has demonstrated its effec-
tiveness and efficiency compared to the state-of-the-arts
inference methods on typical HIN tasks (link prediction
and similarity search).
The rest of this paper is organized as below. We first in-
troduce the HIN inference framework in Section 2. Next in
Section 3, we present HIN inference as meta-path tree con-
strained random walk inference. The experimental results are
shown in Section 4. We finally discuss the related work and
conclude in Section 5 and Section 6 respectively.
2 Problem Definition
In this section, we introduce a general HIN inference frame-
work that could specifically take meta-paths in HIN into con-
2
sideration, and perform large scale inference according par-
ticular task. Before that, some basic concepts of HIN are in-
troduced as below.
2.1 Heterogeneous Information Network
Definition 1 A heterogeneous information network (HIN) is
a graph G = (V,E) with an entity type mapping φ: V → A
and a relation type mapping ψ: E → R, where V denotes the
entity set, E denotes the link set, A denotes the entity type set,
and R denotes the relation type set, and the number of entity
types A > 1 or the number of relation types R > 1.
Notice that, in large scale HINs, such as YAGO, the rela-
tion type mapping is an one-to-one mapping, while the entity
type mapping could be one-to-N mapping. For example, a
specific triplet (Larry Page, alumniOf, Stanford) in YAGO,
the relation type mapping ψ(< LarryP age, Stanf ord >
) = {alumniOf }, while the entity mapping φ(Stanf ord) =
{Organization, University}. The reason why an entity
could be mapping to multiple entity types in YAGO or Free-
base is that, the entities types are often organized in a hierar-
chical manner. For example, as shown in Figure 1, University
is a subtype of Organization, Politician is a subtype of Per-
son. All the types or attributes share a common root, called
Object. The hierarchy of the entity organization raises an-
other challenge on how to infer useful information in HINs.
Figure 1: Hierarchy of entity types.
Definition 2 Given an HIN G = (V,E) with the entity type
mapping φ: V → A and the relation type mapping ψ:
E → R, the network schema for network G, denoted as
TG = (A,R), is a graph with nodes as entity types from A
and edges as relation types from R.
The network schema provides a high-level description of
a given heterogeneous information network. Another impor-
tant concept, meta-path [Sun et al., 2011b], is proposed to
systematically define relations between entities at the schema
level.
Definition 3 A meta-path P is a path defined on the graph of
network schema TG = (A,R), and is denoted in the form of
R2−−→ . . . RL−−→ AL+1 , which defines a composite
A1
relation R = R1 · R2 · . . . · RL between types A1 and AL+1,
where · denotes relation composition operator, and L is the
length of P. When L = 1, we specifically call it as first-order
meta-path; when L > 1, we call it as multi-order meta-path.
We say a path p = (v1 − v2 − . . . − vL+1) between v1 and
vL+1 in network G follows the meta-path P, if ∀l, φ(vl) = Al
R1−−→ A2
l
and each edge el = (cid:104)vl, vl+1(cid:105) belongs to each relation type
Rl in P. We call these paths as path instances of P, denoted
as p ∈ P. R−1
represents the reverse order of relation Rl.
For example, in the YAGO network, the composite relation
two Person co-founded an Organization can be described as
meta-path P = Person found−−−→ Organization found−1
−−−−−→ Person. A
path instance of P is p = Larry Page found−−−→ Google found−1
−−−−−→
Sergey Brin.
2.2 Heterogeneous Information Network Inference
Framework
Meta-paths carry rich information about the semantic rela-
tionships between entities, thus capture subtle proximities in
HINs via several meta-path based similarity measures, such
as Path Count [Sun et al., 2011a], Random Walk [Lao et al.,
2011a], and Pathsim [Sun et al., 2011a]. The proximities
are very important and useful for inference problems, such
as link prediction. Assume that direct links (first-order meta-
paths) are missing between two entities with types Person and
−−−−→
Profession, if there is a multi-order meta-path Person workAt
Organization employ−−−−→ Profession between the entities, and the
meta-path based similarity (e.g., Path Count) of two entities
is large (i.e., number of path instances between two entities
satisfying the meta-path is large), we will have higher confi-
dence in inferring the missing link between the entities based
on the multi-order meta-path.
Traditional
inference framework in machine learn-
ing [Chang et al., 2012] aims to model relevant inference
problems as stochastic processes involving both output vari-
ables and input or observed variables. The framework mainly
includes a model parameter vector θ, corresponding to a
set of representation or feature functions {Φ}. For an in-
put instance x and an output assignment y, the "score" of
the instance can be expressed as a model function f with
the parameter vector and representation functions: score =
f (θ, Φ(x, y)). When the model is evaluated on test instance
x, the inference framework aims to find the best assignment
to the output variables,
y∗ = arg max
y
f (θ, Φ(x, y))
(1)
Notice that a representation function φ ∈ Φ usually focuses
on producing homogeneous flat features and ignores the link
or structure information in both input variables x and output
variables y. As we know, the multi-typed links (meta-paths)
is very important for HIN inference. However, the framework
doesn't model meta-path information; besides it is not trivial
to efficiently incorporate the extra information provided by
meta-paths into the framework. We therefore formally define
the HIN inference framework as below.
Definition 4 Heterogeneous Information Network Infer-
ence (HINI) aims to enable inference with meta-paths in
large scale HINs. To be more specific, HINI aims to infer the
best assignment to the output variables given an HIN model
with meta-paths. HINI is formalized as:
y∗ = arg max
y
f (θ, Φ(x, P, y))
(2)
3
where f is the model or score function and θ is the model pa-
rameters. Different from Eq. 1, Φ aims to leverage the prox-
imities carried by the meta-paths P regarding to the input
instance x or output variable y.
HINI (Eq. 2) could support many mining tasks in HINs,
such as link prediction and similarity search in HINs, as we
will see later. If meta-paths set P is empty. HINI will de-
generate to traditional inference as shown in Eq. 1. We find
there are mainly two challenges in HINI: 1) how to efficiently
generate useful meta-paths P from large-scale HINs consist-
ing of millions of entities and billions of relations? And 2)
how to model the meta-path based proximities to improve the
representation power of Φ? In next section, we will describe
our proposed inference method that could efficiently gener-
ate meta-paths as well as compute meta-path based similarity
simultaneously.
3 HIN Random Walk Inference with
Meta-Path Dependency Tree Search
In this section, we first introduce meta-path constrained HIN
random walk inference with weak supervision,
then talk
about how to leverage the supervision to conduct efficient
HIN random walk inference via a carefully designed data
structure.
3.1 Weakly Supervised HIN Random Walk
Inference
As aforementioned, meta-paths are very important since they
infer important semantic relationships between entities in
HINs, thus capture semantic proximities of entities, which
is very useful for HIN inference as shown in Eq. 2. Most
of existing inference methods are focusing on enumerating
the meta-paths within a fixed length in the full underlying
network [Lao et al., 2011a]. However, this solution is im-
practical in large scale HINs, since it has been proven that
the number of possible meta-paths grows exponentially with
the length of a meta-path. As we find, for a particular infer-
ence task, it is not necessary to do inference in the full HIN,
since only a part of the network or meta-paths relevant to the
inference. To copy with these challenges, we propose a meta-
path constrained random walk method to infer with weak hu-
man supervision in HINs. Weak supervision here provides
guidance to the random walk process, together with the meta-
path generation process by pruning the searching space. Our
method aims to copy with the two HINI challenges in the pre-
vious section.
Given an HIN G = (V,E), similar to [Meng et al.,
2015], we ask users to provide example entity pairs Λ =
{(si, ti)i = 1,··· , N} as supervision to imply meta-paths.
Formally, we are aiming to find a meta-paths set PΛ =
{Pii = 1,··· , M} that could infer high proximities between
entities in Λ. An efficient way to generate PΛ given Λ will
be described in next section.
Now assume when we have Λ, we also get PΛ. For a meta-
path P = A1
Ri−−→ . . . RL−−→ AL+1 ∈
PΛ, we define the following meta-path constrained random
walk starting from s and reaching at s following only path in-
stances p ∈ P. It defines a distribution f (s, tP) recursively
R2−−→ . . . Ai
R1−−→ A2
as below.
f (s, tP1,··· ,L) =
f (s, t
(cid:48)P1,··· ,L−1)
(cid:88)
t(cid:48)∈χ(Ai+1ei;Ri)
·
χ(Ai+1ei; Ri) ,
s.t. f (s, tP) = 1, if s = t; otherwise, f (s, tP) = 0,
1
(3)
1
where χ(Ai+1ei; Ri) indicates the entity set where each entity ei
can be linked via relation type Ri to at least one entity with type
Ai+1.
χ(Ai+1ei;Ri) means a one-step random walk starting from
an entity t(cid:48) ∈ χ(Ai+1ei; Ri) via relation type Ri.
For example, consider a path instance from s to t following a
−−−−→ Organization employ−−−−→ Profession. Sup-
meta-path Person workAt
pose a random walk starts at an entity s (e.g., s=Larry Page). If Org
is the set of Organizations in the HIN that Larry Page has worked
at, after one step, the walker will have probability
1Org of being at
any entity e ∈ Org (e.g., e= Google). Similarly, if Pro is the set of
Professions in the HIN that Google has employed, the walker will
Org·Pro being at any entity t ∈ Pro (e.g., t
have probability of
= Computer Science). It is useful that proximity provided by meta-
path constrained random walk infers the prior probability of t being
the Profession for Person s.
To be more general, we then propose a linear model to com-
bine the single meta-path constrained random walk scores. i.e., for
each meta-path P ∈ PΛ, the HIN inference with joint random walk
model (HINI-JRW) is formalized as below.
1
f (s, tPΛ; ) = θ1f (s, tP1) + θ2f (s, tP2) + ···
+θif (s, tPi) + ··· + θM f (s, tPM ),
(4)
where is the model parameter vector, each element means the weight
or importance of certain meta-path based inference score. The pa-
rameter vector can either be explicitly set by users or implicitly
learned according different HIN inference tasks. By tuning the pa-
rameters, we can avoid the bias induced by certain meta-path(s) and
ensure the model's robustness and stability.
Now let's revisit the relationship between the joint HIN inference
model (Eq. 4) and HINI framework (Eq. 2). In short, the input in-
stance of HINI x = [s, t], the output variable of HINI y is the ran-
dom walk based probabilities or proximities, and meta-paths set P
of HINI equals to PΛ. Similarly, other inference models can also be
unified into HINI framework.
3.2 Efficient Inference via Meta-Path Dependency
Tree Search
To copy with the main challenge of large scale HIN inference, i.e.,
to do efficient inference, we carefully design a tree structure that sig-
nificantly accelerate the above meta-path constrained random walk
inference in HINs. More importantly, the new tree structure enables
doing inference through model f in Eq. 4 and automatically gen-
erate meta-path based inference rules PΛ given the user provided
samples Λ simultaneously.
Even we do not need to enumerate the meta-paths in an HIN, it
is still intractable to find the optimal meta-path set PΛ given the
examples pairs in Λ [Van Horn and Martinez, 1994]. Selecting rel-
evant meta-paths is NP-hard even when some of the path instances
are given. It has been shown as an NP-hard problem [Amaldi and
Kann, 1998; Van Horn and Martinez, 1994]. Inspired by forward
selection algorithms [Koller and Sahami, 1996], we introduce meta-
path dependency tree search algorithm, and enable efficient infer-
ence of HINI-JRW via doing search in the meta-path dependency
tree. Since we do not know the relevant paths beforehand, thus PΛ
4
is empty. The meta-path dependency tree search algorithm is thus
used to insert meta-paths into PΛ and compute the inference score
while searching. There are three differences compared to the method
proposed in [Meng et al., 2015]: 1) our tree search algorithm doesn't
leverage random generated negative examples, which could induce
biased inference; 2) our algorithm doesn't require iterative process
of the algorithm to guide the tree search, which is more efficient; and
3) we do not set hard threshold to terminate the algorithm, which is
very sensitive and could prevent the algorithm from generating more
meaningful meta-paths.
Figure 2: Meta-path dependency tree node structure.
Let's first introduce the meta-path dependency tree structure.
Each tree edge is annotated with an relation type, and each tree node
represents a list of entities pairs with their HINI-JRW scores and a
priority score S. The node structure is shown in Figure 2. The node
stores tuple in the form < (s, t), f (s, tPi) >, where (s, t) repre-
sents a path instance in the graph by its starting and current graph
entities, respectively; Pi is the meta-path starting from s to t. The
edges of the tree are relation types in the HIN. f (s, tPi) is the RW
score for this meta-path according to Eq. 3. The priority score S
determines which is the next tree node to search. We compute the
value of S as follows:
(cid:80)
(cid:80)
u
u
max(ru)·(cid:80)
(cid:80)
max(ru)·(cid:80)
(cid:80)
c(u)
v max(ru)
c(u)
v max(ru)
v f (u,vP)
v f (u,vP)
S =
· βL
(cid:64)(u, v) ∈ Λ
· βL + 1 ∃(u, v) ∈ Λ
(5)
where (u, v) is the entity pairs in the tree node, max(ru) is the
maximum weight among entity pairs starting with u, and c(u) is
the number of example pairs starting with u. From Eq. 5, we can
see that the priority score S is defined as a weighed combination of
random walk score referred as Eq. 3 of the given entity pair in the
tree node. This means that if the entity pair in the tree node exhibit
higher proximity along the corresponding meta-path, they will have
higher probability to be more similar along the new meta-paths gen-
erated through the searching process. This is the reason why the
search algorithm will be introduced soon guide its search as random
walker to the tree node with larger S. Because of this, the priority
score proposed in [Meng et al., 2015] will have the possibility to
generate meta-paths that do not contain example entity pairs in Λ.
This will guide search to find leaf tree node with the largest S among
all the leaf nodes, which will lead to no meta-path generated during
the search, thus suffer from relatively low efficiency. To avoid this,
considering the value of S is normally smaller than one, this slight
modification will ensure the target node t could be reached through
the search, and generate the meta-paths in a more efficient way. In
addition, in order to avoid generating meta-path with infinite length,
we follow [Meng et al., 2015] to add a decay factor β ranging from
0 to 1.
Through searching the meta-path dependency tree, inference and
meta-path generation in HINs can be done at the same time. We
present the details of our proposed meta-path dependency tree search
(MPDTS) in Algorithm 1. In short, we search the tree by moving to
out-neighbor nodes on the graph until a meta-path can be found or
the graph is completely traversed. We first target the tree node with
weights r, Meta-Path Dependency Tree T
corresponding meta-path P
largest priority and examine whether its tuples are example entity
pairs.
If so, we store random walk scores computed by Eq. 3 in
the corresponding tree node, and add the meta-paths into PΛ. If
no example pairs are encountered, then we extend each entity pair
by moving to an out-neighbour. We insert this new pair with its
random walk score to a child node, and compute its priority score.
Notice that, the generated meta-path only contains relation types,
similar to [Meng et al., 2015], we also leverage the Lowest Common
Ancestor (LCS) of type hierarchy of the entities in the HIN to fill
the entity types in the generated meta-path P, to form the complete
meta-paths.
Algorithm 1 MPDTS(G, Λ,T )
Require: an HIN G(V,E), example pairs Λ, example pairs
Ensure: Meta-path P, random walk scores vector y of the
1: while T is not fully traversed do
2: Q ← the largest S score node in T 's child node set L;
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
end if
end for
if y is not empty then
P ← the path from root node to Q;
Remove Q from L;
break;
for Each entity pair (u, v) ∈ Q do
R ←relation type of edge from v to w;
if There is no edge of type R from Q to child
node Z then
y ← {0, . . . , 0};
for Each entity pair (u, v) ∈ Q do
y(u, v) ← f (u, vP) computed based on Eq. 3;
for Each out-neighbour w of v in G do
if (u, v) ∈ Λ then
else
18:
19:
20:
21:
22:
23:
Create child node Z for Q;
Insert child node Z into L;
end if
P ←the path from root node to Z;
Insert a new tuple (cid:104)(u, v), f (u, vP)(cid:105) into Z,
where f (u, vP) is computed based on Eq. 3;
Compute the priority scoreS of Z based on
Eq. 5;
Remove Q from L;
end for
end for
24:
25:
26:
27:
28: end while
29: return P, y
end if
Finally, we obtain the joint random walk scores/probabilities
along the meta-paths PΛ according to HINI-JRW (Eq. 4). The joint
version of MPDTS (Algorithm 1) is shown as JMPDTS in Algo-
rithm 2. We can simply obtain the joint random walk score for cer-
tain entity pair (s, t) ∈ Λ by summing over the rows of Y, since
each entry in the row means a random walk score from s to t fol-
lowing one meta-path. By doing so, we have done inference and
automatic meta-path generation in large scale HINs, which copies
with the two challenges of HINI framework.
There are several advantages for using the tree structure for ran-
5
Algorithm 2 JMPDTS(G, Λ, r)
Require: an HIN G(V,E), example pairs Λ, example pairs
weights r
matrix Y
Ensure: Meta-path set PΛ with size M, random walk scores
1: m ← 0; PΛ ← ∅; Y ← ∅
2: while True do
3:
4:
5:
6:
7:
8:
9: end while
10: return PΛ, Y
if T is not fully traversed then
Pm, ym ← MPDTS(G, Λ, r);
PΛ ← PΛ
m ← m + 1;
Y ← Y(cid:83) ym;
(cid:83)Pm;
end if
dom walk inference we proposed. Firstly, during the search process
of the tree, the node to expand is selected by applying supervision
provided by user examples and the search space can be reduced.
Secondly,during the process, path instances that represent the same
meta-paths are gathered in a single tree node, yet avoid duplicate
calculation. Thirdly, the whole meta-path dependency tree structure
is preserved in the memory and can be reused in the whole iterative
process while traversing the tree.
4 Experiments
In this section, we evaluate our approach using two typical HIN
tasks, link prediction and similarity search.
4.1 HINI for Link Prediction
In this subsection, we validate our algorithm's efficiency and effec-
tiveness by performing link prediction tasks. We chose link pre-
diction for the evaluation because it provides quantitative way to
measure the performance of different methods.
Task Description
For link predication, we propose to use logistic regression model
to leverage the random walk score of each meta-path [Lao et al.,
2011a]. Formally, Given a relation R and a set of entity pairs
{(si, ti)}, we can construct a training dataset D = (xi, ri), where
xi is a vector of all the meta-path based features for the pair
(si, ti) -- i.e., the j-th component of x is f (si, tiPj), and ri indi-
cates whether R(si, ti) is true. Parameter θ is estimated by max-
imizing the regularized objective function proposed in [Lao et al.,
2011a]. Then the link predication model is defined as below:
p(ri = 1xi; θ) =
exp(θT xi)
1 + exp(θT xi)
.
(6)
Dataset
We perform link prediction experiments on a representative dataset
for large scale HIN: YAGO.
YAGO: YAGO1 is a semantic knowledge base, derived from
Wikipedia, WordNet and GeoNames. Currently, YAGO2 has knowl-
edge of more than 10 million entities (like persons, organizations,
cities, etc.) and contains more than 120 million facts about these en-
tities. It contains 350,000 entity types organized in type hierarchy,
and 100 relation types.
1http://www.mpi-inf.mpg.de/departments/
databases-and-information-systems/research/
yago-naga/yago/
Efficiency Study
Figure 3 presents the running time of our method compared to
PCRW models with fixed length, and when varying the number of
example pairs given as input. It can be observed that generally, the
algorithm running time increases sub-linearly in the number of ex-
ample pairs. The increase is due to the PCRW random walks which
need to be performed concurrently for each example pair, but the
number of meta-paths in the model does not increase at the same
rate.
In particular, the algorithm performs better than models of
paths longer than 2 by a factor of up to two orders of magnitude.
However, the models of short path length have limited predictive
power, despite their better running time. In comparison, our algo-
rithm is capable of generating long and meaning meta-paths and
performing efficient inference.
Effectiveness Study
For a certain type of link in YAGO, for instance citizenOf, we re-
move all such links and try to predict them with the above logistic
regression model that leverages the random walk scores as features.
We randomly select a number of pairs of entity pairs with the ac-
cording relation labels as training data, and validate the model using
a test set of an equal number of pairs. We compared our link pre-
diction model with PCRW [Lao et al., 2011a] based models which
generate paths of finite length in 1,2,3,4. The PCRW models also
use the logistic regression model to combine these meta-paths. Be-
sides, we also compare our model with the state-of-the-art FSPG
based prediction model in [Meng et al., 2015]. Following [Meng et
al., 2015], we set β as 0.6 in meta-path dependency tree to avoid
producing meta-paths with infinite length. We use Area Under ROC
Curve (AUC) as the evaluation measure. AUC calculate the area
under Receiver Operating Characteristics (ROC) curve. The x-axis
of ROC is false positive rate, and the y-axis of ROC represents true
positive rate. Thus a large AUC value, a large accuracy in predica-
tion.
Table 1 presents the results for link prediction for three types of
links: citizenOf andadvisorOf in YAGO. For each of these links, we
generated 100 training and 100 test pairs, as described above. The
result shows that fixed-length PCRW suffers from several issues.
When the maximum length is too small (1 or 2), meta-paths can-
not connect example pairs, and as such the model is not better than
a random guess and the model will have low recall. When the max-
imum length is too big, the model introduces too many meta-paths.
For length 3, there are 135 meta-paths, and over 2,000 for length 4.
Notice that our method outperforms FSPG, the reason is that FSPG
incorporates randomly generated negative examples which may lead
to biased inference, while our method does not use.
Table 1: Comparison of link prediction performance (AUC)
Advisor
Our method
FSPG
PCRW-1
PCRW-2
PCRW-3
PCRW-4
citizenOf
0.854
0.822
0.594
0.752
0.567
0.525
advisorOf
0.654
0.647
0.498
0.613
0.545
0.569
Figure 3: Running time of our method (MPG)
4.2 HINI for Similarity Search
In this subsection, we mainly focus on user study of our HINI frame-
work via a typical HIN task, i.e., similarity search, to get a better
understanding of why our method is effective.
Our model is clearly better in predicting the links, and it generates
only a limited number of meta-paths. For instance, the advisorOf
link in YAGO has a model of only 13 meta-paths. Moreover, these
meta-paths are highly relevant and serve as good explanation of the
proximity links. For example, we find the meta-path illustrating the
fact that a person is a strong influencer to another person is the best
predictor of advisorOf, but other, longer, paths are also highly rele-
vant. Compared with PCRW with maximum length 2, it has higher
recall because it also detected longer important meta-paths, for in-
stance, Person win−−→ Award win−1−−−−→ Person. Knowledge bases, such
as YAGO, often suffer from incompletion problem. We thus can use
these multi-order meta-paths as inference rules to predict the direct
links between entities.
Whereas considering direct links is widely done when querying
large scale networks, multi-order meta-paths can improve query re-
sult accuracy. For instance, in advisorOf prediction, only one type of
direct links between people and country exists, namely, influencer.
By leveraging multi-order meta-paths, it will provide us a better
chance to still perform good link prediction when the direct links
are missing. The result shows the power of our HINI framework in
doing large scale inference in HINs.
Task Description
In general, similarity search aims to find similar entities given a
query entity in HINs. Formally, similarity search can be regarded
as random walk, but implemented by commuting matrix manipula-
tion for a meta-path, to generate the similarity score of each targeting
entity. The community matrix is defined as below.
Definition 5 Commuting matrix. Given a network G = (V,E)
and its network schema TG, a commuting matrix MP for a
meta-path P = (A1 − A2 − . . . − AL+1 ) is defined as MP =
WA1A2WA2A3 . . . WALAL+1, where WAiAj is the adjacency
matrix between types Ai and Aj. MP (i, j) represents the num-
ber of path instances between objects xi and yj, where φ(xi) = A1
and φ(yj) = AL+1, under meta-path P.
Given f (s, tP), we can infer:
(cid:88)
MP =
θi · MPi .
(7)
i
Thus in general, if the entry of MP is large, the similarity between
two entities based on meta-path P will be large.
6
Table 2: Meta-path examples generated based on example pairs
Venue publishIn−1
−−−−−−→Paper authorOf−1
−−−−−−→Author authorOf
−−−−→Paper publishIn
−−−−−→Venue
Meta-paths
Venue publishIn−1
Venue publishIn−1
−−−−−−→Paper cite−−→Paper publishIn
−−−−−−→Paper cite−1−−−→Paper publishIn
−−−−−→Venue
−−−−−→Venue
Table 3: Case study on similarity search results
Ranking (Query)
1
2
3
4
5
6
7
8
9
10
KDD
ICML
SIGMOD
ICDM
CIKM
VLDB
SIGIR
WWW
TKDE
ICDE
Machine Learning
Computational Linguistics
ACL
COLING
NAACL
EACL
LREC
EMNLP
SIGIR
IJCAI
AAAI
INTERSPEECH
VLDB
SIGMOD
ICDE
TODS
TKDE
PODS
EDBT
PVLDB
CIKM
IS
VLDB Journal
Dataset
Most of the similarity search research in HINs are evaluated on
DBLP dataset, we also use it for our experiments. DBLP is a bibli-
ographic information network which is frequently used in the study
of heterogeneous networks. We use a subset of DBLP used in [Sun
et al., 2011a] containing scientific papers in four areas: databases,
data mining, artificial intelligence, and information retrieval. The
dataset has four classes of nodes: Paper, Author, Topic, and Venue.
It also has four edge types between different entity types. In totally,
the subset contains 14,376 papers, 14,475 authors, 8,920 topics, and
20 venues. Besides, there are 170,794 links.
Case Study of Similarity Search Results
We select ten groups of similar venues according to different areas
in DBLP, and use them as input example pairs to generate the meta-
paths. Then we compute the commuting matrices for the meta-paths.
When given a query venue, we rank the venues based on the scores
in the row of certain commuting matrix for meta-path. By doing
this, we could find the most similar venues to the query.
In Table 2 and Table 3, we show the generated meta-paths, and
the search results based on the meta-paths. From the results, we
find that 1) meta-paths carry rich proximity information between
the given pairs; and 2) all the search results are very relevant to the
query. This shows the insight of why our HIN inference framework
is able to handle automatic meta-path generation and effective infer-
ence. By providing these two HIN tasks, we can conclude that the
proposed inference framework is not limited to use only in these two
tasks, it will be effective in other HIN tasks, such as recommenda-
tion, classification, etc.
5 Related Work
5.1 Knowledge Base/Network Inference
Although there is a great deal of recent research on extracting knowl-
edge from text [Agichtein and Gravano, 2000; Etzioni et al., 2005;
Snow et al., 2006; Pantel and Pennacchiotti, 2006; Banko et al.,
2007; Yates et al., 2007], much less progress has been made on
the problem of drawing reliable inferences from this imperfectly
extracted knowledge.
In particular, traditional logical inference
methods are too brittle to be used to make complex inferences
from automatically-extracted knowledge, and probabilistic infer-
ence methods [Richardson and Domingos, 2006] suffer from scal-
ability problems, since they cannot generate inference rules directly
and use the full rule set to perform inference. Recently, Ni Lao
et al. [Lao et al., 2011a] provides a probabilistic way to perform
random walk inference, however, it is still costful since it needs to
enumerate the network to generate the meta-paths. However, we
carefully design a tree structure which would efficiently generate
the meta-paths.
5.2 Meta-Path Generation
The first approach regarding to automatic meta-path generation is
proposed by [Lao et al., 2011a]. Their solution enumerate all the
meta-paths within a fixed length L. However, it is not clear how
L should be set. More importantly, L can significantly affect the
meta-paths generated: (a) if L is large, then many redundant meta-
paths may be returned, leading to curse-of dimensionality effects;
and (b) if L is small, important meta-paths with length larger than
L might be missed. Our experiments have shown that the run-
ning time of the meta-path generation process grows exponentially
with length L. Moreover, the accuracy can also drop with increase
in L. Recent studies also propose efficient meta-path generation
algorithms based on localized random walk [Wang et al., 2015a;
Wang et al., 2015b; Wang et al., 2016a; Wang et al., 2016b]. Re-
cently, Meng et al. [Meng et al., 2015] propose to leverage positive
and negative examples to generate meta-paths. However, our so-
lution does not leverage the negative examples, which is randomly
generated and would lead to biased inference.
5.3 Link Prediction
Since the main focus of our work is to generate meta-paths, we only
use link prediction to quantify our advantage compared with the ex-
isting meta-path generation methods. Compared with [Nickel et al.,
2012], our method predicts the relationship between different en-
[Nickel et al.,
tities, rather than predicting the types of entities.
2011] use the factorization of a three-way tensor to perform rela-
tional learning; it only considered simple node types and is not ap-
plicable to our experiment which has both complex node classes and
7
edge types. Besides, our method is aiming to support efficient link
prediction via HINI with random walk.
5.4 Similarity Search
Similarity measures have been a hot research topic for years. They
can be categorized broadly into two types: entity similarity mea-
sures and relation similarity measures. Similarity measures, such as
SimRank [Jeh and Widom, 2002], P-Rank [Zhao et al., 2009], Path-
Sim [Sun et al., 2011a], PCRW [Lao et al., 2011a] and RoleSim [Jin
et al., 2011] capture entity similarity. Recent studies also focus
on similarity search in schema-rich networks [Wang et al., 2015c;
Wang et al., 2016c; Wang et al., 2017c; Wang et al., 2018]. Recent
studies on entity similarity also find rules/meta-paths very useful.
Path ranking algorithm [Lao et al., 2011a], rule mining [Gal´arraga
et al., 2013] and meta-path generation [Meng et al., 2015] have
demonstrated the effectiveness of using the mined rules or meta-
paths for link prediction-like tasks based on entity similarity, while
our work is for leveraging the power of efficient HINI random walk
inference method to perform similarity search.
6 Conclusion
In this paper, we study the problem of large scale HIN inference,
which is important and has broad applications (e.g., link predica-
tion, similarity search). We propose a meta-path constrained in-
ference framework, where many of the existing inference methods
can be unified under this framework.
In particular, we propose a
efficiency-optimized meta-path tree constrained random walk in-
ference method for HIN inference, which is weakly supervised to
model the inference target and is approximated in time independent
of the network size. We conduct experiments on two large scale
HIN, and our proposed inference method has demonstrated its ef-
fectiveness and efficiency compared to the state-of-the-arts inference
methods on typical HIN tasks (link prediction and similarity search).
The effectiveness of the inference method is not limited to the two
tasks, in the future, we plan to apply our method to more real world
applications (e.g., NLP tasks [Wang et al., 2013; Wang et al., 2017b;
Wang et al., 2017a; Wang et al., 2019]).
References
[Agichtein and Gravano, 2000] Eugene Agichtein and Luis Gra-
vano. Snowball: Extracting relations from large plain-text col-
lections. In Proceedings of the fifth ACM conference on Digital
libraries, pages 85 -- 94, 2000.
[Amaldi and Kann, 1998] Edoardo Amaldi and Viggo Kann. On
the approximability of minimizing nonzero variables or unsatis-
fied relations in linear systems. Theoretical Computer Science,
209(1):237 -- 260, 1998.
[Banko et al., 2007] Michele Banko, Michael J Cafarella, Stephen
Soderland, Matthew Broadhead, and Oren Etzioni. Open infor-
mation extraction from the web. In IJCAI, volume 7, pages 2670 --
2676, 2007.
[Chang et al., 2012] Ming-Wei Chang, Lev-Arie Ratinov, and Dan
Roth. Structured learning with constrained conditional models.
Machine Learning, 88(3):399 -- 431, 2012.
[Etzioni et al., 2005] Oren Etzioni, Michael Cafarella, Doug
Downey, Ana-Maria Popescu, Tal Shaked, Stephen Soderland,
Daniel S Weld, and Alexander Yates. Unsupervised named-entity
extraction from the web: An experimental study. Artificial intel-
ligence, 165(1):91 -- 134, 2005.
[Gal´arraga et al., 2013] Luis Antonio Gal´arraga,
Christina
Teflioudi, Katja Hose, and Fabian Suchanek. Amie: asso-
ciation rule mining under incomplete evidence in ontological
knowledge bases. In WWW, pages 413 -- 422, 2013.
[Jeh and Widom, 2002] Glen Jeh and Jennifer Widom. Simrank: a
measure of structural-context similarity. In KDD, pages 538 -- 543,
2002.
[Ji et al., 2010] Ming Ji, Yizhou Sun, Marina Danilevsky, Jiawei
Han, and Jing Gao. Graph regularized transductive classifica-
In ECML, pages
tion on heterogeneous information networks.
570 -- 586, 2010.
[Jin et al., 2011] Ruoming Jin, Victor E Lee, and Hui Hong. Ax-
iomatic ranking of network role similarity. In KDD, pages 922 --
930, 2011.
[Koller and Sahami, 1996] Daphne Koller and Mehran Sahami. To-
ward optimal feature selection. 1996.
[Lao et al., 2011a] Ni Lao, Tom Mitchell, and William W Cohen.
Random walk inference and learning in a large scale knowledge
base. In EMNLP, pages 529 -- 539, 2011.
[Lao et al., 2011b] Ni Lao, Tom Mitchell, and William W Cohen.
Random walk inference and learning in a large scale knowledge
base. In EMNLP, pages 529 -- 539, 2011.
[Lu and Zhou, 2011] Linyuan Lu and Tao Zhou. Link prediction in
complex networks: A survey. Physica A: Statistical Mechanics
and its Applications, 390(6):1150 -- 1170, 2011.
[Meng et al., 2015] Changping Meng, Reynold Cheng, Silviu Ma-
niu, Pierre Senellart, and Wangda Zhang. Discovering meta-paths
in large heterogeneous information networks. In Proceedings of
the 24th International Conference on World Wide Web, pages
754 -- 764. International World Wide Web Conferences Steering
Committee, 2015.
[Nickel et al., 2011] Maximilian Nickel, Volker Tresp, and Hans-
Peter Kriegel. A three-way model for collective learning on
In Proceedings of the 28th international
multi-relational data.
conference on machine learning (ICML-11), pages 809 -- 816,
2011.
[Nickel et al., 2012] Maximilian Nickel, Volker Tresp, and Hans-
Peter Kriegel. Factorizing yago: scalable machine learning for
linked data. In Proceedings of the 21st international conference
on World Wide Web, pages 271 -- 280. ACM, 2012.
[Pantel and Pennacchiotti, 2006] Patrick Pantel and Marco Pennac-
chiotti. Espresso: Leveraging generic patterns for automatically
harvesting semantic relations. In Proceedings of the 21st Inter-
national Conference on Computational Linguistics and the 44th
annual meeting of the Association for Computational Linguis-
tics, pages 113 -- 120. Association for Computational Linguistics,
2006.
[Richardson and Domingos, 2006] Matthew Richardson and Pedro
Domingos. Markov logic networks. Machine learning, 62(1-
2):107 -- 136, 2006.
[Snow et al., 2006] Rion Snow, Daniel Jurafsky, and Andrew Y Ng.
Semantic taxonomy induction from heterogenous evidence.
In
Proceedings of the 21st International Conference on Computa-
tional Linguistics and the 44th annual meeting of the Association
for Computational Linguistics, pages 801 -- 808. Association for
Computational Linguistics, 2006.
[Sun et al., 2011a] Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S
Yu, and Tianyi Wu. Pathsim: Meta path-based top-k similarity
search in heterogeneous information networks. VLDB, 2011.
8
sim: Relation similarity search in schema-rich heterogeneous in-
formation networks. In SDM, pages 621 -- 629, 2016.
[Wang et al., 2017a] Chenguang Wang, Alan Akbik, Laura Chiti-
cariu, Yunyao Li, Fei Xia, and Anbang Xu. CROWD-IN-THE-
LOOP: A hybrid approach for annotating semantic roles.
In
EMNLP, pages 1913 -- 1922, 2017.
[Wang et al., 2017b] Chenguang Wang, Laura Chiticariu, and Yun-
yao Li. Active learning for black-box semantic role labeling with
neural factors. In IJCAI, pages 2908 -- 2914, 2017.
[Wang et al., 2017c] Chenguang Wang, Yangqiu Song, Haoran Li,
Yizhou Sun, Ming Zhang, and Jiawei Han. Distant meta-path
similarities for text-based heterogeneous information networks.
In CIKM, pages 1629 -- 1638, 2017.
[Wang et al., 2018] Chenguang Wang, Yangqiu Song, Haoran Li,
Ming Zhang, and Jiawei Han. Unsupervised meta-path selection
for text similarity measure based on heterogeneous information
networks. Data Min. Knowl. Discov., 32(6):1735 -- 1767, 2018.
[Wang et al., 2019] Chenguang Wang, Mu Li, and Alexander J.
CoRR,
Language models with transformers.
Smola.
abs/1904.09408, 2019.
[Yates et al., 2007] Alexander Yates, Michael Cafarella, Michele
Banko, Oren Etzioni, Matthew Broadhead, and Stephen Soder-
land. Textrunner: open information extraction on the web. In Pro-
ceedings of Human Language Technologies: The Annual Confer-
ence of the North American Chapter of the Association for Com-
putational Linguistics: Demonstrations, pages 25 -- 26. Associa-
tion for Computational Linguistics, 2007.
[Zhao et al., 2009] Peixiang Zhao, Jiawei Han, and Yizhou Sun. P-
rank: a comprehensive structural similarity measure over infor-
mation networks. In CIKM, pages 553 -- 562, 2009.
[Sun et al., 2011b] 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, pages
992 -- 1003, 2011.
[Taskar et al., 2003] Ben Taskar, Ming-Fai Wong, Pieter Abbeel,
and Daphne Koller. Link prediction in relational data. In NIPS,
2003.
[Van Horn and Martinez, 1994] Kevin S Van Horn and Tony R
Martinez. The minimum feature set problem. Neural Networks,
7(3):491 -- 494, 1994.
[Wang et al., 2013] Chenguang Wang, Nan Duan, Ming Zhou, and
Ming Zhang. Paraphrasing adaptation for web search ranking. In
ACL, pages 41 -- 46, 2013.
[Wang et al., 2015a] Chenguang Wang, Yangqiu Song, Ahmed El-
Kishky, Dan Roth, Ming Zhang, and Jiawei Han. Incorporating
world knowledge to document clustering via heterogeneous in-
formation networks. In SIGKDD, pages 1215 -- 1224, 2015.
[Wang et al., 2015b] Chenguang Wang, Yangqiu Song, Haoran Li,
Ming Zhang, and Jiawei Han. Knowsim: A document similarity
measure on structured heterogeneous information networks. In
ICDM, pages 1015 -- 1020, 2015.
[Wang et al., 2015c] Chenguang Wang, Yangqiu Song, Dan Roth,
Chi Wang, Jiawei Han, Heng Ji, and Ming Zhang. Constrained
information-theoretic tripartite graph clustering to identify se-
mantically similar relations. In IJCAI, pages 3882 -- 3889, 2015.
[Wang et al., 2016a] Chenguang Wang, Yangqiu Song, Haoran Li,
Ming Zhang, and Jiawei Han. Text classification with heteroge-
neous information network kernels. In AAAI, 2016.
[Wang et al., 2016b] Chenguang Wang, Yangqiu Song, Dan Roth,
Ming Zhang, and Jiawei Han. World knowledge as indirect super-
vision for document clustering. TKDD, 11(2):13:1 -- 13:36, 2016.
[Wang et al., 2016c] Chenguang Wang, Yizhou Sun, Yanglei Song,
Jiawei Han, Yangqiu Song, Lidan Wang, and Ming Zhang. Rel-
9
|
1907.09395 | 2 | 1907 | 2019-11-11T02:40:35 | Mining Temporal Evolution of Knowledge Graph and Genealogical Features for Literature-based Discovery Prediction | [
"cs.SI",
"cs.LG"
] | Literature-based knowledge discovery process identifies the important but implicit relations among information embedded in published literature. Existing techniques from Information Retrieval and Natural Language Processing attempt to identify the hidden or unpublished connections between information concepts within published literature, however, these techniques undermine the concept of predicting the future and emerging relations among scientific knowledge components encapsulated within the literature. Keyword Co-occurrence Network (KCN), built upon author selected keywords (i.e., knowledge entities), is considered as a knowledge graph that focuses both on these knowledge components and knowledge structure of a scientific domain by examining the relationships between knowledge entities. Using data from two multidisciplinary research domains other than the medical domain, capitalizing on bibliometrics, the dynamicity of temporal KCNs, and a Long Short Term Memory recurrent neural network, this study proposed a framework to successfully predict the future literature-based discoveries - the emerging connections among knowledge units. Framing the problem as a dynamic supervised link prediction task, the proposed framework integrates some novel node and edge-level features. Temporal importance of keywords computed from both bipartite and unipartite networks, communities of keywords, built upon genealogical relations, and relative importance of temporal citation counts used in the feature construction process. Both node and edge-level features were input into an LSTM network to forecast the feature values for positive and negatively labeled non-connected keyword pairs and classify them accurately. High classification performance rates suggest that these features are supportive both in predicting the emerging connections between scientific knowledge units and emerging trend analysis. | cs.SI | cs |
MINING TEMPORAL EVOLUTION OF KNOWLEDGE GRAPHS AND
GENEALOGICAL FEATURES FOR LITERATURE-BASED
DISCOVERY PREDICTION
A PREPRINT
Nazim Choudhury
Department of Computer Science and Engineering
University of South Florida
Tampa, FL 33620
[email protected]
Fahim Faisal∗
Dept. of Computer Science and Engineering
Islamic University of Technology
Dhaka, Bangladesh
[email protected]
Matloob Khushi
School of Computer Science
The University of Sydney
Sydney, NSW 2006, Australia
[email protected]
November 12, 2019
ABSTRACT
Literature-based discovery process identifies the important but implicit relations among information
embedded in published literature. Existing techniques from Information Retrieval (IR) and Natural
Language Processing (NLP) attempt to identify the hidden or unpublished connections between
information concepts within published literature, however, these techniques overlooked the concept
of predicting the future and emerging relations among scientific knowledge components encapsulated
within the literature. Keyword Co-occurrence Network (KCN), built upon author selected keywords
(i.e., knowledge entities), is considered as a knowledge graph that focuses both on these knowledge
components and knowledge structure of a scientific domain by examining the relationships between
knowledge entities. Using data from two multidisciplinary research domains other than the medical
domain, capitalizing on bibliometrics, the dynamicity of temporal KCNs, and a recurrent neural
network, this study develops some novel features to successfully predict the future literature-based
discoveries - the emerging connections among knowledge units. Temporal importance extracted from
both bipartite and unipartite networks, communities defined by genealogical relations, and the relative
importance of temporal citation counts were used in the feature construction process. Both node
and edge-level features were input into a recurrent neural network to forecast the feature values and
predict the future relations between different scientific concepts represented by the author selected
keywords. High performance rates suggest that these features are supportive both in predicting the
future literature-based discoveries and emerging trend analysis.
Keywords Literature-based Knowledge Discovery · Dynamic Supervised Link Prediction · Keyword Co-occurrence
Network (KCN) · Genealogical Community · Weighted Temporal Citation.
1
Introduction
Statistical bibliography or bibliometrics [1] has been supporting researchers to address the challenges related to
the rapid growth of scholarly publications and scientific knowledge. Employing two network-based methods (i.e.,
∗Corresponding Author.
co-citation and keyword co-occurrences network), bibliographic coupling [2] method in bibliometrics enabled us
to explore the structure of scientific and technical knowledge. While the co-citation focuses on the structure of
scientific communication by analyzing citation links, keyword co-occurrence network (KCN) or co-word network,
focuses on knowledge components and knowledge structure by examining co-appearances of keywords found in the
literature. Therefore, it is also known as knowledge graph [3]. Associated as metadata information within the scholarly
publications, the author selected keywords are considered as the carriers of knowledge units, or knowledge entities
[4]. These hand-picked signal words/terms/themes/conceptual keywords denote authors' understandings of their work,
the thematic context of their research, and facilitate knowledge discovery [5]. Further, these topical concepts or
themes are also used for indexing purpose in digital libraries. The co-appearance of two author selected keywords in an
article defines a certain intrinsic relationship between them whereas multiple such instances denote the strength of their
relationships [6]. Representing such co-occurring relationships between the knowledge entities, KCN is bears both
theoretical and practical implications including literature-based discovery [7].
Literature-based discovery [8], widely spelled as Literature Based Discovery (LBD), is the process that seeks to
discover new knowledge (innovation) from the information embedded in published literature. In both automated and
semi-automated ways, it identifies the hidden (implicit) and important connections among knowledge components.
Rapid growth of scholarly publications and associated scientific knowledge prompted scholars to restrict themselves
within their narrow specialties (i.e., fragments within a broad domain) and cite only from the related articles [9].
Consequently, useful connections between fragmented information remain unnoticed due to the lack of awareness of
work from mutually exclusive fragments. LBD addresses this problem of knowledge overspecialization and strives to
identify the implicit and novel connections from the concepts explicitly published in the scientific literature. Preiss
et al.
[10] reported different applications of this approach including identification of treatments for diseases, drug
re-purposing, disease candidate gene discovery, and drug side effect prediction. Similarly, Henry et al. [11] reported
application areas outside the biomedical domain such as efficient water purification systems, development acceleration
of developing countries, potential bio-warfare agents categorization, climate change studies, and identification of
promising research collaborations.
Most methodologies addressing LBD are based on Swanson's 'ABC' model [12] and predominantly rely on text
analysis [13]. Further, most studies in LBD used MEDLINE as the literature database. Sebastian et al. [14] divided
existing techniques of literature-based knowledge discovery from scientific literature into two categories: (i) traditional,
and (ii) emerging approaches. The traditional approach dominates the current research landscape. These are mainly
knowledge-based and comprised of lexical statistics or graph-theoretic methods that require domain knowledge. On the
other hand, emerging approaches prompt new trends and unprecedented paradigm shifts in the knowledge discovery
process. These methods consider the occurrence and co-occurrence frequencies of keywords, probability distributions,
association rule mining, graph data structure, temporal features, relational attributes, and supervised classification
approach. A literature review on these methods is presented in the Related Works section. Being a knowledge network
representation technique, KCN inherently fits best in the LBD process. However, despite their success in mapping
scientific knowledge structure, the evolutionary dynamics of KCN, and metadata information (e.g.authors, citations)
associated with scholarly articles [15] are underutilized in the existing LBD techniques. Temporal changes of metadata
information, and communities of keywords extracted from dynamic KCN can also be considered as non-trivial factors
than mere frequencies or connectivity information in predicting the underlying complex and implicit relations between
keywords. Most importantly, existing methods in LBD concentrates on identifying the implicit or undisclosed relations
between keywords in the published literature instead of predicting their future relations those are yet to emerge. This
can facilitate emerging trend analysis which can be supportive for both researchers and science policymakers. To
this end, by formulating the LBD prediction problem as a dynamic supervised link prediction problem, we propose
a recurrent neural network (RNN)-based method to predict the emerging associations (i.e., co-occurrences) between
author selected keywords found in the scientific literature. This prediction process incorporated features constructed by
leveraging the temporal dynamics of KCNs, evolutionary metadata information (authors, citation) and genealogical
communities of keywords Due to the temporal nature of the constructed features and binary supervised classification
task, we considered the Long Short Term Memory (LSTM) - an artificial RNN architecture, suitable for sequence
classification. The research objective of this study is to predict the co-evolution of knowledge entities (author selected
keywords) found in the scientific literature to facilitate the prediction of future LBD. Since the prediction problem was
framed as a supervised learning process, the scientific contribution of this study is the development of representative
features to describe and successfully classify instances (non-connected keyword pairs). By considering dynamic KCNs,
this study develops some novel features by leveraging the temporal dynamics of KCNs, metadata information (i.e.,
authors, articles, citations), and finally, communities defined via deriving ancestral lineages of keywords extracted
from the dynamic KCNs . The prediction performances of the proposed features, applied over scholarly datasets on
2
two topics 2 ( i.e., Obesity, Sleep Apnea), are also reported. These generic features are domain independent and can be
applied to any multidisciplinary scientific disciplines to predict the emerging trends.
Related Works
Literature-based discovery (LBD) identifies the implicit relations from the explicit information. It is principally used in
biomedical research where running experiments is expensive. This knowledge discovery process broadly encompasses
lexical and semantic text analyses of articles found in the MEDLINE database. Most of the LBD techniques employed
frequency-based approaches. The underlying assumption is that discoveries are likely to arise from logical connections
among source, intermediate, and target concepts (keywords) based on either their frequent/infrequent (co)occurrences
in the literature, or common/rare connections to a knowledge base [16]. Gordon and Dumais [17] took the advantages
of both co-occurrence frequencies and Latent Semantic Indexing (LSI) used in analyzing relationships between a set
of documents. Hristovski et al [18] used association rule mining and Unified Medical Language Systems (UMLS)
to discover the relationships between medical concepts. Yetisgen-Yildiz and Pratt developed [19] 'LitLinker' that
incorporated knowledge-based methodologies with statistical approach considering the background distribution of
keyword probabilities. With the help of Fuzzy Set Theory and probabilistic model of relationships, Wren et al. [20]
developed a computational method to identify large sets of relationships between unrelated items within scientific
reports.
Ensemble-based approaches combined both statistical and temporal features to find the intermediate keywords con-
necting both source and target keywords. These approaches were explored to find meaningful links between two
disparate sets of articles in MEDLINE [21]. Relational techniques [22, 23] used the explicit semantic relationships (i.e.,
predicates) between concepts where such predicates were typically obtained from structured background knowledge
or known a priori via domain experts. Few approaches focused on graph data structures to discover knowledge from
the literature by creating subgraphs based on the binary relationships between literary concepts. These relationships
were also drawn from semantic predications which were extracted directly from assertions in MEDLINE literature.
By weighting links using degree centrality, Wilkowski et al. [24] developed such a graph-theoretic approach. Their
approach used an iterative greedy strategy to create the 'best' subgraph with the help of semantic predications. Cameron
et al. [16] also used semantic predication graph and introduced a method to automatically find clusters of contextually
similar paths in the graph. These clusters were then used to identify the latent associations between disjoint concepts in
the literature to reconstruct eight scientific discoveries. Like most LBD methods, these graph-based data structures were
also primarily constructed in the biomedical domain using predicative relations extracted from MEDLINE literature.
Contemporary approaches developed heterogeneous networks capable of encoding richer information and better
semantics between various real world objects [25]. Known as Heterogeneous Bibliographic Information Network
(HBIN), these networks represent a collection of scientific publications as a network of heterogeneous bibliographic
objects (e.g., keywords, authors etc.). HBIN allowed various information to flow across different types of objects and
links to capture the previously unknown associations between research articles. It also harnessed the meta-path features
found in HBIN networks to discover the latent associations. Sebastian et al. [26] used lexico-citation features of HBIN
networks to predict the co-citation links between articles from previously disconnected research areas. Similarly, Ding
et al. [15] used information from the literature in the form of an 'entitymetrics' graph. The objective was to model
the latent relationships among biomedical entities (e.g. diseases, drugs) based on the existing citation relationships
among the corresponding articles. Apart from predicting the implicit and/or hidden relationships between disjoint
sets of articles, researchers [27, 28] also used HBIN to predict the citation count. HBIN-based LBD used simple
statistics and does not require sophisticated and domain-specific NLP tools and ontologies. It also facilitated the
utilization of metadata information in constructing features for prediction task. However, to predict future links between
author selected keywords instead of links between articles, processing HBIN and calculating meta-path features will be
computationally intensive.
One of the emerging approaches [29] conjectured the task of predicting implicit relationships as a classification task
by leveraging the link prediction methodology of network science. It describe the associations between different
concepts/keywords/topics using networks where the links represent their semantic or co-occurrence relationships.
This approach was primarily used in predicting the implicit links within a co-occurrence network of Medical Subject
Headings (MeSH) terms. Crichton et al. [30] recently investigated how inputs from four node representation algorithms
affect the performance of a neural link predictor on random- and time-sliced biomedical graphs (i.e., Drug-Target
Interactions, Protein-Protein Interaction and LBD) containing information relevant to drugs, protein and literature.
Katukuri et al. [31] used manually-created features in a supervised link prediction task on a large-scale biomedical
network of concepts co-occurrences. The objective was to predict links those represented scientific hypotheses in
2datasets and codes are available in https://github.com/faisal-iut
3
a time-sliced corpus. The authors extracted relevant information from the biomedical corpus to generate a concept
network and concept-author map. They also developed a set of heterogeneous features by exploiting random walk,
network neighborhood and common authorship. Wang and Zeng [32] used two-layer graphical model, called restricted
Boltzmann machine, to perform similar link prediction task on Drug-Target Interactions (DTI) network. Their objective
was to predict multiple types of DTIs, unknown drug-target relationships or drug's modes of action. However, Lu et
al. [33] pointed out the limitation of such machine learning models in DTI predictions despite their high performance.
This included the absence of additional information about the characteristics of drugs, targets and DTIs, (e.g.,chemical
structure, genome sequence, binding types, causes of interactions, etc.). Therefore, the authors used topological
similarity indices, such as Common Neighbours and Katz [34] index from complex network theory, to predict links in
a DTI network. Our study is probably closer to this set of machine-learning based studies however it differs in different
ways which is described later in this section.
In summary, existing research used combinations of lexical distributional statistics, graph-theoretic measures, hetero-
geneous network-based methods, and machine learning models. Predominantly capitulated on MEDLINE database,
most existing methods require domain knowledge to interpret relations between knowledge entities (drug, target,
protein, gene etc.). Graph-theoretic methods considered static networks which is unable to capture the temporal aspects
of network evolution such as 'recency', genealogical traits (origins of relationships), and time-variant frequencies
crucial for predicting the emerging associations in evolving networks. Despite the merits of existing approaches, there
exists lack of generalized predictive features applicable to any research domain. These features should be free from
any domain-specific knowledge-base for predicative relationships or relax the requirements of domain experts for
interpretation. It should accommodate metadata information which is commonly associated with any scientific literature
like authors, keywords, affiliation etc, and integrate the temporal information since scientific knowledge structure is
inherently dynamic. These facts are the motivation behind this study. Further, in most cases, these methods focused
only on identifying implicit(hidden) relationships between scientific concepts which are yet to be recognized from their
explicit relations instead of predicting the future relations those are yet to emerge. To this end, this study develops a
recurrent neural network-based method to predict emerging LBD instead of identifying the hidden relations between
keywords/concepts. It manifests the knowledge evolution as a dynamic process and leverages the evolutionary aspects
of knowledge graphs, temporal recency and citation information to develop hand-crafted features suitable for the
prediction task.
2 Scholarly Datasets
We extracted our scholarly datasets on two topics and the source of these datasets is 'Scopus' -the largest abstract and
citation database of peer-reviewed literature. The first search keyword was 'sleep apnea', also spelled as sleep apnoea,
and the second keyword was 'obesity'. The first topic is related to the serious sleep disorder that occurs when a person's
breathing is interrupted or a person experiences periods of shallow breathing during sleep. The second topic is related
to overweight and represents a complex disorder involving the accumulation of excessive body fat to an extent such that
it may have negative impact on health of individuals. Article metadata information including publication year, article
title, authors, affiliations, author selected keywords and citation count were extracted from the Scopus digital library by
considering the following constraints: (i) article published in English journals, (ii) the search keywords are present in
the article's titles and abstracts, and (iii) articles published within the duration 2007-2016. For the sake of brevity, Gs
and Go will be used to denote the dataset related to sleep apnea and obesity respectively for the rest of the article.
Dataset
Gs
Go
# Articles
29203
107745
# Keywords
12721
11643
Duration
2007-2016
2007-2016
Table 1: Statistics of two datasets
The author selected keywords are crucial to know about the thematic context, topics and related concepts of the
corresponding scholarly articles. Recent advances in network science [35] have prompted researchers to address the
mapping and understanding of scientific knowledge evolution via different types of bibliometric networks. These
networks consist of nodes representing different scholarly items like publications, journals, researchers, or keywords
and edges indicating the relations between pairs of nodes. According to Van Eck and Waltman [36], the most commonly
studied types of relations are: (i) citation relations, (ii) keyword co-occurrence relations, and (iii) co-authorship relations.
To predict the emerging relations between different scientific concepts/keywords, we took the advantage of temporal
keyword co-occurrence relations and constructed dynamic KCNs in both Gs and Go for each year. Therefore, as an
integral part of dynamic KCN construction. we extracted author selected keywords appeared in more than one articles.
4
Keywords failed to gain such minimum attention from the research community were considered to be irrelevant to the
corresponding research and thus discarded. Keyword extraction phase was followed by the text processing phase that
includes text cleaning and transformation. Authors used different spellings and acronyms to represent their keywords
those represent same semantic meaning. Firstly, any unwanted space and other discriminators were eliminated from the
keyword list with the help of NLTK text pre-processing tools [37]. This step also included lower-casing all keywords,
removing singular-plural differences (e.g., epidemiological studies ↔ epidemiological study, dilator muscles ↔ dilator
muscle) and lemmatization of some commonly used keywords (e.g., dreaming ↔ dream). Secondly, semantically
related common keywords were uniformly presented by using one representative word for all (e.g., aspect ↔ feature).
Most widely used abbreviated keywords were kept unchanged however others were changed to it's full forms (e.g.,
bmi ↔ body mass index). There were some abbreviations with different full forms (e.g., egfr ↔ (estimated glomerular
filtration rate, epidermal growth factor receptor)). In these cases, the content of the corresponding article was verified to
identify the right full form. Besides, all numbers and their corresponding roman forms were used in number format.
Basic dataset statistics including the number of these cleaned and transformed keywords are presented in Table 1.
It is noteworthy that although the data collection period includes 2007, however, in the experiment, we considered
2008-2014 duration as our actual training period. This fact will be explained in the later section.
Table 2: Yearly statistics of nodes and edges in keyword co-occurrence networks (KCNs) in
Gs=Sleep Apnea and Go=Obesity
Gs
Go
Nodes
Edges
Nodes
Edges
2007
388
689
2109
9569
2008
459
773
2356
10282
2009
470
688
2092
10477
2010
496
786
2298
12001
2011
536
977
2408
14271
2012
584
1054
2579
16158
2013
634
1280
2582
15851
2014
583
1144
3217
18238
3 Background
3.1 Knowledge Evolution and Dynamic KCN
Currently, scientific knowledge creation is dynamic and interdisciplinary in nature where different avenues of research
converge and new connections emerge among disjoint and existing areas of science [38]. This knowledge is generally
incremental except few revolutionary and fundamental changes. The development and evolution of science and
technology create new knowledge from the previously accumulated and ubiquitous information [39]. New hypotheses
are being postulated by encompassing existing scientific concepts from multiple domains. Agusti Canals [40] pointed
out that the diffusion of scientific knowledge can be mapped into a network structure where knowledge propagates via
interactions among networked agents. A knowledge graph (network) is a specific kind of knowledge representation
technique that uses a semantic network structure where nodes are keywords and links represent their causal relationships
[3] Statistically significant and non-trivial co-occurrence patterns of keywords not only represent their semantic affinity
[41] and relatedness [42] but also their causal relationships [43]. Therefore, using the extracted, cleaned, and processed
keywords, as described above, dynamic Keyword Co-occurrence Networks (KCNs) were constructed for each year in
2008-2015. KCN is defined as an undirected network G(V, E) where nodes V is the set of author selected keywords
and E is the set of edges where each edge represents the co-appearance of two keywords in the same article. Multiple
such co-appearances define the edge weights which is ignored in this study. Dynamic KCNs are the temporal networks
Gt = (vt, et) for time period t = 1, 2, ..., T where vt is the set of nodes or keywords and et is the set of edges
connecting the set of keywords vt ∈ V at time t. These edges at time t can be new or recurring. Table 2 provides
In Table 3, we
statistics of the number of nodes (keywords) and edges (co-occurrences) per year in both datasets.
present the basic statistics of the evolutionary patterns observed in dynamic KCN for both Go and Gs. The dynamicity
of keyword co-occurrences denotes that new research topics, hypotheses, or directions emerge over time through
co-appearances of existing keywords.
Three different scenarios can be observed in dynamic KCNs. As observable in Table 3, firstly, new edges emerge
each year when new keywords co-appear in articles. Secondly, new keywords form edge with old (existing) keywords
(appeared in previous years). Finally, edges are formed in a year between two old keywords from the previous year(s),
where these old keywords appeared in different articles but not co-appeared in the same article. The term 'old' in any
particular year t denotes the set of keywords appeared in year(s) prior to t. e.g., in Gs for the year 2010, the number
of Vo is 668 which denotes that out of Vt = 1000 keywords in 2010, 668 keywords appeared within 2007-2009. It
is observable from the table that the number of new keywords normally decreases. When a new hypothesis gains
considerable attentions in the subsequent years, the related keywords become significant. This fact prompts expansion
of these keywords' degree through new or recurring relations. Besides, new relations between old keywords Eo↔o were
5
Table 3: Evolutionary statistics of nodes (author selected keywords) and edges (co-occurrences in the same article) for two datasets
in this study. Vt = keywords in year t, Vn = new keywords each year, Vo = old keywords from the previous year(s), E = edges,
Eo↔o = recurring edges between v ∈ Vo, E(cid:48)
o↔o = new edges between v ∈ Vo, En↔o = edges between v ∈ Vo and v ∈ Vn, En↔n
= edges between v ∈ Vn. The term 'old' in a particular year denotes the set of keywords appeared in previous year(s)
Gs
Go
year
2007
2008
2009
2010
2011
2012
2013
2014
2015
2007
2008
2009
2010
2011
2012
2013
2014
2015
Vt
857
968
978
1000
1172
1240
1328
1367
1262
3207
3515
4206
4696
5059
5426
5418
4283
4466
Vn
857
572
438
332
345
279
272
291
157
3207
1217
1709
1152
658
364
77
738
367
Vo
0
396
540
668
827
961
1056
1076
1105
0
2298
2497
3544
4401
5062
5341
3545
4099
E
1486
1690
1587
1648
2264
2399
2715
2728
2347
15725
16826
21122
24093
29261
32663
32076
22410
23606
Eo↔o
0
144
202
264
387
271
424
500
512
0
3167
4515
6578
6514
9944
11702
6522
8043
E(cid:48)
o↔o
0
530
609
800
1166
1524
1692
1605
1530
0
8995
9091
13154
19962
20900
19987
12386
13860
En↔o
0
777
636
483
599
531
539
502
235
0
4336
6695
4094
2698
1744
382
3251
1548
En↔n
1486
239
140
101
112
73
60
121
70
15725
328
821
267
87
75
5
251
155
found to be dominating which is the generic trend in inter-disciplinary research. Further, the growth of edges in E(cid:48)
o↔o
implies that most new hypotheses emerge across existing keywords, topics, and/or concepts. Conversely, sporadic
nature of new hypothesis generation across new keywords can be observed through the decreasing number of edges in
En↔n. Delayed consumption of new concepts by the scientific communities can be attributed as a cause. In case of
En↔o, we observed that a lot of edges are formed with the most central keywords within the research domain which is
true for knowledge network evolution (preferential attachment).
3.2 Preferential Attachment and Recursive Centrality
The generation of author selected keywords is governed by the inherent rules of scholarly communication which
is also known as preferential attachment
[44]. In citation-based scholarly communication, few articles within a
research area will draw most attention and subsequently acquire most citations. These articles are considered as the
representatives of the corresponding research area. Essentially, authors select keywords for their new articles either from
the existing pool of keywords previously used or newly generated keywords conjoined with the existing ones. Therefore,
few representative keywords having high degree centrality can always be found in KCNs [7]. This phenomenon is
demonstrated in Figure 1 using two keywords Life course and Endothelial dysfunction (green coloured) from Go
(a-c) and Gs (d-f) respectively in three different times. In these figures, the red and navy coloured keywords are the
existing keywords where the former denote the representative keywords in the respective research domain. The size
variances of the keywords represent their degree of connections in the KCN of the corresponding time. The edge
thickness represents the edge weights. From these network snapshots, it is observable that although new keywords
emerge in association with less-acquainted (relatively low profile) keywords in the corresponding research area at the
beginning, however, gradually, they tend to co-appear with the representative keywords (e.g., cardiovascular disease in
Go and metabolic syndrome in Gs). For example, the endothelial dysfunction formed association with the keyword
erectile dysfunction in the year 2008. In the year 2011, it form associations with the high degree keywords such as
inflammation, hypoxia, metablic syndrome. In the year 2014, it acquire connections with more divergent high degree
keywords of the domain (e.g., hypertension, sleep). Consequently, the rich club phenomena is also evident in dynamic
KCN which can be captured by the degree centrality of the corresponding representative keywords. Alongside the
classical degree centrality measure, researchers attempted to define custom metrics to denote nodal importance. In one
such study, Klimek et al. [45] developed a bipartite relations using term -- document matrix and proposed a recursively
defined document centrality measure to denote the importance of scientific documents. According to their assumption, a
list of documents is considered central if these consume a large number of central terms those in turn also are consumed
6
(b) G0: 2011 (Life course)
(c) Go: 2014 (Life Course)
(a) Go: 2008 (Life course)
(d) Gs: 2008 (Endothelial
dysfunction)
(e) Gs: 2011 (Endothelial dysfunction)
(f) Gs: 2014 (Endothelial dysfunction)
Figure 1: Temporal evolution of associations between new and existing keywords demonstrated by two example keywords Life
course and Endothelial dysfunction in Go and Gs respectively. All network snapshots are timestamped and present the evolving
connection patterns of the example keywords (green coloured) with the existing keywords (red and navy coloured) of the respective
research domain. The red coloured keywords are the representative keywords of the research domain and acquire more connections
over time. The size of the node represents to the corresponding keyword's degree of connections. The thickness of the edges
represents corresponding edge weights.
Notation
kgp, kp, kc, kg
vau
t
, vat
t
N au
t
, N at
t
Table 4: Frequently Used Notations
Description
Four keyword types: Grandparent, Parent, Children, Guest
Two variants of recursive centrality values for keyword v at time t (year)
calculated from keyword-author and keyword-article bipartite relations
Sets of top N central keywords in year t
considering the bipartite network centrality measures
by a large number of central documents. Following their conceptualization of recursive document centrality measures,
two recursive keyword centrality measures were developed for each keyword v in each year t of the training period.
Before jumping into the detail of this recursive centrality definition, Table 4 provides descriptions of some notations
those will be used frequently in the sections below.
The first centrality measure considers the keyword-author bipartite relations whereas the second, considered the
keyword-article relation. In case of the first bipartite relations, the intuition behind a keyword's importance is defined
by the number of central authors who, in turn, also use a large number of central keywords. After the initial textual
cleaning and data processing, we filtered out keywords which appeared in only one article. This category denotes
the specialized keywords not related to the wide majority of the scientific articles. With the remaining keywords, we
7
hysterectomylife courseobesityobesityepidemiologyepidemiologypreventionpreventionbody weightbody weightusausamortalitymortalitysocioeconomic statussocioeconomic statusraceracecardiovascular disease risk factorcardiovascular disease risk factordisabilitydisabilitylife courselife coursefamilyfamilychildhoodchildhoodosteoarthritisosteoarthritisknee painknee painsocial classsocial classbirth cohortbirth cohortsocial mobilitysocial mobilitymexicomexicounited kingdomunited kingdomlatent class analysislatent class analysisembodimentembodimentobesityobesitydiabetesdiabetesblood pressureblood pressuredepression symptomdepression symptommortalitymortalityepidemiologyepidemiologyhypertensionhypertensionethnicityethnicityinflammationinflammationdepressiondepressioncohortcohortfertilityfertilityeducationeducationgendergenderpregnancypregnancyagingagingdisabilitydisabilityadolescent healthadolescent healthcohort studycohort studyagedagedunderweightunderweightobesity paradoxobesity paradoxchinachinametabolic risk factormetabolic risk factorsocioeconomic statussocioeconomic statusbrazilbrazilalspacalspacgestational weight gaingestational weight gainusausarace/ethnicityrace/ethnicitylife courselife courseself-rated healthself-rated healthsocioeconomic positionsocioeconomic positionbirth cohortbirth cohorttrajectorytrajectorysocial mobilitysocial mobilitydisadvantagedisadvantageerectile dysfunctionendothelial dysfunctionmetabolic syndromemetabolic syndromehypoxiahypoxiabiomarkerbiomarkerinflammationinflammationendothelial dysfunctionendothelial dysfunctionmandibular advancementmandibular advancementatherosclerosisatherosclerosisdiseasediseaseinflammatory markerinflammatory markerinflammationinflammationsleepsleepmetabolic syndromemetabolic syndromehypertensionhypertensionatherosclerosisatherosclerosisinsulin resistanceinsulin resistancehypoxiahypoxiacerebrovascular circulationcerebrovascular circulationintermittent hypoxiaintermittent hypoxiabaroreflex sensitivitybaroreflex sensitivityendothelial dysfunctionendothelial dysfunctionreviewreviewendotheliumendotheliuminflammatory markerinflammatory markercarbon monoxidecarbon monoxideultrasonographyultrasonographyasymmetric dimethylarginineasymmetric dimethylargininemicroalbuminuriamicroalbuminuria constructed a keyword-author adjacency matrix MAxK for each year t in the training period. It is worth mentioning
that we considered the first name and last name of the authors and manually verified with the affiliation information
to perform author name disambiguation. Although the objective of this manual author name disambiguation was to
find all publications that belong to a given author and distinguish them from publications of other authors who share
the same name, however, exploration of other methods for the same purpose is left for future studies. The adjacency
matrix M is a binary matrix of size A × K where A and K denote the set of authors and keywords respectively
in year t. An entry in matrix M is 1 if the author a ∈ A uses the keyword k ∈ K in his/her article. Starting with an
example author ai in year t, for each keyword ki used by this author, we can reach all other authors who also used ki in
their article. Iterating this procedure twice, we reach all the authors in two-hop distance from ai. These authors used
keyword(s) that is also used by author(s) sharing some keywords with author ai. Figure 2 demonstrates the iteration
procedure with some example keywords. This measure accumulates the number of different paths available among
authors through their keyword usages. Higher number represents that the corresponding keyword is pervasive across
(cid:80)
the research domain and thus more central and familiar to the key authors. To calculate recursive keyword centrality in
this way, two vectors ψk and ψa are defined recursively:
a M (t)ψa(n − 1, t)
(cid:80)
k M (t)ψk(n − 1, t)
(1)
as ψk(0, t) = (cid:80)
ψk(n, t) =
a∈A M (t) and ψa(0, t) = (cid:80)
with n represents the number of iterations as portrayed in Figure 2. ψk(0, t) and ψa(0, t) denote the initial conditions
k∈K M (t). The values of ψk can have different interpretations. The
ψk(0, t)
, ψa(n, t) =
ψa(0, t)
Figure 2: Computation of recursive keyword centrality measure using keyword-author relations. The computation is an iterative
procedure that starts at an author, a0. The first iteration counts the number of authors that used a keyword which is also used by a0.
Then this iteration continues until it reaches the terminating condition
initial condition ψk(0, t) denotes the degree centrality of a keyword. For different values of n = 1, 2, 3..., values of ψk
assign weights to individual keywords considering the number of authors consuming them in their articles. Therefore,
high values of ψk corresponds to the keywords selected by a large number of authors those used large number of such
keywords. Conversely, low values represent the keywords' specificity and relevance to limited number of research
issues. Similar observations are also true in case of psia. The number of iteration converged and found stable at n = 20.
We also attempted higher value of n however. found similar recursive centrality values returned by the algorithm. Thus,
for each year t in the training period, we calculated a vector of z-score normalized recursive keyword centrality values
t by following the algorithm, from the keyword-author bipartite relations. Likewise, another recursive keyword
C au
centrality measure C at
t was constructed from the keyword-article bipartite relations in each year t during the training
interval. In addition to these recursive centrality measures, to compare the results with a traditional centrality measure,
we also calculated degree centrality values of keywords C d
t , extracted from the unipartite keyword-keyword relations in
KCN for each year t. Future studies can calculate other centrality values to compare the results further.
8
diabetescortisolobesityacculturationestrogendiabetes,cortisol,obesitydiabetes,insulin,obesityestrogen,acculturation,cortisolinsulindiabetescortisolobesityacculturationestrogeninsulindiabetes,cortisol,obesitydiabetes,insulin,obesityestrogen,acculturation,cortisol 3.3 Genealogical Typology
As mentioned earlier, due to prevailing preferential attachment in KCN, there exists a set of most central keywords
which exercise greater influence over the structure of the network. It was observed that other keywords generally tend
to form relations by co-appearing with these rich keywords. To predict the future co-evolution of keywords, this fact is
needed to take into consideration. In this study, we capitalize the idea of preferential attachment, but in a different way.
Figure 3: Genealogical communities of keywords by considering a set of keywords with high recursive centrality values (i.e., N au
t−1)
computed from keyword-author bipartite relationships. Keywords with high centrality values at timestamp t − 1 are designated as
grandparents (kgp at timestamp t. Direct neighbours of these grandparents belong to parents (kp) community, and two-hop
neighbours are designated as children (kc).Rest of the keywords are designated as guests (kg).
We defined a keyword's genealogical community membership in each year that denotes it's current ancestral relation
with the central keywords of the previous year. The underlying objective here was to capture the evolution of the a
keyword's lineage to its ancestry. Genealogy is the study of family tree and a genealogy graph is used to portray the
complex evolving relation originating from the ancestors. In the same way, we defined a keyword's family relationship
in a particular year t based on it's relation with the central keywords from previous year t − 1. The centrality of
keywords was measured by considering two recursive centrality measures defined above and one degree centrality. This
approach would label keywords according to the type of their ancestral relationships (e.g., grandparents, parents, child
or guest) and help us understanding the impact of these relationship types in their co-occurrences. The construction of
the genealogical communities is described below:
In our dynamic KCN, we defined four types of communities for keywords, namely, grandparents (kgp), parents (kp),
children (kc) and guests (kg). Firstly, we defined the grandparents. If we consider identifying community membership
of keywords in year t, then grandparents kgp keywords are the top-N central keywords from year t − 1. Here, the
word 'central' denotes higher values of a chosen centrality measure from the three aforementioned centrality measures
(i.e., two recursive and one degree centrality). In other words, these are the most frequent and influential keywords
both in regards to the contents of metadata information (e.g.,author and article) and in forming relations with other
keywords. We experimented with the number of grandparents ranging from 10 − 200 by considering two recursive
centrality measures. Interestingly, we found that having N (cid:62) 20 grandparents in each year does not change the
prediction performance (described later) significantly in both datasets. Therefore, we designated the top 20 keywords as
grandparents by considering the chosen centrality measures. In this way, we got three sets of top-20 central keywords
based on our three centrality measures (i.e., C au
t−1). Secondly, we define the parents, children and
guests. Keywords having direct relations (i.e., an edge is observed between them) with kgp are labeled as parents (kp)
in the year t. Subsequently, children (kc) are the direct neighboring nodes of kp keywords but not Kgp. Alternatively,
kp keywords are considered as common neighbours between kgp and kc. Once these three communities of keywords
(i.e. kgp, kp, kc) are labeled, rest of the are designated as guests (kg). These guests keywords had maximum distance
t−1, and C d
t−1, C at
9
asfbdlkngprtvmhjwueyear tacfszdcpcagrqgggtpl from the grandparents. In this way, we defined the genealogical communities of all keywords appearing in a particular
year. In our experimental setup, we use top-20 keywords from year (t = 0) (i.e., 2007) for the first training year (t = 1)
(i.e., 2008), though this year (t = 0) is not included within our training period. This fact was mentioned in the earlier
section Scholarly Datasets. For the purpose of constructing link prediction features, we assigned each keyword a score
(a) Gs: 2011 (keyword-article)
(b) Gs: 2011 (keyword-author)
(c) Gs: 2011 (degree)
(d) Gs: 2014 (keyword-article)
(e) Gs: 2014 (keyword-author)
(f) Gs: 2014 (degree)
Figure 4: Temporal variations of genealogical traits demonstrated by keywords depending on the centrality measures used in Go
dataset. All network snapshots are timestamped including the centrality measure used to identify the genealogical traits. The color
codes represent the genealogical communities of keywords: red (grandparent), green (parent) and blue (child). The size of the node
represents to the keyword's corresponding centrality measures.
according to its genealogical community membership. The successive order of keyword's genealogical community
score is: score(kgp) >> score(kp) > score(kc) > score(kg).
Being most influential in the previous year, the grandparents (i.e., kgp) were assigned the highest score in a year.
Scores for the other three keyword types (i.e., kp,kc,kg) were defined according to their distance from the grandparents
kgp. Being the most distant ones, kg (guests) are assigned with lowest community score. The process of assigning
10
diagnosisbehaviourmorbid obesitycomplianceexcessive daytime sleepinessatherosclerosisdiagnosisbehaviourmorbid obesitycomplianceexcessive daytime sleepinessatherosclerosisdiagnosisbehaviourmorbid obesitycomplianceexcessive daytime sleepinessatherosclerosiscomplianceexcessive daytime sleepinessmorbid obesitybehaviourdiagnosisatherosclerosiscomplianceexcessive daytime sleepinessmorbid obesitybehaviourdiagnosisatherosclerosiscomplianceexcessive daytime sleepinessmorbid obesitybehaviourdiagnosisatherosclerosis t
t
t
, vat
t and vd
extracted from keyword-author relations.
a keyword's genealogical community scores is depicted in figure 3. The relative size of each keyword in this figure
represents the weight of the score (i.e. greater size for higher value). In this figure, N au
t−1 denotes the set of ton-N
central keywords by considering recursive centrality values vau
Considering three sets of centrality measures of keywords (i.e., vau
t ), a keyword belonging to one genealog-
ical community (e.g., grandparent) in one set, might belong to different community(e.g., parent) in another. This
effectively means that if a keyword becomes grandparent by considering the vau
, it may not necessarily be true if the
other centrality measures (i.e., vat
t or degree centrality) were considered. Likewise, considering the same centrality
measure, a keyword can belong to different genealogical community in different year(s). In Figure 4, a comparative
representation of these variations is presented for some keywords in Gs datasets. In this figure, top two rows present
network snapshots in Gs considering six keywords (i.e., atherosclerosis, morbid obesity, diagnosis, excessive daytime
sleepiness, behavior, and compliance). (It is evident that there exists some commonality of keywords in both datasets
since some of these keywords resemble to those in the Go dataset. Since the commonality is out of the scope of this
study, we leave it for future studies where domain experts can contribute towards literature based discovery related
to these two domains together.) The first row in this figure presents the network snapshots in the year 2011 and the
second presents snapshots in the year 2014. In each row, the left snapshot presents the genealogical typologies (i.e.,
communities) of keywords identified by considering the recursive centrality (i.e., vat
t ) from keyword-article bipartite
relations. The snapshots in the middle column present the genealogical typologies of keywords identified by considering
the recursive centrality (.i.e., vau
) from keyword-author bipartite relations. The right column presents the similar
by considering the keyword's degree centrality. Color codes represent different types of the keywords: grandparent
(red), parent (green), and child (blue). The sizes of the nodes and labels denote corresponding centrality values. It is
observable that keywords can belong to different communities in different years by considering different centrality
measures used in this study. Consider the genealogical communities of the keyword diagnosis. In the year 2011, it
was designated as 'grandparent' by considering vat
t , whereas considering the same centrality measure, it belonged
to 'parent' community in 2014. Surprisingly, the later is also true ('parent') when recursive centrality measure vau
was considered in the year 2011. Similar observations are evident in case of the keyword behaviour There are some
domain-specific keywords (e.g.sleep, apnea, obesity) which were found as 'grandparents' in all centrality variants since
they are the representative keywords of the respective domain.
t
t
Table 5: Number of nodes (keywords) according to different genealogical typologies (i.e., communities) defined in this study and
edges (keyword pairs) between different typologies of keywords. GP represents the grandparents keywords, P represents the
parents, C represents the children and G represents the guests keywords. Similarly, EP denotes the edges between parent
keywords, and EGP ↔ P denotes the edges between grandparent and parent keywords.
C
EP EP↔C EP↔G EC EC↔G EG
91
89
97
112
151
156
141
3729
3626
4072
4628
5723
5582
6903
70
62
48
64
94
121
140
1083
968
1148
925
1024
1375
2200
86
95
101
115
104
114
66
283
242
302
258
190
211
310
15
13
9
13
19
23
31
72
63
69
37
48
73
193
46
30
34
41
49
49
64
38
44
25
34
12
30
54
26
32
26
25
23
24
27
3
0
5
10
1
1
4
Year
P
2008
2009
2010
2011
2012
2013
2014
2008
2009
2010
2011
2012
2013
2014
197
179
189
225
243
273
235
1765
1614
1752
1963
2109
2117
2252
95
84
98
116
127
167
156
430
360
409
334
368
370
793
27
32
29
29
51
45
53
66
65
70
81
103
99
100
G EGP↔P EGP↔C EGP↔G
Gs
52
56
58
52
49
52
35
Go
67
32
43
27
26
41
51
679
664
556
665
757
726
869
185
157
153
125
131
148
323
147
187
189
175
194
174
172
141
98
117
91
82
75
155
3.4 Genealogical Community and KCN Evolution
In Table 5, basic statistics of nodes (keywords) and different types of edges (keyword co-occurrences) formed between
keywords belonging to different genealogical communities are presented. In this table, we ignored the grandparent
keywords since for each year it was the top-20 keywords from the previous year. Consequently, edges among the
11
Figure 5: Percentage of different types edges among keywords from different genealogical communities with respect to the total
number of edges in the test year (i.e., 2015)
grandparent keywords were also ignored since they were trivial in comparison to the other types of edges. In both
datasets, we found that edges among keywords belonging to the parent community dominate in numbers. This fact is
correlated with the increase of number of edges between grandparents and parents over time. Considering the sizes (i.e.,
number of articles and keywords) of both datasets, Gs harbours more guest keywords and edges. However, the ratio
of edges between children and guest keywords are similar in both datasets. The children and guest keywords tend to
form edges with parents more than the grandparents. This chronology of descendants not only helped us understand the
temporal trends of topics and research hypotheses developments but also attribute specificity in preferential attachment.
For example, in absence of such typologies, we would consider grandparents and parents keywords not only in the
same category and but also the richest since these would acquire most of the emerging edges. From this table, we can
also observe contrasting phenomenon which signifies the fact that the richest does not always remain so as they grow
oldest. Such classification on genealogical traits will allow us to comprehend the evolutionary growth of KCNs better
and thus help in science mapping.
In Figure 5, we observe how this ancestral relationships helped us to classify the edge types in the test year (i.e., 2015).
In this figure, we present the percentage of different types of edges, as presented in Table 5, with respect to the total
number of edges. It is evident from this figure that the keywords belonging to the parent community play crucial roles
in emerging relations than any other keyword types. In both datasets, the parent keywords dominate in attracting both
descendants and antecedents to form emerging relations.
4 Research Methodology
We frame the future LBD prediction problem as a learning-based link prediction task. Predicting co-evolution of
author selected keywords or identifying implicit relationships between them can be mapped to the dynamic supervised
link prediction model from network science. In this section, we describe our research methodology to construct such
a predictive model that incorporates both nodal and edge features, and a recurrent neural network for both feature
forecasting and classification.
4.1 Dynamic Supervised Link Prediction
For the purpose of link prediction, the total duration in each dataset was split into two non-overlapping intervals, T and
T + 1, known as the training and test phase. The primary objective of dynamic supervised link prediction framework is
to analyse the temporal topological structure and nodes' attributes in the network of training phase GT (VT , ET ) to
predict the emergence of future edges in the network of test phase GT +1(VT , ET +1). In dynamic network perspective,
12
EGPEGPPEPEPCECECGEG /070,9439508
7,.94341949,0/08 GsGo GT is sampled using an aggregation granularity to generate a time series of network snapshots. Due to the data
collection granularity this study used yearly window size to sample the network GT and generate time series of network
snapshots. It is impractical to predict the edges between nodes those are absent in the training network. Therefore, VT
is the set of nodes appeared in both phases. Then, we prepared our classification dataset consists of the instances of
non-connected keyword-pairs in the training phase. Each instance is labeled either positive or negative based on it's
presence as a true link in the test network. Supervised method for link prediction problems needs to predict the emerging
edges by successfully discriminating the positive and negatively labelled keyword-pairs within the classification dataset.
Hence, supervised link prediction is considered as a binary classification task by learning positive and negative instances
with the help of interesting features describing each instance. Classification model requires effective features for the
training purposes. To perform this task, we constructed novel features (described later) by incorporating temporal
information, network structure, and genealogical community memberships of author selected keywords. The feature
values for both positive and negative instances were computed by considering each network snapshot at timestamps
t0, t1, t2, ...tn ∈ T . Here, each timestamp represents a year in the training phase T (i.e., 2008-2014) . A recurrent
neural network (described below) was also employed to forecast the feature values for each instance in the test phase.
These forecasted feature values were input into a classifier for the classification purpose.
To measure performance of the classification, there are mainly two categories of evaluation metrics in a supervised
learning problem: (i) Fixed threshold metrics like accuracy, precision, recall (ii) k-equivalents and Threshold curves
like precision-recall (P -- R) curve and the area under the ROC curve (AUC) [46]. The threshold curves like ROC curve
and P-R curve are two dimensional curves. In ROC curve, true positive rates are plotted on the Y-axis and false positive
rates are plotted on the X-axis. In P-R curve, Y-axis is for precision and X-axis is for recall. P-R curve provides
better measurement in case of class distribution with large skewness. It also performs better in providing more insight
regarding the exposure of class differences.
4.2 Feature Engineering
To support supervised link prediction task, we constructed different features by taking advantages of different recursive
centrality measures and genealogical communities extracted from temporal KCN. These features assessed the influence
of keyword in regards to authors and articles and leveraged the temporal significance of evolving networks. Before
generating features for keyword pairs (edge-level), we first identified different characteristics of individual keyword's
importance. The rationale behind this is to contemplate different levels of significance each keyword carries in regards
to both network importance and ancestral lineages. We also considered accumulated citation score over time for each
keyword to construct citation-based features. Since, these features only pertain to one keyword, some aggregation
functions are needed to combine/aggregate the individual feature value of the corresponding keywords in a keyword pair.
To illustrate further, consider the fact that if either(or both) of the keywords are prolific or belong to the same community,
it is more likely that they will co-occur. Before aggregation, the individual measure denotes the proliferation rate of
individual keyword, its community membership, or temporal activeness (number of co-occurrences). Aggregation of
these individual features yields the aggregated features which are meaningful for the pair of keywords in dynamic link
prediction. Therefore, keyword-specific features were then aggregated to develop features for each keyword pairs. In
this study's context, we assume that the higher the aggregated feature values for a keyword pair, the more likely that
those two keywords will co-occur. In the following section, we describe the features used in the classification task.
4.2.1 Keyword centrality
For each keyword, we considered three centrality measures as node-level features. Two recursive centrality measures
were based on the keyword-author (vau
t ) bipartite relations as defined in section Recursive
centrality. The final centrality measure was the corresponding keyword's degree centrality (vd
t ).
) and keyword-article (vat
t
4.2.2 Temporal community importance
This is the first edge-level aggregated feature we computed for each non-connected keyword pair for the link prediction
purpose. The objective of this feature was to capitalize both the origin information (genealogical community) and
current activeness (centrality measures) of both keywords in a keyword pair. Chenhao Tan [47] investigated how the
origin of a community connects to its future growth and demonstrated that the history of the early members allows us to
understand the emergence of communities (i.e., tracing the origin of a community). In regards to the future community
growth, the author believed that the emerging process of a community is analogous to complex contagion that requires
dense connections between early adopters. In our case the early adopters are the members of the grandparent and parent
communities. Therefore, we first multiply individual keyword's community score with its corresponding centrality
value for each year in the training period. The former denotes the keyword's relation with previous year's central
keyword (i.e., keyword's family lineage) and the later denotes its importance which was originally used to determine its
13
community membership. This will augment the community importance scores of grandparent and parent keywords
more than children and guests, and subsequently determine the individual temporal community importance score of a
keyword in each training year. Finally, for each keyword pair, the temporal community importance score was computed
by summing their aggregated individual score in each year t. Thus, the first edge-level feature at each timestamp t in
the training period, known as temporal community importance score for a keyword pair, can be computed as:
t · va
t · vb
t
scoreH
t + gb
t and va
t (a, b) = ga
(2)
where ga
t denote the genealogical community score and the corresponding centrality score of keyword a.
By considering the temporal sequences of individual community score and centrality values of both keywords, we
built a time series of this aggregated feature for each instance of non-connected keyword pair. Since we considered
t ) for each keyword v to determine its community score, the following
three centrality values (i.e., vau
notations will be used to denote three variants of this feature value: scoreH
au(a, b) to denote the temporal community
importance score for an edge (a, b) where recursive centrality measures of keywords a and b were extracted from
at(a, b) will denote the same feature value where recursive centrality
the keyword-author relations. Similarly, scoreH
measures were extracted from the keyword-article relations. Finally, scoreH
d (a, b) will be used to denote the same
feature value where degree centrality values of keywords a and b were considered.
t and vd
, vat
t
4.2.3 Citation-weighted Recency
An individual keyword's citation count represent its influence factor over other keywords. However, domain-specific
and representative keywords can achieve higher citation counts then the others. Relative influence factor should also be
brought into consideration. For example, old keywords accumulate more citation over time then new keywords. In
contrast, some new keywords can gain significance quickly and acquire relatively high citation then other insignificant
but old keywords. Therefore, we considered the temporal factors of keyword's appearances such as current or distant
appearances over time. Considering our training period (e.g., 2008-2014), an article published in 2008 gets more time
to be cited than a article published in 2014. Despite, the later one carries more impact, the former one is weighted more
based on the numbers. Further, training year closer to the test period (e.g., 2014) is more significant than earlier years
since recently co-appeared keywords are more likely to come together in near future, a fact known as 'recency' in link
prediction task [48].
Considering the aforementioned facts on temporal significance and citation count, we constructed our second edge-level
feature for keyword pairs by accommodating both temporal recency and citation counts over time. Since citation count
and temporal recency are applicable to individual keywords, it requires to aggregate these counts for keyword-pairs. For
each keyword pair in year t, we aggregated their total citation counts. Further, we assigned temporal recency score
depending on the appearances of both keywords in a particular year. The recent their appearances are, or alternatively,
the closer their co-appearances towards the test year, the higher the recency score. Considering the keyword-pair (a, b),
let t denotes the training year t ∈ [1, 2, 3, ...T ], hu
t denote the total number of citation counts of the articles in
year t where keywords a and b appears respectively, then the weighted-citation recency is calculated as follows:
t and hv
scoreW
(3)
Here, γ amplifies the recency effect and its value is two if both keywords in a keyword-pair appear in a year, one if
either of the keyword appear, or zero (0) otherwise. ha
t will be assigned zero value if keyword a does not appear in an
article that received any citation. However, the feature value will be augmented depending on the current appearances
of the keywords (i.e., multiplied by value of t). The assumption here is that, keywords having high citation in the recent
years will have high probability to appear together since they represent the recent trends.
t (a, b) = (ha
t + hb
t) · γt
4.3 Feature Forecasting and Classification
Considering temporal KCN Gt(Vt, Et), constructed for each year t during the training period, we constructed time
series of features for non-connected keywords pairs. Like other supervised dynamic link prediction study [49], we
employed deep-learning framework to forecast the future values of constructed features during the test phase (i.e.,
2015). We use a Long Short Term Memory network [50] which is a special kind of Recurrent Neural Network (RNN)
for both forecasting and classification tasks. An LSTM takes sequential data as input and is considered well-suited
in classifying temporal sequence. The LSTM used in this study consisted of two blocks of memory-cells with two
different layers of hidden units. A simple LSTM cell unit takes three inputs (Xt, ht−1, Ct−1). Xt is the input of the
current time step, ht−1 is the output from the previous LSTM unit and Ct−1 is the "memory" of the previous unit. As
for outputs, ht is the output of the current unit and Ct is the memory of the current unit.
14
In our experiment, time series of feature values was input into a 2-layer LSTM network. The number of timesteps for
each training sample is seven since the training period is seven years (2008-2014) long. To forecast numerical features
values, linear activation function and MSE (Mean Squared Error) loss function were used. Categorical features were
encoded as one-hot vectors. In this case, softmax activation function and categorical cross-entropy (loss function) were
used. In all cases, adam optimization technique was used. Further, these forecasted feature values were then fed as
the training samples for our classification task. This classification task was implemented by adding another LSTM
netwrok in the pipeline which includes a dense layer with a single neuron as output layer. A dense layer is a fully
connected layer which means each neuron here receives input from all the neurons in the previous layer, thus densely
connected. In this layer, we used a logistic activation function named sigmoid which is ideal for assisting in binary
mutual exclusive classification problem. This output layer takes the forecasted feature values and predict the class
(positive/negative) of the keyword pairs.
5 Results
In this section we present our feature forecasting and link prediction results. We first present the feature forecasting
performance followed by the classification performance demonstrated by the LSTM.
5.1 Feature Forecasting Performance
Table 6: Normalized RMSE values (0-1) calculated on forecasted feature values against true feature values in the test year. Features
include both node-level (keyword features) and edge-level (aggregated features for keyword-pairs). scoreP A(a, b) represents the
preferential attachment score. vat
t denotes degree centrality
t denote two recursive centrality values of keywords and vd
t , and vau
values of keywords.
Datasets
Number of Iteration
scoreH
at(a, b)
scoreH
au(a, b)
scoreH
d (a, b)
scoreW
t (a, b)
scoreP A(a, b)
vat
t
vau
t
vd
t
Citation
100
0.027
0.025
0.017
0.006
0.016
0.136
0.119
0.027
0.003
Gs
500
1000
100
Aggregated Features (Keyword pairs)
0.023
0.020
0.010
0.004
0.015
0.024
0.022
0.010
0.005
0.015
Individual Features (Keyword)
0.135
0.112
0.021
0.002
0.135
0.110
0.020
0.002
0.002
0.002
0.002
0.001
0.005
0.077
0.070
0.003
0.0004
Go
500
0.003
0.002
0.001
0.001
0.004
0.071
0.063
0.002
0.0004
1000
0.003
0.002
0.001
0.001
0.004
0.060
0.057
0.002
0.0004
We used LSTM to forecast the constructed feature values (Feature Engineering) for the test period. For each keyword pair,
time series of both keyword features (Keyword centrality) and aggregated features (Temporal community importance
and Citation-weighted Recency), computed during the training period, were input in LSTM. To evaluate the performance
of this forecasting model, we computed the RMSE (root mean squared error) values by considering the actual feature
values in the test year against the forecasted values. Table 6 present the RMSE values to measure the performance
of LSTM in forecasting. We trained the LSTM model for different number of iterations (i.e. epochs). The effect of
iteration number is visible in the rmse values. In most cases, higher number of iteration results in more robust model
training. The range of actual values for different features can be varied significantly. So we presented the normalized
RMSE values in the range of 0-1. RMSE values in Go denote that in this dataset, the forecasting errors were smaller
than those in Gs which can be attributed to the size of the dataset. In this table, forecasting performance for both the
keyword feature and aggregated features are presented. For comparison's sake, we also calculated RMSE in forecasting
the preferential attachment score which is a widely used metric in link prediction task. It is noteworthy that preferential
attachment is an aggregated network feature which is computed by multiplying the number of neighbours of each
keyword in a keyword pair.
5.2 Link Prediction Performance
The supervised link prediction framework is subject to highly imbalanced data with a very large number of instances
with negative label. In practice, only a few pairs of keywords participate in true emerging edges out of every possible
pairs. Therefore, following other studies like in [7], the ratio of positive and negative class instances was set to 1:10.
15
Figure 6: P-R and ROC curves in both datasets (Gs and Go) to demonstrate the classification performance of the LSTM classifier
using the feature values constructed in this study. Traditional topological similarity metric 'Preferential Attachment' was also used to
compare the performance with the constructed features. 'Heredity (author)' denotes the aggregated feature temporal community
importance scoreH
au(a, b). Similarly,'Heredity (article)' denotes the temporal community importance scoreH
at(a, b).
We also used 30% of the instances for validation and evaluation purposes. Similar to forecasting performance, Go
dataset was found to demonstrate high performance in comparison to Gs. The AUCROC (i.e., Area Under ROC Curve)
is commonly used for evaluating such imbalanced classification problems. It introduces a probability value to quantify
the uncertainty associated with the classifiers. In case of binary classification, AUCROC enforces larger weight on
smaller class by using this threshold value. In Table 7, we present both the accuracy and AUCROC scores computed in
both Gs and Go datasets. For comparison's sake, we also present the performance of Preferential Attachment metric
which is a well-known topological similarity metric widely used in supervised link prediction task. In comparison to
this topological similarity metric, it is evident from Table 7 that the features constructed in this study outperformed the
traditional and widely prevalent metric for link prediction in KCNs, the preferential attachment. Better performance
was observed in Go rather than Gs which can be attributed to the greater number of instances. It is also evident that the
first aggregated feature (temporal community importance) performed better than the citation weighted recency scores.
Table 7: Dynamic supervised link prediction performance using AUCROC (AUC) and Accuracy (Acc%) values.
Gs
Go
scoreH
at(a, b)
scoreH
au(a, b)
scoreH
d (a, b)
scoreW
t (a, b)
scoreP A(a, b)
AUC
0.749
0.759
0.765
0.680
0.698
Acc(%)
89.3
88.7
89.5
88.9
88.8
AUC
0.783
0.800
0.794
0.776
0.753
Acc(%)
90.9
91.0
91.0
90.7
90.2
Considering other evaluation metrics, the P-R (precision - recall) curve depicts the precision-recall trade-off for a
classifier. This measurement is widely used in information retrieval. Reviewing both precision and recall is useful in
cases where there is an imbalance in the instances between the two classes. The reason for this is that typically the
16
!70.843 , !#.:7;0(Gs)
%7:0!489;0#,90 - # .:7;0(Gs)070/9 ,:947 070/9 /0700 9,943090/70.03.!70107039,,99,.2039
#0.,
!70.843 . !#.:7;0(Go)
,80!489;0#,90
%7:0!489;0#,90 / # .:7;0(Go)070/9 ,:947 070/9 /0700 9,943090/70.03.!70107039,,99,.2039 large number of negative class instances means we are less interested in the skill of the model at predicting negative
class instances correctly, (i.e., high true negatives). The most crucial thing of P-R curve is that this calculation does not
make use of the true negatives. It is only concerned with the correct prediction of the minority class, the positive class
instances. A P-R curve is a plot of the precision in y-axis and the recall in x-axis. In ROC curve, the goal is to have a
model be at the upper left corner, which is basically getting no false positives. Whereas, in P-R curve, the goal is to
have a model be at the upper right corner, which is basically getting only the true positives with no false positives and
no false negatives. In Figure 6, we present both the ROC and P-R curves of our features (solid lines) including the
Preferential Attachment (dashed lines) metrics. The top row represents P-R and ROC curves in Gs and the bottom
row represents the same in Go. Like the aforementioned performance measurements, in most cases of this figure, the
features constructed in this study outperformed the preferential attachment metric in both datasets.
5.3 Distribution of Aggregated Feature Values
In Figure 7, we present the distributions of feature values for the constructed aggregated features (i.e., temporal
community importance scoreH
au(a, b) and citation weighted recency). For all positive and negatively labelled keyword
pairs (samples), the normalized feature values for both aggregated features are presented as kernel density plots.
Distribution of the feature values for the positive samples (true keyword pairs in the test year) are presented in green
color and the negative samples are presented in red color. It is noteworthy that, for temporal community importance
score, we only consider the recursive centrality values extracted from the keyword-author relations.
Figure 7: Positive and negative class density of two aggregated features: temporal community importance score scoreH
where the recursive centrality measures were extracted from the keyword-author relations and Citation weighted recency
au(a, b)
scoreW
t (a, b) in both datasets
Two observations are evident from these figures. Firstly, since the overlap between the red and green colored regions
may trigger classification errors, the lower the overlap, the higher the classification performance. The amount of overlap
signifies that the aggregated features constructed in this study have non-trivial discriminatory characteristics. Secondly,
in contrast to topological similarity metrics where high value of the metrics corresponds to higher similarity between
17
0389/897-:943%02547,.422:3925479,3.0
9,943090/70.03. Gs 0,9;08,2508!489;08,2508
472,0/10,9:70;,:0
0389/897-:943
472,0/10,9:70;,:0
Go Figure 8: Two different centrality values (i) recursive centrality values extracted from the keyword-author relations (left) and (ii)
degree centrality of the keywords participating in the positively labeled edges during the test year in both datasets.
a pair of nodes (e.g., having more common neighbours between two nodes in a network), we found that positively
labelled keyword pairs have high density in lower feature values. This denotes that instead of higher values of the
features, comparatively lower features values have high probability in forming emerging relationships. Although, in Go,
we observe high feature values for positively labeled edge instance in the test period. This fact is contrary to our initial
assumption, mentioned in Feature Engineering that higher values denote higher probability of keyword co-occurrences.
Although, exploring the reason behind this is out of the scope of this study, however, in Figure 8, we present the
normalized keyword feature values (recursive centralities and degree centrality) computed for the test year (i.e., 2015)
in both datasets. The lower feature values can be attributed to the number of children and guest keywords including the
variances in normalized centrality values of the parent keywords. It is noteworthy that from the aforementioned figures
and tables, we have observed that the parent community dominates the merging link formations.
6 Discussion and Conclusion
Scientific progress depends on formulating verifiable and deductible hypotheses generation. This requires both
understanding and informed inferences from existing knowledge and information [51]. Rapid growth of scientific
knowledge and over specializations (domain-specific fragmentation) may engender opportunities to derive solutions
from one domain to address problems in another, although the underlying relationship may remain implicit or the
concerned groups from both domains are unaware of the work of each other [52]. However, the continuous surge
in published scientific literature limits the scope of analyses an individual can accomplish to extract these novel
and implicit relationships between disjoint concepts, topics and domains [20]. Proliferation of scientific production
inhibits scientists and policy makers to detect trending subject areas and the linkages among these areas in their
research fields, and mapping the dynamics of science to plan for research progress [53]. human comprehension of
such massive information and knowledge is challenging when it exceeds the scale of human analytical skills. For
example, as mentioned by Spangler et al.
[54], it is inconceivable for a scholar to possibly assimilate, recall and
accurately process all the known facts relevant to protein functions, relationships between proteins and identifications
of roles of a particular protein related to a disease while there are over 70000 articles on a single protein - the tumour
18
472,0/.0397,9;,:0#0.:78;0.0397,9
0700.0397,9 Gs7,3/5,7039!,7039/:089
4 41047/8
472,0/.0397,9;,:0
4 41047/8
Go suppressor 'p53'. Thus, there is a great difference between what is known, and what we know as individuals from
the collective and multidisciplinary knowledge within a given domain. Further, the specialization or fragmentation
of literature may promote poor communication between specialties since scientists tend to communicate more within
their fragments than the broader community engaged within the domain [55]. According to Ganiz et al. [9] Literature
Based Discovery (LBD) addresses the challenge of seemingly boundless increases in scientific knowledge including
knowledge overspecialization faced by the scientific communities. However, existing LBD models suffers from the
lack of generalized predictive model to successfully predict the emerging trends in such discoveries. Despite their
success, LBD techniques including text analysis, information retrieval and natural language processing are deprived
of the benefits of bibliometrics, specially the analytical advantages of temporal keyword co-occurrence networks
(KCNs). These KCNs and network analysis methods are found to be supportive in identifying technological trends,
analyse research topics and follow their evolution and track the development of innovation system research [56].
Further, temporal dynamics of KCN and community-aware features are underutilized in the process of literature-based
hypotheses generation. To this end, this study proposed a predictive framework that integrates temporal evolution of
KCN, genealogical communities, citation counts of keywords, and an LSTM based forecasting and prediction model.
The KCNs in this study are comprised on author selected keywords which best describe the research themes referred by
the corresponding authors and are also considered as knowledge entities. This framework includes feature engineering
process that builds novel features for both keywords (node) and keyword pairs (edge defined by co-occurrence). In this
study, we developed two recursive centrality measures by considering two types of bipartite relations: keyword-author
and keyword-article. We also considered the degree centrality of keywords extracted from temporal unipartite KCNs.
These centrality measures were used to define temporal communities based genealogical relationships among keywords.
Temporal citation counts of keywords were also used as keyword features. Bipartite recursive centrality values including
the degree centrality, genealogical community information and relative importance of temporal citation counts were
used to construct edge-level (keyword pairs) features. Seven years of training period was used to built time series of both
node and edge-level features those were input into an LSTM network to forecast the feature values in the test year. If
we forecasted only the edge-level features, then the individual node-level significance of keywords would have lost due
to the aggregation of mixed data types (eg. numerical, categorical) in edge-level. For example, in case of the temporal
community importance score, originating from the recursive centrality values, extracted from keyword-author bipartite
relations, the feature vector includes both centrality values of two keywords as well as their corresponding community
and the aggregated feature value. The centrality values are numerical, however, the community is categorical which
was represented as one-hot vector. This can easily be handled in LSTM instead of other forecasting technique. With
relatively trivial forecasting error, the forecasted feature values were used in supervised link prediction to classify both
positive and negatively labeled non-connected keyword pairs. The performance of the LSTM classifier was measure
using well-known performance metrics and also compared against well-known network topological similarity metric
used in link prediction. High performance of the constructed feature indicates that these features are not only supportive
in dynamic supervised link prediction but also can be beneficial in predicting literature based knowledge discovery or
emerging trend detection.
Despite the better performance measurements, this study is not free from its limitation. The first limitation comes
from the lack of accommodating domain experts in identifying semantic similarity of keywords and various forms of
abbreviations. Secondly, in this study we used manual name disambiguation instead of any standard methods to clean
the author names including which can further be explored in future studies. The semantic relationships between textual
contents were ignored in this study. The reason behind is that we only considered the author selected keywords those are
not semantically presented in the scientific literature but only co-appears if the corresponding authors recognize them as
relevant. Finally,lack of any such deep learning-based predictive framework in literature based discovery to compare
the performance of the proposed framework instead of relying on tradition link prediction metric. This study can further
be extended in various ways. Firstly, this study considered previous year's information to identify communities of
keywords in the current year. However, future studies can consider more historical information can be used (more than
one previous year information) to determine such communities. Other centrality measures and network community
detection algorithm can used to compare the performance of link prediction using community-aware features. Further,
some weighting strategies can be followed to assign weights on edges between keywords belonging to either same
community or different communities.
7 References
References
[1] A. Pritchard, et al., Statistical bibliography or bibliometrics, Journal of documentation 25 (4) (1969) 348 -- 349
(1969).
19
[2] M. M. Kessler, Bibliographic coupling between scientific papers, American documentation 14 (1) (1963) 10 -- 25
(1963).
[3] R. Popping, Knowledge graphs and network text analysis, Social Science Information 42 (1) (2003) 91 -- 106
(2003).
[4] H.-N. Su, P.-C. Lee, Mapping knowledge structure by keyword co-occurrence: a first look at journal papers in
technology foresight, Scientometrics 85 (1) (2010) 65 -- 79 (2010).
[5] M. Song, N.-G. Han, Y.-H. Kim, Y. Ding, T. Chambers, Discovering implicit entity relation with the gene-citation-
gene network, PloS one 8 (12) (2013) e84639 (2013).
[6] Y. Yang, M. Wu, L. Cui, Integration of three visualization methods based on co-word analysis, Scientometrics
90 (2) (2011) 659 -- 673 (2011).
[7] N. Choudhury, S. Uddin, Time-aware link prediction to explore network effects on temporal knowledge evolution,
Scientometrics 108 (2) (2016) 745 -- 776 (2016).
[8] R. N. Kostoff, Method for data and text mining and literature-based discovery, uS Patent 6,886,010 (Apr. 26 2005).
[9] M. C. Ganiz, W. M. Pottenger, C. D. Janneck, Recent advances in literature based discovery, Journal of the
American Society for Information Science and Technology, JASIST (Submitted) (2005).
[10] J. Preiss, M. Stevenson, R. Gaizauskas, Exploring relation types for literature-based discovery, Journal of the
American Medical Informatics Association 22 (5) (2015) 987 -- 992 (2015).
[11] S. Henry, B. T. McInnes, Literature based discovery: models, methods, and trends, Journal of biomedical
informatics 74 (2017) 20 -- 32 (2017).
[12] D. R. Swanson, Fish oil, raynaud's syndrome, and undiscovered public knowledge, Perspectives in biology and
medicine 30 (1) (1986) 7 -- 18 (1986).
[13] D. R. Swanson, N. R. Smalheiser, An interactive system for finding complementary literatures: a stimulus to
scientific discovery, Artificial intelligence 91 (2) (1997) 183 -- 203 (1997).
[14] Y. Sebastian, E.-G. Siew, S. O. Orimaye, Emerging approaches in literature-based discovery: Techniques and
performance review, The Knowledge Engineering Review 32 (2017).
[15] Y. Ding, M. Song, J. Han, Q. Yu, E. Yan, L. Lin, T. Chambers, Entitymetrics: Measuring the impact of entities,
PloS one 8 (8) (2013) e71416 (2013).
[16] D. Cameron, R. Kavuluru, T. C. Rindflesch, A. P. Sheth, K. Thirunarayan, O. Bodenreider, Context-driven
automatic subgraph creation for literature-based discovery, Journal of biomedical informatics 54 (2015) 141 -- 157
(2015).
[17] M. D. Gordon, S. Dumais, Using latent semantic indexing for literature based discovery, Journal of the American
Society for Information Science 49 (8) (1998) 674 -- 685 (1998).
[18] D. Hristovski, J. Stare, B. Peterlin, S. Dzeroski, Supporting discovery in medicine by association rule mining in
medline and umls, Studies in health technology and informatics (2) (2001) 1344 -- 1348 (2001).
[19] M. Yetisgen-Yildiz, W. Pratt, Using statistical and knowledge-based approaches for literature-based discovery,
Journal of Biomedical informatics 39 (6) (2006) 600 -- 611 (2006).
[20] J. D. Wren, R. Bekeredjian, J. A. Stewart, R. V. Shohet, H. R. Garner, Knowledge discovery by automated
identification and ranking of implicit relationships, Bioinformatics 20 (3) (2004) 389 -- 398 (2004).
[21] N. R. Smalheiser, V. I. Torvik, W. Zhou, Arrowsmith two-node search interface: a tutorial on finding meaningful
links between two disparate sets of articles in medline, Computer methods and programs in biomedicine 94 (2)
(2009) 190 -- 197 (2009).
[22] C. B. Ahlers, D. Hristovski, H. Kilicoglu, T. C. Rindflesch, Using the literature-based discovery paradigm
to investigate drug mechanisms, in: AMIA Annual Symposium Proceedings, Vol. 2007, American Medical
Informatics Association, 2007, p. 6 (2007).
[23] X. Hu, G. Li, I. Yoo, X. Zhang, X. Xu, A semantic-based approach for mining undiscovered public knowledge
from biomedical literature, in: 2005 IEEE International Conference on Granular Computing, Vol. 1, IEEE, 2005,
pp. 22 -- 27 (2005).
[24] B. Wilkowski, M. Fiszman, C. M. Miller, D. Hristovski, S. Arabandi, G. Rosemblat, T. C. Rindflesch, Graph-based
methods for discovery browsing with semantic predications, in: AMIA annual symposium proceedings, Vol. 2011,
American Medical Informatics Association, 2011, p. 1514 (2011).
20
[25] Y. Sun, J. Han, Mining heterogeneous information networks: principles and methodologies, Synthesis Lectures on
Data Mining and Knowledge Discovery 3 (2) (2012) 1 -- 159 (2012).
[26] Y. Sebastian, E.-G. Siew, S. O. Orimaye, Learning the heterogeneous bibliographic information network for
literature-based discovery, Knowledge-Based Systems 115 (2017) 66 -- 79 (2017).
[27] X. Ren, J. Liu, X. Yu, U. Khandelwal, Q. Gu, L. Wang, J. Han, Cluscite: Effective citation recommendation by
information network-based clustering, in: Proceedings of the 20th ACM SIGKDD international conference on
Knowledge discovery and data mining, ACM, 2014, pp. 821 -- 830 (2014).
[28] X. Liu, Y. Yu, C. Guo, Y. Sun, L. Gao, Full-text based context-rich heterogeneous network mining approach for
citation recommendation, in: IEEE/ACM Joint Conference on Digital Libraries, IEEE, 2014, pp. 361 -- 370 (2014).
[29] A. Kastrin, T. C. Rindflesch, D. Hristovski, Link prediction on a network of co-occurring mesh terms: towards
literature-based discovery, Methods of information in medicine 55 (04) (2016) 340 -- 346 (2016).
[30] G. Crichton, Y. Guo, S. Pyysalo, A. Korhonen, Neural networks for link prediction in realistic biomedical graphs:
a multi-dimensional evaluation of graph embedding-based approaches, BMC bioinformatics 19 (1) (2018) 176
(2018).
[31] J. R. Katukuri, Y. Xie, V. V. Raghavan, A. Gupta, Hypotheses generation as supervised link discovery with
automated class labeling on large-scale biomedical concept networks, in: BMC genomics, Vol. 13, BioMed
Central, 2012, p. S5 (2012).
[32] Y. Wang, J. Zeng, Predicting drug-target interactions using restricted boltzmann machines, Bioinformatics 29 (13)
(2013) i126 -- i134 (2013).
[33] Y. Lu, Y. Guo, A. Korhonen, Link prediction in drug-target interactions network using similarity indices, BMC
bioinformatics 18 (1) (2017) 39 (2017).
[34] L. Katz, A new status index derived from sociometric analysis, Psychometrika 18 (1) (1953) 39 -- 43 (1953).
[35] K. Börner, S. Sanyal, A. Vespignani, Network science, Annual review of information science and technology
41 (1) (2007) 537 -- 607 (2007).
[36] N. J. Van Eck, L. Waltman, Visualizing bibliometric networks, in: Measuring scholarly impact, Springer, 2014, pp.
285 -- 320 (2014).
[37] S. Bird, E. Klein, E. Loper, Natural Language Processing with Python, 1st Edition, O'Reilly Media, Inc., 2009
(2009).
[38] R. K. Pan, S. Sinha, K. Kaski, J. Saramäki, The evolution of interdisciplinarity in physics research, Scientific
reports 2 (2012) 551 (2012).
[39] S. Lee, B. Yoon, Y. Park, An approach to discovering new technology opportunities: Keyword-based patent map
approach, Technovation 29 (6-7) (2009) 481 -- 497 (2009).
[40] A. Canals, Knowledge diffusion and complex networks: a model of high-tech geographical industrial clusters, in:
Proceedings of the 6th European conference on organizational knowledge, Learning, and Capabilities, 2005, pp.
1 -- 21 (2005).
[41] M. A. Montemurro, D. H. Zanette, Keywords and co-occurrence patterns in the voynich manuscript: An
information-theoretic analysis, PloS one 8 (6) (2013) e66344 (2013).
[42] S. Schulz, C. M. Costa, M. Kreuzthaler, J. A. Miñarro-Giménez, U. Andersen, A. B. Jensen, B. Maegaard,
Semantic relation discovery by using co-occurrence information, Proceedings of BioTxtM 220 (2014).
[43] C. McNorgan, R. A. Kotack, D. C. Meehan, K. McRae, Feature-feature causal relations and statistical co-
occurrences in object concepts, Memory & Cognition 35 (3) (2007) 418 -- 431 (2007).
[44] W. Zhao, J. Mao, K. Lu, Ranking themes on co-word networks: Exploring the relationships among different
metrics, Information Processing & Management 54 (2) (2018) 203 -- 218 (2018).
[45] P. Klimek, A. S. Jovanovic, R. Egloff, R. Schneider, Successful fish go with the flow: citation impact prediction
based on centrality measures for term -- document networks, Scientometrics 107 (3) (2016) 1265 -- 1282 (2016).
[46] Y. Yang, R. Lichtenwalter, N. Chawla, Evaluating link prediction methods, Knowledge and Information Systems
45 (3) (2015) 751 -- 782 (12 2015). doi:10.1007/s10115-014-0789-0.
[47] C. Tan, Tracing community genealogy: How new communities emerge from the old, CoRR abs/1804.01990
(2018). arXiv:1804.01990.
URL http://arxiv.org/abs/1804.01990
21
[48] Y. Yang, N. Chawla, Y. Sun, J. Hani, Predicting links in multi-relational and heterogeneous networks, in: 2012
IEEE 12th international conference on data mining, IEEE, 2012, pp. 755 -- 764 (2012).
[49] P. R. da Silva Soares, R. B. C. Prudêncio, Time series based link prediction, in: The 2012 international joint
conference on neural networks (IJCNN), IEEE, 2012, pp. 1 -- 7 (2012).
[50] S. Hochreiter, J. Schmidhuber, Lstm can solve hard long time lag problems, in: Advances in neural information
processing systems, 1997, pp. 473 -- 479 (1997).
[51] B.-K. Choi, T. Dayaram, N. Parikh, A. D. Wilkins, M. Nagarajan, I. B. Novikov, B. J. Bachman, S. Y. Jung, P. J.
Haas, J. L. Labrie, et al., Literature-based automated discovery of tumor suppressor p53 phosphorylation and
inhibition by nek2, Proceedings of the National Academy of Sciences 115 (42) (2018) 10666 -- 10671 (2018).
[52] D. Hristovski, B. Peterlin, S. Dzeroski, Literature-based discovery support system and its application to disease
gene identification, in: Proceedings of the AMIA Symposium, American Medical Informatics Association, 2001,
p. 928 (2001).
[53] Q. He, Knowledge discovery through co-word analysis (1999).
[54] S. Spangler, A. D. Wilkins, B. J. Bachman, M. Nagarajan, T. Dayaram, P. Haas, S. Regenbogen, C. R. Pickering,
A. Comer, J. N. Myers, et al., Automated hypothesis generation based on mining scientific literature, in: Proceed-
ings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, ACM, 2014,
pp. 1877 -- 1886 (2014).
[55] D. R. Swanson, Asist award of merit acceptance speech: On the fragmentation of knowledge, the connection
explosion, and assembling other people's ideas, Bulletin of the American Society for Information Science and
Technology 27 (3) (2001) 12 -- 14 (2001).
[56] F. Dotsika, A. Watkins, Identifying potentially disruptive trends by means of keyword network analysis, Techno-
logical Forecasting and Social Change 119 (2017) 114 -- 127 (2017).
Author contributions
N.C. conceived the manuscript idea, collected data, design experiments and contributed in writing, and F.F cleaned the
datasets, conducted the experiment(s), developed features, analysed the results, and wrote the research methodology
section. M.K. contributed in writing and reviewed the manuscript.
8 Dataset and Code
The dataset and programming codes used in this study is available a https://github.com/faisal-iut/linkPrediction
22
|
1612.06115 | 2 | 1612 | 2016-12-24T11:40:36 | Complex Network Tools to Understand the Behavior of Criminality in Urban Areas | [
"cs.SI",
"physics.soc-ph"
] | Complex networks are nowadays employed in several applications. Modeling urban street networks is one of them, and in particular to analyze criminal aspects of a city. Several research groups have focused on such application, but until now, there is a lack of a well-defined methodology for employing complex networks in a whole crime analysis process, i.e. from data preparation to a deep analysis of criminal communities. Furthermore, the "toolset" available for those works is not complete enough, also lacking techniques to maintain up-to-date, complete crime datasets and proper assessment measures. In this sense, we propose a threefold methodology for employing complex networks in the detection of highly criminal areas within a city. Our methodology comprises three tasks: (i) Mapping of Urban Crimes; (ii) Criminal Community Identification; and (iii) Crime Analysis. Moreover, it provides a proper set of assessment measures for analyzing intrinsic criminality of communities, especially when considering different crime types. We show our methodology by applying it to a real crime dataset from the city of San Francisco - CA, USA. The results confirm its effectiveness to identify and analyze high criminality areas within a city. Hence, our contributions provide a basis for further developments on complex networks applied to crime analysis. | cs.SI | cs | Complex Network Tools to Understand the
Behavior of Criminality in Urban Areas
Gabriel Spadon, Lucas C. Scabora, Marcus V. S. Araujo, Paulo H. Oliveira,
Bruno B. Machado, Elaine P. M. Sousa, Caetano Traina-Jr, Jose F. Rodrigues-Jr
{spadon, lucascsb, pholiveira, araujo}@usp.br
{brandoli, parros, caetano, junio}@icmc.usp.br
University of Sao Paulo, Brazil
6
1
0
2
c
e
D
4
2
]
I
S
.
s
c
[
2
v
5
1
1
6
0
.
2
1
6
1
:
v
i
X
r
a
Abstract -- Complex networks are nowadays employed in sev-
eral applications. Modeling urban street networks is one of them,
and in particular to analyze criminal aspects of a city. Several
research groups have focused on such application, but until now,
there is a lack of a well-defined methodology for employing
complex networks in a whole crime analysis process, i.e. from
data preparation to a deep analysis of criminal communities.
Furthermore, the "toolset" available for those works is not
complete enough, also lacking techniques to maintain up-to-date,
complete crime datasets and proper assessment measures. In
this sense, we propose a threefold methodology for employing
complex networks in the detection of highly criminal areas within
a city. Our methodology comprises three tasks: (i) Mapping of
Urban Crimes; (ii) Criminal Community Identification; and (iii)
Crime Analysis. Moreover, it provides a proper set of assessment
measures for analyzing intrinsic criminality of communities,
especially when considering different crime types. We show our
methodology by applying it to a real crime dataset from the city of
San Francisco -- CA, USA. The results confirm its effectiveness to
identify and analyze high criminality areas within a city. Hence,
our contributions provide a basis for further developments on
complex networks applied to crime analysis.
I. INTRODUCTION
Complex networks have long been used to model social
behavior, spatial patterns, spreading of epidemics, and urban
structures. They are capable of representing neural connec-
tions, data transmission schemes, electric power grids, airports,
as well as street, rail and subway networks [1]. In the context
of street networks, which are of particular interest for this
work, complex networks are able to describe flow relationships
and social behavior in an urban zone [2].
As dynamic organisms, cities evolve (shrink or enlarge)
with respect to their space and population, causing changes in
their structures. Such structures are of high relevance as they
directly influence the flow within the cities and, consequently,
the behavior of their citizens. Nevertheless, violence and social
disorder may also be related to the urban structure of a given
city. For instance, low-flow areas tend to be propitious for
crime events, since such areas end up being less surveilled
when compared to high-flow areas. For that reason, crime
events usually occur within regions that can be detected
based on their structural and behavioral properties [3]. In this
context, it is our contention that, by means of computational
techniques, it is possible to identify and understand the crime
dynamics in an urban network. As so, this paper provides a
methodology for aiding the analysis of urban criminality --
e.g. how specific crimes occur in certain areas and how their
types are related to each other by being in neighboring areas,
that is, by being similar.
The hypothesis of this work is that by employing network
mapping techniques, allied to distance-based properties of
graphs, it is possible to identify and trace the relationship
between areas that are highly criminal within a city. Our
assumption is that similar crimes occur in adjacent regions.
Such crimes emerge from, among other factors, the urban
organization, and their inter-similarity carries correspondence
to the distance from one region to the other. Technically,
our proposal is based on: (i) community detection algorithms
applied to datasets of isolated crime types; (ii) pattern and
similarity analyses, which compare crimes occurred in distinct
regions within the same city. Based on those techniques, we
propose a threefold methodology consisted of: (i) Mapping
of Urban Crimes -- we describe how to combine a city's
georeferenced urban structure with crime records into a com-
plex network with potential for analytical tasks; (ii) Criminal
Community Identification -- we show how to detect criminal
communities by exploring geographical as well as structural
properties; (iii) Crime Analysis -- we methodologically com-
pare crime patterns by identifying relationships between their
communities, especially in the face the of diverse crime types.
Our methodology contributes to the understanding of crimi-
nality in the context of urban organization. We work on issues
related to: (i) an approach to identify highly criminal areas;
(ii) the characterization of the city space by inferring the
homogeneity of its crimes; (iii) the identification of city sectors
that present behaviors related to the sparsity of their crimes;
and (iv) the delineation of the city space by identifying regions
that are related, acting as crime spots for different crime types.
The remainder of the paper is structured as follows. Sec-
tion II presents the related work, discussing how our method-
ology stands out from theirs. Section III describes the dataset
used in this work and the methodology proposed. Section IV
discusses the results obtained from the application of our
methodology to the crimes dataset, which describes a real
scenario from the street network of the North American city
of San Francisco. Finally, Section V presents the conclusion.
II. RELATED WORK
Several works in the literature have dealt with the topics of
urban networks and criminality. This section describes some of
these works organized into two categories related to the phases
of our proposal: Mapping of Urban Crimes, which focuses
on representing raw crime data with a complex network,
and Crime Analysis, which refers to identifying criminal
behaviors and patterns by analyzing urban street networks.
Mapping of Urban Crimes. Spicer et al. [4] describe a
theoretical framework capable of mapping urban crimes into
related locations in a city. Their work discusses the pros and
cons of mapping methods. To represent a city, they use street
networks and GIS software, describing methods for mapping
georeferenced elements into a complex network. However,
their methods associate a crime with the nearest node or
edge based on address geocoding, not having any relation
with graph measures. Moreover, all methods are superficially
introduced, not presenting any formal fundaments.
Shinode and Shinode [5] aim at
introducing a search-
window method to analyze urban crimes in street networks.
They represent a search window as a subarea that concentrates
a high number of crimes in small regions; besides that, the
crimes are used to identify clusters spatially and temporally.
The authors validate their methods through an empirical case
study. However, their criminal data derive from 911-calls made
in 1996; that is, they used a dataset that carries few details and
that is outdated, especially when one considers the conclusions
claimed by the authors. Such information implies a superficial
analysis of the crime behavior and the ways to prevent it.
Those previous works aim at identifying highly criminal
areas to predict future criminal events. They use several
properties of the data to map them into a node, an edge or a
combination of both. However, they use non-graph measures
to build a city scenario, which fails in characterizing the
real world. Our approach, on the other hand, benefits from
modeling the real-world space via a georeferenced complex
network; a better representation of the reality. It allows us to
merge two georeferenced grids, one describing crimes and the
other outlining city elements, which enables us to characterize
crimes and extract their patterns.
Crime Analysis. White et al. [6] represent criminal activities
via a georeferenced complex network, through which they
compare criminal demographics and identify criminal com-
munities. The demographic regions are created by clustering
the urban areas using socioeconomic borders, whereas the
crime network is formed by linking crime locations that are
close according to the Euclidean distance. To identify the
criminal communities, the authors employ a label propagation
technique. Their results show that
the crime network can
be constructed without requiring information about individual
crimes. However, their crime network does not take the urban
topology into account, so their comparison is based only on
socioeconomic information.
Rey et al. [7] employ spatiotemporal techniques to represent
and analyze urban crimes. Their approach provides a way
to quantify neighborhood criminality and is able to identify
patterns from the data. Moreover, they discuss spatial crime
analyses as a way to determine the influence of a crime on
its surroundings, analyzing an entire city and providing an
evaluation of its crimes. However, their work is based on
outdated data, derived from a police district in the period from
2005 to 2009. Besides, their methods rely on black-box GIS
technologies, used to summarize criminal areas by geocoding
each crime into a quarter-mile grid cell. Furthermore, they
analyze only city cells having residential units.
Other works include the proposals of Galbrun et al. [8],
who focus on crime mapping to estimate the probability of a
crime occurrence within a street segment; of Fitterer et al. [9],
who use a statistical model to predict crimes in the city of
Vancouver; and of Bogomolov et al. [10], who predict crime
hotspots in London by using mobile data. Our work differs
from theirs because (i) we consider crimes within a more
complete and up-to-date period; (ii) we map crimes directly,
i.e. there is no space summarization in the crime mapping
phase; and (iii) we benefit from complex networks, which are
a reliable toolset to design georeferenced data.
Considering real crime events and a real-world city, like
those presented in this section, we introduce ways to map and
analyze crimes, supported by criminal communities extracted
from complex networks, allowing outcomes far more detailed
and advanced than the former works. We not only show how to
map crimes into georeferenced complex networks but also how
to analyze their impact on the city, as well as the relationships
among different crime types. Mainly, we describe results from
various points of view, for multiple crime types, and based on
several interrelated metrics.
III. PROPOSED METHODOLOGY
This work aims at analyzing the impact of the urban
structure on city criminal events, providing insights to enhance
the city planning in highly criminal areas. To achieve that goal,
we propose a methodology with three phases: (i) Mapping
of Urban Crimes; (ii) Criminal Community Identification;
and (iii) Crime Analysis. This section describes such phases
according to an application example.
A. Background and Datasets
We refer to an urban street network as a directed georef-
erenced graph G = {V, E} composed of a set E of E
edges (street segments) and another set V of V nodes (streets
intersections). We refer to an edge e ∈ E as an ordered pair
(cid:104)i, j(cid:105), i ∈ V and j ∈ V , in which i is named source and j is
named target. Each node v ∈ V has coordinates lat and lon,
where lat denotes the node's latitude and lon the longitude. We
refer to a set C, C ∩ V = ∅, of crimes, each with coordinates
lat and lon within the area of the city. Each crime c has a
value that belongs to a domain of 39 crime types -- the most
common types are assault, theft, and minor crimes.
Graph Source. Our approach uses electronic cartographic
maps extracted from the OpenStreetMap (OSM)
[11] plat-
form. OSM provides maps representing spatial elements, e.g.
points, lines and polygons, as objects. Such objects, which are
abstractions from the real-world geographical space, are rep-
resented by a triple of attributes that uniquely identifies them
ij = R × cos−1( sin(li
dE
cos(li
at)sin(lj
at)cos(lj
at) +
at)cos((cid:52)lon
ij ) )
(1)
in their set, granting each object an identification number and
georeferenced coordinates. These objects allow to represent
OSM data as a georeferenced graph.
Crime Source. The crime dataset C comes from the San
Francisco OpenData (SFO) initiative 1, a central clearinghouse
for public information about the city of San Francisco -- CA,
USA. SFO provides data concerning public issues. We use a
set of events that corresponds to the period between Jan 1st,
2003 through May 5th, 2016. It has 1,916,911 instances of 39
crime types.
B. Mapping of Urban Crimes
The first phase of our methodology receives the crime
dataset as input, denoted as set C, where each element c ∈ C
corresponds to a crime event occurred in a specific latitude
and longitude. We proceed by mapping the crime events to
the nodes of the urban street network, represented as the graph
G = {V, E}. Given a crime c, its corresponding network node
is the closest node v, v ∈ V , as given by the Euclidean distance
calculated from the coordinates of c and v. This distance refers
to the real length between elements thought the Earth's surface,
which is derived from the law of cosines and defined as:
ij
at are latitudes, (cid:52)lon
at and lj
on and lj
is the difference between
where li
on, R is the earth's radius (6,371km), and
longitudes li
ij denotes the Euclidean distance (E) between points i and j.
dE
All values are represented in radians [12]. With respect to the
georreferenced network, dE
ij also represents an edge's weight,
which refers to the street length (in meters) that connects the
streets' intersections (i.e. nodes) i and j.
Next, we enrich the network representation with a new prop-
erty, making it reflect the geographical distribution of crimes.
The result is a new set of nodes V (cid:48), with each node v ∈ V (cid:48)
having a set with three properties prop(v) = {lat, lon,Cv},
where lat is the latitude, lon is the longitude, and Cv is the
quantity of crimes mapped to node v. Formally, the set of
crimes mapped to a node v, denoted as Cv, is defined as:
Cv = {c ∈ C dE
wc,∀ w ∈ V, v (cid:54)= w}
vc < dE
(2)
Note that this phase is performed separately for each kind
of crime. For example, if performing the mapping for crime
type assault, we get a version of the graph in which the
nodes have a property indicating the number of crimes that
are categorized as an assault. After processing each kind
of crime, we get a new property in the nodes. In a map-
ping perspective, each processing yields a new graph that
differs with respect to the properties of the nodes; formally,
G = {G0, G1, . . . , Gn}, where each G = {V (cid:48), E}, ∀ G ∈ G.
C. Criminal Community Identification
Intuitively, a community is understood as a group of nodes
that have a probability of connecting to each other that is
greater than the probability of connecting to nodes out of the
group [13]. Formally, consider G = {V (cid:48), E} as a city graph
and p(i, j) as the probability of nodes i and j to define a
connection; a community set is understood as a subset of nodes
C. A node i is considered a member of the community c ∈ C if
its probability of connecting to a node j ∈ c is greater than the
probability of connecting to a node j(cid:48) in another community
c(cid:48) ∈ C, as follows:
p(i, j) ≥ p(i, j(cid:48)) i (cid:54)∈ c, i (cid:54)∈ c(cid:48), ∀ j ∈ c, ∀ j(cid:48) ∈ c(cid:48), c (cid:54)= c(cid:48)
The elements of a community define a connected induced
subgraph G(c) in which each node is reachable from all other
nodes of the same community. Formally, G(c) = {V (cid:48)(cid:48), E(cid:48)},
V (cid:48)(cid:48) ⊆ V (cid:48) ⇐⇒ {∀ i ∈ c, ∀ j ∈ c dij < ∞}, where
dij : V (cid:48) × V (cid:48) → R is an arbitrary distance function which
returns the distance from any given pair of nodes.
Next, we proceed with the identification of communities.
To identify them concerning each crime type, we opted for
the tool Nerstrand 2, which carries out a fast multi-thread
detection. This tool, which yields results broadly accepted in
the literature, identifies communities without any information
about how many of them exist in the city. The input for this
phase is an undirected graph, whereas the output is a set
of number-labeled communities. The tool allows the user to
attach weights to the graph elements; for our application, the
edges' weights are the distance between two distinct nodes,
and nodes' properties are the number of mapped crimes.
Our dataset comprises 39 crime types, however, processing
and analyzing all of them would exceed the available space of
this work. Besides, our techniques and results are generalizable
for any number of crime types. For those reasons, we worked
only with the three most recurrent ones, which comprise
i.e.
almost 40% of our dataset and represent: (i) assault,
to inflict injury on a person intentionally; (ii) theft, i.e. to
take a property off a person's possession by stealth, with no
brute force; and (iii) minor crimes, such as offenses that do
not involve any loss or injury. They represent, respectively,
167,832 (8.76%), 392,338 (20.47%) and 204,451 (10.67%)
crimes. As previously explained, for each crime type, we
perform a mapping over the graph, each one with a different
number of crimes mapped into their nodes. Further on, we
discuss that each crime-mapped graph can produce a specific
criminality scenario; that is, when considering each crime type,
we get a different set of communities, each one characterized
by that kind of crime and by the topology of the network.
The next step, after identifying communities, is to filter
out
the less relevant ones for each type of crime. From
the perspective of criminality analysis, relevance here means
choosing the highly criminal communities. This approach
reduces the excessive number of identified communities, as
well as characterizes the most relevant ones. Accordingly,
we worked with the top five highly criminal communities,
which were identified by using two variables: (i) the highest
criminal average, i.e. the average number of crimes per node;
1 Available on "data.sfgov.org".
2 Available on "www-users.cs.umn.edu/ lasalle/nerstrand/".
(ii) community size threshold, which prioritizes communities
containing at
least a specified number of nodes. For the
purpose of this work, the threshold has been set to 100 nodes,
since communities with at least 100 nodes are spatially well
distributed, and can better represent the regions of the city.
This filtering process is depicted in Figure 1, which shows the
average criminality on the x-axis and the number of nodes
on the y-axis. In the figure, we describe the less relevant
communities as the mischaracterized ones, and we show where
the most relevant ones are, considering the previously defined
threshold.
Fig. 1: Filtering process used to identify the relevant commu-
nities while discarding the mischaracterized ones. We sort the
communities by their average criminality, and we eliminate
the spatially-small ones (having less than 100 nodes). The
highlighted area indicates the communities that are, at the
same time, highly populated and highly criminal.
D. Crime Analysis
In this last phase, with the communities already detected,
we proceed to analyze their interrelationship and interaction.
Therefore, here, we consider the communities in pairs of types;
for example, we analyze communities of assault and theft to
verify whether they occur together or if one type of crime
makes the other more intense. The idea is to obtain insights
about their geospatial dependency, pointing to regions that
share space of more than one crime type and, consequently,
are potentially more dangerous. To this end, this phase has
two stages. The first stage obtains the geospatial similarity
between pairs of crime types, which allows determining how
close are their communities and how related are their criminal
activity. The second stage computes the Homogeneity and
Completeness scores, which are useful
to understand the
crimes' behaviors by evaluating specific aspects of each com-
munity set. Combined, both stages lead to a more meaningful
analytical scenario, accomplishing the contributions mentioned
at the end of Section I.
Measuring the Similarity of Communities
A similarity value enables us to characterize the relationship
of distinct events based on their spatial behavior, more specifi-
cally their spatial distribution. Such similarity is based on their
distances and considers the possibility of a non-criminal region
containing crimes as well. This is because a non-criminal area
may have its behavior affected by a close criminal area.
We proceed with the computation of the similarity measure
by considering the previously detected top five highly criminal
communities. The similarity of crime events is based on
the distance between elements belonging to any two sets of
communities E and F, where each one represents a distinct
crime type. Note that, at this point, after communities were
detected based on the topology of the city, we proceed using
only the nodes and their positioning -- i.e. we do not consider
the edges nor the city topology for this processing. Such
additional
information is not necessary since we consider
only the position of highly criminal communities in order to
characterize the geospatial relationship of crime types. Hence,
we define the similarity SK(E, F) : C x C → [0, 1]; the closer
to 0, the more spatially apart are the sets of nodes:
(cid:80)∀u∈E (cid:80)∀v∈F dE
uv
E + F
SK(E, F) = 1 −
Such measure identifies, at
the similarity
between intra-and inter-community elements. The result 1
denotes that the elements within community E are highly close
to each other and to the elements of community F, intra-and
inter-similar respectively, and 0 denotes the opposite.
the same time,
(3)
Identifying the Behavior of Crimes
The characterization of communities identified in a city can
be complemented by the Homogeneity and Completeness
scores. They are derived from an entropy-based cluster evalua-
tion [14], which works by analyzing the presence and absence
of criminality in the network's communities. Such scores
quantify how intensely each crime type manifests across a
community set C. To this end, the Homogeneity score evalu-
ates the criminality in a community by identifying the criminal
distribution among their nodes. Further,
the Completeness
score is used to quantify how intrinsic is the criminality in
a community by measuring their distribution among all the
communities, both in terms of a specific crime type.
Homogeneity. Regarding the top-five criminal communities,
it is desired to know what are their predominant aspects.
For the purpose of achieving a better assessment of criminal
communities, the Homogeneity score is able to measure how
uniform, i.e. homogeneous, a community is. It is defined by
H(C) : C x C → [0, 1] as follows:
where V (cid:48) is the number of nodes in all the communities
of set C (the top five in our example) of a crime type, and
Q = {1, 0}, which, in our scenario, indicates the nodes with
and without crimes, respectively. We used C to denote the
set that contains all communities and Ci ∈ C to indicate a
single community. Finally, Cij, j ∈ Q, represents nodes in
community i in which crimes are present or absent, whereas
Ck, k ∈ Q, represents the same idea but among all commu-
(cid:16)Cij
(cid:80)Q
(cid:80)C
(cid:16)Ck
(cid:17)
(cid:80)Q
Cij
V (cid:48) × log2
Ck
V (cid:48) × log2
V (cid:48)
Ci
k=0
j=0
i=1
(cid:17)
(4)
H(C) = 1 −
10-210-11001011021030100200300400500600700Number of nodesAverage crimeAssaultMinor CrimesTheftThresholdCommunities withhigh criminal activityand high populationHighly criminalcommunityHighly populatedcommunityMischaracterizedcommunities nities. For instance, Ck, k = 0, represents all nodes in all
communities that have no crime, while Cij, i = 1, j = 1,
indicates the nodes from community 1 that have crime(s).
The Homogeneity score equals 0 when the presence and
absence of crimes are proportionally similar across all the
communities. Contrarily, the value 1 denotes that all commu-
nities in C have crimes in all their nodes or no crime at all.
It is noteworthy that the Homogeneity score solely is insuffi-
cient to provide the needed comprehension about the criminal
behavior in a city. This is because a single community can
have all its nodes characterizing either crime occurrences or no
crime at all, which, in turn, could yield the same Homogeneity
score. For this reason, we also employ the Completeness
score that, combined with the Homogeneity score, can better
describe the distribution of crimes across a city. By analyz-
ing both scores, we can effectively compare different areas,
regardless of their high or low criminality. Their purpose is
to identify the communities' behaviors, mainly their tendency
for high/low scores to a given crime type.
Completeness. If crimes do not concentrate in a community,
it is likely that there are multiple communities with crimi-
nal nodes spread across the city, demanding surveillance at
multiple criminal spots. Alternatively, if crimes concentrate,
they tend to be grouped within a single community, requiring
police patrols on specific regions. To determine if crimes are
more or less concentrated, we employ the Completeness score,
which is directly obtained from the whole community set. The
Completeness score equals 0 when the presence of crimes is
totally scattered, i.e. equally absent across all communities --
their occurrence is totally distributed. On the other hand, the
higher the value, the more concentrated are the criminal and
safe zones. This is defined in C(C) : C x C → [0, 1] as:
(cid:16)Cij
(cid:80)Q
(cid:80)C
(cid:17)
(cid:16)Ck
(cid:80)C
Cij
V (cid:48) × log2
i=1
Ck
V (cid:48) × log2
Cj
V (cid:48)
k=1
j=0
(cid:17)
C(C) = 1 −
(5)
Hence, if safe and criminal zones are totally apart, the Com-
pleteness score of a community set equals 1. Conversely, if
criminality has a proportionally equivalent participation in all
communities, the score is 0.
IV. RESULTS AND DISCUSSIONS
This section presents and discusses the results of the exper-
iments regarding the three phases of our methodology.
Influence of Crime Mapping. The first experiment aimed at
analyzing the impact of the urban crime mapping (first phase
of our methodology) on the criminal community identification
process. This process was carried out twice: once considering
the topological graph, i.e. the graph that considers only nodes
and edges, and once over the network after the urban crime
mapping, regarding the newly-included edges' weights (dis-
tances in meters) and nodes' weights (number of crimes). As
a result, we obtained the number of 134 communities from
the topological graph and, when considering the information
introduced by the crime mapping phase,
the numbers of
identified communities were: 210 for assault, 215 for theft
and 211 for minor crimes.
As a first result, we noticed the discrepancy between
the numbers of criminal communities for each graph. Such
difference is due to the presence/absence of crime-related
data derived from the crime mapping phase. The Nerstrand
algorithm considers the crime-related data (the nodes' and
edges' weights), ensuring that the nodes within each identified
community share similar characteristics. This behavior is not
the same when considering the topological graph. Since there
is no crime information attached to its nodes and edges,
the algorithm examines only the connections between nodes.
Therefore, considering only the network's topology, there is no
guarantee that the crimes will be optimally grouped among the
identified communities. These findings show that identifying
communities through a topological graph is an inaccurate way
to represent crime relationships in a city.
Communities' Design Aspects. The second experiment aimed
at comparing the results from the previous analysis by consid-
ering their design aspects. Table I presents the most highly
criminal communities obtained directly from the network's
topology and from the three graphs generated via the Crime
Mapping process. The table shows twelve communities sorted
by crime average, among which are the top five criminal ones.
TABLE I: Communities with the highest crime average, iden-
tified from: (i) topology, (ii) assaults, (iii) theft and (iv) minor
crimes. Column Avg denotes the crime average and # denotes
the number of nodes in each community.
Communities identified from:
Topology
Assault
Theft
Avg
1063.00
157.67
116.30
110.75
108.80
77.93
77.50
65.86
44.17
43.88
36.28
35.88
#
2
788
802
4
5
515
2
539
744
1129
635
518
Avg
362.00
230.59
55.26
47.44
41.00
34.69
34.06
29.53
27.42
22.14
15.52
14.20
#
2
17
316
226
2
87
149
238
142
248
281
54
Avg
4030.00
245.00
201.00
60.15
49.74
47.08
37.88
32.25
25.92
25.77
21.79
21.10
#
3
2
17
281
277
156
66
4
296
342
307
371
00
01
02
03
04
05
06
07
08
09
10
11
Minor Crimes
Avg
#
12
545.92
2
457.00
17
356.41
67
281.91
142
140.13
110.15
188
5
73.80
143
72.99
286
71.37
4
68.50
58.72
32
144
54.43
The table describes a higher crime average in communities
derived from topology only. Such higher average occurs be-
cause, in this context, there is a lower number of communities,
each one containing a higher number of criminal nodes.
As shown in the table, the number of crimes per community
still stands out when comparing the graph from topology and
the "complete" graphs -- i.e. the graphs generated by the
Crime Mapping process. This is because the complete graph
used no attribute that characterizes a crime in the community
identification process, so there is no way to assure that they
are related just by being in the same community. Therefore,
hereafter we discuss only the graphs with crime-related data.
Regarding such graphs, Table I shows a significant number
of communities with few nodes. This is evidence that the
Fig. 2: The most highly criminal region from San Francisco -- CA, USA, identified by evaluating the crime density in the city
from 01/01/2003 to 05/18/2016. Each colored area represents the occurrence of a crime type or the intersection of a crime
type with another type of crime. In the legend, each CM (from 1 to 3) denotes a type -- assault, theft and minor crimes.
crimes are not evenly distributed in the network. Therefore, if
someone eventually attempted to eradicate or prevent crimes in
the entire city based on characteristics of a local region, such
attempt would fail because the crimes are not concentrated in
a single criminal community.
San Francisco's Crime Neighborhood. To allow analyzing
a group of communities containing a significant number of
crimes, we have selected the region from the city of San
Francisco that comprises the highest crime indices. This region
contains the five communities, as mentioned earlier, of each
crime type, which are among the twelve communities shown
in Table I. The top five communities of each type also meet
the requirement regarding the number of nodes, previously
determined as a threshold of 100 nodes (see Section III-D).
Specifically, considering the table, the top five communities for
assault are in lines 2, 3, 6, 7 and 8; the top five for theft are
in lines 3, 4, 5, 8 and 9; and the top five for minor-crimes are
in lines 4, 5, 7, 8 and 11. Figure 2 depicts the entire selected
area of San Francisco, as well as highlights each community
(and each geospatial intersection of communities).
To quantify the relationship between communities in the city
space, we used Equation 3 to obtain the similarity between
communities. Through this measure, we compared each pair
of crime types, computing the similarity between community
sets E of one type and community sets F of another type. By
doing so, we achieved a similarity value SK(E, F) of 0.66 for
assault vs. theft, 0.65 for assault vs. minor crimes and 0.58
for theft vs. minor crimes. Such results suggest that there is a
relationship between crimes of different types, especially when
comparing assault vs. theft and assault vs. minor crimes, since
the values 0.65 and 0.66 can be considered significant enough
if one ponders over what these numbers represent -- that the
majority of the area of these two communities might be more
dangerous due to the occurrence of more than one crime type.
More specifically: (i) it is very likely that criminal activities
occur at intersections of multiple crime types, which we call
crime hubs -- regions that attract or disseminate criminality;
(ii) those crimes are spatially related, so it is possible that they
have been caused by the same people or gang; (iii) places,
where more than one crime type occurs, are high-priority
regions for public safety improvement.
Communities' Intrinsic Criminality. The previous section
discussed the geospatial disposition of different crime types
with respect to intersections between criminal regions. This
section, in turn, brings results concerning the Homogeneity
and Completeness scores for the communities in the most
highly criminal area of San Francisco. This experiment aimed
at measuring how intrinsic was the criminality within the
neighborhoods of San Francisco during the studied period.
To understand this measure, recall that the number of crimes
of some nodes is zero. For the assault type, there are 1,071
nodes, of which 132 (12.33%) contain at least one crime and
939 (87.68%) contain none. For theft, there are 903 nodes, of
which 133 (14.73%) are criminal and 770 (85.27%) are not
criminal. For minor crimes, there are 1,352 nodes, of which
177 (13.09%) present crimes and 1,175 (86.91%) do not.
These values and percentages refer to the entire community
set for each crime type.
The presence and absence of crimes were considered to
calculate the scores presented in Table II, which shows that
assault has the lowest values for both measures. Regarding
Homogeneity, the value indicates that this crime type presents
a reduced concentration of criminal nodes within each com-
munity. This is straightly inferable because, besides the low
Homogeneity value indicating an inexpressive occurrence of
crimes, the actual number of crimes is very low; a pattern that
holds for other types of crime as well.
Since the value for Completeness is also the closest to
0, we can reliably assume the occurrence of crimes to be
scattered across all the assault communities. The theft type has
values close to 0 as well, but the highest ones when compared
to the values of other crime types. The greater difference
Highly Criminal AreaSan Francisco MapGraph CommunitiesCM1 - AssaultCM2 - TheftCM3 - Minor CrimesCM1 ∩ CM2CM1 ∩ CM3CM2 ∩ CM3CM1 ∩ CM2 ∩ CM3LegendN interventions were made by the governors were effective to
prevent them, using time windows and comparing the results
with socioeconomic data.
Acknowledgments. We are grateful to CNPq (National Coun-
sel of Technological and Scientific Development) by the sup-
port (grants 9254-601/M, 444985/2014-0, and 147098/2016-
5), to FAPESP (Sao Paulo Research Foundation) by the assis-
tance (grant 2015/15392-7, 2016/02557-0, 2016/17330-1), to
Capes (Brazilian Federal Agency for Support and Evaluation
of Graduate Education) and to the Databases and Images
Group (GBdI) by the resources ceded to this research.
REFERENCES
[1] S. Boccaletti, V. Latora, Y. Moreno, M. Chavez, and D. U. Hwang,
"Complex networks: Structure and dynamics," Phys. Rep., vol. 424, no.
4-5, pp. 175 -- 308, feb 2006.
[2] S. Porta, E. Strano, V. Iacoviello, R. Messora, V. Latora, A. Cardillo,
F. Wang, and S. Scellato, "Street centrality and densities of retail and
services in Bologna, Italy," Environ. Plan. B Plan. Des., vol. 36, no. 3,
pp. 450 -- 465, 2009.
[3] R. Deryol, P. Wilcox, M. Logan, and J. Wooldredge, "Crime Places
in Context: An Illustration of the Multilevel Nature of Hot Spot
Development," J. Quant. Criminol., vol. 32, no. 2, pp. 305 -- 325, jun
2016.
[4] V. Spicer, J. Song, P. Brantingham, A. Park, and M. A. Andresen, "Street
profile analysis: A new method for mapping crime on major roadways,"
Appl. Geogr., vol. 69, pp. 65 -- 74, apr 2016.
[5] S. Shiode and N. Shiode, "Network-based space-time search-window
technique for hotspot detection of street-level crime incidents," Int. J.
Geogr. Inf. Sci., vol. 27, no. 5, pp. 866 -- 882, may 2013.
[6] H. Serrano, M. Oliveira, and R. Menezes, "The spatial structure of crime
in urban environments," Lect. Notes Comput. Sci. (including Subser.
Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 8852, pp. 102 --
111, 2015.
[7] S. J. Rey, E. A. Mack, and J. Koschinsky, "Exploratory Space-Time
Analysis of Burglary Patterns," J. Quant. Criminol., vol. 28, no. 3, pp.
509 -- 531, sep 2012.
[8] E. Galbrun, K. Pelechrinis, and E. Terzi, "Urban navigation beyond
shortest route: The case of safe paths," Inf. Syst., vol. 57, pp. 160 -- 171,
apr 2016.
[9] J. Fitterer, T. A. Nelson, and F. Nathoo, "Predictive crime mapping,"
Police Pract. Res., vol. 16, no. 2, pp. 121 -- 135, 2015.
[10] A. Bogomolov, B. Lepri, J. Staiano, N. Oliver, F. Pianesi, and A. Pent-
land, "Once Upon a Crime," Proc. 16th Int. Conf. Multimodal Interact.
- ICMI '14, vol. 1, pp. 427 -- 434, 2014.
[11] M. Haklay and P. Weber, "OpenStreet map: User-generated street maps,"
IEEE Pervasive Comput., vol. 7, no. 4, pp. 12 -- 18, oct 2008.
[12] T. Konstantopoulos, Introduction to projective geometry. Dover Publi-
cations, 2012.
[13] E. Estrada, M. Fox, D. J. Higham, and G. L. Oppo, Network science:
Complexity in nature and technology. Cambridge university press, 2010.
[14] A. Rosenberg and J. Hirschberg, "V-measure: A conditional entropy-
based external cluster evaluation measure," Proc. Jt. Conf. Empir.
Methods Nat. Lang. Process. Comput. Nat. Lang., vol. 1, no. June, pp.
410 -- 420, 2007.
TABLE II: The Homogeneity and Completeness scores used
to evaluate the quality of each community set and to quantify
how intrinsic is the criminality within them.
Crime type
Theft
Minor Crimes
Assault
Homogeneity
0.015588
0.014014
0.013557
Completeness
0.004159
0.003441
0.003234
corresponded to the Completeness score, whose value suggests
a geospatial crime concentration more elevated. Regarding
Completeness (see Section III-D), such higher concentration
tends to occur in a smaller set of communities rather than
across all of them, which is more likely for the assault type.
A higher crime concentration means that a particular crime
type is less likely to intersect with another crime type. This is
because, in such a case, the criminal nodes would occupy a
smaller area in the city space. Furthermore, such a crime type
could be considered easier to prevent, since its occurrences
would be in a more compact area of the city. On the other
hand, a lower concentration of a crime type means that it exerts
a stronger influence on the network as a whole, being easier to
propagate to remote areas and demanding global approaches
both to eradicate it and to prevent it.
V. CONCLUSION
In this paper, we propose a threefold methodology for
identifying and analyzing criminal spatial patterns in urban
street networks. Our methodology is based on the hypothesis
that by employing network mapping techniques, allied to
distance-based properties of graphs, it is possible to identify
and trace the relationship between areas that are highly
criminal within a city. To demonstrate our methodology, we
analyzed criminal communities from real crime data rep-
resenting assaults,
thefts and minor crimes in the city of
San Francisco -- CA, USA. The methodology comprises the
following phases: (i) Mapping of Urban Crimes, (ii) Criminal
Community Identification and (iii) Crime Analysis. The latter
employs the well-established Homogeneity and Completeness
scores to analyze the identified criminal communities more
deeply. Our main achievement confirm the hypothesis of our
work, allowing us to state that different crime types share
common spaces, characterizing areas that lack strategies for
crime prevention, and that particular crime types are sparser
than others in the city space.
The highlighted contributions are: (i) the use of a complex
network to represent the real-world space, enabling a complete
analysis of the city; (ii) independence from socioeconomic
information, allowing the analyses of crimes based on solely
their spatial disposition; (iii) the assessment of the impact of
criminal regions, considering the similarity between distinct
crime types and their Homogeneity and Completeness scores.
Finally, as future work, we intend to continue expanding
our methodology developing tools to analyze crimes and their
types considering a temporal perspective. Our aim is to provide
insights about the criminal behaviors in a city, allowing to
determine how a crime spreads throughout the city and when
|
1502.02401 | 1 | 1502 | 2015-02-09T09:04:11 | Random Preferential Attachment Hypergraphs | [
"cs.SI",
"physics.soc-ph"
] | The random graph model has recently been extended to a random preferential attachment graph model, in order to enable the study of general asymptotic properties in network types that are better represented by the preferential attachment evolution model than by the ordinary (uniform) evolution lodel. Analogously, this paper extends the random {\em hypergraph} model to a random {\em preferential attachment hypergraph} model. We then analyze the degree distribution of random preferential attachment hypergraphs and show that they possess heavy tail degree distribution properties similar to those of random preferential attachment graphs. However, our results show that the exponent of the degree distribution is sensitive to whether one considers the structure as a hypergraph or as a graph. | cs.SI | cs | Random Preferential Attachment Hypergraphs∗
Chen Avin †§
Zvi Lotker †
David Peleg ‡
November 11, 2018
Abstract
The random graph model has recently been extended to a random preferential attachment
graph model, in order to enable the study of general asymptotic properties in network types
that are better represented by the preferential attachment evolution model than by the ordinary
(uniform) evolution lodel. Analogously, this paper extends the random hypergraph model to a
random preferential attachment hypergraph model. We then analyze the degree distribution
of random preferential attachment hypergraphs and show that they possess heavy tail degree
distribution properties similar to those of random preferential attachment graphs. However, our
results show that the exponent of the degree distribution is sensitive to whether one considers
the structure as a hypergraph or as a graph.
Keywords: Random Hypergraphs, Preferential attachment, Social Networks, Degree Distri-
bution.
5
1
0
2
b
e
F
9
]
I
S
.
s
c
[
1
v
1
0
4
2
0
.
2
0
5
1
:
v
i
X
r
a
†Department of Communication Systems Engineering, Ben Gurion University of the Negev, Beer-Sheva, Israel.
E-mail: {avin,zvilo}@cse.bgu.ac.il.
§Part of this work was done while the author was a long term visitor at ICERM, Brown University
‡Department of Computer Science and Applied Mathematics, The Weizmann Institute of Science, Rehovot, Israel.
E-mail: [email protected]. Supported in part by the I-CORE program of the Israel PBC and ISF (grant
4/11).
∗Supported in part by the Israel Science Foundation (grant 1549/13).
1
Introduction
Random structures have proved to be an extremely useful concept in many disciplines, including
mathematics, physics, economics and communication systems. Examining the typical behavior of
random instances of a structure allows us to understand its fundamental properties. The founda-
tions of random graph theory were first laid in a seminal paper by Erdos and R´enyi in the late
1950's [7]. Subsequently, several alternative models for random structures, often suitable for other
applications, were suggested. One of the most important alternative models is the preferential
attachment model [2], which was found particularly suitable for describing a variety of phenomena
in nature, such as the "rich get richer" phenomena, which cannot be adequately simulated by the
original Erdos-R´enyi model. It has been shown that the preferential attachment model captures
some universal properties of real world social networks and complex systems, like heavy tail degree
distribution and the "small world" phenomenon [12].
One limitation of graphs is that they only capture dyadic (or binary) relations. In real life, how-
ever, many natural, physical and social phenomena involve k-ry relations for k > 2, and therefore
can be more accurately represented by hypergraphs than by graphs. For example, collaborations
among researchers, as manifested through joint coauthorships of scientific papers, may be bet-
ter represented by hyperedges and not edges. Figure 1(a) depicts the hypergraph representation
for coauthorship relations on four papers: paper 1 authored by {a, b, e, f}, paper 2 authored by
{a, c, d, g}, paper 3 authored by {b, c, d} and paper 4 authored by {e, f}. Likewise, wireless com-
munication networks [1] or social relations captured by photos that appear in Facebook and other
social media also form hyperedges [14]. Affiliation models [10, 13], which are a popular model for
social networks, are commonly interpreted as bipartite graphs, where in fact they may sometimes
be represented more conveniently as hypergraphs. Figure 1(b) presents the bipartite graph repre-
sentation of the hypergraph H of Figure 1(a). Sometimes, one can only access the observed graph
G(H) of the original hypergraph H, that is, only the pairwise relation between players is available
(see Figure 1(c)). In some cases this structure may be sufficient for the application at hand, but in
many other cases the hypergraph structure is more accurate and informative/
The study of hypergraphs, and in particular random hypergraph models, has its roots in a 1976
paper by Erdos and Bollobas [3], which offers a model analogous to the Erdos-R´enyi random graph
model [7]. Recently, several interesting properties regarding the evolution of random hypergraphs
in this model were studied in [5, 6, 8].
The current paper is motivated by the observation that, just as in the random graph case, the
random hypergraph model is not suitable for studying social networks. Our first contribution is in
extending the concept of random preferential attachment graphs to random preferential attachment
hypergraphs. We believe the this natural model will turn out to be useful in the future study of
social networks and other complex systems.
The main technical contribution is that we analyze the degree distribution of random preferen-
tial attachment hypergraphs and show that they possess heavy tail degree distribution properties,
similar to those of random preferential attachment graphs. However, our results show that the ex-
ponent of the degree distribution is sensitive to whether one considers the structure as a hypergraph
or as a graph.
As a reference point, we consider the random preferential attachment graph model of Chung
and Lu [4]. In that model, starting from an initial graph G0, at any time step there occurs an
event of one of two possible types: (1) a vertex-arrival event, occuring with probability p, where a
new vertex joins the network and selects its neighbor among the existing vertices via preferential
2
(a)
(b)
(c)
Figure 1:
of H. (c) The observed graph G(H).
(a) A hypergraph H with 7 vertices and 4 edges. (b) A bipartite graph representation
attachment, or (2) an edge-arrival event, occuring with probability 1−p, where a new edge joins the
network and selects its two endpoints from among the existing vertices via preferential attachment.
It is shown in [4] that the degree distribution of the random preferential attachment graph follows
a power law, i.e., the probability of a random vertex to be of degree k is proportional to k−β, with
βG = 2 + p
2−p . A similar result can be shown in a setting where, at each time step, d edges join the
graph instead of only one (in either a vertex event or an edge event)[12]. This result holds even if
at each step a random number of edges join the network, so long as the expected number of new
edges is d and the variance is bounded.
The model proposed here extends Chung and Lu's [4] model to support hypergrpahs. That is,
the process starts with an initial hypergrpah, and at each time step a random hyperedge joins the
network. With probabilty p this new random hyperedge includes a new vertex, and with probabilty
1 − p it does not. Our model allows the hyperedge sizes to be random (with some restrictions) and
the members of each edge are selected randomly according to preferential attachment.
We show that the degree distribution of the resulting hypergraph (as well as the observed
µ−p , where µ is the expected size of
graph) follows a power law, but with an exponent βH = 2 + p
an hyperedge.
Our results indicate that one should be careful when studying an observed graph of a general
k-ry relation. In particular, it makes a difference if the observed graph was generated by a graph
or by a hypergraph evolution mechanism, since the two generate observed graphs with different
degree distributions.
In the next sections we describe in more detail the preferential attachment model of a hyper-
graph, and then analyze the resulting degree distribution.
2 Preliminaries
A hypergraph H is an ordered pair (V,E), where V is a set of n vertices and E ⊆(cid:83)n
Given a set V and a natural k > 1, let V (k) be the set of all unordered vectors (or multisets) of
k elements from V . A finite undirected graph G is an ordered pair (V, E) where V is a set of n
vertices and E ⊆ V (2) is the set of graph edges (unordered pairs from V , including self loops).
i=2 V (i) is a
set of hyperedges connecting the vertices (including self loops). The rank r(H) of a hypergraph H
is the maximum cardinality of any of the hyperedges in the hypergraph. When all hyperedges have
the same cardinality k, the hypergraph is said to be k-uniform. A graph is thus simply a 2-uniform
3
abcdefg1234c2dga1efb34cdbgaef hypergraph. The degree of a hyperedge e ∈ E is defined to be δ(e) = e. The set of all hyperedges
that contain the vertex v is denoted E(v) = {e ∈ E v ∈ e}. The degree d(v) of a vertex v is the
number of hyperedges in E(v), i.e., d(v) = E(v). H is d-regular if every vertex has degree d.
In the classical preferential attachment graph model [2], the evolution process starts with an
arbitrary finite initial network G0, which is usually set to a single vertex with a self loop. Then this
initial network evolves in time, with Gt denoting the network after time step t. In every time step
t a new vertex v enters the network. On arrival, the vertex v attaches itself to an existing vertex
u chosen at random with probability proportional to u's degree at time t, i.e.,
dt(u)
,
w∈Gt
dt(w)
P [u is chosen] =
where dt(x) is the degree of vertex x at time t.
(cid:80)
3 The nonuniform preferential attachment hypergraph model
Similar to the classical preferential attachment graph model [4], the evolution of the hypergraph
occurs along a discrete time axis, with one event occurring at each time step. We consider two types
of possible events on the hypergraph at time t: (1) a vertex arrival event, which involves adding a
new vertex along with a new hyperedge, and a hyperedge arrival event, where a new hyperedge is
added.
We consider a nonuniform, random hypergraph where self loops (i.e., multiple appearance of a
vertex in a hyperedge) are allowed. We consider self loops as contributing 1 to the vertex degree.
Similar to [4], our preferential attachment model, H(H0, p, Y ), has three parameters:
• A probability 0 < p ≤ 1 for vertex arrival events.
• An initial hypergraph H0 given at time 0.
• A sequence of random independent integer variables Y = (Y0, Y1, Y2, . . . ), for Yi ≥ 2, which
determine the cardinality of the new hyperedge arriving at time t.
The process by which the random hypergraph H(H0, p, Y ) grows in time is as follows.
• We start with the initial hypergraph H0 at time 0.
• At time t > 0, the graph Ht is formed from Ht−1 in the following way:
-- Randomly draw a bit b with probability p for b = 0.
-- If b = 0, then add a new vertex u to V , select Yt − 1 vertices from Ht−1 (possibly
with repetitions) independently in proportion to their degrees in Ht−1, and form a new
hyperedge e that includes u and the Yt − 1 selected vertices∗.
-- Else, select Yt vertices from Ht−1 (possibly with repetitions) independently in proportion
to their degrees in Ht−1, and form a new hyperedge e that includes the Yt selected
vertices.
Hereafter, we consider an initial H0 consisting of a single hyperedge of cardinality Y0 over a
single vertex (recall that self-loops are considered as contributing 1 to the vertex degree).
∗note that as the hypergraph gets larger, the probability of adding a self-loop is vanishing.
4
4 Degree Distribution Analysis
To ensure convergence of the degree distribution we first need to set some conditions on the dis-
tribution of the hyperedge cardinalities. These are somewhat mild conditions that seems to agree
with real data (see Fig. 4 in Section 5). Let Yt be independent (not necessarily identical) random
3 †. Under
variables with constant expectation E[Yt] = µ and bounded support s.t. 2 < Yt < t
these conditions we can show the following.
Theorem 4.1. The degree distribution of a hypergraph H(H0, p, Y ) where E[Yt] = µ follows a
power law with β = 2 + p/(µ − p).
1
Proof. We start with properties of Yt. Let St =(cid:80)t
Lemma 4.2. P(cid:104)St − E[St]) ≥ t
of St from its expected value can be bounded.
2 log t
(cid:105)
√
2
3
< O(1/t4).
1 Yt, so E[St] = µt and St < t
4
3 . The deviation
Proof. By Hoeffding's inequality [9], assuming the random variable Yi satisfies P[Yi ∈ [ai, bi]] = 1
for some reals ai and bi,
P [St − E[St]) ≥ x] ≤ 2 exp
(cid:33)
(cid:32)
(cid:80)t
i=1(bi − ai)2
3 and(cid:80)t
i=1(bi − ai)2 < t
2x2
−
.
2
Taking x = t
√
2
3
2 log t and noting that (bi − ai)2 < t
4
3 yields the result.
To bound the degree distribution of a non-uniform random hypergraph we closely follow Chung
and Lu's analysis on preferential attachment graphs [4]. Let mk,t denote the number of vertices of
degree k at time t. Note that m1,0 = 0 and m0,t = 0. We derive the recurrence formula for the
expected value E[mk,t]. The main observation here is that a vertex has degree k at time t if either
it had degree k at time t− 1 and was not selected into a hyperedge at time t, or it had degree k − 1
at time t − 1 and was selected into a hyperedge at time t. Letting Ft be the σ-algebra associated
with the probability space at time t, we have for any t > 0 and k > 0:
E[mk,tFt−1] = mk,t−1
(cid:32)
(cid:34)(cid:18)
pEYt
(cid:32)
pEYt
(cid:34)(cid:32)
+mk−1,t−1
+ (1 − p)EYt
(cid:34)(cid:18)
(cid:19)Yt−1(cid:35)
1 − k
(cid:34)(cid:18)
St−1
1 − (1 − k − 1
(cid:18)
1 − k − 1
St
1 −
St
+ (1 − p)EYt
(cid:35)
(cid:19)Yt−1
(cid:19)Yt(cid:33)(cid:35)(cid:33)
)
,
(cid:19)Yt(cid:35)(cid:33)
1 − k
St−1
†The exponent 1
3 is chosen somewhat arbitrarily; the result can be extended to any constant 0 ≤ α < 1
2 .
5
hence
E[mk,tFt−1] = mk,t−1
(cid:32)
p − EYt
St
+1 − p − EYt
(cid:32)
St
(cid:21)
− O
(cid:20) (Yt − 1)kp
(cid:32)
(cid:20) (1 − p)Ytk
(cid:21)
(cid:20) (Yt − 1)p(k − 1)
(cid:21)
(cid:20) (1 − p)Yt(k − 1)
(cid:18) k2
St
St
+EYt
EYt
+ O
S2
t
+mk−1,t−1
(cid:19)(cid:19)
(cid:32)
+mk−1,t−1
(cid:18)
1 − (µ − p)k
(cid:32)
St
or
E[mk,tFt−1] = mk,t−1
Using the bound on St we can find the expectation E[mk,t].
E[mk,t] = (1 − 1/t4)
E[mk,t−1]
1 −
+E[mk−1,t−1]
= E[mk,t−1]
1 −
(µ − p)k
√
2
3
+ O
(cid:32)
µt ± t
(µ − p)(k − 1)
√
µt ± t
2 log t
(µ − p)k
2
3
√
µt ± t
2
3
(cid:32)
2 log t
(µ − p)(k − 1)
µt ± t
2 log t
√
2
3
+ O
(cid:18)
(cid:32)
(cid:32)
For t > 0 and the special case of k = 1 we have
E[m1,tFt] = m1,t−1
1 − (µ − p)k
St
+ O(
(cid:19)2(cid:35)(cid:33)(cid:33)
(cid:19)2(cid:35)(cid:33)(cid:33)
(cid:18) (k − 1)2
S2
t
(cid:19)(cid:19)
.
(cid:34)(cid:18) Ytkp
(cid:32)
(cid:32)
EYt
St
(cid:19)2(cid:35)(cid:33)
(cid:34)(cid:18) (1 − p)Ytk
(cid:34)(cid:18) Ytpk
(cid:19)2(cid:35)(cid:33)
(cid:34)(cid:18) (1 − p)k
St
EYt
− O
− O
(cid:21)
St
+ O
(cid:19)(cid:33)
1
t4 · t4/3
− O
St
St
EYt
EYt
(cid:32)
(cid:18) (µ − p)(k − 1)
(cid:18) k2
(cid:19)(cid:33)(cid:33)
(cid:18) k2
(cid:19)(cid:33)
(cid:18) k2
(cid:19)(cid:33)
(cid:18) k2
(cid:19)
2 log t
+ O
t2
t2
t2
t2
+
k2
S2
t
)
(cid:33)
+ p ,
+ p + O(1/t2) .
+E[mk−1,t−1]
+ O
+ O(1/t2) .
thus
E[m1,t] = E[m1,t−1]
1 −
(µ − p)k
√
µt ± t
2
3
2 log t
+ O(
k2
t2 )
We use the following lemma of [4].
[4] Let (at), (bt), (ct) be three sequences such that at+1 =(cid:0)1 − bt
Lemma 4.3.
b > 0 and limt→∞ ct = c. Then limt→∞(at/t) exists and equals c/(1 + b).
t
(cid:1) at +ct, limt→∞ bt =
We show by induction that limt→∞ E[mk,t]/t exists and has a limit Mk for each k. For k = 1,
apply Lemma 4.3 with
bt =
µ ± t
µ − p
√
2
3
2 log t/t
+ O(k2/t)
and ct = p + O(1/t2)
6
lim
t→∞ bt =
µ − p
µ
and
lim
t→∞ ct = p,
M1 = lim
t→∞
E[m1,t]
t
=
µp
2µ − p
We now assume that limt→∞E[mk−1,t]/t exists and apply Lemma 4.3 again with
bt =
(cid:32)
(µ − p)k
√
µ ± t
2
3
2 log t/t
(µ − p)(k − 1)
µ ± t
2 log t/t
√
2
3
+ O
+ O
ct =
E[mk−1,t−1]
t
.
(cid:19)
(cid:19)(cid:33)
(cid:18) k2
(cid:18) k2
t
t
(cid:18) 1
(cid:19)
t2
.
+ O
and hence
to get
and
Then
(µ − p)k
t→∞ bt = b =
lim
µ
and
t→∞ ct = c = Mk−1(µ − p)(k − 1)/µ ,
lim
and by Lemma 4.3 we get that limt→∞E[mk,t]/t exists and satisfies
Mk = Mk−1
(µ − p)(k − 1)
µ(1 + k(µ − p)/µ)
= Mk−1
(k − 1)
k + µ
µ−p
.
(1)
Recall that a power law distribution has the property that Mk ∝ k−β for large k.
Now if Mk ∝ k−β, then
(cid:18)
(cid:19)β
Mk
Mk−1
=
k−β
(k − 1)−β =
1 − 1
k
= 1 − β
k
+ O
(cid:19)
(cid:18) 1
(cid:18) 1
k2
+ O
k2
.
(cid:19)
,
By Eq. (1),
Mk
Mk−1
=
k − 1
k + µ
µ−p
= 1 − 1 + µ
µ−p
k + µ
µ−p
= 1 − 1 + µ
µ−p
k
so the exponent β of the power law satisfies
β = 1 +
µ
µ − p
= 2 +
p
µ − p
.
A special case of H(H0, p, Yt) is when Yt is the constant function d and the hypergraph becomes
a d-uniform hypergraph denoted as H(H0, p, d).
Corollary 4.4. The degree distribution of a d-uniform hypergraph H(H0, p, d) follows a power law
with β = 2 + p/(d − p).
Figure 2 illustrates the difference in exponents β between preferential attachment graphs (i.e.,
2-uniform hypergraphs) and 3-uniform hypergraphs as a function of p.
In many cases one can only observe the graph G[H] that results of the underlying hypergrph H.
That is, the set of vertices of G(H) is identical to the set of vertices of H and for every hyperedge
e ∈ H we create edges in G(H) to form a clique between all the vertices in e. Now we can prove
the following.
7
Claim 4.5. The degree distribution of the observed graph G(H(H0, p, d)) that results from a d-
uniform hypergraph follows a power law with β = 2 + p/(d − p).
Figure 2: The exponent β of a preferential attachment graph and a 3-uniform hypergraph as a
function of p (the probability of an edge arrival event). In graphs it is between 2 and 3, whereas in
3-hypergraphs it is between 2 and 2.5.
Note that the expected degree of vertices in G(H) in this case is d(d − 1)/2.
Interestingly,
if we generate a new graph G(cid:48) with expected degree d(d − 1)/2 according to the classical graph
preferential attachment model, then its degree distribution will be β(cid:48) = 2 + p/(2 − p). Hence the
observed degree distribution of G(H) and G(cid:48), β and β(cid:48) respectively, will be different. On the other
hand, it we generate G(cid:48) (using the classical preferential attachment model) so that it agrees with the
degree distribution of G, then the average degree will be different. This observation is supported
by simulation results depicted in Figure 3.
This discussion seems to indicate that, in some sense, "the blanket (i.e., of the model) is too
short" and one should be careful in deciding what is the right model that captures the observed
degree distribution, and in particular, if the generative model is of a hypergraph or the classical
graph model.
5 Example
To test the above observations empirically, we studied a coauthorship hypergraph of researchers
in computer science, extracted from DBLP [11], a dataset recording most of the publications in
computer science. This hypergraph consists of hundreds of thousands of vertices (representing
authors) and hyperedges (representing papers). Figure 4 shows the degree distribution of hyperedge
sizes in DBLP for hyperedges sizes at least 3. The hyperedge size distribution closely fits a power
law degree distribution with exponent β = 4.66. This means that the hyperedge size is (with high
probabilty) smaller than m1/3, where m is the number of papers (hyperedges). For the example of
8
H(cid:42),p,3:Β(cid:61)2(cid:43)p3(cid:45)pPAgraph:Β(cid:61)2(cid:43)p2(cid:45)p0.00.20.40.60.81.02.02.22.42.62.83.0pΒ Figure 3: Example of the cumulative degree distribution of three networks with n = 10, 000: (1)
A graph G(H(∗, 1, 3)) derived from a 3-uniform hypergraph H(∗, 1, 3), (2) A graph G(H(∗, 1
2 , 3))
derived from a 3-uniform hypergraph H(∗, 1
2 , 3), and (3) A preferential attachment graph with
average degree d(d − 1)/2 = 3. Graphs derived from hypergraphs have lower exponent, also as a
function of p.
DBLP, where the number of papers is m = 2420879, the number of authors on a paper (i.e., the
hyper-edge size) will be with high probability below 134.
Acknowledgements
The authors thank Eli Upfal for helpful discussions about the core ideas of the paper.
9
GH(cid:42),1,3GH(cid:42),12,3PAGraphp(cid:61)1,(cid:60)deg(cid:62)(cid:61)3Β(cid:61)2.43Β(cid:61)2.24Β(cid:61)2.91110100100010(cid:45)40.0010.010.11NodedegreeProbabilty Figure 4: The degree distribution of hyperedge sizes in DBLP for hyperedge sizes at least 3. The
distribution closely fits a power law degree distribution with exponent β = 4.66.
10
EdgesizefrequencyΒ(cid:61)4.66097510205010010(cid:45)710(cid:45)610(cid:45)510(cid:45)40.0010.010.1EdgeSizenumberofauthorsProbability References
[1] Avin, C., Lando, Y., and Lotker, Z. Radio cover time in hyper-graphs. Ad Hoc Networks
12 (2014), 278 -- 290.
[2] Barab´asi, A.-L., and Albert, R. Emergence of scaling in random networks. Science 286,
5439 (1999), 509 -- 512.
[3] Bollob´as, B., and Erdos, P. Cliques in random graphs. In Mathematical Proc. Cambridge
Philosophical Soc. (1976), vol. 80, Cambridge Univ Press, pp. 419 -- 427.
[4] Chung, F. R. K., and Lu, L. Complex graphs and networks. No. 107. AMS, 2006.
[5] Cooper, C., Frieze, A., Molloy, M., and Reed, B. Perfect matchings in random r-
regular, s-uniform hypergraphs. Combinatorics, Probability and Computing 5 (1996), 1 -- 14.
[6] Ellis, D., and Linial, N. On regular hypergraphs of high girth.
arXiv preprint
arXiv:1302.5090 (2013).
[7] Erdos, P., and R´enyi, A. On the evolution of random graphs. Publ. Math. Inst. Hungar.
Acad. Sci 5 (1960), 17 -- 61.
[8] Ghoshal, G., Zlati´c, V., Caldarelli, G., and Newman, M. Random hypergraphs and
their applications. Physical Review E 79, 6 (2009), 066118.
[9] Hoeffding, W. Probability inequalities for sums of bounded random variables. J. Amer.
Statistical Assoc. 58, 301 (03 1963), 13 -- 30.
[10] Lattanzi, S., and Sivakumar, D. Affiliation networks. In Proc. 41st ACM Symp. on Theory
of computing (2009), ACM, pp. 427 -- 434.
[11] Ley, M. Dblp: some lessons learned. Proc. VLDB Endowment 2 (2009).
[12] Newman, M. Networks: an Introduction. Oxford University Press, 2010.
[13] Newman, M. E., Watts, D. J., and Strogatz, S. H. Random graph models of social
networks. Proc. Nat. Acad. Sci. 99, suppl 1 (2002), 2566 -- 2572.
[14] Zhang, Z.-K., and Liu, C. A hypergraph model of social tagging networks. J. Statistical
Mechanics: Theory and Experiment 2010, 10 (2010), P10005.
11
|
1609.02839 | 1 | 1609 | 2016-09-09T15:48:50 | Where is the Goldmine? Finding Promising Business Locations through Facebook Data Analytics | [
"cs.SI",
"cs.IR"
] | If you were to open your own cafe, would you not want to effortlessly identify the most suitable location to set up your shop? Choosing an optimal physical location is a critical decision for numerous businesses, as many factors contribute to the final choice of the location. In this paper, we seek to address the issue by investigating the use of publicly available Facebook Pages data---which include user check-ins, types of business, and business locations---to evaluate a user-selected physical location with respect to a type of business. Using a dataset of 20,877 food businesses in Singapore, we conduct analysis of several key factors including business categories, locations, and neighboring businesses. From these factors, we extract a set of relevant features and develop a robust predictive model to estimate the popularity of a business location. Our experiments have shown that the popularity of neighboring business contributes the key features to perform accurate prediction. We finally illustrate the practical usage of our proposed approach via an interactive web application system. | cs.SI | cs | Where is the Goldmine? Finding Promising Business
Locations through Facebook Data Analytics
Jovian Lin, Richard Oentaryo, Ee-Peng Lim, Casey Vu, Adrian Vu, Agus Kwee
Living Analytics Research Centre, Singapore Management University, 80 Stamford Road, Singapore
[email protected], {roentaryo, eplim, caseyanhthu, adrianvu, aguskwee}@smu.edu.sg
6
1
0
2
p
e
S
9
]
I
S
.
s
c
[
1
v
9
3
8
2
0
.
9
0
6
1
:
v
i
X
r
a
ABSTRACT
If you were to open your own cafe, would you not want to ef-
fortlessly identify the most suitable location to set up your shop?
Choosing an optimal physical location is a critical decision for nu-
merous businesses, as many factors contribute to the final choice
of the location.
In this paper, we seek to address the issue by
investigating the use of publicly available Facebook Pages data-
which include user "check-ins", types of business, and business
locations-to evaluate a user-selected physical location with re-
spect to a type of business. Using a dataset of 20,877 food busi-
nesses in Singapore, we conduct analysis of several key factors in-
cluding business categories, locations, and neighboring businesses.
From these factors, we extract a set of relevant features and develop
a robust predictive model to estimate the popularity of a business
location. Our experiments have shown that the popularity of neigh-
boring business contributes the key features to perform accurate
prediction. We finally illustrate the practical usage of our proposed
approach via an interactive web application system.
Keywords
Location analytics, Facebook, feature extraction, machine learning
1.
INTRODUCTION
Motivation. Location is a crucial factor of retail success, as 94%
of retail sales are still transacted in physical stores [24]. To increase
the chance of success for their stores, business owners require not
only the knowledge of where their potential customers are, but
also their surrounding competitors and complementary businesses.
From the property owners' standpoint, it is also important to assess
the potential success values of their property locations so as to de-
termine the appropriate businesses to lease the locations to and for
the right amounts. However, assessing and picking a store location
is a cumbersome task for both business and property owners.
To carry out the above tasks well, many factors need to be taken
into account, each of which requires gathering and analyzing the
relevant data. Traditionally, business and property owners conduct
surveys to assess the value of store locations [3]. Such surveys,
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].
HT '16, July 10-13, 2016, Halifax, NS, Canada
c(cid:13) 2016 ACM. ISBN 978-1-4503-4247-6/16/07. . . $15.00
DOI: http://dx.doi.org/10.1145/2914586.2914588
however, are costly and do not scale up well. With fast chang-
ing environments (e.g., neighborhood rental, local population size,
composition, etc.) and emergence of new business locations, one
also needs to continuously reevaluate the value of store locations.
Fortunately, in the era of social media and mobile apps, we have
an abundance of online user-generated data, which capture both
activities of users in social media as well as offline activities at
physical locations. Facebook is one of the world's largest social
media platforms, with more than 1 billion active users everyday
[21]. From the business standpoint, the massive availability of user,
location, and other behavioral data in Facebook is attractive, and
has changed the way people do businesses. For instance, many
small/medium business owners are now setting up Facebook Pages
to: (i) allow customers to find their businesses on Facebook; (ii)
connect with customers via "likes" and "check-ins"; (iii) reach out
to more customers through advertising their business pages on Face-
book; and (iv) conduct analytics of their pages to get a deeper un-
derstanding of their customers and marketing activities.
Consumers are also adapting both their online and offline behav-
iors to the introduction of Facebook Pages for businesses. Other
than "liking" businesses on their Facebook Pages, they can do a
"check-in" whenever they physically visit the respective business
stores. Facebook Pages have turned many offline signals into on-
line behavior that can be analyzed for business insights. In particu-
lar, features such as "likes" and "check-ins" can be used as indica-
tors of popularity, and by extension, success. Similarly, Instagram,
Twitter, and Foursquare also have variants of these quantitative sig-
nals that can be retrieved from their geotagged photos, tweets, and
tips. These data allow us to study the dynamics of brick-and-mortar
stores and discover meaningful patterns and insights that will help
retail and property owners make better decisions.
Objective.
In this work, we make use of data collected from
Facebook Pages to answer important research questions such as:
"Where should an owner set up his physical retail store at, so as to
optimize the store's popularity?", "What are the important factors
influencing a store's popularity?", and "Is there a "local" effect,
whereby businesses can benefit from the presence of more popu-
lar/established neighbors?" To this end, we propose a new location
analytics framework that operates on top of Facebook Pages data.
The centerpiece of our current framework is the following predic-
tion task: Given a target location that a business/property owner
wants to hypothetically set his/her store at, how can we extract the
relevant data of businesses within the vicinity of the target location
and use them to estimate the popularity of the target location?
As an illustration, Figure 1 shows a visualization of our web ap-
plication system that realizes our location analytics framework. For
the system's input, a business/property owner first drops a blue pin
on the map that indicates the hypothetical location of his/her new
• To concretely realize our idea, we have built an interactive
web application that allows a business/property owner to drop
a pin on the map and obtain a predicted "check-in" score
for that location. The application is available at http://
research.larc.smu.edu.sg/bizanalytics/.
Paper outline. Section 2 first provides an overview of related
works. In Section 3, we describe the Singapore Facebook Pages
data we use in our experiments. We subsequently elaborate our pro-
posed location analytics approach in Section 4. Section 5 presents
our experimental setup, followed by the results and analyses in Sec-
tion 6. We present our web application prototype in Section 7, and
finally conclude the paper in Section 8.
2. RELATED WORK
Our work can be viewed as a new type of location analytics [11],
which is an emerging area related to business intelligence (BI) [2].
In recent years, organizations have relied on BI tools to delve into
their data and reveal key insights that can aid their decision-making
processes. With these tools, businesses have been able to make in-
formed decisions based on what happened and when - typically
pertaining to sales figures and supplier transactions. Lately, there
is an important trend for organizations to address the question of
where. Conventional BI systems, however, lack location-related
analytics capabilities, and thus do not consider geographic and de-
mographic factors crucial for consumer analysis, e.g., where to set
up stores, warehouses, or marketing campaigns.
Previously, combining separate BI and location-based approaches
such as geographic information systems (GIS), was privileged only
to large enterprises such as oil/gas-exploration companies, trans-
portation companies, or government agencies [2]. These technolo-
gies involve costly data acquisition processes and specialized la-
bor skills. Moreover, their integration requires complex and time-
consuming implementation. A recent survey by ESRI and IT Me-
dia firm TechTarget [4] discovered that many organizations now
believe that it is important to look at business data in a geograph-
ical context. Today, location-based data are abundant, thanks to
the large volumes of user traces available from social media (such
as Foursquare and Facebook) as well as mobile devices. However,
many organizations are still unaware of the value of location-based
data and struggle to put them to effective use.
Using data from social media to understand the dynamics of a
society has always been a popular research theme, particularly, in
recommending a new location to a user. For example, Facebook
researchers Chang and Sun [1] analyzed Facebook users' "check-
ins" data to develop models that predict where users will "check-in"
next. They were able to predict user-to-user friendships (i.e., friend
recommendation) just by the "check-in" data alone. Gao et al. [9]
explored the use of Foursquare "check-ins" and temporal effects for
the task of location-recommendation; subsequently, the data were
also used to predict a user's location [10].
Recent works on social media-based location analytics largely
focus on detecting events and predicting user mobility patterns, al-
though their use for BI applications are still limited so far. For
instance, Li et al. [14] presented a machine learning method to dis-
cover and profile the user's location based on their following net-
work and tweet contents. Noulas et al. [17] used Foursquare data
to study the problem of predicting the next venue a mobile user
will visit, by exploiting transitions between types of venues, mo-
bility flows between venues, and spatio-temporal patterns of user
"check-ins". Also based on Foursquare data, Karamshuk et al. [13]
demonstrated the power of geographic (e.g., types and density of
nearby places) and user mobility (e.g., transitions between venues
Figure 1: Prediction of "check-ins" based on the location indicated
by the blue pin. The red pins represent nearby existing businesses.
store. Our system then retrieves the relevant information about the
area nearby, which are also occupied by the existing businesses as
indicated by the red pins. Based on these inputs, the system ex-
tracts a set of features and invokes a machine learning algorithm to
predict the "check-in" score for the target location, which in turn
serves as an indicator for the potential popularity of that location.
Contributions. In this paper, we show how publicly available
Facebook Pages data can be analyzed and used to predict the poten-
tial popularity of a business location. To our best knowledge, this
is the first work that demonstrates the feasibility of using Facebook
data for business location analytics and, in particular, for aiding
business and property owners to evaluate the value of a store loca-
tion. It is also worth noting that our work presents a fine-grained
approach that allows the business/property owners to estimate the
popularity of any point on the city map. Our approach can be eas-
ily extended to predict multiple points simultaneously as well. We
summarize our main contributions as follows:
• We present a new study on business location analytics using
Facebook Pages data. Specifically, we conduct detailed anal-
yses on 20,877 Facebook Pages of food-related businesses in
Singapore, which constitute one of the largest business types
in the city with generally healthy visitor traffic. Based on the
analyses, we identify key features that can be used to extract
insights, as well as suitable metric for business popularity.
• We develop a location analytics framework that includes a
rich feature extraction module as well as a fast and accu-
rate predictive model based on gradient boosting machine
(GBM) [8]. Unlike the previous close work on optimal store
placement [13] that outputs a ranked list of discretized ar-
eas (circles) with fixed radius, our approach is much more
fine-grained. That is, our model can estimate on the fly the
popularity of any arbitrary point on the map-which can be
the location of an existing or a new/hypothetical business-
without needing the locations to be discretized a priori.
• Based on our (trained) predictive model, we analyze the con-
tribution of key features that are crucial in predicting the
popularity of a retail business-at both chunk (i.e., a group
of features) and individual feature levels. In particular, we
discover that distance-dependent features such as the total
"check-ins" of businesses within certain radius are of utmost
importance. We then provide an in-depth investigation on
whether businesses, particularly smaller ones, benefit from
the existence of other popular businesses within its vicinity.
or incoming flow of users) features in predicting the best placement
of retail stores. In a similar vein, Georgiev et al. [12] conducted a
study to predict the rise and decline of popularity of the local re-
tail shops during the 2012 London Olympic Games. Most recently,
Zhang et al. [25] extracted traffic and human mobility features from
Manhattan restaurants data and studied how static and dynamic fac-
tors affect the economic outcome of local businesses in the city.
Our approach. Our work differs from all the above studies in
several ways. Firstly, to our best knowledge, our work is the first
to explore the use of Facebook data in business-location analyt-
ics. With 1.55 billion monthly active users and 50 million busi-
ness pages [21], Facebook can provide a more comprehensieve
database of crowdsourced locations than other platforms (by com-
parison, Foursquare only has 55 million monthly active users and
1.3 million business pages [22]). Secondly, instead of recommend-
ing places for users to establish retail stores or analyze on how
unique events will affect businesses,we predict the popularity score
of a user-selected venue, giving the user more freedom to choose
anywhere he/she wants to set up his/her business. Thirdly, among
all the works, Karamshuk et al.'s [13] is the closest to ours. But the
key difference is that their work discretized the city into multiple
circles with fixed radius and treated the issue as a "ranking prob-
lem", i.e., producing a ranked list of discretized circles. In contrast,
we view it as a "prediction problem" and provides a much more
fine-grained approach of estimating the popularity of any point on
the map. Our method also works robustly on a range of radius val-
ues, instead of relying on a single predefined radius as in [13].
3. FACEBOOK PAGES DATASET
In this section, we first provide an overview of the data that we
collected from Facebook, and then describe the important attributes
found in the data. We then conduct a simple analysis on the two
popularity measures-"check-ins" and "likes"-to determine the
better metric for quantifying the popularity of a business.
3.1 Data Harvesting
In this paper, we focus our studies on food-related businesses
found in the Facebook of Singapore. We choose food because it
constitutes one of the largest business types in Singapore with gen-
erally healthy visitor traffic ("check-ins" and "likes"). The food-
related businesses were defined based on a manually-curated list
that consists of 133 food-related categories of business, such as
those containing the words "restaurant", "pub", "bar", etc. In par-
ticular, we consider food-related businesses in Singapore Facebook
Pages that explicitly specify latitute-longitude coordinates, and these
coordinates must be within the physical boundaries of Singapore.
Using Facebook's Graph API [6], we obtained a total of 82,566
business profiles within Singapore boundaries, of which we cat-
egorically filtered 20,877 (25.2%) profiles that are food-related.
All business data were analyzed in aggregate, and no personally-
identifiable information was used.
Figure 2 shows an example of one such business profile, Wimbly
Lu Chocolates, with important attributes (highlighted in bold) such
as: (i) ID, (ii) category (i.e., the primary-category), (iii) category
list (i.e., the sub-categories), (iv) "check-in" count, (v) "like" count,
and (vi) location (including latitude and longitude). Figure 3 shows
the corresponding Facebook Page of the business profile.
3.2 Categories Data
From the 20,877 food-related businesses, we retrieved a total of
357 unique categorical labels (as standardized by Facebook) from
the attribute "category list", which represents the sub-categories of
a business. These categories contain not only food-related labels
Figure 2: Facebook Graph API provides a JSON-formatted data
of a business. Shown here is the Facebook Page of the Wim-
bly Lu Chocolates cafe in Singapore. Source: https://graph.
facebook.com/200823339955298/
Figure 3: The Facebook page of Wimbly Lu Chocolates that shows
key popularity metrics such as 8.131 "likes" and 22K "check-ins".
(i.e., "bakery", "bar", "cafe", "coffee shop"), but also non-food la-
bels such as "movie theatre", "shopping mall", and "train station."
The existence of non-food related labels within food businesses is
Facebook's way of allowing business owners to choose more than
one categorical label for their business profile. For example, a Star-
bucks outlet located at a train station in an airport would likely have
a mixture of food and non-food labels, such as "airport", "cafe",
"coffee shop", and "train station." In addition, there is an intimate
relationship between the categories of a target business and those
of its neighbors. For instance, a family-run cafe will unlikely set it-
self next to an established coffee franchise like Starbucks, whereas
a dessert shop may be located near complementary dining places.
Table 1 shows the top 25 categories of the food businesses in
Singapore, their expected "check-ins", and the percentage of busi-
nesses that perform better than expectation. The proportion of busi-
nesses performing better than expectation ranges from 6 to 28%.
The largest category is "food and restaurant", which is the most
common category-type. From the low percentages of those that
actually perform better than the expectation, we can tell that busi-
nesses obey a long-tail distribution, with the majority of businesses
being unable to achieve the expected "check-ins" or more.
{ "id": "200823339955298", "attire": "Casual", "can_post": false, "category": "Restaurant/cafe", "category_list": [ { "id": "197871390225897", "name": "Cafe" } ], "checkins": 22811, "cover": { "cover_id": 465179196853043, "offset_x": 0, "offset_y": 24, "source": "https://fbcdn-sphotos-g-a.akamaihd.net/hphotos-ak-xaf1/t31.0-8/s720x720/471918_465179196853043_134634759_o.jpg", "id": "465179196853043" }, "description": "Chocolate cafe.", "general_manager": "Rachel Foo", "has_added_app": false, "hours": { "tue_1_open": "12:30", "tue_1_close": "22:30", "wed_1_open": "12:30", "wed_1_close": "22:30", "thu_1_open": "12:30", "thu_1_close": "22:30", "fri_1_open": "12:30", "fri_1_close": "23:00", "sat_1_open": "09:00", "sat_1_close": "23:00", "sun_1_open": "09:00", "sun_1_close": "22:30" }, "is_community_page": false, "is_published": true, "likes": 8134, "link": "https://www.facebook.com/WimblyLuChocolates", "location": { "city": "Singapore", "country": "Singapore", "latitude": 1.3474290978105, "longitude": 103.86750392653, "street": "15-2, Jalan Riang", "zip": "358987" }, "name": "Wimbly Lu Chocolates", "parking": { "lot": 0, "street": 1, "valet": 0 }, "payment_options": { "amex": 0, "cash_only": 0, "discover": 0, "mastercard": 1, "visa": 1 }, "phone": "(65) 62891489", "price_range": "$ (0-10)", "public_transit": "Nearest MRT Serangoon: Take bus 105 from bus interchange or take bus 58 from the bus stop. Alight on the second stop and turn into Jalan Girang road. Take the first left and the second right.\nNearest MRT Lorong Chuan: Exit the MRT and walk towards NYJC and turn into Jalan Girang.\nSee you! :)", "restaurant_services": { "delivery": 0, "catering": 0, "groups": 0, "kids": 0, "outdoor": 0, "reserve": 0, "takeout": 1, "waiter": 1, "walkins": 1 }, "restaurant_specialties": { "breakfast": 0, "coffee": 1, "dinner": 0, "drinks": 1, "lunch": 0 }, "talking_about_count": 185, "username": "WimblyLuChocolates", "website": "www.wimblylu.com", "were_here_count": 22811 } Table 1: Top 25 categories of food-businesses, their expected "check-ins", and those that perform better than the expectation.
Categories
Food & Restaurant
Restaurant
Cafe
Shopping Mall
Coffee Shop
Fast Food Restaurant
Food & Grocery
Bakery
Chinese Restaurant
Food Stand
Bar
Japanese Restaurant
Train Station
Nightlife
Movie Theater
Cafeteria
Seafood Restaurant
Italian Restaurant
Thai Restaurant
Ice Cream Parlor
Sushi Restaurant
Pub
Night Club
Indian Restaurant
Businesses
count
6,758
5,233
3,126
3,101
2,959
2,840
1,449
1,338
1,157
1,099
956
922
879
744
717
661
629
459
437
413
380
369
361
350
Total "check-ins"
5,771,148
8,195,356
3,799,849
5,772,105
2,395,000
3,447,999
1,055,175
394,982
2,376,432
1,454,820
3,234,860
1,297,228
429,740
866,619
1,015,632
421,774
1,786,933
735,620
593,546
744,514
741,305
513,009
1,416,278
538,624
Expected "check-ins" % of Businesses that have "check-ins"
per business
above the expected "check-ins"
853.97
1,566.09
1,215.56
1,861.37
809.40
1,214.08
728.21
295.20
2,053.96
1,323.77
3,383.74
1,406.97
488.90
1,164.81
1,416.50
638.08
2,840.91
1,602.66
1,358.23
1,802.70
1,950.80
1,390.27
3,923.21
1,538.93
13.36%
16.09%
19.10%
15.90%
13.99%
19.12%
6.21%
11.58%
23.16%
10.92%
20.08%
22.02%
26.17%
10.62%
9.09%
11.20%
22.58%
27.67%
26.32%
18.40%
26.58%
20.33%
14.40%
18.00%
3.3 Location Data
Each business profile has a location attribute that contains the
physical address and latitude-longitude coordinates (hereafter known
as "lat-long"). Knowing the location of every business allows us
to calculate the neighborhood of a selected business through the
spatial distribution of other businesses around the vicinity. Specif-
ically, we consider the set Pl = {pdist(p, l) ≤ r} of places p
that lie in a radius r around a target location l. The term dist(p, l)
denotes the Haversine distance [20] between two locations p and l.
We can then create a two-dimensional distance matrix containing
the distance between every pair of business. For efficiency, we only
consider a maximum radius of 1km (i.e., r ≤ 1km). This allows us
to quickly retrieve the k nearest neighbors of any location.
3.4 Visitor Data
Ideally, we would like to analyze customer information, such
as who commented on or "liked" a business' Facebook post, and
match/recommend some user profiles to some businesses. How-
ever, due to privacy concerns, Facebook does not allow us to iden-
tify who has checked-in or liked a particular business. Although
one can still crawl the posts on a business' wall, as of Facebook's
Graph API v2.0 [5] (released in 2014), Facebook no longer supplies
a user's actual ID. Instead, Facebook uses the concept of "app-
scoped user IDs", whereby a user's ID is unique to each app and
cannot be used across different apps. As our crawler is considered
an app, and Facebook limits the number of user posts that an app
can query in a day, we are unable to gather enough posts-and by
extension user IDs-to cover all (food-related) businesses in Sin-
gapore. Having multiple crawlers will not work either, as the same
user ID will be different for any two crawlers.
3.5 Popularity Indicator: "Check-in" vs "Like"
Facebook provides two possible indicators for a business page's
popularity: "check-in" and "like". The "check-in" metric is com-
mon in location-based social media like Facebook and Foursquare.
Meanwhile, the "like" metric (shown as a "thumbs up" button) is
(a) Target business' total "check-
ins" vs. neighbors' total checkins
(b) Target business' total "check-
ins" vs. neighbors' total likes
Figure 4: Pearson correlation coefficient (PCC) of neighboring to-
tal "check-ins" (in red) and neighboring total "likes" (in blue).
more unique to Facebook, allowing users to express their recom-
mendation/support for an entity. A "check-in" is the action of reg-
istering one's physical presence, and the total number of "check-
ins" received by a business gives us a rough estimate of how pop-
ular and well-received it is. In contrast, the number of "likes" lit-
erally reflects an online vote for the business.
Intuitively, there-
fore, a "check-in" should be a more suitable measure of a physical
store's popularity, as it indicates a physical presence. Furthermore,
"check-in" can be repeated, i.e., a user could "check-in" to a place
on Monday and do so again on Tuesday. By contrast, "likes" cannot
be done repeatedly-it is a one-time event.
To prove this point, we compute the Pearson correlation coeffi-
cient (PCC) on two pairs: (i) the target business' "check-ins" w.r.t.
its neighbors' total "check-ins", and (ii) the target business' "check-
ins" w.r.t. its neighbors' total "likes". We only use the number of
"check-ins" for the target business because we are only interested
in the physical presence of customers for the target business. But
for the target business's neighbors, we use both "check-ins" and
"likes", as they reflect the popularity-physical or metaphysical-
of the area in which the target business is located. For the neigh-
bors' total "check-ins" and total "likes", we further partitioned them
based on the relative distance from the target business. Specifically,
for every target business, we calculate the PCC between its "check-
ins" and the total "check-ins" or "likes" of its neighbors within ra-
dius r, where r = {50, 100, 150, 200, . . . , 500}.
Figure 4 shows the PCC of the two popularity indicators, broken
down by the relative distance. It is evident that, between the neigh-
bors' "check-ins" and the neighbors' "likes", the "check-in" fea-
ture is the better indicator as it has a higher PCC score than "likes."
Furthermore, nearer "check-ins" (e.g., 50 meters) have better PCC
than further "check-ins", which suggests that the nearer a target
business is to a popular neighbor, the more "check-ins" it reaps.
On the contrary, the PCC score for "likes" increases as the distance
between a target business and its surrounding neighbors increases.
This can be attributed to the nature of "likes", which reflects an
online support for the business and is not limited to physical prox-
imity, whereas "check-ins" represent the registration of a person's
physical presence, which is determined by physical proximity.
4. PROPOSED FRAMEWORK
Our location analytics framework, as illustrated in Figure 5, con-
sists of two phases: training and prediction. The training phase
involves extracting a set of features from the existing business pro-
files and feeding them to a machine learning algorithm (i.e., gradi-
ent boosting [8]; see Section 4.3) in order to a predictive model for
the "check-in" count. In turn, the prediction phase involves extract-
ing features from a target business profile and invoking the trained
predictive model to generate a (predicted) "check-in" count for that
profile. Note that the training phase is carried out offline, whereas
the prediction phase is done on the fly for a (new) target profile.
The modules in our proposed framework consist of three main
types: (i) input profiles, (ii) feature extractor, and (iii) predictive
model. We shall describe each module type in turn.
4.1
Input Profile
The input profile represents a physical business, and is used in
both training (Figure 5(a)) and prediction (Figure 5(b)) phases. An
input profile contains several attributes of a business, namely:
• The lat-long coordinate of the business. This is used in
both training and prediction phases. Note that, during train-
ing, we only use the lat-long of the existing business profiles,
whereas during prediction the lat-long being queried can be
at an arbitrary (new or existing) location.
• The business categories. Examples include "bar", "cafe",
"dining", "train station", etc.. This information is also used
in both training and prediction phases.
• The "check-in" counts. In the training phase, we feed the
actual "check-in" counts of the existing businesses as the tar-
get variable of our algorithm. In the prediction phase, the
"check-in" counts of the queried locations are assumed to
be unknown and our algorithm is supposed to predict them,
whether it is for an existing or a new location.
All input profiles of the existing businesses are stored in a database
of place profiles (cf. Figure 5). Using this database, we can extract
a set of features for a given business, which include features derived
from its own (input) profile as well as features from its neighbors
(computed based on a range of radius as described in Section 3.3).
The next section describes our feature extraction procedure.
Figure 5: Our location analytics framework.
Target's category
[v][asian restaurant]
[v][bakery]
[v][bar]
[v][cafe]
[v][chinese restaurant]
...
Neighbors' category
[v neighbors][asian rest.]
[v neighbors][bakery]
[v neighbors][bar]
[v neighbors][cafe]
[v neighbors][chinese rest.]
...
Category related
C1
C2
C3
C4
C5
C6
Total "check-ins" of Food Neighbors
[Total "check-ins"][Food neighbors][distance],
where distance ∈ {50m, 100m, ..., 1000m}
Average "check-ins" of Food Neighbors
[Average "check-ins"][Food neighbors][distance],
where distance ∈ {50m, 100m, ..., 1000m}
Total "check-ins" of All Neighbors
[Total "check-ins"][All neighbors][distance],
where distance ∈ {50m, 100m, ..., 1000m}
Average "check-ins" of All Neighbors
[Average "check-ins"][All neighbors][distance],
where distance ∈ {50m, 100m, ..., 1000m}
Hotspots related
Figure 6: A break down of the feature vector into six chunks.
4.2 Feature Extraction
The feature extraction module serves to construct a feature vec-
tor representing a particular business. In this work, we divide our
feature vector into six chunks, which represent different aspects of
a target business. Figure 6 summarizes our feature chunks. The
first two are associated with categorical data, while the remaining
four are about hotspots (i.e., location and "check-ins") data. Table 2
summarizes the unique identifier (ID), description, and the number
of features of each chunk. We describe each chunk below.
Chunk C1: The categories of the target business. This chunk
is represented using a binary feature vector. For example, a cat-
egorical variable with four possible values: "A", "B", "C", and
"D" is encoded using four binary features: [1, 0, 0, 0], [0, 1, 0, 0],
[0, 0, 1, 0], and [0, 0, 0, 1], respectively. To represent multiple cat-
egories, we simply use "0" and "1" to indicate the absence and
presence of each category label respectively. For example, we rep-
resent a profile with categories "A, C" and another with categories
"A, B, C, D" as [1, 0, 1, 0] and [1, 1, 1, 1], respectively. In other
words, we use a one-vs-all scheme where we convert multi-class
labels to binary labels (i.e., belong or does not belong to the class).
As there are a total of 357 unique categories in the dataset of food
venues, the binary feature vector will have 357 elements.
Chunk C2: The categories of the target business' neighbors.
We first select-from our database of place profiles-the neigh-
boring food businesses within r meters from the target business,
after which we extract and sum up the category feature vectors of
the neighbors. To define category neighborhood, we use r = 200
meters, which we found to give optimal performance in our exper-
iments. Similar to C1, chunk C2 is also a 357-long feature vector
that corresponds to the same number of unique categories, except
that each feature value is now an integer. Returning to our toy ex-
ample of the four categories "A", "B", "C", and "D", if a profile
Database
of
Place
Profiles
Feature
Extractor
features
actual
popularity
Gradient
Boos6ng
loca6on
info
neighborhood
info
Feature
Extractor
input
profile
loca6on
info
neighborhood
info
features
predicted
popularity
(a)
Training
(b)
Predic6on
input
profile
input
profile
input
profile
input
profile
input
profile
input
profiles
Predic6on
Model
Table 2: Feature chunks used in our location analytics work.
Chunk ID Chunk Description
#Features
C1
C2
C3
C4
C5
C6
Categories of the profile
Categories of the profile's neighbors
Total "check-ins" of food-related hotspots
Average "check-ins" of food-related hotspots
Total "check-ins" of all hotspots
Average "check-ins" of all hotspots
357
357
20
20
20
20
only has 5 neighbors of category "A" and 7 neighbors of category
"B", its integer feature vector will be [5, 7, 0, 0].
Chunks C3 and C4: Food-related hotspots. The two chunks
are related in that both only use food-related neighbors. In other
words, they exclude neighbors that have no relevance to food, such
as clothing and electronic stores. For each chunk, we are interested
in "hotspots", which are circular areas with the profile in the center
and each area is quantified by the "popularity" of stores within it.
We define 20 hotspots around the profile whereby each hotspot is
demarcated by a maximum distance of r meters, of which r ∈
{50, 100, 150, . . . , 1000}. Finally, the only difference between C3
and C4 is in how "popularity" is defined; the former computes the
(natural) logarithm of the total "check-ins" within a hotspot, while
the latter computes the logarithm of the average "check-ins".
It must be noted that the total and average "check-ins" include
only the "check-in" counts of the neighbors and not the count of the
target business itself (which is assumed to be unknown). Also, the
purpose of applying logarithmic transformation to the "check-in"
counts is to reduce the skewness in the counts distribution (i.e., most
businesses have small "check-ins" counts, but there is a handful
number of businesses with unusually large "check-ins"). In other
words, applying logarithm transformation would allow us to miti-
gate the impact of (unusually) high "check-ins" for popular busi-
nesses. Previously, we conducted an experiment that used the raw
"check-ins" instead of the logarithmic values. Indeed, we observed
that using the logarithm values yielded lower prediction errors than
using the raw counts. As such, we shall focus on the results of the
logarithm-scaled "check-ins" throughout the rest of this paper.
Chunks C5 and C6: All (food + non-food) hotspots. These
chunks are similar to C3 and C4, respectively. The only differ-
ence is that, instead of solely using food-related neighbors, chunks
C5 and C6 use food and non-food neighbors together. The non-
food neighbors include bookshops, transportation facilities like bus
and train stations, furniture stores, universities, etc. We include
non-food hotspots so as to capture the complementary (non-food)
businesses within the neighborhood of a target business.
4.3 Predictive Model
In order to learn the association between the extracted features
and "check-in" scores of a given business. we train a supervised re-
gression model called gradient boosting machine (GBM) [8]. GBM
is a machine learning algorithm that iteratively constructs an en-
semble of weak decision tree learners through boosting mechanism.
Specifically, the boosting procedure consists of training weak learn-
ers and adding them into a final strong model in a forward stage-
wise manner. By combining many weak learners that have high
bias (i.e., high prediction error), GBM yields an accurate and ro-
bust predictive model that has a lower bias than its constituent weak
learners [16]. The GBM allows for the optimization of arbitrary dif-
ferentiable loss functions for classification and/or regression task.
For the purpose of "check-in" regression, however, we shall focus
on the least square loss function [8] in this work.
Another major benefit of using GBM is that it can automatically
derive the so-called feature importance metric [8, 16]. This pro-
vides an important mechanism to interpret the trained model and
identify the key features that contribute substantially to the predic-
tion of the target variable (i.e., "check-in" score).
In particular,
each decision tree in the GBM intrinsically performs feature selec-
tion by choosing the appropriate split points. This information can
then be used to measure the importance of each feature. That is,
the more often a feature is used in the split points of a tree, the
more important that feature is. This notion can be extended to the
tree ensemble by averaging the feature importance of each tree. We
further elaborate our feature importance analysis in Section 6.3.
5. EVALUATION PRELIMINARIES
We preface our evaluation proper by detailing the evaluation met-
rics and procedure, the baseline models against which we compare
GBM, as well as the model variations we considered in our study.
5.1 Evaluation Metrics and Procedure
To measure how accurate our predicted "check-in" scores dif-
fer from the actual (observed) scores, we use two popular regres-
sion quality metrics: mean-squared logarithmic error (MSLE) and
mean absolute logarithmic error (MALE). The MSLE and MALE
metrics are respectively defined as:
n(cid:88)
n(cid:88)
i=1
i=1
M SLE =
M ALE =
1
n
1
n
(log(pi + 1) − log(ai + 1))2
log(pi + 1) − log(ai + 1)
(1)
(2)
where n is the number of samples in the test set, p is the predicted
"check-ins", and a is the actual "check-ins". The MSLE metric
measures the averaged squared errors, which gives a higher penalty
to large logarithmic differences log(pi + 1) − log(ai + 1). On
the other hand, the MALE metric measures the averaged absolute
errors, whereby all the individual differences are weighted equally.
To assess the performance of our predictive model, we perform
a 10-fold cross-validation procedure whereby the dataset is ran-
domly partitioned into 10 equal sized subsamples. A single sub-
sample is retained as the validation data for testing our models,
while the remaining 9 subsamples are used as training data. The
cross-validation process is then repeated 10 times, with each of the
10 subsamples used exactly once as the validation data. We then
report the averaged performance.
Finally, to test for the statistical significance of our results, we
utilize the independent two-sample t-test [19].
In particular, we
look at the p-value of the t-test involving two performance vectors,
at a significance level of 0.01. If the p-value is less than 0.01, we
can conclude the performance difference is statistically significant.
5.2 Baselines
We compare GBM with several regression baseline algorithms.
To foster reproducibility of this work, our implementations of all
these algorithms (including GBM) are based on the scikit-learn li-
brary [18]. The following baselines are used in this work:
• Distance-based nearest neighbors (DNN). This is a simple
baseline that takes the logarithm of the average "check-ins"
of the neighbors that reside within some radius r of a target
business location. DNN works based on a simple intuition:
"the more popular the neighborhood, the more popular the
target location is going to be, all else being equal". We test
on r ∈ {50, 100, . . . , 500} and found that DNN with r =
100m brings about the best results.
• Support vector regression with linear kernel (SVR-Linear)
[7]. This method produces a linear regression model that de-
pends only on a subset of the training data, since the cost
function for building the model ignores any data points close
to the model prediction. For this method, we set the cost
parameter to C = 1 and the epsilon parameter (for control-
ling epsilon-insensitive loss) to = 0.1, which give the best
performance in our experiments.
• Support vector regression with radial basis function ker-
nel (SVR-RBF) [23]. This is the same as SVR-Linear, ex-
cept that now it uses a radial basis function (Gaussian) ker-
nel. As with SVR-Linear, we use C = 1 and = 0.1, which
again constitute the optimal configuration for SVR-RBF.
Last but not least, we configure our GBM algorithm using the
"least squares" loss function, a learning rate of 0.1, a maximum
tree depth of 10, and a maximum tree width of "sqrt" (i.e., the
square root of the total number of features). For the number of
boosting iterations N, we perform an exhaustive grid search on
N ∈ {100, 200, . . . , 5000} and found that N = 1000 produces
the best results. Note that, in each boosting iteration, a new tree
is created and added into the ensemble. As such, the number of
boosting iterations N is equal to the number of trees constructed.
5.3 Model Variations
To evaluate the contributions of different feature chunks, we con-
struct a variation of the predictive models by enumerating all pos-
sible combinations of the six chunks (see Section 4.2). That is, we
construct all possible 26 − 1 = 63 chunk combinations and build a
predictive model for each combination. We represent a model vari-
ant using a binary array of length six, where chunk Ci maps to the
ith element in the array. We use the notation "[model_name]xxxxxx"
to represent a particular model variant, where x ∈ {0, 1}. For ex-
ample, a GBM model using C1, C2, and C4 is denoted as GBM110100.
Note that DNN does not use this notation, since it works based on
spatial distance only, instead of feature chunks. For SVR-Linear,
SVR-RBF and GBM, we run experiments on all 63 variants and
report the best results for each of the three methods.
6. RESULTS AND ANALYSIS
We now present our main experimental results. Our experiments
seek to answer several key research questions (RQs):
RQ1: How well can our predictive model (GBM) estimate the
popularity (i.e., "check-in" scores) of business locations?
RQ2: What are the contributions of different feature chunks? How
robust is our model against different feature combinations?
RQ3: Do the important features found by our model make sense?
What can we learn/conclude from them?
6.1 Performance Assessment (RQ1)
Table 3 compares the cross-validation performances (i.e., av-
eraged MALE and MSLE) of different regression methods. For
SVR-Linear and SVR-RBF, we show both the "full variant" (i.e.,
SVR-Linear111111 and SVR-RBF111111) as well as the variants that
give the best results for the same method (i.e., SVR-Linear111000 and
SVR-RBF100011). We observe that GBM consistently and signifi-
cantly outperforms other models (at p < 0.01), particularly against
SVR-RBF100011, which is the best among all the baselines.
Table 3: Performance comparisons of different models.
Feature Chunks
-
{C1, C2}
Model
DNNr=100m
SVR-Linear111000
SVR-Linear111111
SVR-RBF100011
SVR-RBF111111
GBM111111
*: significant at 0.01 with respect to SVR-RBF100011
{C1, C2, C3, C4, C5, C6}
{C1, C2, C3, C4, C5, C6}
{C1, C2, C3, C4, C5, C6}
{C1, C5, C6}
MALE
1.99305
1.59072
2.12345
1.47518
1.53067
1.16362*
MSLE
7.27499
4.25301
7.35446
3.61863
3.92219
2.56924*
We can explain the results in terms of model complexity. For
instance, we can expect the simplest nearest neighbor method (i.e.,
DNN) to be beaten by other methods, as it only uses spatial dis-
tance. We can also anticipate that SVR-RBF would outperform
SVR-Linear, as the RBF kernel maps the original features into
a high-dimensional space. This expanded feature space provides
SVR-RBF with a greater representation power to model a much
more complex relationship than SVR-Linear. Finally, as GBM
combines weak learners into a strong learner, the aggregate predic-
tion of the ensemble is more accurate than the prediction of any of
its constituent learners. This aggregation also provides GBM with
more robustness to data overfitting, as compared to SVM-RBF.
Additionally, the results in Table 3 suggest that the two SVR
methods are more sensitive to the variation of feature chunks, par-
ticularly to the presence of less relevant (or irrelevant) features.
This can be attributed to the fact that each tree in GBM intrinsically
performs feature selection, for which less important features are un-
likely to be chosen and used in the ensemble. Indeed, we can see
that SVR-Linear with all six chunks (i.e., SVR-Linear111111) is out-
performed by the simpler SVR-Linear variant (i.e., SVR-Linear111000)
that uses only three chunks. Surprisingly, the former is also out-
performed by even the DNN method. The same conclusion can
be made by comparing SVR-RBF111111 and SVR-RBF100011. On
the contrary, GBM is more robust against inconsequential features.
In fact, GBM generally improves its performance as we add more
chunks, as we will see shortly in Section 6.2.
6.2 Contribution of Feature Chunks (RQ2)
The partitioning of the feature vectors into six chunks allows us
to investigate the contribution of each feature group. Table 4(a)
lists the top 10 GBM variants (out of 63 possible variants), sorted
in an ascending order of their MALE scores. Similarly, Table 4(b)
lists the top 10 GBM variants, sorted in ascending order of their
MSLE scores. Note that the top 10 GBM variants happen to be
the same for the two tables, except that they have slightly different
ordering. From the results, it is evident that GBM111111 does not
significantly outperform the other nine variants (at a significance
level of 0.01). This shows that GBM is robust against the variation
of feature chunks. We can also see that the performance of the
GBM improves as we add more feature chunks. Again, this can
be attributed to the feature selection mechanism of each tree in the
ensemble, which helps exclude irrelevant features.
Based on the binary representation of the six chunks, we can
also calculate the relative significance of a chunk by counting the
number of times in which it is present (i.e., when the chunk is as-
signed the value of 1). The sum of each chunk's presence in the 10
GBM variants is shown at the last row of Tables 4(a) and 4(b), enti-
tled "Count". We see that the categories of the target business (i.e.,
chunk C1) is present in all the top 10 GBM variants, indicating that
it is an essential feature. This may seem to suggest that the nature
Table 4: Cross-validation results of the top 10 GBM variants.
(a) MALE results of the top 10 GBM variants.
Feature Chunks
C5
C4
C3
C2
C1
Model
GBM111111 Yes Yes Yes Yes Yes Yes
–
–
GBM111100 Yes Yes Yes Yes
Yes
GBM111010 Yes Yes Yes
–
–
–
GBM110011 Yes Yes
Yes Yes
–
Yes Yes Yes Yes
–
GBM101111 Yes
–
Yes Yes
–
GBM101100 Yes
–
Yes Yes
–
GBM100011 Yes
–
–
GBM111000 Yes Yes Yes
–
–
–
–
Yes
–
–
GBM110010 Yes Yes
–
Yes
–
Yes
–
GBM101010 Yes
Count
10
6
7
4
7
4
C6 MALE
1.163618
1.172693
1.173910
1.175062
1.177136
1.182053
1.184053
1.184895
1.189258
1.191831
(b) MSLE results of the top 10 GBM variants.
–
–
Feature Chunks
C5
C4
C3
C2
C1
Yes
Model
C6
GBM111111 Yes Yes Yes Yes Yes Yes
GBM111010 Yes Yes Yes
–
Yes Yes Yes Yes
GBM101111 Yes
–
GBM111100 Yes Yes Yes Yes
–
Yes Yes
–
GBM110011 Yes Yes
–
–
Yes Yes
–
GBM101100 Yes
–
GBM100011 Yes
–
–
–
Yes Yes
–
–
–
GBM111000 Yes Yes Yes
–
Yes
–
–
GBM110010 Yes Yes
–
Yes
–
Yes
–
GBM101010 Yes
Count
10
6
7
4
7
4
MSLE
2.569236
2.608927
2.609254
2.610255
2.615818
2.627101
2.628505
2.653032
2.660369
2.667292
of the business itself plays a pivotal role. However, as described
in Section 3.2, food businesses on Facebook may contain non-food
labels such as "airport" and "shopping mall" (e.g., for a cafe lo-
cated in the shopping mall of an airport). In turn, this suggests that
the "environment" around a selected business is also a key factor.
The method of chunk counting presented in this section is a coarse-
grained analysis and is not sufficient to validate this conjecture. We
will further analyze this in Section 6.3, where we employ a more
fine-grained analysis of the individual feature's importance.
Moving on, we also notice that the total "check-ins" chunks (i.e.,
chunks C3 and C5) are ranked higher than the average "check-ins"
(i.e., C4 and C6), i.e., the counts are 7/10 vs. 4/10. This suggests
that the total "check-ins" have more discriminatory power than the
average "check-ins", which could be due to the averaging failing to
account for the number of business nearby. On the other hand, total
"check-ins" (of an area) gives a more accurate reflection of the po-
tential human traffic that an area has. Finally, we see no substantial
performance difference between food-related hotspots and all (i.e.,
food + non-food) hotspots (both have a count sum of 7 + 4 = 11).
This implies that the presence of non-food-related categories does
not contribute significantly to the prediction quality.
6.3 Analysis of Feature Importance (RQ3)
The analysis of the six chunks of the feature vectors in the previ-
ous section represents a coarse-grained analysis. To perform a more
fine-grained analysis, we look into the full feature vector (with all
six chunks included) in the GBM111111 model and try to compute
the relative importance of each individual feature. GBM derives
this automatically, by measuring how many times a feature is used
in the split points of a tree [8] (see also Section 4.3).
Figure 7(a) shows the relative importance of the top 20 features
in descending order of importance, while Figure 7(b) shows the
relative importance of the 71st to the 90th features. (We do not
include the results for the 21st to 70th features here, as the changes
in the feature importance score are fairly smooth.) Accordingly, we
can make the following observations:
• Chunks C3 to C6 (black bars in Figure 7) dominate the top
80 feature importance positions (not fully shown in Figure 7),
and it is not until the 81st top feature that chunks C1 and C2
show up. This suggests that hotspot features play a very cru-
cial role: the more "check-ins" a target business' neighbors
have, the more popular the target business is likely to be.
• From Figure 7(a), 14 out of 20 hotspot features are below
500m, suggesting that nearer "check-ins" are used as a strong
signal to make a split in the decision tree. This is not surpris-
ing, as it may be physically tiring for customers to travel far-
ther than 500m, and most will settle all their outdoor needs
in a specific area, such as a shopping mall.
• Comparing the total and average "check-ins" in Figure 7(a),
14 out of 20 features belong to the former. This indicates
that total "check-ins" is a better input feature/signal for split
points in the GBM's trees. This finding is generally in agree-
ment with what we have found in Section 6.2.
• Figure 7(a) also shows that the type of neighbors (i.e., "food-
only" or "all") are equally matched with 10 counts each.
Again, this finding conforms with the earlier finding in Sec-
tion 6.2. Note that, despite the different approaches in Sec-
tions 6.2 and 6.3, both arrived at the similar conclusion with
regard to the type of "check-in" and the type of neighbor.
• From the colored bars (i.e., chunks C1 and C2) in Figure 7(b),
we see that C1 is dominated by the C2. This suggests that
categories of the neighboring businesses are more important
than those of the target business (C1). Together with the
"hotspot" features in Chunks C3 to C6, this reinforces the
idea of a "local effect" whereby business benefit by being
close to more established neighbors.
• Finally, we notice a significant and faster drop in the impor-
tance scores from the 81st to 90th features (as compared from
the 1st to 80th). In this case, places or franchises that typi-
cally attract general (and larger) crowd, such as "restaurant",
"coffee", or "shopping mall"-related categories, take the first
top spots among the neighbors' categories. This suggests that
food-related categories (of the neighbors) are more important
than the non-food categories.
7. WEB APPLICATION PROTOTYPE
We implement our location analytics framework as an interac-
tive web application service, which can be accessed at: https:
//research.larc.smu.edu.sg/bizanalytics/.
Technologies. We employ the following technologies to build
our web application: (i) Python (implementing the predictive model
and feature extraction), (ii) RabbitMQ (a messaging passing sys-
tem that allows querying the predictive model), (iii) Node.js (for
processing users' queries and returning the prediction results to the
front-end), (iv) ElasticSearch (a distributed search engine for quer-
ing the database of place profiles), and (v) Google Maps (for visu-
alization at the front-end). This configuration provides an efficient
and scalable way to process a user's location query (via Node.js and
(a) Feature importance of the top 20 features descending order
of importance. The top 20 happen to consists of Chunks C3 to C6.
(b) Feature importance of the top 71st to 90th features. Besides Chunks C3 to
C6, it also contains Chunks C1 and C2, in red and blue, respectively.
Figure 7: Top features from the feature importance of the GBM111111 model.
Figure 8: Our online interactive location analytics system.
RabbitMQ), retrieve the relevant neighbors (via ElasticSearch), in-
volve feature extraction and predictive models (in the Python com-
ponent), and finally display the prediction results to the users (again
via Node.js and RabbitMQ, along with Google Maps).
User interaction. Figure 8 shows an example of how our web
application works. A user drops a pin (i.e., the blue pin in the mid-
dle of the screen) to indicate where his (hypothetical) store location
would be. Depending on the location, the interface also dynami-
cally shows the neighboring businesses on the right panel and their
respective information, such as (i) the distance from the drop-pin,
and (ii) the number of physical "check-ins" and "likes". When the
user is ready, he/she may click the "Calculate Predicted Check-ins"
button, which will then calculate the predicted "check-in" score on
the fly. After presenting the predicted "check-in" score, the user
can also open a new panel, showing a ranked list of his/her target
location relative to the nearby businesses (Figure 9). The rank-
Figure 9: After computing the predicted "check-in" score, users
can compare their target location with the surrounding businesses.
ing allows users to understand how their target location would fare
against the neighboring businesses. The panel also shows the high-
est, lowest, and the median scores of these neighbors.
Qualitative study. Figures 8 and 9 demonstrate the on-the-fly
prediction of our web application, where it is able to predict the
"check-in" score of a hypothetical, inexistent target business. For
this example, the score of 802.24 in Figure 9 represents a fairly
conservative prediction of the potential "check-ins" in the target
location (i.e., the blue pin) and the selected type of business. This
hypothetical business is ranked 160th among 213 businesses, with
the lowest "check-ins" being 106. This is a reasonable prediction.
On the one hand, because the place is near places with consistent
human traffic, such as the Hilton Hotel and several other shopping
malls, it should garner a decent amount of check-ins. On the other
hand, as there are many other businesses in the area (the area is a
renowned shopping paradise in Singapore), it may be challenging
for the hypothetical business to compete with these businesses.
8. DISCUSSION AND FUTURE WORK
In this work, we investigate whether businesses can benefit from
other (popular) businesses within its vicinity. Our results show
not only a positive correlation between the popularity of a target
0
0.002
0.004
0.006
0.008
0.01
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[100m]
[ALL]
[AVERAGE
CHECKINS]
[350m]
[ALL]
[AVERAGE
CHECKINS]
[550m]
[ALL]
[TOTAL
CHECKINS]
[100m]
[ALL]
[AVERAGE
CHECKINS]
[150m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[600m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[150m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[450m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[50m]
[ALL]
[TOTAL
CHECKINS]
[350m]
[ALL]
[AVERAGE
CHECKINS]
[250m]
[ALL]
[TOTAL
CHECKINS]
[550m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[750m]
[ALL]
[AVERAGE
CHECKINS]
[50m]
[ALL]
[TOTAL
CHECKINS]
[200m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[550m]
[ALL]
[TOTAL
CHECKINS]
[600m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[200m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[300m]
[FOOD--‐ONLY]
[TOTAL
CHECKINS]
[250m]
20
19
18
17
16
15
14
13
12
11
10
9
8
7
6
5
4
3
2
1
Feature
Importance
Score
0
0.002
0.004
0.006
0.008
0.01
[Profile's
Neighbors]
Food
Stand
[Profile's
Neighbors]
Food
&
Grocery
[Profile's
Neighbors]
Shopping
Mall
[Profile's
Neighbors]
Bakery
[Profile]
Bakery
[Profile's
Neighbors]
Fast
Food
Restaurant
[Profile's
Neighbors]
Restaurant
[Profile's
Neighbors]
Cafe
[Profile's
Neighbors]
Food
&
Restaurant
[Profile's
Neighbors]
Coffee
Shop
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[800m]
[ALL]
[AVERAGE
CHECKINS]
[800m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[700m]
[ALL]
[AVERAGE
CHECKINS]
[750m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[900m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[850m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[500m]
[FOOD--‐ONLY]
[AVERAGE
CHECKINS]
[1000m]
[ALL]
[TOTAL
CHECKINS]
[950m]
[ALL]
[AVERAGE
CHECKINS]
[450m]
90
89
88
87
86
85
84
83
82
81
80
79
78
77
76
75
74
73
72
71
Feature
Importance
Score
business and its neighbors, but also the critical importance of the
"hotspot" features: the nearer a target location is to a popular place
with larger "check-ins", the more successful it would be. This find-
ing conforms with our intuition. But more importantly, it demon-
strates that ubiquitous online data (such as Facebook Pages) can be
used to gauge the socioeconomical values. We also show how our
predictive model can be used to accurately estimate the "check-in"
score of a particular location, allowing us to identify the best loca-
tions that would bring popularity, and by extension, success.
Despite the promising potentials of our approach, there remains
room for improvement. For instance, our current work has not
taken into account the temporal aspects of the business popular-
ity, such as modeling the trend of the "check-in" scores over time.
Further quantitative and qualitative studies may also be needed in
the future to compare our work with other location-based services
such as Foursquare. To facilitate more comprehensive location ana-
lytics, we can extend our approach by building a two-level location
recommendation system, whereby we first (coarsely) recommend a
city district [15] and then pinpoint (multiple) promising locations
within that district. As we include more data, such as non-food cat-
egories and auxiliary data that reflect the human flow of different
areas of an urban city, we will be able to further improve on our cur-
rent model and findings. To address all these, we plan to develop a
new spatiotemporal predictive model that integrates a richer set of
residential, demographics, and other social media data.
Acknowledgements. This research is supported by the National
Research Foundation, Prime Minister's Office, Singapore under its
International Research Centres in Singapore Funding Initiative.
9. REFERENCES
[1] J. Chang and E. Sun. Location3: How users share and
respond to location-based data on social networking sites. In
Proceedings of the International AAAI Conference on
Weblogs and Social Media, pages 74–80, 2011.
[2] H. Chen, R. H. L. Chiang, and V. C. Storey. Business
intelligence and analytics: From big data to big impact. MIS
Quarterly, 36(4):1165–1188, 2012.
[3] N. Cohen. Business location decision-making and the cities:
Bringing companies back. Technical report, Brookings
Institution Center on Urban and Metropolitan Policy, 2000.
[4] ESRI. Revealing the "where" of business intelligence using
location analytics.
http://www.esri.com/library/whitepapers/pdfs/
business-intelligence-location-analytics.pdf,
2012.
[5] Facebook. Facebook platform upgrade guide. https:
//developers.facebook.com/docs/apps/upgrading,
2016.
[6] Facebook. Graph API reference. https://developers.
facebook.com/docs/graph-api/reference/page, 2016.
[7] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J.
Lin. LIBLINEAR: A library for large linear classification.
Journal of Maching Learning Research, 9:1871–1874, 2008.
[8] J. H. Friedman. Greedy Function Approximation: A
Gradient Boosting Machine. The Annals of Statistics,
29:1189–1232, 2001.
[9] H. Gao, J. Tang, X. Hu, and H. Liu. Exploring temporal
effects for location recommendation on location-based social
networks. In Proceedings of the ACM Conference on
Recommender Systems, pages 93–100, 2013.
[10] H. Gao, J. Tang, X. Hu, and H. Liu. Modeling temporal
effects of human mobile behavior on location-based social
networks. In Proceedings of the ACM Conference on
Information and Knowledge Management, pages 1673–1678,
2013.
[11] L. Garber. Analytics goes on location with new approaches.
IEEE Computer, 46(4):14–17, 2013.
[12] P. Georgiev, A. Noulas, and C. Mascolo. Where businesses
thrive: Predicting the impact of the Olympic Games on local
retailers through location-based services data. In
Proceedings of the International AAAI Conference on
Weblogs and Social Media, pages 151–160, 2014.
[13] D. Karamshuk, A. Noulas, S. Scellato, V. Nicosia, and
C. Mascolo. Geo-spotting: Mining online location-based
services for optimal retail store placement. In Proceedings of
the ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pages 793–801, 2013.
[14] R. Li, S. Wang, and K. C.-C. Chang. Multiple location
profiling for users and relationships from social network and
content. Proceedings of the International Conference on Very
Large Data Bases, 5(11):1603–1614, 2012.
[15] J. Lin, R. J. Oentaryo, E.-P. Lim, C. Vu, A. Vu, A. T. Kwee,
and P. K. Prasetyo. A business zone recommender system
based on Facebook and urban planning data. In Proceedings
of the European Conference on Information Retrieval, pages
1–7, 2016.
[16] A. Natekin and A. Knoll. Gradient boosting machines: A
tutorial. Frontiers in Neurorobotics, 7(21):1–21, 2013.
[17] A. Noulas, S. Scellato, N. Lathia, and C. Mascolo. Mining
user mobility features for next place prediction in
location-based services. In Proceedings of the IEEE
International Conference on Data Mining (ICDM), pages
1038–1043, 2012.
[18] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss,
V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau,
M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn:
Machine learning in python. Journal of Machine Learning
Research, 12:2825–2830, 2011.
[19] W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T.
Vetterling. Numerical Recipes in C: The Art of Scientific
Computing. Cambridge University Press, 1988.
[20] R. W. Sinnott. Virtues of the haversine. Sky and Telescope,
68(2):159, 1984.
[21] C. Smith. 200+ amazing facebook user statistics.
http://expandedramblings.com/index.php/
by-the-numbers-17-amazing-facebook-stats/, 2016.
[22] C. Smith. By the numbers: 17 important Foursquare stats.
http://expandedramblings.com/index.php/
by-the-numbers-interesting-foursquare-user-stats/,
2016.
[23] A. J. Smola and B. Schölkopf. A tutorial on support vector
regression. Statistics and Computing, 14(3):199–222, 2004.
[24] B. Thau. How big data helps chains like starbucks pick store
locations-an (unsung) key to retail success.
http://onforb.es/1iijr2o, 2015.
[25] Y. Zhang, B. Li, and J. Hong. Understanding user economic
behavior in the city using large-scale geotagged and
crowdsourced data. In Proceedings of the International
World Wide Web Conference, 2016.
|
1710.00364 | 1 | 1710 | 2017-10-01T15:12:13 | Activating the "Breakfast Club": Modeling Influence Spread in Natural-World Social Networks | [
"cs.SI",
"cs.MA"
] | While reigning models of diffusion have privileged the structure of a given social network as the key to informational exchange, real human interactions do not appear to take place on a single graph of connections. Using data collected from a pilot study of the spread of HIV awareness in social networks of homeless youth, we show that health information did not diffuse in the field according to the processes outlined by dominant models. Since physical network diffusion scenarios often diverge from their more well-studied counterparts on digital networks, we propose an alternative Activation Jump Model (AJM) that describes information diffusion on physical networks from a multi-agent team perspective. Our model exhibits two main differentiating features from leading cascade and threshold models of influence spread: 1) The structural composition of a seed set team impacts each individual node's influencing behavior, and 2) an influencing node may spread information to non-neighbors. We show that the AJM significantly outperforms existing models in its fit to the observed node-level influence data on the youth networks. We then prove theoretical results, showing that the AJM exhibits many well-behaved properties shared by dominant models. Our results suggest that the AJM presents a flexible and more accurate model of network diffusion that may better inform influence maximization in the field. | cs.SI | cs | Activating the "Breakfast Club": Modeling Influence Spread in Natural-World
Social Networks
Lily Hu 1, Bryan Wilder 2, Amulya Yadav 2, Eric Rice 2, Milind Tambe 2
1Harvard University, Cambridge, MA, 02138, USA
{lilyhu}@g.harvard.edu
2University of Southern California, Los Angeles, CA, 90089, USA
{bwilder, amulyaya, ericr, tambe}@usc.edu
7
1
0
2
t
c
O
1
]
I
S
.
s
c
[
1
v
4
6
3
0
0
.
0
1
7
1
:
v
i
X
r
a
Abstract
While reigning models of diffusion have privileged the struc-
ture of a given social network as the key to informational ex-
change, real human interactions do not appear to take place
on a single graph of connections. Using data collected from a
pilot study of the spread of HIV awareness in social networks
of homeless youth, we show that health information did not
diffuse in the field according to the processes outlined by
dominant models. Since physical network diffusion scenarios
often diverge from their more well-studied counterparts on
digital networks, we propose an alternative Activation Jump
Model (AJM) that describes information diffusion on physi-
cal networks from a multi-agent team perspective. Our model
exhibits two main differentiating features from leading cas-
cade and threshold models of influence spread: 1) The struc-
tural composition of a seed set team impacts each individual
node's influencing behavior, and 2) an influencing node may
spread information to non-neighbors. We show that the AJM
significantly outperforms existing models in its fit to the ob-
served node-level influence data on the youth networks. We
then prove theoretical results, showing that the AJM exhibits
many well-behaved properties shared by dominant models.
Our results suggest that the AJM presents a flexible and more
accurate model of network diffusion that may better inform
influence maximization in the field.
1
Introduction
Research in influence maximization lies at the crossroads
of the who, what, and when of information diffusion. In
their seminal paper, Kempe, Kleinberg, and Tardos (2003)
formalized the problem by imagining influencing agents as
seed nodes in a network initialized to propagate informa-
tion first to neighbors and eventually throughout the net-
work as influence spreads. Along with theoretical work in
the field, the wide availability of large-scale digital data has
positioned internet networks-web link traces, email com-
munication, social media platforms-at the center of the dis-
cussion of the influence maximization problem. Even so,
empirical work that compares predictions by the dominant
Independent Cascade (ICM) and Linear Threshold (LTM)
models with realized diffusion at the node level is limited.
Prevailing methodologies for estimating diffusion model pa-
rameters often achieve low accuracy in replicating observed
Copyright c(cid:13) 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
behavior even when applied to well-defined online networks
with temporal information flow data (Goyal, Bonchi, and
Lakshmanan 2011; Wang et al. 2013).
Although the ICM and LTM were originally formulated
to describe social influence in natural environments (Gra-
novetter 1978; Schelling 1978), there is a dearth of high-
quality data that support the theories on networks in these
physical settings. Moreover, multiple challenges exacerbate
these deficiencies in physical settings. First, both the ICM
and LTM assume that the topology of social ties is iden-
tical with the mesh of connective channels through which
information spreads. While on many social media platforms
a user's social network delimits her space of communica-
tion, in the natural world, an individual's total space of
social navigation dwarfs the space of those she calls her
"friends," and there exists a multiplicity of information av-
enues within the network that do not coincide with one's
social ties (Shafie 2015). Further, previous work has shown
that models with strong assumptions about a particular con-
structed graph topology are more prone to error and inaccu-
racy in their predictions of information spread (Butts 2003;
Krackhardt and Kilduff 1999).
In this paper, we analyze information diffusion data from
the first empirical study of influence maximization in the
physical world (Yadav et al. 2017), which tracked 173 indi-
viduals across 3 distinct networks over a multi-year time pe-
riod. We delved into this significant corpus of natural world
network data to investigate influence spread at an individ-
ual node level rather than a network-wide volume level. We
found that information did not diffuse from seed nodes to
the greater network according to processes suggested by
the ICM or LTM. Most strikingly, we found that across
all three networks in the study, 50 % of informed nodes
lacked any path to a seed node and moreover, a node's de-
gree of connectivity-both generally and specifically to seed
nodes-exhibited no correlation with likelihood of becom-
ing informed. These results directly contradict predictions
put forth by the ICM and LTM and call into question the
suitability of these leading models of diffusion for approxi-
mating information spread on physical networks.
These negative results against the ICM and LTM can be
generalized to apply to other graph-based models of diffu-
sion: The high proportion of informed yet isolated nodes
cannot be explained by models that rely on edge-based prop-
agation as the sole avenue for influence spread (DeGroot
1974). On the other hand, more flexible approaches such as
Hawkes contagion processes (Zhou, Zha, and Song 2013),
which allow for influence to spread over greater social dis-
tances, cannot explain the differing spread outcomes across
the networks. In general, appealing only to non-graph fea-
tures of the diffusion scenarios cannot reconcile the diver-
gences in observed information diffusion in the three differ-
ent Yadav et al. studies. In response to the shortcomings of
existing models, we develop a new model of targeted peer-
to-peer information spread on natural networks that does
not rely on strong tie assumptions and instead incorporates
an understanding of influencing as a "team" behavior.
Our model features two distinguishing characteristics that
are aligned with real-world information diffusion in phys-
ical spaces: 1) Nodes exchange information beyond their
immediate social ties, and 2) Seed nodes act as a multi-
agent team to spread information, where their overall in-
fluencing efficacy is a function of both individual and team
attributes. In the proposed Activation Jump Model (AJM),
team-based influence spread in a network is driven by acti-
vating the "Breakfast Club," where individuals from differ-
ent social contexts band together for a common cause and
form a united team for information diffusion. These features
confer a flexibility to our forecasts of information flow and
allow our model to achieve a 60% to 110% improvement
over the best ICM and LTM predictions in its predictions of
which nodes will be influenced.
We also point to a methodological pitfall of research in
influence maximization that focuses solely on achieving a
particular level of information diffusion within a network.
Namely, matching magnitude of influence spread under sim-
ulations to observed influence spread is insufficient evidence
for determining the underlying diffusion process. We show
that optimal seeding under one model achieves near-optimal
(> 90%) influence spread under another diffusion process
on three natural-world networks. In fact, any magnitude of
influence spread can be explained by varying ICM and LTM
parameters, pointing to a fundamental ambiguity in identi-
fying the true diffusion process based on this metric alone.
Even when seeding strategies achieve high levels of influ-
ence spread, leading models' failures to predict node-level
influence can limit their applicability. In domains of sus-
tainability, network interventions can generate knowledge or
promote behavioral changes within a community. These pro-
grams typically identify individuals and groups that may es-
pecially benefit from the intervention. For example, school
network-based suicide prevention programs aim to increase
general awareness about signs of suicidal behavior but espe-
cially seek to reach high-risk adolescents and their social cir-
cles (Kalafat and Elias 1994). Similarly, peer-led HIV pre-
vention programs akin the to the fieldwork by Yadav et al.
hope to reach a diverse set of individuals but especially those
who participate in risky behaviors (Broadhead et al. 1998).
As such, most social interventions have the dual purpose of
maximizing influence coverage while also targeting vulner-
able individuals. With its superior performance in predicting
node-level influence, the AJM may serve as a more desirable
framework to guide these intervention strategies in the field.
2 Network Intervention Data Analysis
Pilot Study Procedure Yadav et al.'s long-standing col-
laboration with homeless youth service providers in a large
urban area sought to improve peer-led heatlh interventions
by leveraging research in influence maximization (2017). To
this end, they conducted a series of head-to-head comparison
studies of seeding strategies to select cohorts of Peer Lead-
ers among the youth that would be trained for the task of
HIV awareness diffusion in their communities. Three stud-
ies took place on three distinct social networks of homeless
youth. Each study recruited youth and gathered social net-
work data using online contacts, field observations, and sur-
veys. A different seeding strategy was then deployed on each
of the generated networks: In two of the pilot studies, Peer
Leaders were chosen via two algorithmic agents for influ-
ence maximization, HEALER and DOSIM, which were de-
signed to optimize network-based intervention strategies for
health providers. The third network was seeded via degree
centrality (DC), the most commonly-used heuristic in net-
work interventions (Valente 2012), such that the most popu-
lar youth were chosen as Peer Leaders.
Each network's Peer Leaders underwent an intensive
training course led by pilot study staff that served to both in-
struct the youth in spreading information about HIV to their
peers as well as bind the members together in their shared
roles as health ambassadors. After Peer Leaders were sent
out into the field, youth were asked in 1-month and 3-month
follow-up surveys about whether they had received infor-
mation about HIV from a Peer Leader. These responses re-
vealed the extent to which information had spread from seed
nodes to the greater network. The post-intervention results
revealed that the HEALER and DOSIM seeding strategies
resulted in greater informational spread compared to DC,
with ∼74% and ∼72% respectively of non-Peer Leaders
reporting having received information about HIV in the 3-
month survey compared to ∼35% in the control study. Since
both HEALER and DOSIM solved the influence maximiza-
tion problem by assuming a model of information spread
based on a generalization of the Independent Cascade, the
success initially seemed to validate the model as an accurate
approximation of information spread in the physical world.
Table 1: Fraction of nodes in each network that were in-
formed, sorted by connectivity status to Peer Leaders. De-
nominator gives total number of nodes of that type; numer-
ator gives number of those that are informed. Direct nodes
have an edge to a PL; indirect nodes are connected via inter-
vening neighbor(s); isolated nodes lack a path to any PL.
Isolated Proportion
Informed
Indirect
Network
HEALER 34
DOSIM 25
26
n Direct
15/21
5/6
5/12
DC
4/7
5/10
1/4
6/6
8/9
3/10
25/34
18/25
9/26
Node-level analysis of information diffusion However,
the empirically observed node-level patterns of information
spread in the three networks wildly diverged from Inde-
pendent Cascade and Linear Threshold predictions. Table 1
gives an overview of the connectivity of nodes that reported
receiving information about HIV from a Peer Leader in the
3-month follow-up survey. In each of the three networks,
nodes lacking a path to any seed Peer Leaders-denoted as
"isolated" in the table- represented a high proportion of
all nodes that were informed. Notably, in both the HEALER
and DOSIM interventions, isolated youth were informed at
a rate higher than even those youth who were directly con-
nected to one or more Peer Leaders, with 100% (6/6 in
HEALER) and 89% (8/9 in DOSIM) informed compared to
∼71% (15/21) and ∼83% (5/6). In the DC network, the ef-
fect is less pronounced, though isolated nodes were still in-
formed at a rate comparable to the general non-Peer Leader
population (30% compared to ∼35%). Nonetheless within
the context of the ICM and LTM, such nodes have a 0 prob-
ability of receiving information. Thus, these results imme-
diately challenge the claim that existing ties are the domi-
nant avenues of informational exchange and also call into
question the premise that information radiates out from seed
nodes first to neighbors and then to the rest of the network.
In order to more finely assess the effect that a node's con-
nectivity had on its likelihood of receiving HIV information,
we calculated Pearson correlation coefficients between two
degree measures and a node's final information status. Our
results in Table 2 show that all such correlations are not sig-
nificantly different from no correlation, thus indicating that
connectivity has no bearing on likelihood of being influ-
enced. This stands in contrast to prevailing models, in which
a node's edges represent its "opportunities" to receive infor-
mation, and thus both Peer Leader degree-the number of
ties a node has to Peer Leaders-and total degree should be
strictly positively correlated with becoming informed.
Table 2: Pearson correlation coefficients r between nodes'
influence statuses and their Peer Leader (PL) and total de-
gree. PL degree counts edges to PLs; total degree counts
edges to all nodes. Positive (negative) r implies a positive
(negative) relationship between reporting hearing HIV infor-
mation and degree. 95% confidence intervals are also given.
Network
n
HEALER 34
DOSIM 25
DC
26
PL Degree
-0.0685
Total Degree
-0.1867
(-0.397, 0.276)
(-0.494, 0.162)
(-0.268, 0.508)
(-0.419, 0.370)
0.1418
0.1547
-0.0290
0.1304
(-0.247, 0.511)
(-0.271, 0.493)
The lower contact rate of directly connected nodes and
the lack of positive correlation between degree and influ-
ence status are even more dissonant with edge-based models
of propagation when considered alongside the high levels
of influence spread achieved in the studies. In the HEALER
network, information was successfully transmitted to ∼74%
of all non-Peer Leaders, corresponding to a most likely prop-
agation probability of p ≈ 0.84. Such a high propagation
probability further suggests that Peer Leader-neighboring
nodes should be even more heavily favored to receive infor-
mation, with simulations predicting that nearly all (∼99%)
would become informed, whereas in reality, only ∼71%
of these nodes received information. Simulations on the
HEALER network with this p value produce correlations of
0.489 and 0.598 between degree and likelihood of being in-
formed (PL and total respectively), indicating a moderate to
strong positive relationship compared to the actual values of
−0.0685 and −0.1867, which indicate negative to no rela-
tionship between degree and influence status. In the DOSIM
study, the graph topology itself, with 5 connected compo-
nents in addition to 7 nodes of degree 0, restricts information
spread under the ICM and LTM to maximally reach 68% of
all non-Peer Leaders. Even under perfect information propa-
gation, as long as nodes are only able to influence neighbors,
simulations under-predict the observed information spread.
While the presence of a single phenomenon such as the
activation of a small proportion of isolated nodes could rep-
resent mere aberrations of data, these multiple contradic-
tions with prevailing models indicate that the results cannot
be dismissed as simply anomalous. Given the unique chal-
lenges and complexities of information diffusion on physi-
cal networks, we must accept that the data's divergence from
predictions by models that have been largely validated only
on digital networks is one step in the uncovering and under-
standing of a qualitatively different influence process. We
thus conclude that there is no evidence that a cascade or
threshold-like process of information diffusion produced the
observed data and move toward developing a new model of
influence dynamics on real-world physical networks.
3 Proposed Model
In this section, we introduce a new model of information
spread for this class of peer-to-peer diffusion phenomena.
Activation Jump Model
Beginning with the premise that instances of social exchange
are not limited to nodes that share a tie, our model of dif-
fusion does not constrain information flow to the edges in
a network. In the Activation Jump Model (AJM), influenc-
ing agents may leave their immediate social neighborhood to
contact and propagate information to other nodes. This ac-
tion of contacting nodes beyond one's first-order ties is sig-
nified as a "jump." We recognize the heterogeneity of active
nodes' social dispositions by differentially modeling each
influencer's jump behavior. A seed node's jump activity has
two main components: 1) activation level, a measure of how
many other nodes it will attempt to influence, and 2) land-
ing distribution, a probability distribution that expresses to
which inactive nodes it will jump. Together, these two fea-
tures describe how often and to whom an influencing agent
contacts as she navigates the network to spread information.
Thus the AJM comprises two stages: First, each seed node
determines its activation level, giving the number of other
nodes to which it will jump. Second, the seed set is deployed
in the network, and the social influence process unfolds in
time. When a given seed jumps at time t, it selects from its
landing distribution a target node uninformed at time t, mod-
eling the process by which seed nodes seek nodes to inform.
Influence is then successfully propagated with probability p.
The AJM takes a multi-agent systems approach to the in-
fluence maximization problem by constructing a model of
node activation that is a function of both individual and
"team" attributes. In contrast to prior models, the seed set
is not a collection of independent influencers, rather nodes
exhibit behavioral dependencies wherein group dynamics ei-
ther contribute to or detract from aggregate activation levels.
Model Formalization
While in this paper and all our results, we use a form of the
AJM tied to the graph's structural properties, we first discuss
the general form of the model to show that it can accommo-
date a broad range of properties and then discuss our specific
form. We return to the generalization in the Discussion.
Consider a team of seed nodes, S, tasked with informa-
tion diffusion on a network G = (V, E). Each seed node
draws its activation level, giving the number of jumps it
will make, from a distribution that is a function of both the
node's individual attributes as well as the seed set's team
attributes. Formally, let ∆Z be the set of distributions over
integers Z ≥ 0. Each node v ∈ V is associated with a func-
tion fv : 2V → ∆Z that maps the set of seed nodes to a
distribution A(v, S) over discrete activation levels. A(v, S)
is a parameterized distribution (e.g., geometric) with mean
µv = h(S)[aT xv] where xv is the node's attribute vector
with coefficients aT . Together aT xv represents the partic-
ular node's maximum activation level, which is modulated
by the team activation level term given by h(S) ≤ 1, a
function of the structural positions of nodes in S that cap-
tures discomplementarities among team members. Figure 1
illustrates this dual-individual and team-composition of
a seed node's activation level distribution.
Figure 1: In the Activation Jump Model, a seed node v is
associated with an activation level distribution that is a func-
tion of individual node as well as seed set S attributes.
Each node v is also associated with a landing distribution,
Lv,T , giving the probabilities with which v jumps to a set of
potential target nodes T . The landing probability is a func-
tion of the attributes of the influencing seed v and particular
targeted node u. Based on these qualities, the node pair is as-
signed a score φ(v, u) ≥ 0, and Lv,T (u) = Γφ(v, u) where
Γ is a normalization factor such that(cid:80)
u∈T Lv,T (u) = 1.
We now instantiate the AJM in a specific form that fea-
tures the concept of "structural diversity," which highlights
groups with members who participate in multiple distinct
social contexts. The team thus acts to unite otherwise dis-
parate nodes, producing the "Breakfast Club" effect, which
has been shown to be a key determinant of diffusion in net-
works (Ugander et al. 2012). Thus we formulate the function
(cid:88)
(u,v)∈E
h(S) = 1 − 1
A
1 [u, v ∈ S]
(1)
where A > 0 is a constant equal to seed set size S. h
illustrates the negative effect of social homogeneity in the
form of between-seed-node edges on a team's effectiveness.
Each pair of connected seed nodes entails a loss of 1S of the
team's effectiveness. Barring negative influence, h(cid:48)(S) =
max(h(S), 0) without loss of generality. To reflect the corre-
lation between degree and propensity towards sociality and
thus activation, we parsimoniously set aT xv = deg(v).
1...tv
Av
The landing distribution score for seed v and target u is
given by φ(v, u) = 1
d(v,u), where d(v, u) is the path-length
distance between the two nodes. When d(v, u) = 0, we set
φ(v, u) = , where > 0 is a small constant. As in the ICM,
the propagation probability p is able to be varied.
Information diffusion thus occurs in two stages. First, dur-
ing the activation stage each node in the seed set v ∈ S is
initialized by drawing an activation level Av from its dis-
tribution A(v, S). Then, the jump stage unfolds over the
time interval [0, 1]. Each seed node v draws a series of jump
times tv
from the uniform distribution over [0, 1]. At
each jump time tv
i , v jumps to an uninfluenced target node u
drawn from Lv,T where T the set of uninfluenced nodes at
i . Finally, u is successfully influenced with probability p.
tv
Model Discussion
The Activation Jump Model's incorporation of seed set
team dynamics follows a line of multi-agent systems re-
search which demonstrates the importance of careful team
formation when agents must collaborate to achieve a goal
(Gaston and DesJardins 2003; Gaston, Simmons, and Des-
Jardins 2004; Liemhetcharat and Veloso 2012; Agmon and
Stone 2012). In particular, previous work has focused on
the importance of creating a diverse team (Balch 2000;
Hong and Page 2004; Marcolino, Jiang, and Tambe 2013).
In the AJM, we computationalize this concept by using the
group effectiveness function h to model network structural
diversity by penalizing seed sets with many within-team
edges. Thus h ∈ [0, 1] is decreasing in the level of con-
nectivity among seed nodes, and influencing nodes are more
active when they occupy distinct neighborhoods of the net-
work rather than when the team is socially homogeneous.
v∈S h(S)[aT xv] >(cid:80)
node w ∈ V such that(cid:80)
It is important to note that under this model set-up, a
node's marginal effect on the aggregate activation level of
a seed set is not guaranteed to be positive. There may exist a
v∈S∪w h(S∪
w)[aT xv], with the effect that the influence function f (·),
giving the expected number of influenced nodes, is non-
monotone. Although this is a significant departure from the
ICM and LTM, we argue that non-monotonicity is a real-
istic feature of team-based influence spread, since a new
seed node may interfere with team dynamics, resulting in a
deleterious effect that outweighs its positive individual con-
tribution. This balance between the quantity and quality of
80 77
75
61
48
44
46
34
C
O
R
U
A
60
40
20
HEALER DOSIM
Network
ICM
AJM
29
20
DC
LTM
Figure 2: Comparison of model classification performances
members in a seed set is an important consideration in team
formation in the real world. As a result, the influence max-
imization problem under the AJM requires examination of
not only a node's individual attributes but also its effect on
the group composition of nodes already in the seed set.
Model Validation on Post-Intervention Data
We evaluate the performance of the Activation Jump Model
by comparing its predictions to Yadav et al.'s dataset of HIV
awareness spread on three distinct social networks of home-
less youth. Standard experiments of diffusion models com-
pare the magnitude of total influence spread under simula-
tions to that observed empirically in order to assess model
accuracy. Here, we perform a finer-grained analysis by eval-
uating and comparing AJM, ICM, and LTM predictions of
node-level influence. We treat each model as a binary clas-
sifier that outputs the predicted probability of each node be-
coming influenced. Each model is then evaluated according
to its AUROC, a standard measure of classification accuracy.
Parameter settings: Physical networks present challenges
in data collection that limit the ability to view multiple cas-
cades, rendering standard methods of inferring diffusion pa-
rameters inoperable. We work under this constraint by fit-
ting the ICM directly to the test data by running simulations
under the propagation value that gives its best classification
performance and then forcing the AJM to also work under
this value. Thus any experimental bias favors the ICM.
For the AJM, the only parameter we set is the small con-
stant φ(v, u) = 0.1 for the landing distribution score when
d(v, u) = 0. By contrast, we present the strongest possible
version of the ICM for each network, fitting it directly to the
test data by selecting the propagation value p that maximized
the ICM's AUROC value. We then used this same probabil-
ity for the AJM. By forcing the AJM to operate under the
ICM's optimal parameterization, we ensure that our experi-
ments truly test the AJM's better suitability for modeling the
data, rather than a better ability to "memorize" the data.
Assessing classification accuracy: Using selected Peer
Leaders in the field experiments as seed nodes, we generated
diffusion instances according to the Activation Jump, Inde-
pendent Cascade, and Linear Threshold models, tracing out
Figure 3: Models' ROC curves on HEALER network
Receiver Operating Characteristic (ROC) curves for each set
of simulations. This evaluation methodology has been used
in previous node-level analyses of information diffusion
models (Wang et al. 2013; Goyal, Bonchi, and Lakshmanan
2010) and has been recognized as superior to Precision-
Recall curves for the binary classification task (Provost et
al. 1998). ROC curves plot a classifier's True Positive Rate
(TPR) against its False Positive Rate (FPR) with each point
on the curve corresponding to a predictive threshold such
that all nodes with a probability of being informed above
(below) the threshold are classified as influenced (not influ-
enced). We used the area under the ROC curve (AUROC)
to evaluate classification performance (Fawcett 2006) where
an AUROC of 1 corresponds to a perfect classifier.
Results: Each model's AUROC values for the three net-
works are shown in Figure 2; the ROC curves for all three
models' predictions on the HEALER network are shown in
Figure 3. The AJM outperforms the ICM and LTM across all
networks, with the model achieving accuracies (measured
via AUROC) of 77% and 75% on the HEALER and DOSIM
networks respectively, while the best ICM and LTM issue
predictions that, on average, perform worse than a random
classifier (20-48%). For DC, one possible explanation for
all three models' lower AUROCs is the overall poor perme-
ation of influence throughout the network, since low base-
rates cause the measure to be sensitive to small classifica-
tion changes. Even so, the AJM is far from a trivial classi-
fier, with an AUROC of 0.61 compared to the ICM and LTM
values of 0.29 and 0.20 respectively.
4 The Influence Maximization Problem
under the Activation Jump Model
given by f (S) = h(S)(cid:80)
We now consider the influencing activity of a coordinated
multi-agent team under the AJM. Since seed agents do not
target nodes that have already been informed, the influence
function is captured by the total number of expected jumps,
v∈S aT xv, where h(S) follows the
form in Equation 1. We show that under natural conditions,
f is a (potentially nonmonotone) submodular function.
Lemma 1. h is monotone-decreasing and submodular.
Proof. h is monotonically decreasing by inspection. For
submodularity, consider the marginal impact of adding a
given node v to an existing seed set S:
(cid:88)
(u,w)∈E
h(vS) = − 1
A
1 [{u, w} (cid:54)⊆ S,{u, w} ⊆ S ∪ {v}] .
Now consider some S(cid:48) ⊇ S, v (cid:54)∈ S(cid:48). The indicator
function in the above sum counts edges where one of the
two nodes is not contained in S, but both are contained in
S ∪ {v}. If S(cid:48) extends S but does not contain v, then the
summation for h(vS(cid:48)) can only include more nonzero terms
than the summation for h(vS). Since each term is nonposi-
tive, h(vS(cid:48)) ≤ h(vS).
In fact, h being monotone-decreasing and submodular is
sufficient for the objective f to also be submodular:
Proposition 1. Whenever h is a monotone-decreasing sub-
modular function, f is submodular.
Proof. Consider the marginal gain of adding a node v to a
given seed set S:
f (vS) = h(S ∪ {v})aT xv + [h(S ∪ {v}) − h(S)]
aT xu
(cid:88)
u∈S
We prove that f is submodular by showing that each cor-
responding term in f (vS(cid:48)) can only decrease for all S ⊆ S(cid:48).
The first term decreases since h is a submodular function
as shown in the lemma. The second term, corresponding to
the individual contribution of v, also decreases because h is
monotonically decreasing. Thus f (vS(cid:48)) ≤ f (vS).
Having shown that f is submodular, a natural approach
to seeding would use the greedy algorithm, giving a 1 − 1
e
approximation for the ICM and the LTM (Kempe, Klein-
berg, and Tardos 2003; Leskovec et al. 2007; Chen, Wang,
and Yang 2009). However, since f is non-monotone, this
approach does not apply. Instead, we adopt the stochastic
greedy algorithm proposed by Feldman, Harshaw, and Kar-
basi (2017). Algorithm 1 runs the normal greedy algorithm
(lines 4-6) but only selects from a limited set of nodes P .
Each node is included in P with probability 1/2. This ran-
dom removal reduces the chance that the greedy method will
prematurely commit to a node that later become problematic
due to non-monotonicity. Feldman et al. (2017) show that
4-approximation to the
this algorithm obtains a guaranteed 1
optimal value and has excellent empirical performance. Our
experiments follow their suggested strategy of running the
algorithm several times (we both use 4).
Algorithm 1 StochasticGreedyAJM (V, f, k)
1: initialize S = ∅, P = ∅
2: for v ∈ V do
3:
4: while S < k and ∃v ∈ P such that f (S ∪ v) ≥ f (S)
2 , P ← P ∪ v
with probability 1
do
5:
6:
7: return S
v = arg maxv∈P f (S ∪ v) − f (S)
S ← S ∪ v
Meta-Analysis of Influence Spread Metrics
The finding that the ICM is a poor predictor of node-level
influence is dissonant with the fact that seeding algorithms
based on the ICM have proved effective in the field (Yadav
et al. 2017). After all, how can an algorithm based on an in-
accurate model of diffusion manage to nevertheless achieve
a high level of influence spread? To address this seeming
conflict, we confront a larger question about the prevailing
methodology of the influence maximization problem. In this
section, we show that appealing solely to the magnitude of
influence spread achieved is a fundamentally inconclusive
method of determining whether a particular diffusion model
underlies an observed instance of spread. This ambiguity is
problematic when using the influence maximization frame-
work to inform the seeding strategies of network interven-
tions in sustainability domains. In many such cases, in ad-
dition to diffusing information generally, programs seek to
target particular individuals or groups, and thus a model's
ability to make node-level predictions is a valuable asset.
Magnitude of Influence Spread Previous research com-
paring information diffusion model predictions to empirical
results has tended to rely on metrics related to volume of
spread-such as minimizing RMSE as a function of actual
spread or recapitulating cascade sizes-to determine the fi-
delity of a model to ground truth processes (Goyal, Bonchi,
and Lakshmanan 2011). However, one cannot extrapolate
processes from such coarse-grain outcomes. The following
experiments use three examples of physical, meso-scale net-
works: Homeless, a network of 142 nodes gathered via in-
terviews with homeless youth, India, a household-level net-
work gathered from a rural village in India (Banerjee et al.
2013), and SBM, a synthetic network of 200 nodes gener-
ated via the Stochastic Block Model, which replicates the
community structure found in real social networks.
We evaluate how seed sets selected under one diffusion
model perform in an influence maximization task under
the other models. Figure 4a examines the consequences of
model misspecification for influence maximization. We set
the parameters equally across all networks-0.1 for propa-
gation probabilities in the ICM and AJM and edge weights
in the LTM. Each table entry shows the percentage of opti-
mal influence spread obtained when a seed set selected ac-
cording to the model on the column is assessed with the
model on the row. For example, the cell (ICM, LTM) in-
dicates that a seed set selected via the greedy algorithm for
the LTM produced influence spread that was 99.8% optimal
when diffusion actually occurred under the ICM. Given that
all entries are greater than 90%, determining model fit by
examining the magnitude of influence spread achieved un-
der its seeding strategy leads to great ambiguity. Since all
of these models result in high influence spread, any model
could account for the "true" underlying process of diffusion.
One explanation for this phenomenon points to the com-
munity structure common in social networks. Algorithms
for influence maximization under the ICM tend to distribute
seed nodes across different communities to avoid the redun-
dancy of seeding the same community multiple times. But
seeding according to the AJM results in a similar recom-
mendation to ensure diversity among seed nodes. Hence, at-
Figure 4
99.8
100
96.1
ICM LTM AJM
98.6
ICM
100
98.8
LTM 99.6
AJM 97.4
100
ICM LTM AJM
100
99.8
ICM
98.9
LTM 99.9
AJM 93.7
100
ICM LTM AJM
99.4
ICM
100
LTM 99.3
99.8
100
AJM 96.3
98.4
100
97.8
98.7
100
93.9
(a) Percentage of optimal influence spread achieved when network
is seeded according to the column model and evaluated according
to the row model. Networks top to bottom: Homeless, India, SBM.
(b) Fraction of all nodes informed under ICM and LTM with vary-
ing parameters. Networks left to right: Homeless, India, SBM.
taining high influence spread is insufficient for identifying
a model as the true diffusion mechanism. On the one hand,
achieving comparable final magnitudes of influence spread
is a handy property for influence maximization tasks, as it
suggests that high-quality results are attainable even when
the true model is uncertain. However, many important in-
fluence maximization tasks require a descriptively accurate
diffusion model, not just one that works by coincidence.
Next, in Figure 4b, we show that common diffusion mod-
els are capable of reproducing any observed level of total
influence spread. Each plot gives the fraction of the graph
influenced by a random set of 10 seed nodes under the ICM
and LTM as we vary a parameter for each model. For the
ICM, we vary the propagation probability p. For the LTM,
deg(v) and vary w. Each
we assign each edge (u, v) a weight
line is an average of 30 draws of the random seeds. Under
both models, any level of influence can be explained by a
parameter choice in either of the models. We conclude that
even if a given model exactly replicates the observed amount
of influence spread in a network, this provides no evidence
that the model truly describes the underlying diffusion pro-
cess. Hence, we must use a finer-grained assessment such as
node-level activations to produce accurate diffusion models.
Toward Node-Level Influence Spread The prevailing
methodology's blind spot to node-level information spread
also engenders severe limitations in the actual deployment
of the influence maximization problem in the real world.
Seeding networks to maximize the scalar volume of influ-
ence spread is unproblematic when one is agnostic about
who is influenced. But in many sustainability domains, in-
fluence maximization in the field is not a crude game of
w
coverage, and the individual identities or attributes of influ-
enced nodes carry import. In the Introduction, we referenced
the variety of network interventions that seek to target par-
ticular individuals or sub-populations. Here, a single metric
capturing the magnitude of influence spread achieved is in-
sufficient in determining the success of a network interven-
tion. Thus even when in cases when seeding according to
the AJM rather than the ICM or LTM would result in simi-
lar levels of influence spread, social programs would prefer
seeding according to the model that makes accurate node-
level predictions as it will also be better equipped to target
specific individuals. We have shown that in these cases, the
AJM significantly outperforms leading models.
5 Discussion
A breadth of research has investigated information diffusion
on online networks, but the problem of influence maximiza-
tion remains under-explored on natural networks. By ana-
lyzing node-level data from the first ever large-scale study
of influence maximization on physical social networkw, we
show that neither of the prevailing models of information
diffusion–the Independent Cascade and Linear Threshold–
could account for the empirical findings. Even after fitting
the best of these models to the data, they perform worse than
a random classifier in predicting a node's influence status.
We approached the shortcomings of the dominant mod-
els with an open mind to related research that may inform
our understanding of information diffusion in this domain.
Our proposed Activation Jump Model (AJM) draws from a
lineage of work in multi-agent systems and social network
theory that suggests that 1) social exchange need not only
occur along network ties and 2) an individual's influencing
behavior is affected by her surrounding community. Of par-
ticular note, we model seed set structural diversity as con-
ferring benefits to each node's influencing level.
The AJM is a more inclusive model of diffusion and supe-
rior to leading models in its predictive prowess. When val-
idated on three real-world networks with information diffu-
sion data, the AJM issues predictions of node-level influence
spread that improve upon the best ICM and LTM predictions
by 60% to 110%. Moreover, as the AJM is submodular and
non-monotone, we adopt a seeding algorithm that achieves a
4-approximation to the optimal influence spread. Thus high-
1
efficacy influence maximization under the AJM is computa-
tionally ready to be deployed in the real world.
It has long been accepted in the social sciences that the
link between individual and group social behaviors is bidi-
rectional (Mead 1934). A multi-agent team perspective is
thus particularly suited to describe peer-to-peer informa-
tion diffusion in the natural world, where a group's social
dynamic impacts how individual members will behave in
spreading information. By modeling team-formation, a cen-
tral component of many network interventions, the AJM sig-
nificantly updates the influence maximization problem for
natural world settings. Its framework, with flexible activa-
tion level and landing distribution functional forms, also
allows for contextually relevant information such as node-
specific attributes like gender and ethnicity to be incorpo-
rated when deployed in real-world network applications.
0.00.51.0p0.51.0Fraction informed0246wICMLTM0.00.51.0p0.00.51.0Fraction informed0246wICMLTM0.00.51.0p0.00.51.0Fraction informed0246wICMLTM References
[Agmon and Stone 2012] Agmon, N., and Stone, P. 2012.
Leading ad hoc agents in joint action settings with multi-
In Proceedings of the 11th International
ple teammates.
Conference on Autonomous Agents and Multiagent Systems-
Volume 1, 341–348.
International Foundation for Au-
tonomous Agents and Multiagent Systems.
[Balch 2000] Balch, T. 2000. Hierarchic social entropy: An
information theoretic measure of robot group diversity. Au-
tonomous robots 8(3):209–238.
[Banerjee et al. 2013] Banerjee, A.; Chandrasekhar, A. G.;
Duflo, E.; and Jackson, M. O. 2013. The diffusion of micro-
finance. Science 341(6144):1236498.
[Broadhead et al. 1998] Broadhead, R. S.; Heckathorn,
D. D.; Weakliem, D. L.; Anthony, D. L.; Madray, H.; Mills,
R. J.; and Hughes, J. 1998. Harnessing peer networks as an
instrument for aids prevention: results from a peer-driven
intervention. Public health reports 113(Suppl 1):42.
[Butts 2003] Butts, C. T. 2003. Network inference, error,
and informant (in) accuracy: a bayesian approach. social
networks 25(2):103–140.
[Chen, Wang, and Yang 2009] Chen, W.; Wang, Y.; and
Yang, S. 2009. Efficient influence maximization in social
networks. In Proceedings of the 15th ACM SIGKDD inter-
national conference on Knowledge discovery and data min-
ing, 199–208. ACM.
[DeGroot 1974] DeGroot, M. H. 1974. Reaching a con-
Journal of the American Statistical Association
sensus.
69(345):118–121.
[Fawcett 2006] Fawcett, T. 2006. An introduction to roc
analysis. Pattern recognition letters 27(8):861–874.
[Feldman, Harshaw, and Karbasi 2017] Feldman, M.; Har-
shaw, C.; and Karbasi, A. 2017. Greed is good: Near-optimal
submodular maximization via greedy optimization. arXiv
preprint arXiv:1704.01652.
[Gaston and DesJardins 2003] Gaston, M., and DesJardins,
M. 2003. Team formation in complex networks. In Pro-
ceedings of the 1st NAACSOS Conference.
[Gaston, Simmons, and DesJardins 2004] Gaston, M.; Sim-
mons, J.; and DesJardins, M. 2004. Adapting network struc-
ture for efficient team formation. In Proceedings of the AAAI
2004 fall symposium on artificial multi-agent learning.
[Goyal, Bonchi, and Lakshmanan 2010] Goyal, A.; Bonchi,
F.; and Lakshmanan, L. V. 2010. Learning influence proba-
bilities in social networks. In Proceedings of the third ACM
international conference on Web search and data mining,
241–250. ACM.
[Goyal, Bonchi, and Lakshmanan 2011] Goyal, A.; Bonchi,
F.; and Lakshmanan, L. V. 2011. A data-based approach
to social influence maximization. Proceedings of the VLDB
Endowment 5(1):73–84.
[Granovetter 1978] Granovetter, M. 1978. Threshold mod-
els of collective behavior. American journal of sociology
83(6):1420–1443.
[Hong and Page 2004] Hong, L., and Page, S. E.
2004.
Groups of diverse problem solvers can outperform groups
of high-ability problem solvers. Proceedings of the Na-
tional Academy of Sciences of the United States of America
101(46):16385–16389.
[Kalafat and Elias 1994] Kalafat, J., and Elias, M. 1994. An
evaluation of a school-based suicide awareness intervention.
Suicide and Life-Threatening Behavior 24(3):224–233.
[Kempe, Kleinberg, and Tardos 2003] Kempe, D.; Klein-
berg, J.; and Tardos, ´E.
2003. Maximizing the spread
In Proceedings
of influence through a social network.
of the ninth ACM SIGKDD international conference on
Knowledge discovery and data mining, 137–146. ACM.
[Krackhardt and Kilduff 1999] Krackhardt, D., and Kilduff,
M. 1999. Whether close or far: Social distance effects on
perceived balance in friendship networks. Journal of per-
sonality and social psychology 76(5):770.
[Leskovec et al. 2007] Leskovec, J.; Krause, A.; Guestrin,
C.; Faloutsos, C.; VanBriesen, J.; and Glance, N. 2007.
Cost-effective outbreak detection in networks. In Proceed-
ings of the 13th ACM SIGKDD international conference on
Knowledge discovery and data mining, 420–429. ACM.
[Liemhetcharat and Veloso 2012] Liemhetcharat, S.,
and
Veloso, M. 2012. Modeling and learning synergy for team
In Proceedings of
formation with heterogeneous agents.
the 11th International Conference on Autonomous Agents
and Multiagent Systems-Volume 1, 365–374.
Interna-
tional Foundation for Autonomous Agents and Multiagent
Systems.
[Marcolino, Jiang, and Tambe 2013] Marcolino, L. S.; Jiang,
A. X.; and Tambe, M. 2013. Multi-agent team formation:
diversity beats strength? In IJCAI, volume 13, 279–285.
[Mead 1934] Mead, G. H. 1934. Mind, self and society.
Chicago University of Chicago Press.
[Provost et al. 1998] Provost, F. J.; Fawcett, T.; Kohavi, R.;
et al. 1998. The case against accuracy estimation for com-
paring induction algorithms. In ICML, volume 98, 445–453.
[Schelling 1978] Schelling, T. 1978. Micromotives and mac-
robehavior. nueva york.
[Shafie 2015] Shafie, T. 2015. A multigraph approach to
social network analysis. Journal of Social Structure 16:0 1.
[Ugander et al. 2012] Ugander, J.; Backstrom, L.; Marlow,
C.; and Kleinberg, J. 2012. Structural diversity in social
contagion. Proceedings of the National Academy of Sciences
109(16):5962–5966.
[Valente 2012] Valente, T. W. 2012. Network interventions.
Science 337(6090):49–53.
[Wang et al. 2013] Wang, Y.; Shen, H.-W.; Liu, S.; and
Cheng, X.-Q. 2013. Learning user-specific latent influence
and susceptibility from information cascades. arXiv preprint
arXiv:1310.3911.
[Yadav et al. 2017] Yadav, A.; Wilder, B.; Rice, E.; Peter-
ing, R.; Craddock, J.; Yoshioka-Maxwell, A.; Hemler, M.;
Onasch-Vera, L.; Tambe, M.; and Woo, D. 2017. Influence
maximization in the field: The arduous journey from emerg-
In Proceedings of the 16th
ing to deployed application.
Conference on Autonomous Agents and MultiAgent Systems,
150–158. International Foundation for Autonomous Agents
and Multiagent Systems.
[Zhou, Zha, and Song 2013] Zhou, K.; Zha, H.; and Song,
L. 2013. Learning triggering kernels for multi-dimensional
hawkes processes. In Proceedings of the 30th International
Conference on Machine Learning (ICML-13), 1301–1309.
|
1804.08419 | 1 | 1804 | 2018-04-20T10:24:31 | A Robust Process to Identify Pivots inside Sub-communities In Social Networks | [
"cs.SI"
] | In this work, we extend a previous work where we proposed a suitable state model built from a Karhunen-Loeve Transformation to build a new decision process from which, we can extract useful knowledge and information about the identified underlying sub-communities from an initial network. The aim of the method is to build a framework for a multi-level knowledge retrieval. Besides the capacity of the methodology to reduce the high dimensionality of the data, the new detection scheme is able to extract, from the sub-communities, the dense sub-groups with the definition and formulation of new quantities related to the notions of energy and co-energy. The energy of a node is defined as the rate of its participation to the set of activities while the notion of co-energy defines the rate of interaction/link between two nodes. These two important features are used to make each link weighted and bounded, so that we are able to perform a thorough refinement of the sub-community discovery. This study allows to perform a multi-level analysis by extracting information either per-link or per-intra-subcommunity. As an improvement of this work, we define the notion of pivot to relate the node(s) with the greatest influence in the network. We propose the use of a thorough tool based on the formulation of the transformation of a suitable probabilistic model into a possibilistic model to extrac | cs.SI | cs |
A Robust Process to Identify Pivots inside
Sub-communities In Social Networks
Joseph Ndong1 and Ibrahima Gueye2
1 Department of Mathematics and Computer Science
University Cheikh Anta Diop, BP 5005 Fann Dakar, Senegal
2 Computer Engineering and Telecom, Polytechnic School, Thies, Senegal
[email protected], [email protected]
Abstract. In this work, we extend a previous work where we proposed a suitable
state model built from a Karhunen-Loeve Transformation to build a new decision
process from which, we can extract useful knowledge and information about the
identified underlying sub-communities from an initial network. The aim of the
method is to build a framework for a multi-level knowledge retrieval. Besides
the capacity of the methodology to reduce the high dimensionality of the data,
the new detection scheme is able to extract, from the sub-communities, the dense
sub-groups with the definition and formulation of new quantities related to the
notions of energy and co-energy. The energy of a node is defined as the rate of
its participation to the set of activities while the notion of co-energy defines the
rate of interaction/link between two nodes. These two important features are used
to make each link weighted and bounded, so that we are able to perform a thor-
ough refinement of the sub-community discovery. This study allows to perform
a multi-level analysis by extracting information either per-link or per-intra-sub-
community. As an improvement of this work, we define the notion of pivot to
relate the node(s) with the greatest influence in the network. We propose the use
of a thorough tool based on the formulation of the transformation of a suitable
probabilistic model into a possibilistic model to extract these pivot(s) which are
the nodes that control the evolution of the community.
Keywords: Social network analysis, community detection, Energy, Pivot, Influencer;
1
Introduction
Social networks describe web-based services that allow users (individuals) to connect
with other users, communicate, share or publish contents within the network [1], [2].
The rise of web 2.0 has come with the ease in the production and sharing of content that
allowed the social networks development. Users themselves become the producers of
web content [1], [3]. In a formal and simple description, social network can be seen as
a graph consisting of nodes (as individuals) and links (as social links) used to represent
social relations on social network sites [4] (as presented in Figure 1).
These last years, Social networks sites have become some important sources of on-
line interactions, contents sharing [6] and communication means. While giving these
opportunities, they are affordable and universally acclaimed. Social network sites are
2
Fig. 1. Social Network Showing Nodes and Links (Source: [5])
commonly known for information dissemination, personal activities posting [7], prod-
uct reviews, online pictures sharing, professional profiling, advertisements, subjectivity
[7], assessments [8], approaches [9], influences [10], observations [11], feelings [3],
opinions and sentiments expressions [12], news, remarks, reactions, or some many other
content type [13]. News alerts, breaking news, political debates and government policy
are also posted and analyzed on social network sites.
Obervations [14] show that more and more people are becoming interested and in-
volved in social networks. Sometimes, users of social networks exploit them for mak-
ing decisons. These decisions can be maked based on know or famialiar users or, in
some case, based on unfamiliar users [12]. Such a situation increases the degree of con-
fidence in the credibility of these sites. The social network has transformed the way
different entities procure and retrieve valuable information regardless of their location.
The social network has also given users the privilege of giving their opinion.
The massive data generated or produced by all the social networks users allows
for thorough analysis to efficiently extract useful knowledge such as trends, opinion
leaders, influencers, feelings, etc. Such an analysis can be done as a whole, which means
the representations of all the social network actors then proceed to the detection of
communities and/or opinion leaders. The main aim of community detection methods is
to partition the network into dense regions of the graph. Those dense regions typically
correspond to entities which are closely related. The determination of such communities
is useful in the context of a variety of applications such as customer segmentation,
recommendations, link inference and influence analysis [15].
In this work, we focus on community influencers identification. We extend our
previous work [16] in which, we derive a suitable state model from Karhunen-Loeve
Transformation [17]. From this state model we built a new decision process, which al-
lows us to extract useful knowledge and information about the identified underlying
sub-communities from an initial network.
This extension is interested in the extraction of the most "important" nodes in given
sub-communities already detected or known. Identifying sub-groups under a social net-
work is a great challenge in the area of social network analysis [18], [19], [20], but
analyzing the intrinsic behavior of each sub-group might be an important need if one
desires to focus on the quality of nodes/actors [21]. If sub-communities are identified,
3
decisions could be taken on a group of nodes regardless the underlying contents of the
group itself since it can be viewed as a single homogeneous entity in which all items
are of same behavior. One could also see the sub-community as a machine where the
dynamic structure and the evolution depend strongly on the nature and quality of its
combined pieces. If this second point-of-view is adopted, we need a tool to identify the
"most important" pieces whose actions are more influential than those of the rest of the
group. We define these pieces as "pivots" and the quality of the network is as impor-
tant as it contains influential pivots with high degree of influence. In the following, we
define pivots and how they could be important for a manager, a advertiser and so on.
Inside a sub-community, nodes share links so, identifying pivots can be helpful in
many situations:
-- (i) if in a sub-community, the pivot(s) is/are linked to a few other nodes, the man-
ager can conclude that this network is of less quality simply because the "most"
important nodes have less interactivity and thus less influence. But in the contrary,
he can pay more attention to the underlying evolution of the group;
-- (ii) the different sub-communities can be classified in an ascending order based on
their quality -- the first group is the one with the greater number of pivots, and so on;
-- (iii) in a sub-community, we can find other sub-communities in the following man-
ner: If the pivots themselves are not linked to each other, each of them can be the
center of a group/entity formed by all the nodes linked to a pivot.
The procedure for sub-communities detection/identification we built in our previ-
ous work [16] consisted in extracting a decision variable from a state-space model we
defined with a Karuhen-Loeve Transformation (KLT). This have served also to reduc-
ing the dimensionality of the dataset in order to maintain the only relevant part of the
data. Then, after defining some new quantities as "energy" of an actor/node and the "co-
energy between" actors, we apply a decision process to identify the sub-communities
and the features inside each of them. The improvement we bring in this present work
is the ability to learn more about the impact of node's energy to see if some nodes
could be selected as those with the highest degree of influence. The notion of highest
degree must be properly define to achieve this aim. We found that a suitable definition
of a probability density function (pdf) related to the defined node's energy can be done.
Then, a second definition of a possibilistic model will be able to discover the relevant
node(s) we classify as pivot(s).
Thes rest of this paper is organized as follow: In section 2 we do a thorough study
of the related work; in section 3 we detail the methodology and aogorithm for the sub-
commumity detection; in section 4 we present our use of possibility theory to identifiy
the pivots; in Section 5 we present the results of our experimental validation of our
propositions, and we conclude this paper in Seciton 6.
2 Background
Social sites have undoubtedly bestowed unimaginable privilege on their users to access
readily available never-ending uncensored information. Twitter, for example, permits its
4
users to post events in real time way ahead the broadcast of such events on traditional
news media. Also, social network allow users to express their views, be it positive or
negative [21]. Organizations are now conscious of the significance of consumers' opin-
ions posted on social network sites to the patronage of their products or services and the
overall success of their organisations. On the other hand, important personalities such
as celebrities and government officials are being conscious of how they are perceived
on social network. These entities follow the activities on social network to keep abreast
with how their audience reacts to issues that concerns them [22].
Opinion of influencers on social network is based largely on their personal views
and cannot be hold as absolute fact. However, their opinions are capable of affecting the
decisions of other users on diverse subject matters. For example, Rihanna's rant has cost
Snapchat $800 million in one Day [23]. Opinions of influential users on Social network
often count, resulting in opinion formation evolvement. Clustering technique of data
mining can be used to model opinion formation by assessing the affected nodes and
unaffected nodes. Users that depict the same opinion are linked under the same nodes
and those with opposing opinion are linked in other nodes. This concept is referred to
as homophily in social network [24]. Homophily can also be demonstrated using other
criteria such as race and gender [25].
Researchers in social network analysis are facing many other research issues and
challenges such as in Linkage-based and Structural Analysis, or in Dynamic Analysis
and Static Analysis.
Linkage-based and Structural Analysis consists of anlyzing of the linkage behaviour
of the social network so as to ascertain relevant nodes, links, communities and imminent
areas of the network - Aggarwal (2011) [21].
Static analysis, such as in bibliographic networks, is presumed to be easier to carry
out than those in streaming networks. In static analysis, it is presumed that social net-
work changes gradually over time and analysis on the entire network can be done in
batch mode. Conversely, dynamic analysis of streaming networks like Facebook and
YouTube are very difficult to carry out. Data on these networks are generated at high
speed and capacity. Dynamic analysis of these networks are often in the area of interac-
tions between entities such as in Papadopoulos et al. 2012 [26] and Sarr et al. 2014 [27].
Dynamic analysis is also covered through temporal events on social networks such as in
Adedoyin-Olowe et al (2013) [28] and Becker et al (2011) [18], and we also have such
dynalysis through the study of evolving communities - Fortunato, (2010) [19], Sarr et
al. 2014 [20].
3 Methodology and algorithm for the sub-communities detection
This methodology tracks and detects sub-communities based on the analysis of a huge
number of features corresponding to events/activities for which a group of actors/nodes
participate. First, we aim at finding the main features, to incorporate in our model,
by means of extended principal component analysis. The second relevant issue of this
methodology is related to the specification of a new detection procedure consisting
of merging all the relevant features into a single process we will label as a "Decision
Variable" (DV). By analyzing this process for the sub-community tracking operation,
5
we can discover subgroups of actors using a multi-level thresholding and the notion of
"energy dissipation" of an actor over the events.
1, . . . ,xi
K) i = 1, . . . ,R.
We consider a community of R actors Ω = (a1, . . . ,aR) which perform activities
on a set of K initial correlated events (e1, . . . ,eK). For each event ek, we have a column
vector of size R containing the amount of participation of all R actors to the correspond-
ing activity. This operation gives us the R× K matrix of correlated random variables
X = (X1, . . . ,XK). In other words, one observes these random variables through R inde-
pendent realization vectors xi = (xi
After extracting the relevant components from the Karhunen-Loeve transformation [16],
we can build our decision variable as a row vector DV = (y1, . . . ,yK). Then we can set
a certain number of concepts for our methodology. We introduce the notion of "energy
dissipation" (Ed) to quantify the degree of importance a given actor puts on a series of
events. This notion is simple and intuitive. When considering the set of events/activities,
the events for which the actor puts a high degree of importance constitutes his energy.
For example, we can consider money as energy. When someone goes to buy some prod-
ucts, we can say that he/she is dissipating a certain amount of his/her energy. In this
case, he/she should buy a "product A", and consequently buy another "product B" nec-
essary to use the product A. Here, we can see the notion of correlation between these
products/variables. When an athlete performs several disciplinary exercises in sport, we
can view his actions as the dissipation of his energy over the different events, in order
to win a medal. The energy of an actor is then quantifiable, its a measure of the strength
of his participation to the a series of activities.
If the actor participates actively to all or most of the activities with a high intensity,
then his energy increases, otherwise we say that this actor has less energy according to
the ensemble of events happening at a given period of time.
Since the DV variable contains the aggregated amount of all actors participation to
all events, the energy dissipation Ed of an actor i is the row vector defined as:
Edi =(cid:8)k, /xi
k ≥ DV[k],∀k = 1, . . . ,K(cid:9)
(1)
Edi contains all the index of events for which the energy dissipation is greater than
the reference DV. Consequently, we can calculate the total energy of the actor i as:
Ei =
Edi
DV
(2)
where . indicates the size of a vector.
We also refer to the notion of "co-energy" dissipation (CED) as the amount of en-
ergy between two actors according to their participation to the same set of activities.
This quantity is a measure of the mean energy produced simultaneously by the two
actors on the same activities:
CEDi j =
(Edi ∩ Ed j)
DV
(3)
Finally, our detection procedure boils down to fixe a threshold α and put a link
between actor i and actor j if the rate of their co-energy exceeds the limit α. This
6
means the following inequality must be held to add the link:
CEDi j ≥ α
(4)
When Eq. 4 holds, the value of CEDi j becomes the weight of the link between actor
i and actor j. And then, this link is bounded by the interval [min(Ei,E j),max(Ei,E j)].
By varying the threshold α ∈ [0;1] , one can build many different sub-communities
with the same dataset, each sub-community with a score α which measures its degree
of realization. The algorithm to achieve our aim is described as follow:
Algorithm 1: Sub-Community Discovering
Input
Output
: Ct, a community
Ω (Ct ), the sets of actors within Ct
xi = (xi
DV = (dv1, . . . ,dvK) the decision variable
α, the link detection threshold
: V, a sub-community
1, . . . ,xi
K) the vector of participation of actor i
1 /* Calculate Co-Energy dissipation between actors and apply
threshold to add link*/
foreach (k,l) ∈ Ω (Ct ), k (cid:54)= l do
/* Apply Eq. (1)*/
2 begin
3
4
Edk =(cid:8)p, /xk
Edl =(cid:8)p, /xl
p ≥ dvp,∀p = 1, . . . ,K(cid:9)
p ≥ dvp,∀p = 1, . . . ,K(cid:9)
5
6
7
8
9
10
11
12
/* Apply Eq. (3)*/
CEDkl =
(Edk∩Edl )
DV
/*Apply threshold to decide to put a link, Eq. (4)*/
if CEDkl ≥ α then
addLink(V,k,l)
return V
4 Using possibility theory to identify pivot(s)
Our pivot identification methodology relies mainly in the feature of energy we defined
for the actors/nodes. Pivots refer to nodes with "high" energy. These nodes have the
opportunity to control the dynamic evolution of the network. If the energy of a pivot
decreases or increases, the structure of the network might evolve towards a new direc-
tion allowing to suppress or add links between nodes. A simple question arises from the
perspective of pivots identification: how much energy is necessary for a node to be clas-
sified as a pivot? We believe that it is very difficult to answer this question by analyzing
only the amount of energy of each node. If someone would like to do so, he/she should
build a kind of threshold and apply the decision to put the label "pivot" on a node if
7
its energy exceeds this limit. This methodology weakens the objectivity of pivots' iden-
tification. To surround this difficulty, we propose a more robust identification scheme
based on a link between probability theory and possibility theory.
The energy property can be viewed as a continuous random variable X : Ω → V ,
where Ω is the set of actors/nodes and V the measurable function giving the real value
of the energy as defined in equation 2. X does not return a probability. But we want to
know, if it were the case, could this probability help achieving our goal.
Probability theory is a valuable quantitative tool to study randomness/uncertainty in
random phenomena. In this area, we can find the probability of occurrence of different
possible outcomes in an experiment. If a random variable returns a probability P, then
V = [0,1] and ∑
P(X = u) = 1. As we defined the energy in equation 2, the energy Ei
of actor i is always between 0 and 1. So, we can take V = [0,1], but we do not have
K
∑
Ei = 1 (K is the number of nodes).
u
i
4.1 Defining probability to characterize energy
u
For the purpose to return a probability from X, we have just to normalize the energy by
the size of E, where E = {Ei,i = 1 . . . ,Ω}. At this point, the random variable X returns
a probability since the equation ∑
P(X = u) = 1 holds. We build this probability to
characterize each energy of a node with a probability of occurrence in [0,1]. Now, using
this probability definition, our scope is to build a test to identify the potential pivots
under each sub-community. Generally, a typical test consists of applying a threshold
on the outcomes (probabilities of energy) to decide to put the label pivot on a node if
the probability distribution takes a value higher than this threshold. The task to build
this threshold is not straightforward even if, for given nodes if the probability of their
respective energy is 0.95, 0.6 and 0.3 for example, how could we decide to label a
node as pivot, based only on the not obvious notion of "high" probability. Is 0.6 a
"high" or "low" probability? By pointing out this example, we simply want to show
that it is not evident to know the "best" value of the probability threshold in order to
conclude if the node is a pivot or not. Nevertheless, one could, for simplicity, build
an heuristic decision process where the threshold is set manually. With probability, the
only evident decision we can take is, when the probability of a node is P(X = u) = 1.
If this case happens, it means that there's only one pivot in the entire sub-community,
since the other nodes have probability 0 and so this sub-community is built with only
one node. Finally, we believe that pivots might be nodes with any probability, different
to zero, "sufficient" to become member of a sub-community and to have the potential to
change the dynamic evolution of the network. To go towards the direction of extracting
the "best" level of probability measure, we think of an alternative related the area of
possibility theory, which gives as another tool to represent uncertainty in a qualitative
fashion. This tool can also helps to learn more about the incompleteness that reflects
the lack of information. We have seen above that, our defined probability doesn't give
us the information about the limit to apply to detect pivots. So, the idea behind the use
8
of this new scheme is to associate to each node, in accordance to its energy, a degree of
possibility which quantifies the level of "importance" of that node among the others.
4.2 Detecting pivot by possibility degree
∀A ⊆ Ω
P(A) ≤ Π (A)
Between probability and possibility, we can state a consistent principle is this terms:
"what is probable should be possible" [29]. This requirement can then be translated as
follow:
(5)
where P and Π are, respectively, the probability and the possibility measure on the
domain Ω. In this case, Π is said to dominate P. With Π, given nodes should have
the maximum possibility degree, i.e. the value 1. In possibility theory the equality
∑
Π (X = u) = 1 is not guaranteed. So, for any node, when its possibility degree reaches
u
the maximum, we can robustly say that this node is a pivot since it is entirely sure that
it might exist in the network. So, we see that the notion of "high" probability can be
defined properly, since it corresponds to the maximum degree of possibility.
: V → [0,1], and is defined by:
A possibility measure, [30], Π on V is characterized by a possibility distribution π
∀A ⊆ V,Π (A) = sup{π(v),v ∈ A}.
(6)
If V is a finite set, thus : ∀A ⊆ V,Π (A) = max{π(v),v ∈ A}. The key concept of a possi-
bility distribution is the preference ordering it establishes on V . Basically, π designates
what one knows about the value of a given variable X, and π(v) > π(v(cid:48)) states that
X = v is more plausible than X = v(cid:48). When π(v) = 0, thus, v is an impossible value of
the variable X while π(v) = 1 means that v is one of the most plausible values of X. For
us, identifying pivots is just a process to searching at these most plausible values.
Transforming a probability measure into a possibilistic one then amounts to choos-
ing a possibility measure in the set ℑ(P) of possibility measures dominating P. This
should be done by adding a strong order preservation constraint, which ensures the
preservation of the shape of the distribution:
pi < p j ⇔ πi < π j
∀i, j ∈ {1, . . . ,q},
(7)
where pi = P({Ei}) and πi = Π ({Ei}),∀i ∈ {1, . . . ,K}. It is possible to search for the
most specific possibility distribution verifying (5) and (7). The solution of this problem
exists, is unique and can be described as follows. One can define a strict partial order
P on Ω represented by a set of compatible linear extensions Λ (P) = {lu,u = 1,L}. To
each possible linear order lu , one can associate a permutation σu of the set {1, . . . ,q}
such that:
σu(i) < σu( j) ⇔ (ωσu(i),ωσu( j)) ∈ lu,
(8)
The most specific possibility distribution, compatible with the probability distribution
(p1, p2, . . . , pK) can then be obtained by taking the maximum over all possible permu-
tations:
(9)
πi = max
u=1,L
∑
p j
{ jσ−1
u ( j)≤σ−1
u (i)}
9
Finally, the vector (π1,π2, . . . ,πK) gives us all the possibility degrees for the K nodes,
corresponding to the (p1, p2, . . . , pK) vector of probability of their energies. And pivots
are nodes for which the possibility degree exceeds a given rate δ . For this study we set
this threshold δ = 1, it corresponds to the maximum value a possibility degree might be
set. A more flexible and non-heuristic method might be to set δ to a value less than the
maximum. But, here, we set δ to 1 in order to show that in all situations or scenario, a
pivot must be found.
For a thorough view of possibility theory, we recommend the reader to [29,30,31].
5 Validation
We validate our approach on the real world collection of data coming from Reddit.com
[32]. We use several samples of different sizes and, build four scenarios A, B, C and D
with dimension (N × K, N the number of actors and K the number of events) 10× 15,
10 × 150, 10 × 500 and 10 × 1200 respectively. In Table .1, we give an idea on the
content of the data, in each column vector, we have the total amount of submissions to
an image by the set of actors.
PPPPPPPP
Actors
1
2
3
4
5
6
7
8
9
10
events e1 e2 e3 e4 e5 e6 e7 e8 e9 e10 e11 e12 e13 e14 e15
0
0
2
0
0
0
0
0
0
0
11 0 11 4 0 2 0 4 18 2
5 0 0 0 0 0 1 0 0 0
1 0 2 0 3 1 1 0 1 0
4 1 0 0 0 1 2 0 7 0
1 0 0 0 0 0 0 0 0 0
1 0 0 0 0 0 0 0 0 0
0 2 0 0 1 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0
2
0
1
0
0
0
0
0
0
6 16 1
0
0
0
2
1
0
0
0
1
0
4
0
0
0
0
0
0
0
2
1
9
0
0
0
0
0
0
Table 1. Activities and amount of actor participation to submissions on events. Scenario A.
Two levels of information can be retrieved from the results. In the first level, we have
the results about the formation of the underlying sub-communities. This result corre-
sponds to the natural clustering of the different nodes according to the energy provided
by each of them. The second level of information refers to the characteristics of links
and nodes inside the given sub-groups. This refinement provides useful information
when one wants to emphasize and explore some parts of the network.
5.1
Information about the formation of sub-communities
One of the main objectives of social network analysis is related to clustering in order
to study the similarities inside the network [33,34]. So, the first result is about the for-
mation of sub-communities. The graphs in Fig. 2 show all the groups we discover with
the different scenarios. In Fig. 2(a), we see a sparse sub-community within two dense
10
Table 2. Total Energy of each actor for the different scenarios.
a1
a2
a3
a4
a5
Energy of the different actors (Eq. 2)
a8
Ei 0.80 0.66 0.53 1 0.26 0.40 0.60 0
0
Ei 0.88 0.25 0.86 0.77 0.46 0.25 0.74 0.62 0.15 0.13
Ei 0.79 0.26 0.79 0.74 0.23 0.19 0.69 0.46 0.11 0.06
Ei 0.86 0.27 0.70 0.68 0.25 0.23 0.74 0.31 0.97 0.55
a6
a7
a9 a10 Number of events
0
15
150
500
1200
sub-communities. The term 'sparse' refers to a group of nodes with different levels of
energy. When we observe a group of linked nodes with the same value for their energy,
we consider this group as an inner dense sub-community; here, we have {1,3,7} and
{2,3,7}.
In graphs of Fig. 3, we draw the co-energy participation between two nodes to em-
phasise the fact the role of the Decision Variable DV have to set a potential link. Con-
sidering all the events at the same time, a link can be put between two nodes if the
amount of participations of both two nodes, for the same set of events, exceeds by far
the reference point given by the decision variable. We use circles to identify the events
where the energy of the actors is higher than the reference point. Clearly, for most of
the given events, if the energy of each node exceeds the value of the DV for that events,
then we put a link between the two nodes.
5.2
Information about the intrinsic behavior inside sub-communities
The second level of information this technique might deliver is about the dynamics of
nodes and their relation inside the detected sub-groups. So, another result is related to
the boundary of each detected link. In Table 2, we have for each node, its total energy,
i.e. the amount of power of this node to participate to all events. By observing carefully
this table and the graphs of the sub-communities, we see that the bounds of a link is the
interval [a,b], where a and b are the respective total energy of the specified nodes. And
so, the score/weight of a link is always inside this interval, as we can observe for all links
detected. For example, in Fig. 2(d), the link between node #4 (with energy 0.68) and
node #7 (with energy 0.74)has a weight of 0.59 and its bounded interval is [0.68,0.75].
As the network evolves, whenever the energy of a node belongs out of the interval, the
link will disappear. By inspecting frequently the evolution of the bounded interval, one
can retrieve useful information about the degree of importance of the different nodes by
analyzing their energy.
Each of the other scenarios (B, C and D) give also a sparse sub-community.
5.3 Results and discussion about the pivots
We apply the detection process to identify pivots to the data samples for scenarios A,
B, C, D, E, F and G. We have arbitrarily chosen the size of the sample and the interval
where data come from. We just want to show cases where there's one pivot or more.
The table 3 resumes the data selection scheme, and in we put the results of the detection
11
sub-
(a) A sparse
community
within
two dense intra sub-
communities. Scenario
A with 15 events.
(b) A sparse
sub-
community. Scenario B
with 150 events.
(c) A sparse
sub-
community. Scenario C
with 500 events.
(d) A sparse
community.
D with 1200 events.
sub-
Scenario
Fig. 2. Results of the sub-community members identification. In scenario A 2(a), we can observe
two dense sub-communities. In the other scenarios (B, C and D), we obtain only a sparse sub-
community. The different scenarios are built with different size of the vector of events.
(a) Co-Energy between actor 1
and actor 2
(b) Co-Energy between actor 1
and actor 3
(c) Co-Energy between actor 2
and actor 7
(d) Co-Energy between actor 4
and actor 7
Fig. 3. Impact of the Co-Energy dissipation between two actors in the link detection phase. The
circles represents events where both the two actors have their co-energy higher than the refer-
ence value in the decision variable DV. These actors are linked since their co-energy dissipation
concern more than half of the events.
246810121402468101214Linkdetection:ImpactoftheCo-Energydissipationbetween:1-2EventsIDNumberofactorparticipation DVActor#1Actor#2246810121402468101214Linkdetection:ImpactoftheCo-Energydissipationbetween:1-3EventsIDNumberofactorparticipation DVActor#1Actor#3246810121400.511.522.533.544.5Linkdetection:ImpactoftheCo-Energydissipationbetween:2-7EventsIDNumberofactorparticipation DVActor#2Actor#72468101214012345678Linkdetection:ImpactoftheCo-Energydissipationbetween:4-7EventsIDNumberofactorparticipation DVActor#4Actor#7 12
procedure in table 4 where the variable NBE refers to the total amount of participation
on events by nodes, p is the probability of the energy and π the corresponding degree of
possibility. The results give many useful information. For the scenarios A, B, C and D,
we extract only one pivot which have the particularity to be the node with the highest
amount of activities in the network. The pivot also has the highest probability of energy
but, the value of this probability seems to be "very low" in a pure point-of-view of the
probability theory. For example, for scenario A, the pivot has a probability p = 0.30 but
its possibility degree reaches π = 1. For scenarios A, B, C and D, we find only one pivot
and it corresponds to the node with the highest probability and which have perform the
greater amount of activities in the network. But this findings are not a generality since
we see, in the scenario G, the pivot has the highest probability but it does not have the
greatest amount of activities. This discover says clearly that, learning only the amount
of activities is not a sufficient process to analyze a network in order to detect links or to
put a level of importance/quality to nodes. In scenarios E and F, we detect twos pivots
which have not the same amount of actions but they have the same probability of energy.
We think that this situation is due to the fact that the quality of node depends not only
to its level of participation on events but, it depends also to its interaction with other
nodes. So one must have two (or more) nodes with different level of actions on events
but they have the same behavior regarding to other nodes.
Scenario A
15
Size
B
150
C
500
D
1200
E
155
F
255
G
555
interval [1:15] [1:150] [1:500] [1:1200] [1100:1254] [1000:1254] [700:1254]
Table 3. Information about the selection of the different data samples
6 Conclusion
In this work, we have extent a new technique related to an extended version of principal
component analysis to build a methodology for the purpose of community detection
in a social network. The initial work built a technique more elaborated to run within
stochastic process than the classical PCA which is designed originally to solve the prob-
lem of dimensionality reduction for univariate dataset. The main innovation of the work
is manifold: (i) we define the notion of co-energy between two nodes to quantify the
intensity of their relation, (ii) we can also extract the proper energy of a given node
to know how it influences the overall community, (iii) technically, the KL-PCA tech-
nique makes possible to build a decision variable and to form a state model from which
we apply a decision process to identify each link. The introduction of the notion of
energy make possible to see potential intra sub-communities (i.e. nodes with the same
co-energy) inside a sub-community; (iv) each detected link is bounded, so we know how
much energy is necessary to maintain a link over time. In this complementary study, we
show that a possibility distribution can be properly defined from the energy to solve an
interesting feature i.e, the problem of identifying pivots which have the main impact in
Table 4. Detection of pivots when degree of possibility π = 1
13
a7
7
a8 a9 a10 scenario
a6
2
0
0 0.10 0
0 0.20 0
0
0
0
A
1
1
1
1
p
π
p
π
p
π
p
π
0 0.11 0
0 0.22 0
a2
10
a5
Nodes i a1
NBE
75
1
0.30 0.10 0.25 0.25 0
0.20 0.70 0.70 0
a3
14
a4
26
0.39 0.11 0.17 0.22 0
0.22 0.38 0.61 0
31 117 53 10
0
0 0.11 0
0 0.17 0
0
0
0
0
pivot Yes No No No No No No No No No
NBE 591 28 218 142 44
8
0
0.42 0.06 0.18 0.23 0
0.06 0.35 0.59 0
0
pivot Yes No No No No No No No No No
NBE 1638 108 668 348 120 63 348 170 27 16
0
0
pivot Yes No No No No No No No No No
NBE 3659 257 1596 773 206 164 890 425 59 40
0
0
pivot Yes No No No No No No No No No
NBE 428 34 204 95
4
0
0
pivot Yes No Yes No No No No No No No
NBE 702 50 321 140 38
8
0.32 0.07 0
0.23 0
54 122 82 13
0 0.07 0.15 0 0.07
0 0.23 0.38 0 0.23
pivot Yes No Yes No No No No No No No
NBE 1569 127 663 335 72
81 405 170 22 16
0
0
pivot No No Yes No No No No No No No
6
0 0.06 0.06 0.06 0.06 0
0 0.25 0.25 0.25 0.25 0
0 0.06 0 0.18 0.06 0
0 0.12 0 0.31 0.12 0
0.39 0.11 0.17 0.22 0
0.22 0.38 0.61 0
0 0.11 0
0 0.22 0
0.32
0.62
0
0
0
0
0
0
0
0
0
0
0.38
0.38
26
40
52
40
0.38
p
π
p
π
p
π
0.32
1
1
1
1
1
B
C
D
E
F
G
the dynamic nature of the network. From this work, we plane to learn the impact of the
number of pivots in a given sub-community and between sub-communities to face the
idea related to their impact on a network distributed in many geographic area.
References
1. Adedoyin-Olowe, M., Gaber, M.M., Stahl, F.T.: A survey of data mining techniques for social
media analysis. JDMDH 2014 (2014)
2. Chen, Z., Kalashnikov, D.V., Mehrotra, S.: Exploiting context analysis for combining multi-
ple entity resolution systems. In: ACM SIGMOD International Conference on Management
of Data. SIGMOD '09, pp. 207 -- 218 (2009)
3. Kaplan, A.M., Haenlein, M.: Users of the world, unite! the challenges and opportunities of
social media. Business Horizons 53(1), 59 -- 68 (2010)
4. Borgatti, S.P.: In: Meyers, R.A. (ed.) Social Network Analysis, Two-Mode Concepts in, pp.
2912 -- 2924. Springer, ??? (2012)
14
5. Amiri, S.E.: Node Classification in Social Networks. https://filebox.ece.
vt.edu/~s14ece6504/projects/esorour_socialnets/, visited 2017-10-22
(2014)
6. Chelmis, C., Prasanna, V.K.: Social networking analysis: A state of the art and the effect of
semantics. In: IEEE, International Conference on Social Computing, pp. 531 -- 536 (2011)
7. Asur, S., Huberman, B.A.: Predicting the future with social media. In: Proceedings of the Intl.
Conf. on Web Intelligence and Intelligent Agent Technology, pp. 492 -- 499. IEEE Computer
Society, ??? (2010)
8. Kim, Y., Hsu, S.-H., de ZÞôsiga, H.G.: Influence of social media use on discussion network
heterogeneity and civic engagement: The moderating role of personality traits. Journal of
Communication 63(3), 498 -- 516 (2013)
9. Korda, H., Itani, Z.: Harnessing social media for health promotion and behavior change.
Health Promotion Practice 14(1), 15 -- 23 (2013)
10. Bakshy, E., Hofman, J.M., Mason, W.A., Watts, D.J.: Identifying â AŸinfluencersâ A ´Z on
twitter. In: Fourth ACM International Conference on Web Seach and Data Mining (WSDM)
(2011)
11. Atkinson, N., Chou, W.-y.S., Hunt, Y.M., Beckjord, E.B., Moser, R.P., Hesse, B.W.: Social
media use in the united states: Implications for health communication. In: Journal of Medical
Internet Research (2009)
12. Pang, B., Lee, L.: Opinion mining and sentiment analysis. Found. Trends Inf. Retr. 2(1-2),
1 -- 135 (2008)
13. Liu, B.: Sentiment Analysis and Opinion Mining. Morgan & Claypool Publishers, ???
(2012)
14. Statista: Number
billions).
(in
number-of-worldwide-social-network-users/, visited 2017-10-09 (2017)
social media
2021
https://www.statista.com/statistics/278414/
users worldwide
of
from 2010
to
15. Khatoon, M., Banu, A.: A survey on community detection methods in social networks. I.J.
Education and Management Engineering (1), 8 -- 18 (2015). doi:10.5815/ijeme.2015.01.02
16. Ndong, J., Gueye, I.: A new decision technique for sub-community and multi-level knowl-
edge extraction in social networks. In: Complex Networks & Their Applications V: Pro-
ceedings of the 5th International Workshop on Complex Networks and Their Applications
(COMPLEX NETWORKS), vol. 693, p. 263. Springer, ??? (2016)
17. Gray, R.M., Davisson, L.D.: An Introduction to Statistical Signal Processing. Cambridge
University Press, ??? (2004)
18. Becker, H., Naaman, M., Gravano, L.: Beyond trending topics: Real-world event identifica-
tion on twitter. In: Proceedings of the Fifth International Conference on Weblogs and Social
Media (2011)
19. Fortunato, S.: Community detection in graphs. Physics Reports 486(3), 75 -- 174 (2010)
20. Sarr, I., Ndong, J., Missaoui, R.: Overlaying social networks of different perspectives for
inter-network community evolution. In: Social Network Analysis-Community Detection and
Evolution, pp. 45 -- 70. Springer, ??? (2014)
21. Aggarwal, C.C.: An Introduction to Social Network Data Analytics, pp. 1 -- 15 (2011)
22. Castellanos, M., Dayal, U., Hsu, M., Ghosh, R., Dekhil, M., Lu, Y., Zhang, L., Schreiman,
M.: Lci: A social channel analysis platform for live customer intelligence. In: ACM SIG-
MOD International Conference on Management of Data, pp. 1049 -- 1058. ACM, ??? (2011)
23. FR, Y.N.: Rihanna's rant has cost Snapchat $ 800 million. https://fr.news.yahoo.
com/coup-gueule-rihanna-fait-perdre-114958692.html, visited 2018-
03-19 (2018)
24. McPherson, M., Smith-Lovin, L., Cook, J.M.: Birds of a feather: Homophily in social net-
works. Annual Review of Sociology 27(1), 415 -- 444 (2001)
15
25. Jackson, M.O.: Social and Economic Networks. Princeton University Press, Princeton, NJ,
USA (2008)
26. Papadopoulos, S., Kompatsiaris, Y., Vakali, A., Spyridonos, P.: Community detection in so-
cial media. Data Min. Knowl. Discov. 24(3), 515 -- 554 (2012)
27. Sarr, I., Missaoui, R.: Temporal Analysis on Static and Dynamic Social Networks Topolo-
gies. Springer. To appear
28. Adedoyin-Olowe, M., Gaber, M.M., Stahl, F.T.: Trcm: A methodology for temporal analysis
of evolving concepts in twitter. In: Artificial Intelligence and Soft Computing, pp. 135 -- 145
(2013)
29. Zadeh, L.A.: Fuzzy sets as a basis for a theory of possibility. In: Fuzzy Sets and Systems, pp.
3 -- 28 (1978)
30. Dubois, D., Prade, H., Sandri, S.: On possibility/probability transformations. In: In Proceed-
ings of the Fourth Int. Fuzzy Systems Association World Congress (IFSA'91), Brussels,
Belgium, pp. 50 -- 53 (1991)
31. Zadeh, L.A.: Fuzzy sets. In: Information and Control 8, pp. 338 -- 353 (1965)
32. Gueye, I., Ndong, J., Sarr, I.: An accurate probabilistic model for community evolution anal-
ysis in social network. In: International Conference on Signal-Image Technology Internet-
Based Systems (SITIS), pp. 343 -- 349 (2015)
33. McGloin, J.M., Kirk, D.S.: An overview of social network analysis. Journal of Criminal
Justice Education 21(2), 169 -- 181 (2010)
34. Wasserman, S., Faust, K.: Social Network Analysis: Methods and Applications vol. 8. Cam-
bridge university press, ??? (1994)
|
1606.08084 | 1 | 1606 | 2016-06-26T20:41:52 | Cyberbullying Identification Using Participant-Vocabulary Consistency | [
"cs.SI",
"stat.ML"
] | With the rise of social media, people can now form relationships and communities easily regardless of location, race, ethnicity, or gender. However, the power of social media simultaneously enables harmful online behavior such as harassment and bullying. Cyberbullying is a serious social problem, making it an important topic in social network analysis. Machine learning methods can potentially help provide better understanding of this phenomenon, but they must address several key challenges: the rapidly changing vocabulary involved in cyber- bullying, the role of social network structure, and the scale of the data. In this study, we propose a model that simultaneously discovers instigators and victims of bullying as well as new bullying vocabulary by starting with a corpus of social interactions and a seed dictionary of bullying indicators. We formulate an objective function based on participant-vocabulary consistency. We evaluate this approach on Twitter and Ask.fm data sets and show that the proposed method can detect new bullying vocabulary as well as victims and bullies. | cs.SI | cs | Cyberbullying Identification Using Participant-Vocabulary Consistency
6
1
0
2
n
u
J
6
2
]
I
S
.
s
c
[
1
v
4
8
0
8
0
.
6
0
6
1
:
v
i
X
r
a
Elaheh Raisi
Virginia Tech, Blacksburg, VA
Bert Huang
Virginia Tech, Blacksburg, VA
Abstract
With the rise of social media, people can now
form relationships and communities easily re-
gardless of location, race, ethnicity, or gender.
However,
the power of social media simulta-
neously enables harmful online behavior such
as harassment and bullying. Cyberbullying is
a serious social problem, making it an impor-
tant topic in social network analysis. Machine
learning methods can potentially help provide
better understanding of this phenomenon, but
they must address several key challenges:
the
rapidly changing vocabulary involved in cyber-
bullying, the role of social network structure, and
the scale of the data. In this study, we propose a
model that simultaneously discovers instigators
and victims of bullying as well as new bullying
vocabulary by starting with a corpus of social
interactions and a seed dictionary of bullying
indicators. We formulate an objective function
based on participant-vocabulary consistency. We
evaluate this approach on Twitter and Ask.fm
data sets and show that the proposed method can
detect new bullying vocabulary as well as victims
and bullies.
1. Introduction
Social media has significantly changed the nature of so-
ciety. Our ability to connect with others has been mas-
sively enhanced, removing boundaries created by location,
gender, age, and race. However,
the benefits of this
hyper-connectivity also come with the enhancement of
detrimental aspects of social behavior. Cyberbullying is
an example of one such behavior that is heavily affecting
the younger generations (Boyd, 2014). The Cyberbullying
2016 ICML Workshop on #Data4Good: Machine Learning in
Social Good Applications, New York, NY, USA. Copyright by the
author(s).
[email protected]
[email protected]
Figure 1. Survey Statistics on Cyberbullying Experiences. Data
collected and visualized by the Cyberbullying Research Center
(http://cyberbullying.org/).
Research Center defines cyberbullying as "willful and
repeated harm inflicted through the use of computers, cell
phones, and other electronic devices." Like traditional
bullying, cyberbullying occurs in various forms. Examples
include name calling, rumor spreading, threats, and sharing
of private information or photographs.1,2 Even seemingly
innocuous actions such as supporting offensive comments
by "liking" them can be considered bullying (Wang et al.,
2009). As stated by the National Crime Prevention Coun-
cil, around 50% of American young people are victimized
by cyberbullying. According to the American Academy of
Child and Adolescent Psychiatry, victims of cyberbullying
have strong tendencies toward mental and psychiatric dis-
orders (American Academy of Child Adolescent Psychia-
try, 2016). In extreme cases, suicides have been linked to
cyberbullying (Goldman, 2010; Smith-Spark, 2013). The
phenomenon is widespread, as indicated in Fig. 1, which
plots survey responses collected from students. These facts
make it clear that cyberbullying is a serious health threat.
Machine learning can be useful in addressing the cyber-
bullying problem. Recently, various studies considered
supervised, text-based cyberbullying detection, classifying
1http://www.endcyberbullying.org
2http://www.ncpc.org/cyberbullying
46
Cyberbullying Identification Using Participant-Vocabulary Consistency
social media posts as 'bullying 'or 'non-bullying'. Training
data is annotated by experts or crowdsourced workers.
Since bullying often involves offensive language,
text-
based cyberbullying detection studies often use curated
swear words as features, augmenting other standard text
features. Then, supervised machine learning approaches
train classifiers from this annotated data (Yin et al., 2009;
Ptaszynski et al., 2010; Dinakar et al., 2011).
We identify three significant challenges for supervised
cyberbullying detection. First, annotation is not an easy
task.
It requires expertise about culture, examination of
the social structure of the individuals involved in each
interaction. Because of the difficulty of labeling these of-
ten subtle distinctions between bullying and non-bullying,
there is likely to be disagreement among labelers, making
costs add up quickly for a large-scale problem. Second,
reasoning about which individuals are involved in bullying
should do joint, or collective, classification. E.g., if we
believe a message from A to B is a bullying interaction,
we should also expect a message from A to C to have
an increased likelihood of also being bullying. Third,
language is rapidly changing, especially among young
populations, making the use of static text indicators prone
to becoming outdated. Some curse words have completely
faded away or are not as taboo as they once were, while new
slang is frequently introduced into the culture. These three
challenges suggest that we need a dynamic methodology to
collectively detect emerging and evolving slurs with only
weak supervision.
In this paper, we introduce an automated, data-driven
method for cyberbullying identification. The eventual goal
of such work is to detect such harmful behaviors in social
media and intervene, either by filtering or by providing
advice to those involved. Our proposed learnable model
takes advantage of the fact that the data and concepts
involve relationships. We train this relational model in a
weakly supervised manner, where human experts provide
a small seed set of phrases that are highly indicative
of bullying.
Then the algorithm finds other bullying
terms by extrapolating from these expert annotations. In
other words, our algorithm detects cyberbullying from key-
phrase indicators. We refer to our proposed method as the
participant-vocabulary consistency (PVC) model; It seeks
a consistent parameter setting for all users and key phrases
in the data that characterizes the tendency of each user to
harass or to be harassed and the tendency of a key phrase
to be indicative of harassment. The learning algorithm
optimizes the parameters to minimize their disagreement
with the training data which are highly indicative bullying
phrases in messages between specific users.
A study by ditchthelabel.org (2013) found that Facebook,
YouTube, Twitter, and Ask.fm are the platforms that have
the most frequent occurrences of cyberbullying. To evalu-
ate the participant-vocabulary consistency method, we ran
our experiments on Twitter and Ask.fm data. From a list
of highly indicative of bullying key phrases, we subsample
small seed sets to train the algorithm. We then examine the
participant-vocabulary consistency method to see how well
it recovers the remaining, held-out set of indicative phrases.
Additionally, we extract the detected most bullying phrases
and qualitatively verify that they are in fact examples of
bullying.
2. Related Work
There are two main branches of research related to our
topic. One of them is online harassment and cyberbullying
detection;
the other one is associated with automated
vocabulary discovery. Various studies have used fully
supervised learning to classify bully posts from non-bully
posts. Many of them focus on the textual features of
post to identify cyberbullying incidents (Dinakar et al.,
2011; Ptaszynski et al., 2010; Hosseinmardi et al., 2015;
Chen et al., 2012; Margono et al., 2014). Some of them
use other features than only textual features, for example
content, sentiment, and contextual features (Yin et al.,
2009),
the number, density and the value of offensive
words (Reynolds et al., 2011), or the number of friends,
network structure, and relationship centrality (Huang &
Singh, 2014). Nahar et al. (2013) used semantic and
weighted features; they also identify predators and victims
using a ranking algorithm. Many studies have been applied
machine learning techniques to better understand social-
psychological issues such as bullying. They used data sets
such as Twitter, Instagram and Ask.fm to study negative
user behavior (Bellmore et al., 2015; Hosseinmardi et al.,
2014a;b).
Various works use query expansion to extend search
queries to dynamically include additional
For
example, Massoudi et al. (2011) use temporal information
as well as co-occurrence to score the related terms to
expand the query. Mahendiran et al. propose a method
based on probabilistic soft logic to grow a vocabulary using
multiple indicators (e.g., social network, demographics,
and time).
terms.
3. Proposed Method
To model the cyberbullying problem, for each user ui,
we assign a bully score bi and a victim score vi. The
bully score measures how much a user tends to bully
others; likewise, victim score indicates how much a user
tends to be bullied by other users. For each feature wk,
we associate a feature-indicator score that represents how
much the feature is an indicator of a bullying interaction.
47
Cyberbullying Identification Using Participant-Vocabulary Consistency
Each feature represents the existence of some descriptor
in the message, such as n-grams in text data. The sum
of senders bullying score and receivers victim score (bi +
vi) specifies the message's social bullying score, which
our model aims to make consistent with the vocabulary-
based feature score. We formulate a regularized objective
function that penalizes inconsistency between the social
bullying score and each of the feature scores.
(cid:0)b2 + v2 + w2(cid:1) +
(cid:0)bs(m) + vr(m) − wk
(cid:1)2
(1)
J(b, v, w; λ) =
(cid:88)
λ
2
(cid:88)
m∈M
k:wk∈f (m)
1
2
Learning is then an optimization problem over parameter
vectors b, v, and w.
The consistency penalties are
determined by the structure of the social data. We include
information from an expert-provided initial seed of highly
indicative bully words. We require these seed features to
have a high score, adding the constraint:
J(b, v, w; λ) s.t. wk = 1.0, ∀k : xk ∈ S.
(2)
min
b,v,w
We refer to this model as the participant-vocabulary consis-
tency model because we optimize the consistency of scores
computed based on the participants of each social interac-
tion as well as the vocabulary used in each interaction. The
objective function Eq. (1) is not jointly convex; However,
if we optimize each parameter vector in isolation, we then
solve convex optimizations with closed form solutions.
The optimal value for each parameter vector given the
others can be obtained by solving for their zero-gradient
conditions. The update for bully score vector b is:
(cid:88)
(cid:88)
(cid:0)wk − f (m)vr(m)
(cid:88)
f (m)
k∈f (m)
(cid:1)
m∈Ms(m)=i
arg min
J =
bi
m∈Ms(m)=i
λ +
where {m ∈ Ms(m) = i} is the set of messages that are
sent by user i, and f (m) is the number of n-grams in the
message m. The closed-form solution for optimizing with
respect to victim score vector v is
(cid:88)
(cid:88)
(cid:88)
(wk − f (m)bi)
arg min
J =
vj
m∈Mr(m)=j
k∈f (m)
λ +
m∈Mr(m)=j
f (m)
where {m ∈ Mr(m) = j} is the set of messages sent to
user j. The word score vector w can be updated with
(cid:88)
(cid:0)br(m) + vs(m)
(cid:1)
arg min
J =
wk
m∈Mk∈f (m)
λ + {m ∈ Mk ∈ f (m)}
.
48
Figure 2. ROC curve of recovered target words for Ask.fm (top)
and Twitter (bottom).
The set {m ∈ Mk ∈ f (m)} indicates the set of messages
that contain the kth feature or n-gram. The strategy
of fixing some parameters and solving the optimization
problem for the rest of parameters known as alternating
least-squares. Our algorithm iteratively updates each of
parameter vectors b, v, and w until convergence. The
output of the algorithm is the bully and victim score of all
the users and the bully score of all the words.
4. Experiments
We ran our experiments on Twitter and Ask.fm data.
To collect our Twitter data set, we first collected tweets
using words from our curse word dictionary; then we use
snowball sampling from these profane tweets to collect
additional messages. Since we want to consider messages
between users, we are interested in @-replies, in which one
user directs a message to another or answers a post from
another user. From our full data set, we remove the tweets
which are not part of any such conversation, all retweets
and duplicate tweets. After the data collection and post-
processing, our Twitter data set contains 180,355 users and
296,308 tweets.
Ask.fm is a social question-answering service, where users
post questions on the profiles of other users. We use part
of the Ask.fm data set collected by Hosseinmardi et al.
(2014b). They used snowball sampling, collecting user
profiles and a complete list of answered questions. We
remove all the question-answer pairs where the identity of
the questioner is hidden. Our Ask.fm data set consists of
Cyberbullying Identification Using Participant-Vocabulary Consistency
Table 1. Identified bullying bigrams detected by participant-
vocabulary consistency from Twitter and Ask.fm data sets.
Data Set
Twitter
Ask.fm
c*nt
sp*nkbucket
Selected High-Scoring Words
sh*tstain arisew,
lying, w*gger,
commi
f*ggot,
lowlife,
f*cking nutter, blackowned whitetrash,
monster
dumb*ss,
*ssface mcb*ober, ignorant *sshat
total d*ck, blaky,
ilysm n*gger, fat sl*t,
pathetic waste, loose p*ssy, c*cky b*stard,
wifi b*tch,
stupid hoee,
sleep p*ssy, worthless sh*t, ilysm n*gger
que*n c*nt,
hatchling,
f*ggot
41,833 users and 286,767 question-answer pairs.
We compare our method with two baselines. The first
is co-occurrence. All words or bigrams that occur in
the same tweet as any seed word are given the score 1,
and all other words have score 0. The second baseline is
dynamic query expansion (DQE) (Ramakrishnan et al.,
2014). DQE extracts messages containing seed words, then
computes the document frequency score of each word. It
then iterates selection of the k highest-scoring keywords
and re-extraction of relevant messages until it reaches a
stable state.
We evaluate performance by considering held out words
from our full curse word dictionary as the relevant target
words. We measure the true positive rate and the false posi-
tive rate, computing the receiver order characteristic (ROC)
curve for each compared method. Fig. 2 contains the ROC
curves for both Twitter and Ask.fm. Co-occurrence only
indicates whether words co-occur or not, so it forms a
single point in the ROC space. Our PVC model and DQE
compute real-valued scores for words and generate curves.
DQE produces very high precision, but does not recover
many of the target words. However, co-occurrence detects
a high proportion of the target words, but at the cost of
also recovering a large fraction of non-target words. PVC
is able to recover a much higher proportion of the target
words comparing DQE. PVC enables a good compromise
between recall and precision.
We also compute the average score of target words, non-
target words, and all of the words.
If the algorithm
succeeds, the average target-word score should be higher
than the overall average. For both Twitter and Ask.fm,
our proposed PVC model can capture target words much
better than baselines. We measured how many standard
deviations the average target-word score is above the over-
all average. For Twitter, PVC provides a lift of around
1.5 standard deviations over the overall average, while
DQE only produces a lift of 0.242. We also observe the
same behavior for Ask.fm: PVC learns scores that have a
defined lift between the overall average word score and the
average target word score (0.825). DQE produces a small
lift (0.0099). Co-occurrence has no apparent lift.
By manually examining the 1,000 highest scoring words,
we find many seemingly valid bullying words. These de-
tected curse words include sexual, sexist, racist, and LGBT
(lesbian, gay, bisexual, and transgender) slurs. Table 1 lists
some of these high-scoring words from our experiments.
The PVC algorithm also computes bully and victim scores
for users. By studying the profiles of highly scored
victims in Ask.fm, we noticed that some of these users do
appear to be bullied. This happens in Twitter as well, in
which some detected high scoring users are often using
offensive language in their tweets. Fig. 3 shows some
bullying comments to an Ask.fm user and her responses,
all of which contain offensive language and seem highly
inflammatory.
Figure 3. Example of an Ask.fm conversation containing possible
bullying and heavy usage of offensive language.
5. Conclusion
In this paper, we proposed the participant-vocabulary con-
sistency method to simultaneously discover victims, insti-
gators, and vocabulary of words indicates bullying. Start-
ing with seed dictionary of high-precision bullying indica-
tors, we optimize an objective function that seeks consis-
tency between the scores of the participants in each interac-
tion and the scores of the language use. For evaluation, we
perform our experiments on data from Twitter and Ask.fm,
services known to contain high frequencies of bullying.
Our experiments indicate that our method can successfully
detect new bullying vocabulary. We are currently working
on creating a more formal probabilistic model for bullying
to robustly incorporate noise and uncertainty.
49
Cyberbullying Identification Using Participant-Vocabulary Consistency
References
American Academy of Child Adolescent Psychiatry.
the american academy of
URL http:
Facts for families guide.
child adolescent psychiatry.
//www.aacap.org/AACAP/Families_and_
Youth/Facts_for_Families/FFF-Guide/
FFF-Guide-Home.aspx.
2016.
Bellmore, Amy, Calvin, Angela J., Xu, Jun-Ming, and Zhu,
Xiaojin. The five w's of bullying on twitter: Who, what,
why, where, and when. Computers in Human Behavior,
44:305–314, 2015.
Boyd, Danah.
2014.
It's Complicated. Yale University Press,
Chen, Ying, Zhou, Yilu, Zhu, Sencun, and Xu, Heng.
Detecting offensive language in social media to protect
International Conference
adolescent online safety.
on Privacy, Security, Risk and Trust
(PASSAT),
and International Conference on Social Computing
(SocialCom), pp. 71–80, 2012.
Dinakar, Karthik, Reichart, Roi, and Lieberman, Henry.
Modeling the detection of
textual cyberbullying.
International Conference on Weblog and Social Media
- Social Mobile Web Workshop, 2011.
ditchthelabel.org. Ditch the label-anti-bullying charity:
The annual cyberbullying survey. 2013. URL http:
//www.ditchthelabel.org/.
Goldman, Russell. Teens indicted after allegedly taunting
2010. URL http://
girl who hanged herself.
abcnews.go.com/Technology/TheLaw/.
Hosseinmardi, Homa, Han, Richard, Lv, Qin, Mishra,
Shivakant, and Ghasemianlangroodi, Amir. Towards
in a semi-
understanding cyberbullying behavior
Advances in Social
anonymous social network.
Networks Analysis and Mining (ASONAM), 2014
IEEE/ACM International Conference on, pp. 244–252,
August 2014a.
Hosseinmardi, Homa, Li, Shaosong, Yang, Zhili, Lv, Qin,
Rafiq, Rahat Ibn, Han, Richard, and Mishra, Shivakant.
A comparison of common users across instagram and
ask.fm to better understand cyberbullying. Big Data
and Cloud Computing (BdCloud), 2014 IEEE Fourth
International Conference on, pp. 355–362, 2014b.
Hosseinmardi, Homa, Mattson, Sabrina Arredondo, Rafiq,
Ibn, Han, Richard, Lv, Qin, and Mishra,
Rahat
Shivakant. Detection of cyberbullying incidents on
Association for the
the instagram social network.
Advancement of Artificial Intelligence, 2015.
50
Huang, Qianjia and Singh, Vivek Kumar. Cyber bullying
SAM
detection using social and textual analysis.
'14 Proceedings of the 3rd International Workshop on
Socially-Aware Multimedia, pp. 3–6, 2014.
Mahendiran, Aravindan, Wang, Wei, Arredondo, Jaime,
Lira, Sanchez, Huang, Bert, Getoor, Lise, Mares,
David, and Ramakrishnan, Naren. Discovering evolving
political vocabulary in social media.
Margono, Hendro, Yi, Xun, and Raikundalia, Gitesh K.
Mining indonesian cyber bullying patterns in social
the Thirty-Seventh Aus-
networks.
tralasian Computer Science Conference (ACSC 2014),
147, January 2014.
Proceedings of
Massoudi, Kamran, Tsagkias, Manos, de Rijke, Maarten,
and Weerkamp, Wouter. Incorporating query expansion
and quality indicators in searching microblog posts.
AIR11. Springer, 15(5):362–367, November 2011.
Nahar, Vinita, Li, Xue, and Pang, Chaoyi. An effective
approach for cyberbullying detection. Communications
in Information Science and Management Engineering.,
3(5):238–247, May 2013.
Ptaszynski, Michal, Dybala, Pawel, Matsuba, Tatsuaki,
Masui, Fumito, Rzepka, Rafal, and Araki, Kenji.
Machine learning and affect analysis against cyber-
bullying. In the 36th AISB, pp. 7–16), 2010.
Ramakrishnan, Naren, Butler, Patrick, Muthiah, Sathap-
pan, Self, Nathan, Khandpur, Rupinder, and et al.
'Beating the news' with EMBERS: Forecasting civil
unrest using open source indicators. KDD, 2014.
Reynolds, Kelly, Kontostathis, April, and Edwards, Lynne.
Using machine learning to detect cyberbullying. 10th
International Conference on Machine Learning and
Applications and Workshops (ICMLA), 2:241–4, 2011.
Smith-Spark, Laura. Hanna Smith suicide fuels calls
2013. URL
for action on ask.fm cyberbullying.
http://www.cnn.com/2013/08/07/world/
europe/uk-social-media-bullying/.
Wang, Jing, Iannotti, Ronald J., and Nansel, Tonja R.
School bullying among us adolescents: Physical, verbal,
relational and cyber. Journal of Adolescent Health, 45:
368–375, 2009.
Yin, Dawei, Xue, Zhenzhen, Hong, Liangjie, Davison,
Brian D., Kontostathis, April, and Edwards, Lynne.
Detection of harassment on web 2.0. Proceedings of the
Content Analysis in the WEB 2.0, 2009.
|
1901.05149 | 2 | 1901 | 2019-12-20T20:56:38 | Beyond Uniform Reverse Sampling: A Hybrid Sampling Technique for Misinformation Prevention | [
"cs.SI",
"cs.DS"
] | Online misinformation has been considered as one of the top global risks as it may cause serious consequences such as economic damages and public panic. The misinformation prevention problem aims at generating a positive cascade with appropriate seed nodes in order to compete against the misinformation. In this paper, we study the misinformation prevention problem under the prominent independent cascade model. Due to the #P-hardness in computing influence, the core problem is to design effective sampling methods to estimate the function value. The main contribution of this paper is a novel sampling method. Different from the classic reverse sampling technique which treats all nodes equally and samples the node uniformly, the proposed method proceeds with a hybrid sampling process which is able to attach high weights to the users who are prone to be affected by the misinformation. Consequently, the new sampling method is more powerful in generating effective samples used for computing seed nodes for the positive cascade. Based on the new hybrid sample technique, we design an algorithm offering a $(1-1/e-\epsilon)$-approximation. We experimentally evaluate the proposed method on extensive datasets and show that it significantly outperforms the state-of-the-art solutions. | cs.SI | cs | Beyond Uniform Reverse Sampling: A Hybrid
Sampling Technique for Misinformation Prevention
Guangmo Tong∗ and Ding-Zhu Du†
∗Department of Computer and Information Sciences, University of Delaware, USA
†Department of Computer Science, University of Texas at Dallas, USA
Email: [email protected], [email protected]
9
1
0
2
c
e
D
0
2
]
I
S
.
s
c
[
2
v
9
4
1
5
0
.
1
0
9
1
:
v
i
X
r
a
Abstract -- Online misinformation has been considered as one
of the top global risks as it may cause serious consequences
such as economic damages and public panic. The misinformation
prevention problem aims at generating a positive cascade with
appropriate seed nodes in order to compete against the misinfor-
mation. In this paper, we study the misinformation prevention
problem under the prominent independent cascade model. Due
to the #P-hardness in computing influence, the core problem is to
design effective sampling methods to estimate the function value.
The main contribution of this paper is a novel sampling method.
Different from the classic reverse sampling technique which treats
all nodes equally and samples the node uniformly, the proposed
method proceeds with a hybrid sampling process which is able to
attach high weights to the users who are prone to be affected by
the misinformation. Consequently, the new sampling method is
more powerful in generating effective samples used for computing
seed nodes for the positive cascade. Based on the new hybrid
sample technique, we design an algorithm offering a (1−1/e−)-
approximation. We experimentally evaluate the proposed method
on extensive datasets and show that it outperforms the state-of-
the-art solutions.
Index Terms -- misinformation prevention, sampling, social net-
work
I. INTRODUCTION
According to the World Economic Forum1, misinformation
has been one of the top global risks due to its potential
for spreading fake news and malicious information. Once
misinformation is detected, one feasible method is to introduce
a positive cascade which is expected to reach the users before
they are affected by the misinformation. The misinformation
prevention (MP) problem aims at selecting effective seed
nodes for the positive cascade such that the spread of mis-
information can be maximally limited. The MP problem is
naturally formulated as a combinatorial optimization problem
and it has drawn great attention [2] -- [8]. Following this branch,
we in this paper study the MP problem and our goal is to
design solutions which are theoretically supported and highly
effective in practice.
Influence Maximization.
In the seminal work [9] of
Kempe, Kleinberg and Tardos, the well-known influence max-
imization (IM) problem was studied under two prominent
operational models, independent cascade (IC) model and linear
threshold (LT) model. The goal of the IM problem is to select
This is the online version of [1] to fix a mistake in Sec. IV-E. In particular,
a new analysis of estimating the lower bound is proposed.
1http://reports.weforum.org/
seed nodes which can maximize the influence resulted by a
single information cascade. There are two core techniques used
for solving the IM problem. One is the study on its combi-
natorial properties and the other one is sampling technique
utilized to speed up the algorithms. From the former, it has
been shown that the IM problem is monotone and submodular,
and therefore the simple greedy algorithm provides a (1−1/e)-
approximation. Unfortunately, computing the influence of the
cascade is a #P-hard problem [10]. As a result, the objective
function cannot be efficiently computed, and we wish to
estimate the function value by sampling techniques. As the
most successful framework, C. Borg et al. [11] proposed the
uniform reserve sampling technique which finds an unbiased
estimator of the objective function via two steps: (1) uniformly
select a node v from the ground set; (2) simulate the diffusion
process from v in a reverse direction until no node can be
further reached, and collect the set Sv of the traversed nodes.
After each sampling, we obtain a sample Sv which is a subset
of the users. The idea is that those in Sv are the nodes that
can influence v. Therefore, if Sv ∩ S (cid:54)= ∅, it implies that v can
be influenced by S. Thus, given a collection of the samples,
intuitively, the seed set S which covers2 the maximum number
of samples should be able to maximize the objective function
[11]. Such a framework was later improved by [12] -- [14] and it
has been widely applied to other problems regarding influence
diffusion in social networks.
Misinformation Prevention. Following the Borg's work
[11], authors in [4] designed a reverse sampling method for
the MP problem. Given the seed set of the misinformation, the
method proposed in [4] proceeds with two analogous steps: (1)
uniformly select a node v from the ground set; (2) simulate the
diffusion process from v in a reverse direction in the manner
of BFS until a seed of misinformation is reached, and collect
the set Sv of the traversed nodes. It is shown in [4] that those
in Sv are in fact the nodes that can protect v from being
affected by the misinformation. Therefore, the nodes that can
maximally cover the samples are able to best limit the spread
of misinformation.
Contribution. Our work is driven by considering the differ-
ence between the IM problem and the MP problem. In the IM
problem, all the nodes are treated equally and we cannot do
better, without any prior information, than uniformly selecting
2We say S covers a sample Sv if S ∩ Sv (cid:54)= ∅
Fig. 1: An illustrative example
a node in the first step of the sampling process. However, in
the MP problem, the ultimate goal is to protect the nodes from
being influenced by the misinformation, so intuitively it would
be better if we could pay more attention to the nodes that
are more likely to be misinformation-influenced. Consider an
illustrative example shown in Fig. 1 where there is one seed
node of the misinformation. If we uniformly select a node
and do the reverse sampling, we may select the nodes that
are far distant from the misinformation and they actually do
not require much effort to be protected (e.g., the safe nodes
in the graph). Instead, because the nodes around the seed of
misinformation are the most likely to be influenced by the
misinformation, we should focus more on the nodes in the
M-area shown in the graph. Furthermore, because a node can
be protected only if the positive cascade can reach that node
before the arrival of misinformation, to protect the nodes in
the M-area, the most effective positive seed nodes should lie
in the P-area in the graph. Motivated by this idea, we present
a hybrid sampling method consisting of two high-level steps:
1) Forward Sampling: simulate the diffusion process of
the misinformation and collect the nodes that are influ-
enced by the misinformation.
2) Reverse Sampling: for each node v collected in the first
step, do the reverse sampling process to obtain the nodes
that can protect that v.
it will be affected by the misinformation. Thus,
In this sampling method, the frequency that a node can be
collected in the first step is proportional to the probability
that
the
samples produced by the second step are more likely to be the
protectors of the nodes which are prone to be misinformation-
influenced. Note that
the pattern of the sample areas is
determined by the nodes collected in the first step. As shown
in Fig. 2, under the uniform reverse sampling the samples
are uniformly distributed to the whole graph, while under
the hybrid sampling the samples tend to be centered around
the seed node of the misinformation. As shown later, the
sample obtained by our hybrid sampling method can be used to
directly estimate the prevention effect of the positive cascade.
Based on the hybrid sampling method, we propose a new
(a) Sample pattern under the
uniform reverse sampling
(b) Sample pattern under the hybrid
reverse sampling
Fig. 2: Sample patterns
randomized approximation algorithm for the MP problem.
In order to evaluate the proposed algorithm, we design ex-
periments which compare the algorithms by evaluating their
performance under the same time constraint. The effectiveness
of our solution is supported by encouraging experimental
results.
Organization. In Sec. II, we introduce the considered
diffusion model and formally formulate the MP problem. We
present the new hybrid sampling technique in Sec. III. Our
algorithm is later given in Sec. IV in which we also provide
theoretical analysis. In Sec. V, we experimentally evaluate the
proposed solution. Sec. VI concludes this paper and proposes
some future problems. Due to space limitation, the related
works will be introduced throughout this paper and we do not
spare a separate section for them.
II. SYSTEM MODEL
We consider the classic independent cascade (IC) model
for influence diffusion. A social network is given by a directed
graph G = (V, E) where V is the set of the user and E denotes
the relationship between users. Let n and m be the number
of nodes and edges, respectively. Associated with each edge
(u, v) ∈ E, there is a real number Pr[(u, v)] ∈ (0, 1] denoting
the propagation probability from u to v. An information
cascade C starts to spread from its seed users S ⊆ V . We
say a user is C-active if they are influenced by cascade C. All
users are initialized as ∅-active, and the seed nodes of cascade
C are the first that are C-active. When a user v becomes C-
active for some cascade C, they have one chance to attempt to
activate their ∅-active neighbor v where the success probability
is Pr[(u, v)]. If the process succeeds, v becomes C-active as
well. We consider the scenario where there are two cascades,
the misinformation Cr and the positive cascade Cp. We assume
that the misinformation has a higher priority. That is, when
a user v is (a) selected as a seed node by both cascades
or (b) simultaneously activated by two neighbors which are
respectively Cr-active and Cp-active, v will be Cr-active.3
3We note that this is not a critical setting and the analysis of this paper can
be applied to the case when the positive cascade has a higher priority.
Seed of MisinformationM - areaP - areaSafe NodeddSample area from vvSample area from vv We assume that the seed set of the misinformation, denoted
as Sr, is known to us, and we aim at introducing the positive
cascade to spread competing against the misinformation cas-
cade. In particular, our goal is to select an appropriate seed set
for the positive cascade such that the misinformation can be
prevented maximally. We say a user is Cr-active if they are
not Cr-active. We use f (S) to denote the expected number
of the Cr-active nodes under a positive seed set S ⊆ V .
The following two problems have been studied in the existing
work.
Problem 1 (Maximum Non-misinformation (MN) Prob-
lem). Given an instance of the IC model, the seed set Sr ⊆ V
of the misinformation, and a budget k ∈ Z+, find a subset
S ⊆ V with S ≤ k such that f (S) is maximized.
Problem 2 (Maximum Prevention (MP) Problem). Given
an instance of the IC model, the seed set Sr ⊆ V of the
misinformation, and a budget k ∈ Z+, find a subset S ⊆ V
with S ≤ k such that f∗(S) := f (S) − f (∅) is maximized.
The MN problem maximizes the expected number of the
nodes which are not activated by the misinformation, while the
MP problem maximizes f (S) − f (∅) where f (∅) is expected
number of the Cr-active nodes when there is no positive
cascade. Because f (S) − f (∅) denotes the exact effect of the
positive cascade in preventing the misinformation, we call the
second problem as the Maximum Prevention problem. These
two problems are equivalent to each other with respect to
the optimal solution, as their objective functions are differed
by a constant f (∅). However, with respect to optimization,
the MP problem is more challenging especially when f (∅) is
relatively large. One can see that, for each α ∈ (0, 1), each α-
approximation to the MP problem is also an α-approximation
to the MN problem. In this paper, we study the MP problem
and use Sopt to denote its optimal solution.
A. Derandomization of the diffusion process
We say an edge (u, v) is live if u can successfully activate
v. According to the IC model, the randomness of the diffusion
process comes from the states of the edges, where each edge
can be live or not live with probability Pr[e] and 1 − Pr[e],
respectively. The diffusion process is equivalent to that first
sample the states of the edges and then run the diffusion
process with the sampled states [9]. The concept of realization
is used to describe the basic event space.
Definition 1 (Realization). Given an instance of IC model,
a realization g is a two-tuple g = (g(Et), g(Ef )) where
g(Et), g(Ef ) ⊆ E and g(Et) ∩ g(Ef ) = ∅. g(Et) (resp.,
g(Ef )) denotes the set of the edges that are live (resp., not
live). In addition, we say g is a full-realization if g(Et) ∪
g(Ef ) = E. Otherwise, it is a partial-realization. For each
realization g, either full or partial, we use Pr[g] to denote
the probability that g happens. As the states of edges are
independent to each to other, we can easily check that
(cid:89)
Pr[e] · (cid:89)
e∈g(Et)
e∈g(Ef )
Pr[g] =
(1 − Pr[e]).
We use G to denote the set of the full-realizations. The
full-realizations provide a way to enumerate all the possible
outcomes, and the partial-realizations are useful to describe an
intermediate state during the diffusion process.
It is useful to interpret a full-realization g as a subgraph of G
with the edge set g(Et). For each g ∈ G, the diffusion process
in g is well-defined because the state of each edge is known
to us, and furthermore, the process is deterministic. Due to
v∈V fv(S)
where fv(S) is the probability that node v is Cr-active under
g∈G Pr[g] · fv(S, g),
the linearity of the expectation, we have f (S) =(cid:80)
S. fv(S) can be further expressed as (cid:80)
where
(cid:40)
fv(S, g) :=
1 v is Cr-active in g under S
0 otherwise
,
and finally, we have
f (S) =
Pr[g] · fv(S, g).
(cid:88)
(cid:88)
v∈V
g∈G
A natural problem is that under which condition that v can
be Cr-active in g. For each V1 ⊆ V , v ∈ V and g ∈ G,
we define disg(V1, v) := minu∈V1 disg(u, v) where disg(u, v)
is the distance4 from u to v in g. According to [2], [4], [15] and
[16], the following lemma gives the necessary and sufficient
condition for a node to be Cr-active in a full-realization.
Lemma 1. Suppose the seed set of the positive cascade is
S ⊆ V . For each v ∈ V and g ∈ G, v is Cr-active in g if and
only if disg(S, v) < disg(Sr, v) or disg(Sr, v) = +∞.
For each v ∈ V and S ⊆ V , we use
G(v, S) :={g ∈ G fv(S, v) = 1}
to denote the set of the full-realization(s) g where v is Cr-
active under S in g. Therefore,
(cid:88)
(cid:88)
Pr[g] −(cid:88)
g∈G(v,S)
v∈V
Pr[g]
(cid:88)
f (S) =
(cid:88)
(cid:88)
and
f∗(S) =
Pr[g].
v∈V
g∈G(v,S)
v∈V
g∈G(v,∅)
According to Lemma 1, it is clear that G(v,∅) ⊆ G(v, S)
and G(v,∅) consists of the realizations where there is no path
from Sr to v. Consequently, the difference between G(v,∅)
and G(v, S) is the set of the realizations g where disg(S, v) <
disg(Sr, v) and meanwhile there is a path from Sr to v. We use
G∗(v, S) to denote the set of such realizations, i.e., G∗(v, S) :=
{g ∈ G disg(S, v) < disg(Sr, v), disg(Sr, v) (cid:54)= +∞}.
Now our objective function f∗(S) can be expressed as
f∗(S) =
Pr[g],
(1)
(cid:88)
(cid:88)
v∈V
g∈G∗(v,S)
which is a useful expression for the analysis later.
4The distant from u to v is measured in terms of the length of the shortest
path from u to v. We denote it as disg(u, v) = +∞ when there is no path
from u to v.
Algorithm 1 Hybrid Sampling
1: Input: G = (V, E) and Sr;
2: Output: a collection P of subsets of V .
3: Simulate the diffusion of the misinformation from Sr. Let
Vr be the Cr-active nodes and g∗ be the realization when
the simulation terminates;
4: P ← ∅;
5: for each v ∈ Vr do
(P, g∗) ← Alg. 2 (G, g∗, Sr, v);
6:
P = P ∪{P};
7:
8: return P;
III. HYBRID SAMPLING TECHNIQUE
It has been shown in [2] that the MP problem is mono-
tone nondecreasing and submodular, and therefore the greedy
algorithm provides a good approximation [17]. However, the
greedy algorithm demands an efficient oracle of the objective
function f∗(S) which is #P-hard to compute [10]. Therefore,
it is hard to directly maximize f∗(S) and, alternatively, we can
first construct a good estimator of f∗(S) by sampling methods,
and then maximizing the obtained estimator. Towards this end,
we design a novel sampling process, shown in Alg. 1, which
includes the following two parts:
Forward Sampling (line 3). First, we ignore the positive
cascade and simulate the diffusion process of the misinforma-
tion from Sr. When the simulation terminates, we output the
immediate realization g∗ as well as the set Vr of the nodes
that are Cr-active. The g∗ produced in this step is usually a
partial realization.
Reverse Sampling (line 4-7). Let Vr and g∗ be the result
of the first step. We consider the nodes in Vr one by one
in an arbitrary order and run Alg. 2 for each v ∈ Vr. In
Alg. 2, given the node v and a realization g, we simulate
the diffusion from v in the reverse direction in the manner of
BFS until a misinformation seed is reached. When an edge e
is encountered, if the state of e is known in g (i.e., e ∈ g(Et)∪
g(Ef )), we keep its state; otherwise, we sample the state of e
and update g. During this process, we collect the nodes that are
traversed before the misinformation seed is reached. We start
with the realization g∗ obtained in the first step and any node
in Vr, and repeat this process for the rest nodes in Vr with the
updated realization. In each run of Alg. 2, a misinformation
node must be reached due to the construction of g∗ and Vr.
It is important to note that for each edge we sample its state
for at most one time during the whole process in Alg. 1.
For each V1, V2 ⊆ V , we use
(cid:40)
y(V1, V2) :=
1 V1 ∩ V2 (cid:54)= ∅
0 otherwise
S ⊆ V , define x(P, S) as x(P, S) :=(cid:80)
to denote that if the intersection of V1 and V2 is empty. Each
run of Alg. 1 gives a family P of subsets of V . For each
P∈P y(P, S). The
following lemma shows that x(P, S) is an unbiased estimator
of f∗(S).
else
if P ∗ ∩ Sr (cid:54)= ∅ then
Algorithm 2 Reverse Sampling from v
1: Input: G = (V, E), g = (g(Et), g(Ef )), Sr and v;
2: Output: P ⊆ V and an updated realization g;
3: P ∗ ← {v}, P ← ∅;
4: while true do
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
return (P , g);
P ← P ∪ P ∗;
E∗ ← {(u1, u2)u2 ∈ P ∗, u1 ∈ V \ P};
P ∗ ← ∅;
for each edge (u1, u2) ∈ E∗ do
P ∗ ← P ∗ ∪ {u2};
if (u1, u2) /∈ g(Ef ) then
if (u1, u2) ∈ g(Et) then
else
distribution
17:
18:
19:
20:
21:
(cid:46) (0,1) uniform
rand ← U(0, 1);
if rand ≤ p(u1,u2) then
P ∗ ← P ∗ ∪ {u2};
g(Et) ← g(Et) ∪ {(u1, u2)};
g(Ef ) ← g(Ef ) ∪ {(u1, u2)};
else
Lemma 2. For each S ⊆ V , E[x(P, x)] = f∗(S).
Proof. The randomness of the sampling process in Alg. 1
again comes from the random states of the edges. Although
Alg. 2 are possibly called several times during Alg. 1, we only
update the realization g∗ obtained in line 3 and no edge is
sampled more than once. Thus, it is equivalent to that we first
sample the states of all the edges to obtain a full-realization
g and then run the sampling process with g. As a result, the
basic event space is again G, and therefore we can express
g∈G Pr[g] · x(P g, S) where
P g is output of Alg. 1 when the realization is fixed as g.
r to denote the
When the realization is fixed as g, we use V g
Vr obtained in line 3 of Alg. 1, and let P g
v be subset returned
by Alg. 2 with an input v. Now we have
E[x(P, S)] as E[x(P, S)] = (cid:80)
E[x(P, S)] =
=
Pr[g] · x(P g, S)
(cid:88)
Pr[g]
y(P g
v , S)
(2)
Note that v is contained in V g
g when there is no positive cascade. That is, v ∈ V g
only if disg(Cr, v) (cid:54)= +∞. Thus, we have
g∈G
r if and only if v is Cr-active in
r if and
v∈V g
r
(cid:88)
(cid:88)
g∈G
E[x(P, S)] =
=
(cid:88)
(cid:88)
g∈G
Pr[g]
(cid:88)
(cid:88)
v∈V g
r
v∈V
g∈G(v,∅)
y(P g
v , S)
Pr[g] · y(P g
v , S)
Furthermore, in Alg. 2 we search the nodes from v in a reverse
direction in the manner of BFS, so P g
v contains exactly the
node(s) u with a distance to v shorter than disg(Sr, v). There-
v , S) = 1 if and only if disg(S, v) < disg(Sr, v),
fore, y(P g
which means g ∈ G∗(v, S). Since G∗(v, S) ⊆ G(v,∅), we
have
(cid:88)
(cid:88)
v∈V
(cid:88)
(cid:88)
g∈G(v,∅)
v∈V
g∈G∗(v,S)
{By Eq. (1)}
= f∗(S).
E[x(P, S)] =
=
Pr[g] · y(P g
v , S)
Pr[g] · 1
Thus, proved.
For convenience, we call each family P of the subsets of
V produced by Alg. 1 as an R-sample. The complexity of
generating one R-sample is given as follows.
Lemma 3. The expected running time of generating one R-
sample is O(m + m · max1) where max1 := max f∗({v}).
Proof. Let T be the expected running time of Alg. 1, Tg be
the running time when the realization is fixed as g. When the
realization is fixed as a full-realizatoin g, for each v ∈ V ,
we use T IM Eg to denote the running time of line 3, and
g to denote the running time of Alg. 2 with input v.
T IM Ev
Following the proof of Lemma 2, we have the notations V g
r
v .
and P g
According to the analysis in proof of Lemma 2, we have
T =
Pr[g] · Tg =
Pr[g] · (T IM Eg +
(cid:88)
g∈G
(cid:88)
g∈G
(cid:88)
v∈V g
r
Because T IM Eg is asymptotically bounded by the number of
the traversed edges before the simulation terminates, we have
(cid:88)
g∈G
Pr[g] · T IM Eg ≤(cid:88)
g∈G
Pr[g] · m = O(m).
g =
y(P g
(u,v)∈E
T IM Ev
v ,{u}),
Furthermore, according to Alg. 2, an edge (u, v) will be
checked in Alg. 2 if and only if u ∈ P v
v ,{u}) = 1.
Thus, we have
g , i.e., y(P g
and therefore,
(cid:88)
Pr[g] · (cid:88)
(cid:88)
(cid:88)
g∈G Pr[g]·(cid:80)
For each u ∈ V , by Eq. (2),(cid:80)
1 + max(cid:0)f∗({u})(cid:1)(cid:17)
T ≤ m ·(cid:16)
v ,{u}) is
equal to E[x(P,{u})] which is actually f∗({u}). As a result,
we have
v ,{u}).
T = O(m) +
(u,v)∈E
g∈G
v∈V g
r
v∈V g
r
y(P g
y(P g
.
Thus, proved.
IV. THE ALGORITHM
In this section, we present the algorithm designed based on
the hybrid sampling method for solving the MP problem and
provide theoretical analysis.
Algorithm 3 Greedy
1: Input: Rl, V , and k ∈ Z+;
2: S∗ ← ∅;
3: while S∗ < k do
v∗ ← arg maxv∈V x(Rl, S∗ ∪ {v});
4:
S∗ ← S∗ ∪ {v};
5:
return S∗;
A. The Idea.
According to Lemma 2, x(P, S) is an unbiased estima-
tor of f2(S). Therefore, if we generate a collection Rl =
{P 1, ...,P l} of l R-samples by running Alg. 1,
(cid:80)P∈Rl
x(P, S)
l
x(Rl, S) :=
should be an accurate estimate of f∗(S) when l is sufficiently
large. As a result, the subset S ⊆ V that can maximize
x(Rl, S) is intuitively a good solution to maximizing f∗(S).
B. Maximizing x(Rl, S)
Let us first consider that how to maximize x(Rl, S). The
following result shows that x(Rl, S) is monotone nondecreas-
ing and submodular.
Lemma 4. For each Rl, x(Rl, S) is monotone nondecreasing
and submodular.
(cid:80)P∈Rl
x(P, S)
l
=
(cid:80)P∈Rl
(cid:80)
P∈P y(P, S)
l
.
x(Rl, S) =
It is monotone nondecreasing because adding a new node
to S does not decrease y(P, S). Since the submodularity is
preserved under addition, it suffices to show that y(P, S) is
submodular with respect to S. That is, for each P ⊆ V ,
S1 ⊆ S2 ⊆ V and v /∈ S2, we have
y(P, S1 ∪ {v}) − y(P, S1) ≥ y(P, S2 ∪ {v}) − y(P, S2).
Because y is either 0 or 1 and it is monotone, it suffices to
prove that y(P, S1 ∪ {v}) − y(P, S1) is equal 1 whenever
y(P, S2 ∪ {v}) − y(P, S2) is equal to 1. When y(P, S2 ∪
{v}) − y(P, S2) is equal 1, we have y(P, S2 ∪ {v}) = 1 and
y(P, S2) = 1. That is, P ∩ (S2 ∪ {v}) (cid:54)= ∅ and P ∩ S2 = ∅,
which implies that v ∈ P . As a result, P ∩ S1 = ∅ and
P ∩ (S1 ∪ {v}) (cid:54)= ∅. Therefore, y(P, S1 ∪ {v}) − y(P, S1) is
also equal to 1.
Since x(Rl, S) is monotone and submodular, due to the
classic result in [17], the greedy algorithm given in Alg. 3
provides a (1− 1/e)-approximation for maximizing x(Rl, S).
Lemma 5. For each Rl = {P 1, ...,P l} and S ⊆ V with
S ≤ k, the S∗ produced by Alg. 3 satisfies that x(Rl, S∗) ≥
(1 − 1/e) · x(Rl, S).
T IM Ev
g ).
Proof. Recall that
Algorithm 4 Framework
1: Input: G = (V, E), Sr, k, and l;
2: Generate l R-samples Rl = {P 1, ...,P l} by Alg. 1;
3: Obtain an S∗ by Alg. 3 with Rl, V and k;
4: Return S∗;
C. The Framework
By taking Alg. 3 as a subroutine, the framework of our
algorithm is shown in Alg. 4. Given a threshold l ∈ Z+, we
first generate l R-samples and then run the greedy algorithm
to obtain a seed set S∗ of the positive cascade. Now the only
part left is to determine the number of R-samples to be used.
We will discuss this problem in the next subsection.
Proof. It can be proved by simple rearrangements. By Eq. (5),
we have
f∗(S∗) ≥ x(Rl, S∗) − 11(1 + )f∗(Sopt).
Due to Lemma 5, we have x(Rl, S∗) ≥ (1− 1/e)x(Rl, Sopt)
and therefore,
f∗(S∗) ≥ (1 − 1/e)x(Rl, Sopt) − 11(1 + )f∗(Sopt).
Finally, by Eqs. (6) and (7), we have f∗(S∗) ≥ (1− 1/e)(1−
12(1 + ))f∗(Sopt) − 11(1 + )f∗(Sopt) ≥ (1 − 1/e −
)f∗(Sopt).
We take the following setting of l to make Eqs. (5) and (6)
satisfied with probability at least 1 − 2/N.
D. Determining l
The goal is to set l sufficiently large in order to ensure
the theoretical guarantees. Since our algorithm is randomized,
given any parameters ∈ (0, 1) and N > 1, the goal is to
produce an S∗ such that
f∗(S∗) ≥ (1 − 1/e − )f∗(Sopt)
where
and
l1 :=
l∗ :=
n(ln(cid:0)n
k
max(l1, l2)
(1 + )2f∗(Sopt)
(cid:1) + ln N )(2 + 11(1 + ))
(11)2
2 · n · ln N
(12)2
.
l2 :=
Pr
Xi − l · µ ≤ −δ · l · µ
),
(4)
= exp
holds with probability at least 1 − 1/N.5 Throughout our
analysis, we assume that and N are fixed, and we aim to
achieve an arbitrary small error . We use the Chernoff bound
[18] to analyze the accuracy of the estimates.
Lemma 6 (Chernoff Bound). Let X1, ..., Xl ∈ [0, 1] be l
i.i.d random variables where E(Xi) = µ. The Chernoff bound
states that
Xi − l · µ ≥ δ · l · µ
),
(3)
Pr
(cid:104)(cid:88)
(cid:104)(cid:88)
and
(cid:105) ≤ exp(− l · µ · δ2
(cid:105) ≤ exp(− l · µ · δ2
2 + δ
2
for each δ > 0.
In particular, we need x(Rl, S) to be the accurate estimates
for the solution S∗ produced by the greedy algorithm as
well as for the optimal solution Sopt. We use 11 > 0 and
12 > 0 to control the accuracy of x(Rl, S∗) and x(Rl, Sopt),
respectively. The following conditions will be sufficient.
x(Rl, S∗) − f∗(S∗) ≤ 11(1 + )f∗(Sopt)
x(Rl, Sopt) − f∗(Sopt) ≥ −12(1 + )f∗(Sopt)
(1 − 12(1 + ))(1 − 1/e) − 11(1 + ) = 1 − 1/e −
11, 12 ∈ (0, 1)
(5)
(6)
(7)
(8)
Lemma 7. Whenever Eqs. (5)-(8) hold, we have
f∗(S∗) ≥ (1 − 1/e − )f∗(Sopt).
5As discussed in [2], 1 − 1/e is the best possible approximation ratio.
(cid:105)
Lemma 8. f∗(S∗) ≥ (1 − 1/e − )f∗(Sopt) with probability
at least 1 − 2/N provided that l ≥ l∗.
Proof. For each S ⊆ V ,
(cid:17)
11(1 + )2 · f∗(Sopt)
2 · n + n · 11(1 + )
(1 + )2 · f∗(Sopt)
}
l1
= Pr
≤ exp
n
Pr[x(Rl, S) − f∗(S) ≥ 11(1 + )f∗(Sopt)]
· 11(1 + ) · f∗(Sopt)
− f∗(S)
≥ f∗(S)
·(cid:0) 11·(1+)f∗(Sopt)
(cid:1)2
(cid:17)
{By the Chernoff bound}
l ·(cid:0)11 · (1 + )f∗(Sopt)(cid:1)2
2 + 11·(1+)f∗(Sopt)
f∗(S)
f∗(S)
f∗(S)
(cid:17)
n
2 · n · f∗(S) + n · 11 · (1 + )f∗(Sopt)
{Since f∗(S) ≤ f∗(Sopt) }
n
n
(cid:104) x(Rl, S)
(cid:16) − l · f∗(S)
(cid:16) −
(cid:16) − l · 2
(cid:1) · N
1(cid:0)n
{Since l ≥
≤ exp
≤
k
Note that S∗ = k and there are at most(cid:0)n
(cid:1) subsets of V with
k
size k. By the union bound, with probability at least 1− 1/N,
we have x(Rl, S∗)− f∗(S∗) ≤ 11(1 + )· f∗(Sopt). Here we
have to consider all the subsets with size k because the solution
produced by Alg. 3 is unknown to us before generating the
R-samples.
that when l ≥
we have x(Rl, Sopt) − f∗(Sopt) ≥ −12(1 + )f∗(Sopt).
Second, applying the Chernoff bound immediately yields
(1+)2f∗(Sopt), with probability at least 1−1/N,
l2
Since l ≥ l∗, by the union bound, f∗(S∗) ≥ (1 − 1/e −
)f∗(Sopt) with probability at least 1 − 2/N.
So far 11 > 0 and 12 > 0 are free parameters as long
as they satisfy that Eqs. (7) and (8). Note that Eqs. (7) and
(8) have at least one feasible solution. Ideally we should
minimize l, so the optimal setting can be obtained by solving
the following problem.
minimize max(l1, l2)
subject to Eqs. (7) and (8)
(9)
Now the only unknown part of l∗ is f∗(Sopt). According to
Lemma 8, for each OP TL with OP TL ≤ (1 + )2f∗(Sopt),
setting l as max(l1, l2)/OP TL will be sufficient. In the next
subsection, we will show how to obtain such a OP TL.
E. Lower Bound of f∗(Sopt)
Note that for each SL ⊆ V with SL = k, f∗(SL) is a lower
bound of f∗(Sopt), and according to Lemma 2, f∗(SL) can be
estimated by the hybrid sampling process. The following result
establishes the relationship between the number of samples
and the accuracy.
Lemma 9 (Dagum et al. [19]). For each 0 ≤ 0 ≤ 1 and
N ≥ 0, there exists an algorithm which, with probability at
least 1 − 1/N, produces a f∗
(SL) such that
(1 − 0)f∗(SL) ≤ f∗
(10)
where the number of the used simulations is asymptotically
bounded by O( 2
0+4(e−2)(1+0) ln(N/2)
(SL) ≤ (1 + 0)f∗(SL)
).
0
0
0· f∗(SL)
n
2
0
0
(cid:17)
Taking 0 as 2 + 2, such a lower bound OP TL = f∗
(cid:16) n·m(1+max1)·ln N
(SL)
(SL) ≤ (1 + 2 + 2)f∗(SL) ≤ (1 + )2f∗(Sopt),
satisfies f∗
and combining Lemma 3, the expected running time of gener-
ating such an estimate is O
for any small
. Since we would prefer a tight lower bound so that the
running time can be minimized, we could select SL through
intuitive heuristics. For example, selecting the our-neighbors
of the nodes in Sr.
F. The HMP Algorithm
2·OP TL
Now we are ready to present our algorithm. The algorithm
is formally given in Alg. 5, termed as Hybrid-sampling based
Misinformation Prevention (HMP) algorithm. As shown in
Alg. 5, we first compute a lower bound OP TL of (1 + )2 ·
f∗(Sopt) according to Sec. IV-E, then solve Eq. (9) to find the
optimal setting of 11 and 12. Finally, we use the framework
Alg. 4 to obtain the solution S∗.
Performance Guarantee. The HMP algorithm has the
following performance bounds.
Theorem 1. With probability at least 1 − 3/N, the HMP
algorithm returns a solution S∗ such that
f∗(S∗) ≥ (1 − 1/e − )f∗(Sopt).
Proof. Since OP TL ≤ (1+)2f∗(Sopt) holds with probability
at
the theorem follows immediately from
Lemma 8.
least 1 − 1/N,
Algorithm 5 HMP Algorithm
1: Input: G, , N, Sr and k;
2: Select any subset SL with SL = k and compute OP TL
by Lemma 9 with 0 = 2 + 2;
3: Compute 11 and 12 by solving Eq. (9).
4: l ← max(l1, l2)/OP TL
5: S∗ ← Alg. 4(G, Sr, k, l);
6: Return S∗;
Running Time. In Alg. 4, there are totally l R-samples
). By Lemma 3, the
generated where l = O( n·(k ln n+ln N )
expected running time of line 2 in Alg. 4 is
(cid:16) m · n(k ln n + ln N )(1 + max1)
2·OP TL
(cid:17)
.
(11)
O
2 · OP TL
Furthermore, Alg. 3 can be implemented to run in time
linear to the total size of its input [20]. Therefore, line 5
in Alg. 5 runs in the same as Eq. (11). The rest part of
Alg. 5 is dominated by line 5 so the whole algorithm runs
asymptotically in the same as line 5 does.
For the one interested in a success probability of 1 − n−l,
the running time would be
(cid:16) m · n · ln n · (k + l)(1 + max1)
(cid:17)
.
O
2 · OP TL
G. Comparing to the Existing Works.
There are three most related existing works, [2], [4], [15],
which aim at solving Problems 1 and 2 by designing ap-
proximation algorithms. C. Budak et al. [2] first proposed
the problem of limiting the spread of information and they
considered Problem 2. In [2], it is shown that Problem 2
is monotone and submodular, and the authors use Monte
Carlo simulation straightforwardly (i.e., the forward sampling
method) to overcome the difficulty in computing the objective
function. G. Tong et al. [4] applied the reverse sampling
technique invented by [11] -- [13] to solve Problem 1. Very
recently, M. Simpson et al. [15] adopted the same technique
to solve Problem 2. In the work [4] and [15], the authors
utilized the same sampling method designed based on the
uniform reverse sampling framework. As mentioned earlier,
selecting a node uniformly is not efficient for the MN or
MP problem, because it cannot attach extra importance to
the nodes that are prone to be Cr-active. Furthermore, as
discussed in [15], for the MP problem, the sample obtained
by the uniform reverse sampling method can be an empty set.
When an empty set is sampled, it corresponds to the case that
the selected node will not be affected by the misinformation
and we do not need to select any positive seed node to protect
that node. As shown in Alg. 3, the empty set cannot provide
any information for selecting effective nodes. On the other
hand, in the hybrid sampling method, every sample we obtain
by Alg. 2 is guaranteed to be non-empty, and therefore our
sampling process is more effective in generating the samples
Dataset
Wiki
HepPh
Epinions
DBLP
Youtube
Node
7K
34K
75K
317K
1.1M
Description
Edge
103K Wikipedia
421K
who-votes-on-whom
network
Arxiv High Energy Physics paper
citation network
508K Who-trusts-whom network
Epinions.com
of
1.04M DBLP collaboration network
6.0M
Youtube online social network
TABLE I: Datasets
that can help us find high-quality positive seed nodes. We have
observed the same in our experiments.
V. EXPERIMENTS
We now the present the experiments done for evaluating
the proposed solution by comparing it to the state-of-the-art
solutions, RBR [4] and RPS [15].
A. Experimental Setting
In most of the existing works,
the performance of an
algorithm is measured by either the running time or the quality
of the produced solution. However, for two algorithms namely
A and B, it happens that A produces a better solution than that
produced by B, but meanwhile, B runs much faster than A
does. Under such a scenario, one can hardly decide that which
algorithm is better. In order to fairly compare the algorithms,
we define the performance of an algorithm as the quality of
the produced solution under a specific time constraint.
Note that all the three algorithms, HMP, RBR and RPS,
follow the framework shown in Alg. 4, and the only difference
is that HMP utilizes a different method to generate samples.
The three algorithms all have the following two properties: (1)
the more samples we use, the better solution we obtain; (2) the
quality of the solution finally converges to an absolute value
upper bounded by the optimal value. Therefore, by controlling
the number of the used samples, there is a natural trade-off
between the running time and quality of the solution. In this
way, we are able to see that with the same running time which
algorithm is the best. As discussed in Sec. IV-G, the RBR
algorithm and the RPS algorithm essentially use the same
method to generate samples, so we only compare HMP with
RBR.
Experimental Setting. We adopt five datasets, from small
to large, Wiki, HepPh, Epinions, DBLP, and Youtube, bor-
rowed from SNAP [21]. A brief description of each dataset
is given in Table. I. We consider three types of propagation
probability: (1) the uniform setting where Pr[e] = 0.1; (2)
the uniform setting where Pr[e] = 0.01; (3) the weighted-
cascade model where Pr[(u, v)] = 1/dv for each (u, v) ∈ E
and dv is the in-degree of v. These settings have been widely
adopted in prior work. We set k = 10 and Sr = 15 in our
experiments, unless otherwise specified. The nodes in Sr are
selected from the nodes with the highest individual influence.
For each algorithm and dataset, we continuously increase the
number of the used samples and record the produced positive
the seed set, until the quality of the solution tends to converge.
Each solution is finally evaluated by 10,000 Monte Carlo
simulations.
B. Experimental Results
The experimental results are shown in Fig. 3. Each figure
gives two curves plotting the results of HMP and RBR
respectively. The minimum unit of the running time is 0.1
second, and it is shown as 0 when the running time is less
than 0.1 second.
Major observations. The most important observation is
that HMP consistently outperforms RBR under all considered
datasets and propagation probability settings. Under the same
time constraint, the solution of HMP is more effective than that
of RBR with respect to maximizing f∗(S). For example, in
Fig. 3d, when the running time is required to be less than one
second, HMP is able to protect more than 500 users while
RBR can only protect 220 users. Furthermore, HMP takes
much less time to reach the maximal performance. As an
extreme example, in Fig. 3f, HMP converges after 5 seconds
while RBR converges after 40 seconds. In other words, a
few samples will be sufficient for HMP to achieve a high
performance. Finally, the curves of HMP is smoother than that
of RBR, which means HMP is more robust than RBR. Such
a scenario suggests that the variance of the solution produced
by RBR is typically higher than that of the solution given by
HMP.
Minor observations. According to the figures, HMP usually
has a breakpoint after which the quality of the solution remains
unchanged. For example, (0.25, 49) in Fig. 3a and (50, 3500)
in Fig. 3i. On the other hand, curves of RBR tend to increase
gradually and there is no significant breakpoint. It would be
great if we can identify the numbers of the samples corre-
sponding to such breakpoints, because they are the minimum
number to reach the maximal performance. Unfortunately, we
find that the theoretical bound given in Sec. IV usually far
exceeds the minimum requirement of the breakpoints. One
possible reason is that the theoretical bounds are dealing with
the worst-case scenario so it is too pessimistic for the average
case. Another reason is that we assume the variance is n2 in the
Chernoff bound, which leads to an overestimate of the number
of the required samples. The variance of the distribution over
the diffusion outcomes can significantly help us to obtain an
accurate estimate. Currently, we are not aware of any existing
work which has considered this issue.
Remark 1. It should be noted that when the objective function
stops increasing, its maximal performance has been almost
reached. When the maximal performance is reached,
the
solution produced at that moment is no worse than a (1−1/e)-
approximation as 1 − 1/e is the worst-case ratio.
VI. CONCLUSION AND FUTURE WORK
In this paper, we present a hybrid sampling method which
is designed particularly for the misinformation prevention
problem. We show that the new sampling method can be used
(a) Wiki under weighted-cascade model
(b) HepPh under weighted-cascade model
(c) HepPh under uniform 0.1
(d) Epinions under weighted-cascade model
(e) Epinions under uniform 0.01
(f) DBLP under weighted-cascade model
(g) DBLP under uniform 0.1
(h) Youtube under weighted-cascade model
Fig. 3: Experimental Result.
(i) Youtube under uniform 0.01
to design an approximation algorithm which outperforms the
state-of-the-art solutions.
We propose some future works which we believe are
interesting. It has been shown that there exist O(k·m·log n/2)
algorithms for the MN and IM problem. Though the MP
problem and the MN problem merely differ by a constant, the
best algorithm for the MP problem is O(k · m · n · log n/2)
(assuming OP TL ≥ (1 + max1)), and it is open that if we
can find a O(k · m · log n/2) algorithm for the MP problem.
In addition,
the existing lower bounds on the number of
the samples used to reach the breakpoints observed in the
experiments are still pessimistic. We are looking for more
practical methods to help us find such breakpoints so that we
can achieve the best performance with the minimum running
time.
ACKNOWLEDGMENT
We thank the reviewer for pointing out the issue discussed
in Remark 1. This work is supported in part by the start-up
grant from the University of Delaware.
REFERENCES
[1] G. A. Tong and D.-Z. Du, "Beyond uniform reverse sampling: A hybrid
sampling technique for misinformation prevention," in IEEE INFOCOM
2019-IEEE Conference on Computer Communications.
IEEE, 2019, pp.
1711 -- 1719.
[2] C. Budak, D. Agrawal, and A. El Abbadi, "Limiting the spread of
misinformation in social networks," in Proc. of WWW, 2011.
[3] L. Fan, Z. Lu, W. Wu, B. Thuraisingham, H. Ma, and Y. Bi, "Least cost
rumor blocking in social networks," in Proc. of ICDCS, 2013.
[4] G. Tong, W. Wu, L. Guo, D. Li, C. Liu, B. Liu, and D.-Z. Du, "An
efficient randomized algorithm for rumor blocking in online social
networks," IEEE Transactions on Network Science and Engineering,
2017.
[5] S. Li, Y. Zhu, D. Li, D. Kim, and H. Huang, "Rumor restriction in
online social networks," in Proc. of IPCCC, 2013.
[6] X. He, G. Song, W. Chen, and Q. Jiang, "Influence blocking maximiza-
tion in social networks under the competitive linear threshold model."
in Proc. of SDM, 2012.
[7] Y. Ping, Z. Cao, and H. Zhu, "Sybil-aware least cost rumor blocking in
social networks," in Proc. of GLOBECOM, 2014.
[8] G. Tong, W. Wu, and D.-Z. Du, "On misinformation containment in
online social networks," in Proc. of NIPS, 2018.
[9] D. Kempe, J. Kleinberg, and ´E. Tardos, "Maximizing the spread of
influence through a social network," in Proc. of SIGKDD, 2003.
0246810303540455055Running time (s)Objective value HMP AlgorithmRBR Algorithm5101520600650700750800850Running time (s)Objective value HMP AlgorithmRBR Algorithm123456789009501000105011001150120012501300Running time (s)Objective value HMP AlgorithmRBR Algorithm020406080200300400500600700800900Running time (s)Objective value HMP AlgorithmRBR Algorithm050100150405060708090100110120130Running time (s)Objective value HMP AlgorithmRBR Algorithm0204060802000300040005000600070008000900010000Running time (s)Objective value HMP AlgorithmRBR Algorithm01002003004005004005006007008009001000110012001300Running time (s)Objective value HMP AlgorithmRBR Algorithm050010001500100020003000400050006000700080009000Running time (s)Objective value HMP AlgorithmRBR Algorithm02004006008001000120014005001000150020002500300035004000Running time (s)Objective value HMP AlgorithmRBR Algorithm [10] W. Chen, C. Wang, and Y. Wang, "Scalable influence maximization for
prevalent viral marketing in large-scale social networks," in Proc. of
SIGKDD, 2010.
[11] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier, "Maximizing social
influence in nearly optimal time," in Proc. of SODA, 2014.
[12] Y. Tang, X. Xiao, and Y. Shi, "Influence maximization: Near-optimal
time complexity meets practical efficiency," in Proc. of SIGMOD, 2014.
[13] Y. Tang, Y. Shi, and X. Xiao, "Influence maximization in near-linear
time: a martingale approach," in Proc. of SIGMOD, 2015.
[14] 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. of SIGMOD, 2016.
[15] M. Simpson, V. Srinivasan, and A. Thomo, "Scalable misinformation
prevention in social networks," arXiv preprint arXiv:1807.01162, 2018.
[16] G. Tong, W. Wu, and D.-Z. Du, "Distributed rumor blocking with
multiple positive cascades," IEEE Transactions on Computational Social
Systems, vol. 5, no. 2, pp. 468 -- 480, 2018.
[17] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, "An analysis of ap-
proximations for maximizing submodular set functionsi," Mathematical
Programming, vol. 14, no. 1, pp. 265 -- 294, 1978.
[18] R. Motwani and P. Raghavan, Randomized algorithms. Chapman &
[21] J. Leskovec and A. Krevl, "SNAP Datasets: Stanford large network
dataset collection," http://snap.stanford.edu/data, Jun. 2014.
Hall/CRC, 2010.
[19] P. Dagum, R. Karp, M. Luby, and S. Ross, "An optimal algorithm for
monte carlo estimation," SIAM Journal on computing, vol. 29, no. 5,
pp. 1484 -- 1496, 2000.
[20] V. V. Vazirani, Approximation algorithms. Springer Science & Business
Media, 2013.
|
1809.06656 | 1 | 1809 | 2018-09-18T12:00:57 | Probing Limits of Information Spread with Sequential Seeding | [
"cs.SI",
"physics.soc-ph"
] | We consider here information spread which propagates with certain probability from nodes just activated to their not yet activated neighbors. Diffusion cascades can be triggered by activation of even a small set of nodes. Such activation is commonly performed in a single stage. A novel approach based on sequential seeding is analyzed here resulting in three fundamental contributions. First, we propose a coordinated execution of randomized choices to enable precise comparison of different algorithms in general. We apply it here when the newly activated nodes at each stage of spreading attempt to activate their neighbors. Then, we present a formal proof that sequential seeding delivers at least as large coverage as the single stage seeding does. Moreover, we also show that, under modest assumptions, sequential seeding achieves coverage provably better than the single stage based approach using the same number of seeds and node ranking. Finally, we present experimental results showing how single stage and sequential approaches on directed and undirected graphs compare to the well-known greedy approach to provide the objective measure of the sequential seeding benefits. Surprisingly, applying sequential seeding to a simple degree-based selection leads to higher coverage than achieved by the computationally expensive greedy approach currently considered to be the best heuristic. | cs.SI | cs | Probing Limits of Information Spread with Sequential Seeding
Jarosław Jankowski1∗, Boleslaw K. Szymanski2,4, Przemysław Kazienko3,
Radosław Michalski3, Piotr Bródka3
8
1
0
2
p
e
S
8
1
]
I
S
.
s
c
[
1
v
6
5
6
6
0
.
9
0
8
1
:
v
i
X
r
a
1 Faculty of Computer Science and Information Technology, West Pomeranian University of Technology,
70-310 Szczecin, Poland
2 Social and Cognitive Networks Academic Research Center and Department of Computer Science,
Rensselaer Polytechnic Institute, Troy NY 12180, USA
3 Faculty of Computer Science and Management, Wroclaw University of Science and Technology, 50-370
Wroclaw, Poland
4 Społeczna Akademia Nauk, Łódź, Poland
Abstract
We consider here information spread which propagates with certain probability from nodes just activated
to their not yet activated neighbors. Diffusion cascades can be triggered by activation of even a small set
of nodes. Such activation is commonly performed in a single stage. A novel approach based on sequential
seeding is analyzed here resulting in three fundamental contributions. First, we propose a coordinated ex-
ecution of randomized choices to enable precise comparison of different algorithms in general. We apply it
here when the newly activated nodes at each stage of spreading attempt to activate their neighbors. Then,
we present a formal proof that sequential seeding delivers at least as large coverage as the single stage seed-
ing does. Moreover, we also show that, under modest assumptions, sequential seeding achieves coverage
provably better than the single stage based approach using the same number of seeds and node ranking.
Finally, we present experimental results showing how single stage and sequential approaches on directed
and undirected graphs compare to the well-known greedy approach to provide the objective measure of the
sequential seeding benefits. Surprisingly, applying sequential seeding to a simple degree-based selection leads
to higher coverage than achieved by the computationally expensive greedy approach currently considered to
be the best heuristic.
Keywords: diffusion, social network, seed selection, influence maximization, sequential seeding
Introduction
Making decisions is often difficult; this is why it is often worth to split it into consecutive stages to reduce
potential risk. Such approach gives the decision maker an opportunity to learn the outcomes of the previous
stages and adjust accordingly the current stage. This approach applies also to influence maximization and
information spread in complex networks. One of the main challenges there is the selection for initial and
activation network nodes referred to as seeds, to maximize the spread of information within the network [1].
∗E-mail:[email protected]
1
Various factors affecting the diffusion and social influence in complex networks were analyzed including
the role of different centrality measures in selection of initial influencers [2], impact of homophily for successful
seeding [3] and others [4]. While most of the relevant research is related to marketing, the problem is more
generally defined as a target set selection in combinatorial optimization in theoretical computer science [5 -- 7].
The influence maximization problem is also explored in physics from the perspective of network structures [8].
Some other studies discuss the role of communities [9] or propose to use optimal percolation [10]. Initial
research has been carried out to identify seeds for temporal [11, 12] and multi-layered social networks [13].
Several comparative studies on seeding strategies were presented in [14,15]. Since the seed selection process
is NP-hard [1], several heuristics have been proposed. Most of them use network structural properties like
degree or eigenvector measure to rank seed candidates [14].
Seeding strategies have been applied to word-of-mouth marketing, social and political campaigns and
diffusion of information in social media. Vast majority of them are based on the assumption that all seeds
are activated at the beginning of the process or campaign and then diffusion starts and continues naturally
without any additional support [14]. Some recent research proposes to apply adaptive approaches with
two-stage stochastic model exploring the potential of neighboring nodes [16], further extended towards more
scalable approach [17]. Using some seeds after the first stage was preliminary proposed in [18, 19] and
potential of multi-period spraying for routing in delay-tolerant networks was discussed as an effective solution
for propagation in computer networks in [20].
On top of that, the concept of sequential seeding was introduced [21]. It takes the advantage of delayed
seeding by interweaving activating seeds with diffusion. Here we firm the foundations of sequential seeding
and address a number of research questions that have not been studied before [21 -- 23]. First, we prove
formally and confirm experimentally on real networks that the sequential seeding is at least as good as the
single stage method and under modest conditions, it is provably better. Then, we evaluate the realistic
benefits of this method by comparing its capabilities to both greedy and optimal (maximal) solutions to
obtain some measure of improvements provided by this approach. Finding optimal seed sets and computing
the maximal coverage for various scenarios allow us to represent gains from sequential seeding as a fraction
of the space between coverage of other algorithms and the maximal coverage. This has been investigated
empirically for directed and undirected networks using independent cascades model [1].
Strategy Description
No. of seeding stages
SN
SQ
Single Stage Seeding -- all seeds are activated in one stage at the
beginning
Sequential Seeding -- one seed activation starts the stage and re-
vives the previously stopped diffusion; each stage ends when dif-
fusion stops
1
n
Table 1: Seeding strategies with n number of seeds
In the first phase, we ran simulations on six real complex networks with sizes in the range of 1.5k-17k
nodes. The visualization of the networks and their basic properties are provided in the Supplementary
Material. The independent cascades model (IC) [1] was used with a given propagation probability P P (a, b)
that node a activates (influences or infects) node b in step t+1 under the condition that node a was activated
at time t [24]. The results for undirected networks assume that P P (a, b) = P P (b, a) for all edges. This
simplification allowed us to obtain the optimal seed set through the complete search even for largest networks
among those used for experimentation. To explore the directed networks without this assumption, another
large set of extensive experiments were performed on the smaller network, for which it was also possible to
compute optimal solutions, see the next section.
2
Figure 1: Undirected networks: (A) The averaged performance of sequential SQ and single stage seeding
SN with greedy based nodes selection as a fraction of the maximum coverage CM ax and as a function of
the network size N, probability of propagation PP across each edge, and the fraction of nodes selected as
seeds (seed selection percentage) SP; see (A1) Performance of sequential SQ and single stage seeding SN
with degree-based ranking in comparison with maximum coverage and the upper bound as a function of
the individual configurations, each defined by N, PP, SP, and the random binary choice made at simulation
initialization for each edge to propagate or not information across this edge; (B) Coverage of sequential
method SQ as percentage of CM ax placed between single stage seeding SN and Max for greedy nodes
selection; (B1), (B2) Sequential seeding performance SQ between single stage seeding SN and Max for
random and degree-based node selection, respectively; (B3) Performance of sequential SQ and single stage
seeding SN represented by percentage of activated nodes within network (coverage) for random seed selection,
degree-based ranking and greedy seed selection in comparison with maximum coverage as a function of the
individual configurations; (C) Gain for different propagation probabilities PP averaged over all cases; (D)
Gain for different seeding percentages SP; (E) Gain for node ranking strategies based on random, greedy
and degree selection; (F) Gain for networks N1 - N6.
With the independent cascades model (IC) even a single seed can induce diffusion (which is crucial for
sequential approaches), while in the linear threshold model [1] (LT), a small seed set could have no effect.
For that reason, experiments were carried out on the IC model.
Results achieved by sequential seeding (SQ) were compared with the single stage approach (SN) and the
maximum coverage (Max) for the same setup: network and parameters: the propagation probability, the
seeding percentage, and the seed ranking strategies: degree, greedy and random. The sequential coverage
CSQ was on average 7.1% better than the single stage coverage CSN. The positive gain of SQ over SN was
also confirmed by statistical tests, see Section Material and Methods.
Results for greedy based ranking for 50,000 simulations represented as percentage of maximum coverage
CM ax averaged over individual configurations and ordered by the ratio CSN / CM ax are presented in Fig. 1
(A). Additionally, the upper bound CGreedySN ∗ e/(e − 1) [1] is for many configurations up to 50% greater
than real maximum value observed, see Fig. 1 (A1). These results demonstrate that upper limit derived
from greedy approach is not tight. Sequential methods always outperform single stage ones, if we consider
averaged coverage, for any strategy as well as for every configuration. Moreover, degree SQ is able to cover
more nodes than greedy SN, especially for configurations with coverage significantly lower than CM ax, as
shown in Fig. 1 (A1).
There were 8,100,000 individual simulation cases defined during simulation initialization, each case by
its N, PP, SP, and binary random choice selected for each edge to propagate information or not across
3
this edge. These cases are ordered by coverage obtained in the single stage method. The resulting plots
demonstrate that CSQ performs better than CSN in almost every case. The greatest increase is observed for
CSN ∈ [30%, 85%]; simply the space for improvement is larger in such cases.
The global results for all networks, strategies, parameters and random binary choices for each edge to
propagate information or not across this edge made at the simulation initialization yield better values of
CSQ than of CSN in 96.7% of cases. The increase over 5% was achieved in 20.2% of cases. The results
were dependent on node ranking strategy with 96.0% better results for random rankings, 100% for degree
based rankings and 93.9% for greedy approach based ranking. The improvement over 5% was observed for
11.2% cases for random strategy and as much as 38.7% cases for degree-based selection and 10.9% for greedy
method. It should be also noted that, surprisingly, the random selection performs slightly better than the
expensive greedy selection.
Results for individual cases with respect to maximum coverage are presented in Fig. 1 (B). The sequential
results are localized above the single stage border, filling the space towards maximum (100%). The dispersion
is the highest for degree based rankings Fig. 1 (B1). The lowest dispersion of results is observed for random
based node selection, see Fig. 1 (B2). Moreover, if the single stage coverage is at least at the level of 90%, all
strategies are able to provide sequential cases very close to maximum. For lower coverage, only degree-based
rankings can improve results so much.
The results revealed an important phenomenon: sequential seeding based on degree selection in 92.2%
of cases outperforms single stage greedy approach. A similar improvement is also observed for the averaged
values. It means that sequential approach almost always is able to boost performance of the simple degree-
based ranking over computationally expensive greedy heuristic. Moreover, CdegreeSQ is greater than CgreedySQ
in 62.6% of cases but for single stage methods such superiority can be observed in only 0.01% of cases.
Hence, the main finding is that the computationally ineffective greedy strategy is suitable for single
stage approach, while degree based selection for sequential seeding is significantly better than other selection
methods.
In general, the obtained results were dependent on the network profile and parameters of the diffusion.
Space between the maximum coverage CM ax and the single stage seeding CSN is an area in which sequential
approaches deliver improvement. This area on average is (CM ax − CSN )/CSN so only 25%, of the all
simulation cases.
To evaluate the improvement, a gain measure G was defined; it is based on average coverage values:
G = ((CSQ − CSN )/(CM ax − CSN )) ∗ 100%. It shows what part of the improvement area is reachable by
sequential approach. Depending on process parameters and network, the gain varies from 30% to 83%. In
general, the greater propagation probability value, the greater coverage and gain, see Fig. 1 (C). The same
phenomenon arises for seeding percentages, as seen in Fig. 1 (D). Regarding node selection strategies, the
highest average gain 74% was observed for degree based selection, while for greedy and random strategies
it is much lower: 27% and 33% respectively, as shown in Fig. 1(E). The gain strongly depends on network
structure, see Fig. 1 (F). The highest gain (83%) was achieved for network N6, whereas the least gain (30%)
was achieved for network N3.
Results
Experimental results for directed networks
As in case of undirected networks, coordinated execution experiments for a real directed network were run
to analyze gain from sequential seeding compared to greedy approach and maximal possible coverage. Since
they requires computationally expensive search for the optimal seed sets, only a small network with 16 nodes
and 58 edges was used [25]. Following the coordinated execution principles, 10k instances of the network were
4
randomly generated to assign binary choices of propagation or not for each directed edge, i.e., independently
for a → b and b → a activation. One of these instances is presented in Fig 2. For each instance and
propagation probability PP, an optimal 4-node seed set was computed to estimate the maximum coverage
for this instance and PP. Results for all probabilities and ranking strategies (degree, greedy, random) show
that the highest increase of coverage was observed for the sequential degree based selection, see Fig. 3 (B),
(D). Moreover, the sequential approach with the degree rankings delivered better results than greedy-based
selection regardless if used in single stage or sequential mode, as shown in Fig. 3 (A), (D). It means that the
degree approach should be among the first choices considered when selecting the node ranking method.
The average performance (gain in coverage) is strongly dependent on propagation probabilities (PP) and
it increases with PP values, see Fig. 3 (C), even though higher probabilities leave less space for improvements,
since they also raise the single stage coverage. This effect is present for all selection strategies. The reason
is that although the processes with high PP reaches more than 80% of nodes for all strategies and the area
for gain is much smaller, the savings from the sequential approaches are even greater than for smaller PP's,
as shown in Fig. 3 (E), benefiting the final results.
In general, the sequential approach enables us to save seeds and allocate them to other network regions.
This gain varies for different propagation probabilities and ranking strategies, as shown in Fig. 3 (E), and it
comes from activation of seed candidates by inter-stage diffusion. The degree ranking reveals its superiority
over other ranking methods also in seed saving, which may, in fact, be the reason why the sequential degree
method outperforms the single stage greedy, as seen in Fig. 3 (A). More than 25% of seeds can be saved even
for smallest propagation probabilities and up to 48% for PP=0.25 and degree-based selection. These results
may guide future research on methods to find the minimal number of seeds used sequentially to achieve the
same coverage as in the single stage seeding.
Overall, the main findings for directed networks are similar to the ones for undirected graphs: (1)
sequential approach almost always significantly increases coverage of spreading and reaches beyond the
current limits (never makes it worse), (2) sequential degree-based selection is commonly better than the
greedy selection and (3) it is often close to the maximum, (4) usually, the greater propagation probability,
the greater gain, and (5) sequential approach enables us to replace some of the initial seeds with the additional
ones thereby increasing coverage, especially for greater PP and degree selection.
Discussion
The presented results lay firm foundation for the earlier studies related to sequential seeding and address
currently unanswered research questions
[21 -- 23]. Averaging the coverage obtained by the typical agent
based simulations delivers improvement in about 90% of the cases
[21], the formal proof relying on the
coordinated execution and corroborated by the new empirical studies shows that sequential seeding delivers
at least as large coverage as the widely used single stage seeding; under modest assumption this coverage is
provably larger. The performance of sequential seeding is compared with both greedy seed selection (which
may be treated as the current limit) and the optimal seed set providing maximal possible coverage for a
given number of seeds and the given network configuration. Experiments were performed for directed and
undirected networks by means of independent cascades model [1]. Sequential usage of seeds selected by the
simplest degree-based heuristics delivers better coverage than the greedy algorithm used in the single stage
mode. However, coverage achieved by the greedy selection can also be further improved, if seeds are used
sequentially instead of being all activated at the beginning of execution, even though the resulting coverage
still remains worse than the one that the sequential degree selection delivers.
Overall, the results confirm that sequential allocation of seeds improves the coverage of information
diffusion in both directed and undirected complex networks. The measure of improvements provided by this
approach is showed. It helps to substitute seeding nodes activated by diffusion by selecting new seeds based
5
Figure 2: One of the experimental cases for the directed network (run 4336) with the number of seeds
n = 4 and the propagation probability P P = 0.05. For clarity and high network density only edges assigned
the random choice of propagating information for them by coordinated execution are shown. Degree- and
greedy-based rankings are computed for the full initial network with all edges. (A) Single stage seeding
(SN) with degree-based ranking, the coverage CSN = 6, a diffusion cascade is visible when node 11 activates
nodes 9 and 10. (B) Diffusion with sequential seeding and degree based selection; the ranking is the same
as in (A); one seed is activated in each of four stages; in the first stage node 6 is activated according to its
highest degree. It activates node 11 which in turn activates nodes 15 and 9; finally, node 9 activates node 10
and diffusion stops. The two seeds used in (A) are already activated by diffusion in this case. In the second
stage node 16 is selected since it has the highest degree among not activated yet nodes. Its only neighbor,
node 11, is already active so the process stops. In the next stage node 1 is selected as a seed but it lacks
active edges so diffusion cannot progress. Fourth seed is node 12 and it activates node 14. Sequential seeding
avoids using nodes 11 and 15 seeds which are activated as seeds in single stage seeding. This allow sequential
seeding to activate two more seeds and three more nodes in total compared to single stage seeding. (C)
Single stage seeding with greedy-based ranking and coverage CSN = 7. (D) Single stage seeding based on
the seed set optimal for this individual case, including knowledge of which edges are active for propagation,
with the resulting coverage CSN = 11.
on node's ranking which may activate new regions, see Fig. 2 (B).
As a result, sequential seeding commonly provides better coverage than the single stage approach, no
matter which initial node selection strategy is used. In the worst case, the results remain the same but the
experimental studies show that it happens in fewer than 10% of all cases.
For single stage mode the greedy approach is commonly better than typical ranking selections based on
structural measures, e.g. degree and it can be treated as the method defining the current limit. However,
6
Figure 3: The directed network: (A) Performance of sequential SQ and single stage seeding SN with degree-
based ranking in comparison with maximum coverage; (B) The coverage achieved by sequential strategies
in comparison to coverage obtained by a single stage approach; (C) Averaged gain for all used propagation
probabilities; (D) Averaged gain for node ranking strategies based on random, greedy and degree selections;
(E) Percentage of saved seeds for degree, random and greedy based selections.
it requires extensive prior simulations, which in practice are hardly possible due to (1) high computational
complexity, and (2) feasibility of simulations for real data -- very rarely we can run thousands of diffusion
processes when the results need to be delivered in the real-time of the campaign.
Sequential seeding outperforms the greedy search method. Moreover, sequential seeding using structural
network measures like degree, for seed selection yields coverage greater than obtained by greedy, without
necessity of any prior simulations. In our experiments, it happened in over 92% of cases. It appears that
using degree-based rankings with sequential strategy is better than using the computationally expensive
greedy approach.
We observed that the main factor that defines performance of sequential seeding versus single stage
seeding is to what extent seed selection method avoids selecting nodes that will be activated through diffusion.
The gain of sequential seeding arises when the selected seed to be activated in the current stage is already
active. In such a case, this seed is removed from the seed list and replaced by a highest ranking node which is
not yet activated. It should be noted that we compare the resulting coverage with the real maximum coverage,
which very often is much lower than the theoretical upper bound for the average maximum coverage suggested
in [1, 26], see red line in Fig. 1 (A1) and 3 (A).
Presented study has several implications for practice. Instead of introducing the product to large number
of customers at the outset of commercial campaign, better strategy is seeding a small fraction of nodes
and giving the chance to natural diffusion driven by social influence mechanisms to spread the content.
Marketing budgets can be optimized if additional seeds are utilized only if campaign fails and revival is
needed. Moreover, the knowledge gain from the initial spreading may improve seed selection for revival.
Increased coverage of spreading might be crucial for campaigns with limited budgets, such as spreading
security information, or disease warnings and awareness. During massive campaigns habituation phenomenon
can arise among customers resistant to marketing messages. It can be avoided by limiting the intensity of
marketing activity. While campaigns with higher intensity can be perceived negatively as unsolicited massive
communication, sequential strategies may avoid making such negative impact on customers. Sequential
seeding is also a low risk strategy with possible high gains because, as we proved, results will never be worse
when compared to single stage seeding.
7
Figure 4: Diffusion processes for the number of seeds n = 3; the classical single stage method (SN) compared
with sequential seeding strategy (SQ).
Methods
Independent cascade model
We consider an independent cascade that is a stochastic diffusion model of information spread in the network
initiated by seeds [1]. A basic, commonly used, single stage seeding (SN) consists of only one activation stage
in which the fixed number of n seeds are activated, see Table 1. They initiate diffusion which runs until
no more nodes can be activated. In the independent cascade model each diffusion step consists of a single
attempt by all nodes activated in the previous step, to activate their direct not yet active neighbors with a
given propagation probability (PP). We measure diffusion time in the number of such steps, assuming that
each lasts a unit of execution time.
Sequential seeding
Let TSN denote the time the whole process of the single stage seeding lasts, see Fig. 4. Often, some seeds
activated at the beginning could have been naturally activated later on. The sequential seeding strategy
(SQ) uses the same number of seeds n as SN [21]. Its idea is to take the same initial ranking of nodes as
in case of single stage seeding, but activate them sequentially in n consecutive stages each with one seed
activated. After each activation, the diffusion proceeds. The next activation is suspended until the current
diffusion process stops, i.e., when the last diffusion step does not increase the coverage. In the independent
cascade model, each recently activated node has only one chance to infect its neighbors. Hence, if the
diffusion stops, the only way to continue activation is activating a seed. In the sequential method, it is the
highest ranking node not yet activated. It means that nodes already activated by diffusion are omitted. The
gain of sequential approach comes from the new areas activated by additional seeds substituting the already
activated ones. Hence, the set of seeds activated in sequential seeding often differs from the one activated
in the single stage approach. Sequential approach may be applied to any initial node ranking computed
once at the beginning of the process. Such ranking may utilize random choice, any structural measures like
commonly used degree or any other heuristic, e.g., greedy [1, 26]. Some questions arise here: what is the
increased coverage CSQ; is CSQ always greater than CSN and what is the gain achieved comparing with the
estimate of the maximum coverage CM ax for a given configuration.
8
Coordinated execution and maximal coverage
The need often arises to compare algorithms which during execution make random choices, e.g., to break
ties, simulate random outcomes according to the given distribution, or are randomized in nature. In such
cases often average values of performance are measured over many runs of each algorithm and used for
comparison between them. Here, we propose coordinated execution as the more direct alternative, in which
we compare performance of runs with the same choices made during executions with different algorithms,
providing run by run comparison of the results. This requires that all random choices are made and recorded
before each run for all compared algorithms. The subsequent runs for comparison of algorithms simply
execute preprocessing of choices with unique seeds for random number generator.
When applied to comparing various seeding methods uses, the coordinated execution requires that each
node activated in any execution makes the same edge diffusion transmission decision. To ensure that all
edges randomly select their activity status before the coordinated executions are run for all variants of single
stage and sequential strategies. It enables us to make fair comparisons between different seeding strategies.
Coordinated execution in the undirected network uses only one edge between two nodes. A propagation
probability between two nodes a → b is the same as for b → a transmission. Coordinated execution in
directed network replaces each edge with two directed edges, and then decides which directed edges will be
activated in diffusion. The edge states are sampled from the binary set {0, 1} with state 1 being chosen with
uniform probability PP.
Edges assigned value of 0 are removed, while edges with value 1 stay, creating an instance of an active
edge graph. Two instances of execution of sequential and single stage methods are coordinated if all edges
originating from node activated in both instances have the same states. Such coordinated instances are used
to compute the maximum coverage CM ax for a given network and active edge choices. Such computation
requires identification of all connected components in the active edge graph, which can be done in O(e) steps,
where e is the number of active edges. Then, the maximum coverage CM ax for a given seed size n is equal
to the total number of nodes within n largest components.
Many instances of active edge graph are then created by simply using a different random number generator
seed for each instance. The number of instances needed to collect the reliable estimate of the average coverage
of the tested seeding methods is provided in [26]. Computing the average of the maximum coverage over
these instances provides a reliable estimate of an upper bound for the maximum average coverage of any
seeding method. As shown in Fig. 1 (A1) and 3 (A), this upper bound is in most cases tighter than the one
derived from the greedy seed selection.
Formal proof of non-decreasing coverage
In a formal proof, we show that the sequential seeding always provides at least as large coverage as the
classical single stage approach does. Using coordinated execution, we also demonstrate that under modest
assumptions the former approach is provably better than the latter when both use the same number of seeds
and node ranking method. To do that, we demonstrate that there exist network configurations for which
sequential seeding coverage is larger than the coverage yielded by the single stage seeding. Configurations
are defined at simulation initialization by choosing randomly the diffusion transmission status for each edge.
Theorem: For arbitrary node ranking and number of seeds, a sequential seeding execution has at least
the same coverage CSQ as the corresponding single stage execution CSN coordinated with it. Moreover, if
there is an initial seed u which is reachable in the original graph from a seed s with rank lower than rank
of u, then there exists a configuration of the original graph for which the coverage for sequential seeding is
larger than it is for single stage seeding.
Proof: For clarity, we assume that seeds are activated in the increasing order of their ranks. We start
by observing that in sequential seeding, the seed ranked r is activated no later than in stage r. It can be
9
activated in the earlier stage either by diffusion or by having a seed with rank lower than r activated by
diffusion, thereby decreasing the ranks of all seeds ranked higher than the activated seed. If neither of these
two cases happens before stage r starts, in this stage, the seed ranked r will be the highest ranking not yet
activated seed and therefore it will be chosen for activation.
We observe also that not yet activated node will be activated by diffusion in a single stage execution of a
given configuration if and only if this node is reachable from a seed by the edges active in this configuration.
Indeed, if a node m is reachable from a seed, this seed will be activated in the first and only stage of
activation. Then, by the rules of diffusion, this seed's activation cascade will reach node m and activate it,
if it is still non-active. Conversely, the seed activated in step t of diffusion in sequential seeding activates
only not yet activated nodes that are distance t from it. The upper bound for the number of diffusion steps
before the diffusion stops is the diameter of the configuration. Hence, no node that is not reachable from any
seed in the given configuration will be activated by diffusion. Since in sequential seeding all original seeds
will be activated no later than by stage n, all nodes reachable from these seeds will also be activated by this
stage, proving the first part of the Theorem.
Regarding the second part of the Theorem, let H be a configuration in which all the edges on the
reachability path from seed s to node u defined in the Theorem are active. At the end of stage n − 1 of
sequential execution of configuration H, all seeds activated in the single stage seeding are activated. Indeed
by definition, all seeds ranked n − 1 has to be activated by then, and the seed of rank n has to be activated
either by diffusion from seed s when rank of u is n or by lowering of its rank by activation of seed u from seed
s when rank of u is less than n. Consequently, diffusion at stage n − 1 comes from all seeds, and therefore
all nodes activated in single stage seeding will be activated in sequential seeding by then. In the next stage
n, not yet activated node will be activated as a seed, making the overall coverage of the sequential seeding
larger for this configuration than the coverage of the single stage seeding; QED.
Greedy seed set search
Greedy method requires many simulations to estimate the potential of a single node, i.e., its ability to
activate other nodes via diffusion [1, 26]. Here, the greedy method is based on approach presented in [26].
Averaged results of the greedy algorithm (coverage Cgreedy - the average total number of nodes activated by
the process) are no worse than CM ax∗(1−1/e), where CM ax is the expected maximum coverage [26]. It means
that the greedy approach also defines the theoretical upper bound for maximum coverage: Cgreedy ∗ e/(e− 1).
Due to its great coverage, the greedy algorithm is currently treated as the benchmark to beat and reference
to. Since it requires prior simulations to compute the node potential, it is also very inefficient thus hardly
applicable in practice, especially for large networks.
Experimental setup: undirected networks
The experiments were carried out on six real complex networks: N1 - Condensed Matter collaboration
[27], N2 - Communication network at University of California [28], N3 - High-Energy Theory collaboration
network [29], N4 - Political blogs [30], N5 - ego-Facebook [31] and N6 - wiki-Vote [32].
The parameters used in experimental configurations define diffusion, networks and seed selection strategy
as shown in Table 2. Three commonly used strategies were exploited: the highest degree, greedy [26]
and random selection. Simulation parameters create configuration space N × P P × SP × R with 162
configurations, each were independently applied to both single stage (SN) and sequential seeding strategies
(SQ) using coordinated execution repeated 50,000 times, resulting in 162*50,000*2=16,200,000 simulation
cases; 8,100,000 cases for each single stage and sequential strategies.
For the greedy approach, finding the seed sets required another 10,000 simulations computed indepen-
dently for each configuration, as defined in [1,26]. The nodes were ranked according to their average coverage
10
Parameter
Network - N
Propagation probability - PP
Seeds percentage - SP
Seed ranking strategy - R
Values Variants
6
3
3
3
N1-N6 real networks
0.05, 0.1, 0.20
1%, 3%, 5%
random, degree, greedy
Table 2: Configuration space of the simulated diffusion processes
over all 10,000 simulations; a separate greedy ranking was computed for each of the 162 parameter combi-
nations.
Statistical tests
The positive gain of sequential approach (SQ) over single stage (SN) for experiments on undirected networks
was also confirmed by the Wilcoxon signed rank test, with p<2.2e-16 and ∆=1.9, with the Hodges-Lehmann
estimator used as a difference measure. Values ∆ > 0 demonstrate significantly larger coverage for CSQ
than CSN. Regarding seeding strategies based on node ranking methods, sequential approach increases
coverage for random ranking on average by 3.2%, with p<2.2e-16 and ∆=1.6. Sequential seeding based on
degree delivered coverage results 15.5% better than single stage seeding, with p-value<2.2e-16 and ∆=4.2.
For sequential seeding and greedy based ranking 2.5% average coverage improvement was achieved with
p-value<2.2e-16 and ∆=0.7.
Experimental setup: a directed network
As much as 10,000 instances of coordinated execution for the small real network of 16 nodes were randomly
selected, to ensure stability of the solutions; further instances would not affect results [1]. Due to a small
number of network nodes, only four seeds were used arbitrarily. Five propagation probabilities (PP) with
values 0.05, 0.1, 0.15, 0.2 and 0.25 were applied. Similarly to the undirected version, three ranking strategies
were analyzed: degree, greedy and random.
References
[1] Kempe D, Kleinberg J, Tardos É (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), pp. 137 -- 146.
[2] Kiss C, Bichler M (2008) Identification of influencers - measuring influence in customer networks. Decis.
Support Syst. 46(1):233 -- 253.
[3] Nejad MG, Amini M, Babakus E (2015) Success factors in product seeding: The role of homophily.
Journal of Retailing 91(1):68 -- 88.
[4] Liu-Thompkins Y (2012) Seeding viral content. Journal of Advertising Research 52(4):465 -- 478.
[5] Ackerman E, Ben-Zwi O, Wolfovitz G (2010) Combinatorial model and bounds for target set selection.
Theoretical Computer Science 411(44 -- 46):4017 -- 4022.
[6] Ben-Zwi O, Hermelin D, Lokshtanov D, Newman I (2011) Treewidth governs the complexity of target
set selection. Discret. Optim. 8(1):87 -- 96.
11
[7] Chiang CY, Huang LH, Li BJ, Wu J, Yeh HG (2013) Some results on the target set selection problem.
Journal of Combinatorial Optimization 25(4):702 -- 715.
[8] Galstyan A, Musoyan V, Cohen P (2009) Maximizing influence propagation in networks with community
structure. Phys. Rev. E 79:056102.
[9] He JL, Fu Y, Chen DB (2015) A novel top-k strategy for influence maximization in complex networks
with community structure. PloS one 10(12):e0145283.
[10] Morone F, Makse HA (2015) Influence maximization in complex networks through optimal percolation.
Nature 524(7563):65 -- 68.
[11] Michalski R, Kajdanowicz T, Bródka P, Kazienko P (2014) Seed selection for spread of influence in
social networks: Temporal vs. static approach. New Generation Computing 32(3-4):213 -- 235.
[12] Jankowski J, Michalski R, Kazienko P (2013) Compensatory seeding in networks with varying avail-
ability of nodes in Advances in Social Networks Analysis and Mining (ASONAM), 2013 IEEE/ACM
International Conference on. (IEEE), pp. 1242 -- 1249.
[13] Michalski R, Kazienko P, Jankowski J (2013) Convince a dozen more and succeed -- the influence in
multi-layered social networks in Signal-Image Technology & Internet-Based Systems (SITIS), 2013 In-
ternational Conference on. (IEEE), pp. 499 -- 505.
[14] Hinz O, Skiera B, Barrot C, Becker JU (2011) Seeding strategies for viral marketing: An empirical
comparison. Journal of Marketing 75(6):55 -- 71.
[15] Libai B, Muller E, Peres R (2013) Decomposing the value of word-of-mouth seeding programs: Accel-
eration versus expansion. Journal of Marketing Research 50(2):161 -- 176.
[16] Seeman L, Singer Y (2013) Adaptive seeding in social networks in Foundations of Computer Science
(FOCS), 2013 IEEE 54th Annual Symposium on. (IEEE), pp. 459 -- 468.
[17] Horel T, Singer Y (2015) Scalable methods for adaptively seeding a social network in Proceedings of
the 24th International Conference on World Wide Web. (International World Wide Web Conferences
Steering Committee), pp. 441 -- 451.
[18] Sela A, Ben-Gal I, Pentland A, Shmueli E (2015) Improving information spread through a scheduled
seeding approach in The international conference on Advances in Social Network Analysis and Mining
2015.
[19] Zhang H, Procaccia AD, Vorobeychik Y (2015) Dynamic influence maximization under increasing re-
turns to scale in Proceedings of the 2015 International Conference on Autonomous Agents and Multiagent
Systems. (International Foundation for Autonomous Agents and Multiagent Systems), pp. 949 -- 957.
[20] Bulut E, Wang Z, Szymanski B (2010) Cost-effective multiperiod spraying for routing in delay-tolerant
networks. Networking, IEEE/ACM Transactions on 18(5):1530 -- 1543.
[21] Jankowski J, Bródka P, Kazienko P, Szymanski BK, Michalski R, Kajdanowicz T (2017) Balancing
speed and coverage by sequential seeding in complex networks. Scientific Reports 7(1):891.
[22] Jankowski J (2017) Dynamic rankings for seed selection in complex networks: Balancing costs and
coverage. Entropy 19(4):170.
12
[23] Jankowski J, Bródka P, Michalski R, Kazienko P (2017) Seeds buffering for information spreading
processes in International Conference on Social Informatics. (Springer), pp. 628 -- 641.
[24] Wang C, Chen W, Wang Y (2012) Scalable influence maximization for independent cascade model in
large-scale social networks. Data Mining and Knowledge Discovery 25(3):545 -- 576.
[25] Read KE (1954) Cultures of the central highlands, new guinea. Southwestern Journal of Anthropology
pp. 1 -- 43.
[26] Kempe D, Kleinberg J, Tardos É (2015) Maximizing the spread of influence through a social network.
Theory of Computing 11(4):105 -- 147.
[27] Newman ME (2001) Scientific collaboration networks. i. network construction and fundamental results.
Physical review E 64(1):016131.
[28] Opsahl T, Panzarasa P (2009) Clustering in weighted networks. Social networks 31(2):155 -- 163.
[29] Newman ME (2001) The structure of scientific collaboration networks. Proceedings of the National
Academy of Sciences 98(2):404 -- 409.
[30] Adamic LA, Glance N (2005) The political blogosphere and the 2004 us election: divided they blog in
Proceedings of the 3rd international workshop on Link discovery. (ACM), pp. 36 -- 43.
[31] Leskovec J, Mcauley JJ (2012) Learning to discover social circles in ego networks in Advances in neural
information processing systems. pp. 539 -- 547.
[32] Leskovec J, Huttenlocher D, Kleinberg J (2010) Predicting positive and negative links in online social
networks in Proceedings of the 19th international conference on World wide web. (ACM), pp. 641 -- 650.
13
Acknowledgment
This work was partially supported by the National Science Centre, Poland, grant no. 2017/27/B/HS4/01216
(JJ), 2016/21/B/ST6/01463 (PK), 2015/17/D/ST6/04046 (RM), and 2016/21/D/ST6/02408 (PB), the Eu-
ropean Union's Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant
agreement No. 691152 (RENOIR); the Polish Ministry of Science and Higher Education fund for supporting
internationally co-financed projects in 2016 -- 2019 (agreement no. 3628/H2020/2016/2); the Army Research
Laboratory under Cooperative Agreement Number W911NF-09-2-0053 (the ARL Network Science CTA);
and the Office of Naval Research Grant No. N00014-15-1-2640.
Author contributions
Author contributions: All authors contributed to the study design and manuscript preparation, BKS con-
tributed proof of sequential seeding properties and the concept of coordinated execution, both JJ and RM
conducted simulations and analyzed the data, PB prepared illustrative simulation cases.
Competing interests
Authors declare no conflict of interest.
Supplementary Information
Experimental results for undirected networks
Each from 162 combinations from configuration space N × P P × SP × R was applied as a simulation
configuration to single stage SN and sequential seeding SQ with the use of coordinated execution. Repeated
simulations resulted in 16,200,000 simulation cases. Results from individual simulation cases are presented in
Fig. 5 (A). They show that CSQ performs better than CSN for most cases with coverage CSQ ∈ [30%, 85%].
Fig. 5 (B) shows coverage for sequential approaches CSQ compared to single stage methods CSN. The
maximum coverage CM ax for optimal seed set is presented as well as the upper bound CGreedySN ∗ e/(e − 1)
for all cases. The results show that sequential seeding based on simple degree-based ranking in 92.2% of cases
outperforms single stage computationally expensive greedy heuristic, see Fig. 5 (B1). Moreover, sequential
seeding based on degree CdegreeSQ is greater than sequential greedy approach CgreedySQ in 62.6% of cases.
For all simulation cases CSQ shows better results than CSN in 96.7% of cases what as presented in Fig. 5
(C). Increase above 5% was achieved in 20.2% of simulation cases. The obtained coverage was dependent on
seed selection strategy with 96.0% better results for random selection, 100% for degree based selection and
93.9% for greedy based selection, see Fig. 5 (C1).
Fig. 5 (D) shows one of real diffusion cases with visible differences between sequential (SQ) and single
stage seeding (SN). As a result of better usage of natural diffusion processes sequential approach outperforms
the single stage approach.
Apart from user rankings the results were dependent on the network characteristics with visible differences
seen in Fig. 5 (E). Network N6 delivered the highest gain (83%) while the least gain (30%) was achieved for
network N3.
14
Figure 5: (A) Coverage for sequential approach CSQ compared with single stage CSN and maximum coverage
CM ax for all configurations and all simulation cases; ordered by CSN; (B) Coverage as percentage of all
activated network nodes for greedy and degree-based single stage and sequential strategies and their relation
to CM ax and the upper bound; ordered by CSN; (B1) Performance of sequential and single stage approaches
with degree-based selection in relation to single stage greedy ranking;(C) Performance of sequential seeding
SQ in the relation to single stage seeding SN, a vertical tick denotes percentage of cases with no gain:
CSQ = CSN; (C1) Performance of random, degree and greedy based sequential seeding compared with
single stage seeding SN; (D) Steps of one sequential and one single stage diffusion process in the coordinated
execution for one configuration; (E) Coverage values for sequential CSQ and single stage seeding CSN as
well as maximum coverage CM ax for networks N1-N6 and all simulation cases.
Experimental results for directed network
Performance of sequential seeding within directed network was analyzed for random node selection, degree
based selection and greedy approach. Detailed results for all used probabilities and strategies for single stage
and sequential seeding are presented in Table 3. The highest increase of coverage was observed for sequential
degree based selection with results better than greedy selection for both single and sequential seeding. The
highest gain for degree based selection was observed for low propagation probabilities 0.05, 0.01, 0.15. For
higher probabilities differences between strategies are smaller.
15
Coverage for random seed selection
PP Single stage % of Max Sequential
0.05
0.1
0.15
0.2
0.25
5.43
7.43
9.65
11.76
13.34
Coverage for degree based seed selection
62.2%
61.5%
67.4%
75.9%
83.9%
5.67
8.17
11.03
13.56
15.07
PP Single stage % of Max Sequential
0.05
0.1
0.15
0.2
0.25
64.0%
62.4%
66.7%
73.9%
81.5%
5.99
8.60
11.39
13.75
15.14
5.59
7.54
9.56
11.45
12.96
Coverage for greedy seed selection
PP Single stage % of Max Sequential
0.05
0.1
0.15
0.2
0.25
65.6%
65.2%
70.8%
78.9%
86.3%
5.88
8.47
11.08
13.59
15.09
5.73
7.88
10.14
12.22
13.72
Increase Gain
7.4%
16.0%
29.5%
48.3%
68.0%
1.04
1.09
1.14
1.15
1.13
Increase Gain
12.5%
23.4%
38.5%
57.0%
74.5%
1.07
1.15
1.21
1.22
1.19
Increase Gain
5.0%
14.2%
22.4%
42.0%
62.9%
1.02
1.08
1.09
1.12
1.11
Table 3: Coverage for single stage and sequential seeding for used seed selection methods and propagation
probabilities
16
No. Network name
1
2
3
4
5
6
7
- Condensed
collabora-
- Communi-
network
of
N1
Matter
tion
N2
cation
at University
California
N3 - High-Energy
Theory collabora-
tion network
N4 - Political blogs
N5 - ego-Facebook
N6 - wiki-Vote
N7 - Social network
of tribes
Ref Nodes Edges Edge type Components CC Diameter
[27]
47,594 Undirected
16,264
726
0.638
18
[28]
1,899
20,296 Undirected
4
0.109
8
[29]
7,610
15,751 Undirected
581
0.486
19
[30]
[31]
[32]
[25]
1,224
4,039
7,115
16
19,090 Undirected
88,234 Undirected
103,689 Undirected
Directed
114
2
1
24
1
0.320
0.606
0.141
0.519
8
8
7
3
Table 4: Description of networks used in the experiments
Figure 6: Degree distribution for networks N1-N6
17
|
1306.4534 | 1 | 1306 | 2013-06-19T13:22:11 | Exploiting Cellular Data for Disease Containment and Information Campaigns Strategies in Country-Wide Epidemics | [
"cs.SI",
"physics.soc-ph"
] | Human mobility is one of the key factors at the basis of the spreading of diseases in a population. Containment strategies are usually devised on movement scenarios based on coarse-grained assumptions. Mobility phone data provide a unique opportunity for building models and defining strategies based on very precise information about the movement of people in a region or in a country. Another very important aspect is the underlying social structure of a population, which might play a fundamental role in devising information campaigns to promote vaccination and preventive measures, especially in countries with a strong family (or tribal) structure.
In this paper we analyze a large-scale dataset describing the mobility and the call patterns of a large number of individuals in Ivory Coast. We present a model that describes how diseases spread across the country by exploiting mobility patterns of people extracted from the available data. Then, we simulate several epidemics scenarios and we evaluate mechanisms to contain the epidemic spreading of diseases, based on the information about people mobility and social ties, also gathered from the phone call data. More specifically, we find that restricting mobility does not delay the occurrence of an endemic state and that an information campaign based on one-to-one phone conversations among members of social groups might be an effective countermeasure. | cs.SI | cs |
Exploiting Cellular Data for Disease Containment and
Information Campaigns Strategies in Country-Wide Epidemics∗
A. Lima, M. De Domenico, V. Pejovic, and M. Musolesi
School of Computer Science, University of Birmingham, United Kingdom
Human mobility is one of the key factors at the basis of the spreading of diseases in a popula-
tion. Containment strategies are usually devised on movement scenarios based on coarse-grained
assumptions. Mobility phone data provide a unique opportunity for building models and defining
strategies based on very precise information about the movement of people in a region or in a coun-
try. Another very important aspect is the underlying social structure of a population, which might
play a fundamental role in devising information campaigns to promote vaccination and preventive
measures, especially in countries with a strong family (or tribal) structure.
In this paper we analyze a large-scale dataset describing the mobility and the call patterns of
a large number of individuals in Ivory Coast. We present a model that describes how diseases
spread across the country by exploiting mobility patterns of people extracted from the available
data. Then, we simulate several epidemics scenarios and we evaluate mechanisms to contain the
epidemic spreading of diseases, based on the information about people mobility and social ties, also
gathered from the phone call data. More specifically, we find that restricting mobility does not delay
the occurrence of an endemic state and that an information campaign based on one-to-one phone
conversations among members of social groups might be an effective countermeasure.
I.
INTRODUCTION
Health and well-being of populations are heavily in-
fluenced by their behaviour. The impact of the habits
and local customs, including patterns of interactions and
mobility at urban and regional scales, on health issues
is remarkable [1]. The diffusion of mobile technology
we are experiencing nowadays gives scholars an unprece-
dented opportunity to study massive data that describe
human behavior [2]. An increasing number of people
carries smart mobile phones, equipped with many sen-
sors and connected to the Internet, for the whole day [3].
Data coming from a large number of people can describe
trends in the macroscopic behavior of populations [4 -- 6].
The results of the analysis of these trends can be directly
applied to a number of real-world scenarios, and, more
in general, to several applications where cultural and lo-
cal differences play a central role. Analyzing this kind of
data can provide invaluable help to support the decision-
making process, especially in critical situations. For this
reason, many public and private organizations are nowa-
days increasingly adopting a data-centric approach in
their decisional process [7]. We believe that this strat-
egy can be particularly useful in developing countries,
which might have a lacking infrastructure1.
∗Appeared in Proceedings of NetMob 2013. Boston, MA, USA.
May 2013.
1 We use the term "developing" to indicate countries that are as-
signed a low Human Development Index (HDI) by United Na-
tions Statistics Division. We are aware of the limitations of this
classification. As reported by UN, the designations "developed"
and "developing" are intended for statistical convenience and do
not necessarily express a judgment about the stage reached by a
particular country or area in the development process [8].
Among the issues that developing countries are facing
today, healthcare is probably the most urgent [9].
In
these countries the effectiveness of campaigns is often re-
duced due to low availability of data, inherent limits in
the infrastructure and difficult communication with the
citizens, who might live in vast and remote rural areas.
As a result, action plans are difficult to deliver. How-
ever, we believe that a data-centric approach can be an
innovative and effective way to address these issues.
In this paper, we focus on containment of epidemics.
We use movement data extracted from the registration
patterns in a cellular network to evaluate the influence
of human mobility on the spreading of diseases in a ge-
ographic area.
In particular, we utilize this model to
investigate how infectious agents might spread to dis-
tant locations because of human movement in order to
identify optimal strategies that can be adopted to con-
trast the epidemics. We also evaluate how the collabo-
rative effort of the population can be crucial in critical
scenarios. For the reasons we mentioned before, in coun-
tries that are facing development challenges, vaccination
campaigns are often hard to advertise to the population.
Lack of communication and information is believed to be
among the main causes of failure for immunization cam-
paigns. The same applies to awareness campaigns that
try to promote prophylaxis procedures that reduce the
occurrence of contagion. However, in these cases, we ar-
gue that a collaborative effort leveraging individual social
ties can be effective in propagating effective information
(i.e., a sort of "immunizing information") to a widespread
audience. Moreover, information received by people who
are socially close can have a higher chance of leading to
an actual action.
A large body of research has been conducted on models
that describe the diffusion of diseases, with a particular
recent interest on the role that human movement plays
in spreading infections in large geographic areas [10 -- 12],
and also on the impact of human behavior on the spread-
ing itself [12, 13]. With respect to the state of the art,
the main contributions of this paper can be summarized
as follows:
• We propose an epidemic model based on a net-
work of geographic metapopulations, which de-
scribes how people move between different geo-
graphic regions and spread the disease.
• We evaluate containment techniques based on the
restriction of mobility of the most central areas.
The centrality of the areas is extracted by building
a movement network between all the geographic ar-
eas based on the mobility patterns of the individu-
als.
• We extend the model with a competing informa-
tion spreading where distance contagion might take
place. In other words, we study the dynamics of the
system considering three characterizing aspects of
the problem: the disease epidemics, human mobil-
ity and information spreading. This epidemics rep-
resents the diffusion of information related to mea-
sures to prevent or to combat the diseases, such
as information about the ongoing vaccination and
prevention campaigns in a certain area or actions
that will help to limit spread of the infection, such
as boiling water or avoiding contacts with people
that are already ill.
• We evaluate the models by using the data provided
by the Orange "Data for Development" [14]. We
discuss the effectiveness of the containment strate-
gies and, in particular, for the information dissem-
ination strategy, we identify the degree of partici-
pation that is required to make it successful.
• We observe that restricting mobility by disallowing
any movement from and to a limited set of subpre-
fectures does not delay the occurrence of the en-
demic state in the rest of the country. We also find
that a collaborative effort of prevention information
spreading can be an effective countermeasure.
2
(a)
(b)
Figure 1: Logarithmic representation of the calls matrix (a)
and the mobility matrix (b). Null values are indicated using
the white color. For both matrices highest values are mostly
concentrated along the diagonal, showing that communication
and movement between sub-prefectures is highly uncommon.
However, the calls matrix is visibly denser than the mobility
matrix, confirming that phone contacts between different sub-
prefectures are more usual than movement.
II. OVERVIEW OF THE DATASET
The data provided for the D4D challenge [14] consist
of four datasets (identified by the labels SET1, SET2,
SET3, SET4), containing information about user mobil-
ity and call patterns at various levels of granularity and
time duration. We will now discuss how these datasets
can be used to build a model which accounts for user
mobility and information spreading.
Two datasets contain information about mobility and
communication patterns at macroscopic level. More pre-
cisely:
• The SET1 dataset contains the number and the
duration of calls between pairs of cell phone tow-
ers, aggregated by hour. This dataset provides
macroscopic information about communication in
the country. We associate cell phone towers with
the sub-prefecture they are located in, by using the
supplied geographic position. Then, we evaluate
the probability of a call being established between
sub-prefectures i and j with:
cij = Cij(cid:80)
k Cik
,
(1)
This paper is organized as follows. In Sec. II we briefly
describe the four different datasets provided by Orange
and we specify how they are used in the present study.
In Sec. III we introduce our two models for epidemics
and information spreading by taking into account hu-
man mobility and call patterns observed in Ivory Coast.
In Sec. IV we present the results obtained by simulating
several epidemics scenarios and evaluating mechanisms
to contain the epidemic spreading of diseases. Finally, in
Sec. V we summarize our main findings and we propose
how the present study can be improved if more detailed
data about mobility and calls will be available.
where Cij is the number of phone calls initiated
from the sub-prefecture i and directed to the sub-
prefecture j, during the entire period of observa-
tion. The term at denominator indicates the to-
tal communication flux between every pair of sub-
prefectures and it is used to normalize the prob-
ability. Using these values we build a calls ma-
trix C, shown in Fig. 1(a). This matrix also shows
high values along the diagonal, but it is distinctly
denser, showing that calls between sub-prefectures
are more common than movement. The vertical
line at x = 60 identifies calls directed to the sub-
prefecture that contains the capital.
050100150200250050100150200250CallsmatrixC10−610−510−410−310−210−1100050100150200250050100150200250MobilitymatrixM10−810−710−610−510−410−310−210−1100 3
(a)
(b)
Figure 2: Geographic network obtained from call logs (a) and mobility traces (b). Color is used to indicate the community
structure: nodes within the same community are represented with the same color.
• The SET3 dataset contains the trajectories of
50,000 randomly-selected individuals, at a sub-
prefecture level resolution, for five months.2 This
dataset can be used to estimate the probability that
an individual moves from the sub-prefecture i to the
sub-prefecture j:
(cid:80)
(cid:80)
(cid:80)
u Mu
u Mu
ij
k
ik
mij =
,
(2)
where Mu
ij is the number of times user u moves
from the sub-prefecture i to j. The numerator
counts how many times users who are in i move
to j; the denominator normalizes this number by
the total number of transitions from i to any sub-
prefecture k. Using these values we build a mobility
matrix M , shown in Fig. 1(b). By using this ma-
trix, we model human mobility in the country as a
Markov process [15]. We observe that the matrix
is quite sparse and the highest values are concen-
trated along the diagonal. As the representation
is in logarithmic scale, this demonstrates that the
movement between sub-prefectures is present, but
rather uncommon.
In Fig. 2(a) and Fig. 2(b) we show the geographic net-
works of calls and mobility, respectively. Nodes are
positioned using the geographic locations of the sub-
prefecture they represent, and their color indicates the
community structure of the network based on [16].
The other two datasets provide microscopic informa-
tion about mobility and communication patterns between
individuals. Although we do not use them for the anal-
ysis in this paper, we now briefly outline how they could
be used:
• The SET2 dataset contains fine-grained individual
trajectories of 50,000 randomly sampled individu-
als over two-week periods. This dataset could be
used to estimate the number of potential connec-
tions that an individual might have in a certain
area, served by a cell phone tower.
• The SET4 dataset contains time-varying ego-
networks of 5,000 users, describing the network of
If two
communication in time-slots of 2 weeks.
users are connected by a link in a time-slot,
it
means that at least one call occurred during the
two weeks under consideration3. The ego-network
2 17 sub-prefectures do not have any cell phone towers and for this
reason do not appear in SET3. We discard these sub-prefectures
from our analysis, since their users will be considered as belong-
ing to nearby sub-prefectures.
3 We have found that 1.31% of the total number of edges in ego-
networks connect pairs of users who are neither egos nor first-
level neighbors: therefore, we do not consider such edges in our
analysis.
respectively. At each time t a person is either infected or
susceptible, therefore Ni[t] = Ii[t] + Si[t].
4
through
Simultaneously,
individuals move
the
metapopulation network according to the mobility
matrix M of dimension n× n extracted from the cellular
traces. The generic element mij of the matrix represents
the probability that a person moves from the metapop-
ulation i to j, as described by Eq. 24. This matrix
describes how the state variables Ni[t] evolve over time:
j=1 mjiNi[t]. Under the assumption that
individuals inside the classes I and S move consistently
we can write the last relation also for the state variables
Ii[t] and Si[t]5. The contagion-mobility combined system
can then be described by the following set of equations:
Ni[t + 1] =(cid:80)n
(cid:20)
(cid:20)
n(cid:88)
n(cid:88)
j=1
j=1
Ii[t + 1] =
Si[t + 1] =
(cid:21)
(cid:21)
mji
Ij[t] + λ
Sj[t]
Nj[t]
Ij[t] − γIj[t]
mji
Sj[t] − λ
Sj[t]
Nj[t]
Ij[t] + γIj[t]
,
for each sub-prefecture i = 1, 2, . . . , n, with λ being
the product of contact rate and contagion probability
and γ being the recovery rate. The formulae inside the
square brackets describe the evolution of n SIS models,
one for each metapopulation. They are multiplied for
the elements of the mobility matrix, which accounts for
individuals moving between metapopulations.
This analytical model describes the expected outcome
of a stochastic model where the following actions occur
at each time step:
1. Each infected person in the sub-prefecture j causes
individuals inside j. This
the infection of new λ Sj
Nj
step is repeated for each sub-prefecture.
2. A new position i is assigned to each individual in
the sub-prefecture j according to the probability
density function [mj1, mj2, . . . , mjn]. This step is
repeated for each sub-prefecture.
B.
Information Spreading
The model we presented in the last section tries to re-
produce the spreading of a disease in a population where
4 In general, this matrix can be time-varying, and it can be ad-
justed according to seasonal trends or real-time data at each step,
for example following estimates based on historical data. In par-
ticular, this matrix can be used to study the impact of policies
in real-time. However, in order to simplify the presentation, we
use a matrix not changing over time. The treatment can be gen-
eralized, also applying the recent theoretical results related to
time-varying networks [18, 19].
5 This assumption can also be relaxed when data about the dif-
ferent classes of individuals is available, i.e., when a matrix for
each class can be defined.
Figure 3: Distribution of friends for the ego-networks ag-
gregated over time. Error bars indicate statistical uncer-
tainty, while the shaded area represent 99% confidence in-
tervals around the observed value.
aggregated over the whole observation time, built
considering every link that is present at least once,
describes the number of people contacted by an
individual during the entire period. This dataset
could be used to estimate the number of potential
social connections that an individual might get in
touch with. The degree distribution of the aggre-
gate ego-network is shown in Fig. 3.
III. SPREADING MODELS
In this section we discuss two models: a model of dis-
ease spreading as a function of the mobility patterns of
individuals between different geographic areas inferred
from the cellular registration records and a model for in-
formation spreading among the same population, consid-
ering the social structure inferred from the call records.
In the following section, we will evaluate the models using
the data provided for the Orange Data for Development
challenge.
A. Epidemic Spreading and Mobility
We will now present a model that represents the evolu-
tion of an epidemic taking place on a network of metapop-
ulations. The aim of the model is to describe how the sys-
tem evolves under the action of two processes, contagion
and mobility. For this dataset, each metapopulation is
composed by the individuals located in a particular sub-
prefecture. Hence, the population is distributed in n dif-
ferent metapopulations, each having Ni[t] individuals at
time t. We make the simplifying assumption that there
are no deaths and births in the considered time window,
i.e., at each time t = 1, 2, . . . T the total population is
constant(cid:80)n
i=1 Ni[t] = N .
We assume that contagion happens
inside each
metapopulation following a standard SIS model [17]. We
indicate the number of infected and susceptible individ-
uals at time t in a sub-prefecture i with Ii[t] and Si[t],
R
ω
ξ
start
S
γ
λ
I
start
U
A
ψ
Figure 4: State machines describing the state transitions of
a person with respect to the disease contagion (R=Resistant,
S=Susceptible and I=Infected) and with respect to the in-
formation spreading (U=unaware, A=aware), respectively. A
person starts in the susceptible and unaware states. We as-
sume that aware individuals spread the information and can-
not go back to the unaware state.
Ri[t + 1] =
Ai[t + 1] =
individuals change locations over time. The aim of this
work is to analyze some scenarios and study the effec-
tiveness of some containment techniques. In particular,
as anticipated, we would like to investigate if a collabora-
tive effort of the population is able, in theory, to reduce
considerably the spread of the disease and what propor-
tions should it have to be effective. More precisely, the
population can disseminate information through personal
social ties immunizing, such as information about pre-
vention techniques, hygiene practises, advertisement of
nearby vaccination campaigns and in general any infor-
mation that can lead to a reduction of the number of
contagion events.
In order to take into consideration these aspects, we
now use a SIR model for each metapopulation, so that
each person either belongs to the susceptible (S), in-
fected (I) or resistant (R) category. At the same time,
another simultaneous epidemic happens on the network
of metapopulations, disseminating information that can
make individuals resistant to the disease.
In fact, a
person also belongs to the category of unaware (U) or
aware (A) individuals, with respect to the immuniz-
ing information. More formally, we have that Ni[t] =
Ii[t] + Si[t] + Ri[t] = Ai[t] + Ui[t].
It is worth noting that this "immunizing epidemic"
goes beyond the boundaries of metapopulations (sub-
prefectures):
in other words, it is a distance contagion.
It is also important to remark that the states "aware"
and "resistant" are substantially different. An unaware
person that receives the information (i.e. has an "infor-
mation contact") becomes aware with rate ψ; since the
person is aware, he or she will start spreading the in-
formation as well. An infected person that receives the
information becomes immune with rate ω. Additionally,
individuals who have acquired immunity through infor-
mation can lose it with rate ξ. The transition rates be-
tween states are summarized in Fig. 4. The model can
be described by the following set of equations, specifying
how state vectors evolve over time:
(cid:20)
(cid:20)
Ij[t] + λ
mji
mji
Sj[t] − λ
n(cid:88)
n(cid:88)
j=1
j=1
Ii[t + 1] =
Si[t + 1] =
5
(cid:21)
Ij[t] − γIj[t]
Sj[t]
Nj[t]
Ij[t] + γIj[t] + ξRj[t]+
(cid:21)
Rj[t] − ξRj[t] + ωSj[t]
k=1 ckjAk[t]
k=1 ckjNk[t]
Aj[t] + ψUj[t]
(cid:20)
(cid:20)
(cid:20)
mji
mji
j=1
n(cid:88)
n(cid:88)
n(cid:88)
j=1
Sj[t]
Nj[t]
− ωSj[t]
(cid:80)n
(cid:80)n
(cid:80)n
(cid:80)n
(cid:80)n
(cid:80)n
k=1 ckjAk[t]
k=1 ckjNk[t]
(cid:80)n
(cid:80)n
k=1 ckjAk[t]
k=1 ckjNk[t]
k=1 ckjAk[t]
k=1 ckjNk[t]
(cid:21)
(cid:21)
(cid:21)
(3)
Ui[t + 1] =
mji
Uj[t] − ψUj[t]
j=1
for every i = 1, 2, . . . , n. The fraction
(cid:80)n
(cid:80)n
k=1 ckj Ak[t]
k=1 ckj Nk[t] rep-
resents the probability that a call from an aware person
occurs in the metapopulation j. It models the distance-
contagion, and it is possible to verify that if the matrix
is identical (absence of contacts between populations) it
reduces to Ak[t]/Nk[t], falling back to a model where con-
tagion occurs only inside metapopulations.
This analytical model describes the expected value of
a stochastic model where the following actions occur at
each time step t:
1. Each infected person in the sub-prefecture j causes
new individuals to get infected, inside j. This
λ Sj
Nj
step is repeated for each sub-prefecture.
(cid:80)n
(cid:80)n
2. Each unaware person in the sub-prefecture j be-
k=1 ckj Ak[t]
k=1 ckj Nk[t] . This
comes aware with probability ψ
step is repeated for each sub-prefecture.
3. Each person in the sub-prefecture j who is
(cid:80)n
(cid:80)n
susceptible, becomes resistant with probability
k=1 ckj Ak[t]
ω
k=1 ckj Nk[t] . This step is reapeated for each sub-
prefecture.
4. A new position i is assigned to each person in the
sub-prefecture j according to the probability den-
sity function [mj1, mj2, . . . , mjn]. This step is re-
peated for each sub-prefecture.
IV. ANALYSIS
We initialize each scenario by allocating 22 mil-
lion individuals (the estimated population size of Ivory
Coast for July 2012 is 21,952,093 [20]) to different sub-
prefectures across the country, according to the data in
SET3. In each scenario we bootstrap the spreading pro-
cess by infecting a fraction of the population (0.1%) dis-
tributed across metapopulations according to different
criteria:
6
Figure 5: Fraction of infected population at the stationary
state (left panel) and time required to reach the stationary
state (right panel), for different values of r0 and for different
initial conditions. Missing values in the curves mean that,
for the corresponding values, no stationary state is reached
during the period of observation.
Figure 6: Fraction of infected population at the stationary
state (left panel) and time required to reach the stationary
state (right panel), for different values of r0 when the epi-
demic starts from a random sub-prefecture, and different lev-
els of geographic quarantine are applied. Missing values in the
curves mean that, for the corresponding values, no stationary
state is reached during the period of observation.
• Uniform distribution: every sub-prefecture gets a
number of infected proportional to their popula-
tion, i.e., every sub-prefecture has the same fraction
of infected population.
• Random: a single sub-prefecture, chosen randomly,
is the origin of the infection.
• Centrality based: the sub-prefectures are ordered
by decreasing centrality values, then the first 1, 5
or 10 highest ranked sub-prefectures are chosen, as
shown in Table I.
We study the evolution of the epidemics for a period
of 6 months. We investigate multiple scenarios using the
analytical model considering a large set of ranges for the
key parameters. We conducted a series of Monte-Carlo
simulations for multiple sets of parameters, confirming
the validity of the analytical models presented in the pre-
vious section. In the following, we present results based
on these models.
A. No Countermeasures
We will firstly explore the evolution of the epidemics
in the case where no countermeasures are taken. In order
Betweeness Closeness Degree Eigenvalue
60
39
89
58
75
144
138
165
212
168
60
58
39
69
138
250
64
144
182
122
60
58
39
69
138
64
144
250
122
182
60
58
39
69
250
138
64
144
122
182
Table I: Highest ranked sub-prefectures, according to different
definitions of centrality. We observe that the sets of the top
10 sub-prefectures ordered by centrality are very similar.
to analyze the evolution of the system more clearly, we
investigate two measures: the fraction of infected popu-
lation i∞ at the stationary state and the time required to
reach the stationary state τ . In Fig. 5 we plot their val-
ues versus r0 = λ
γ , which is the basic reproductive ratio
of a classic SIS model [17]. As a future work, we plan to
derive the analytical form of the basic reproductive ra-
tio of our models, which take into account mobility and
information spreading. We observe that for r0 = λ
γ < 1
there is no endemic state (i.e., the final fraction of in-
fected population is zero), whereas for r0 > 1 a non-null
fraction of population is infected. Values for r0 = 1 are
missing since no stationary state is reached within our
observation window. In other words, for this particular
scenario, experimental results show that the basic repro-
ductive ratio of our model is very close to r0; we expect
this to be a consequence of the low inter-subprefectures
mobility. We can also notice that the initial conditions
do not affect i∞ at all. Before the critical point (i.e.,
r0 = 1) the choice of the initial conditions has also no
impact on the delay time, whereas for r0 > 1 it slightly
affects the delay: epidemics that initially involves more
sub-prefectures are slightly faster than the others.
B. Geographic Quarantine
We now analyze the effects of curbing on the mobility
between sub-prefectures, i.e., forbidding all the incoming
and outgoing movement of a group of sub-prefectures.
In order to do so, we calculate the centrality values of
each sub-prefecture in the mobility matrix. We present
the results for eigenvalues centrality. As it is possible to
observe in Tab. I, the ranking based on other centralities
is very similar. Then, for the quarantine operations, we
select those with the highest centrality values. From a
practical point of view, this is achieved by simply chang-
ing the i−th row and column in the mobility matrix, so
that all the elements mij and mji are null, except for
the elements mii = 1. For these scenarios, we randomly
012345r00.00.10.20.30.40.50.60.70.80.9i∞uniformtop10top5top1random0.00.51.01.52.02.5r00100200300400500600τuniformtop10top5top1random012345r00.00.10.20.30.40.50.60.70.80.9i∞notcurbedcurbedtop10curbedtop5curbedtop10.00.51.01.52.02.5r00100200300400500600τnotcurbedcurbedtop10curbedtop5curbedtop1 7
Figure 7: Fraction of infected population at the stationary state (top row) and time required to reach the stationary state
(bottom row), for different values of r0 = λ
γ (from left to right 1, 2, 4, 10, respectively). White spaces show that no stationary
state is reached during the period of observation.
choose a single sub-prefecture where the initial individu-
als are infected, and then we average i∞ and τ over all
runs. As shown in Fig. 6, the fraction of the infected
population is sensibly affected by this measure, as the
population inside the quarantined areas is protected from
contagion. However, contrary to the intuition, the delay
is not affected by the quarantine, even when the coun-
termeasures involves 10 sub-prefectures, which account
for almost half population. This suggests that such an
invasive, expensive and hard to enforce measure reduces
considerably the endemic size, but does not slow down
the disease spreading in the rest of the country. For this
reason, we now investigate a radically different approach
to protect the population.
C.
Information Campaign (Social Immunization)
We now show how a collaborative information cam-
paign could help in contrasting the spread of the disease,
following the model we presented in the last section. We
initialize the scenario by distributing the immunizing in-
formation to 1% of the population, randomly chosen re-
gardless of their location. These people will be informed
and will be instructed to spread the information. In other
words, we assume that they will contact their social con-
nections, according to the call matrix.
In Fig. 7 we show the density plots describing i∞ and τ
for various values of r0 (, for a subset of scenarios where
ω = ψ, i.e., when the information that spreads among
the population has the same chance to immunize a per-
son and to involve the person in the spreading process.
This is consistent with a scenario where the same set of
people who become aware also become immunized by the
information they have received. Blank squares show that
a stationary state was not reached for the correspond-
ing set of parameters. The figure shows how contagious
(ω=ψ) the immunizing information has to be with re-
spect to how often people "forget" (ξ) in order to slow
down the disease considerably and to reduce the endemic
cases. When ω = ψ = 0 we fall back to the model with-
out information spreading, and the value of ξ does not
affect i∞ and τ . For ω = ψ > 0 and ξ = 0 the fraction
of infected population goes to zero in all cases, because
the number of people aware of the information does not
decrease, thus increasing the number of new immunized
individuals at each step. We can notice that even for low
values of participation ω and for information that gives
temporary immunization (ψ > 0), the final fraction of in-
fected individuals is considerably lower than in the case
where no countermeasures are taken.
In Figs. 8 and 9 we show the density plots for ω and ψ
when ξ is constant. In particular, we analyze the scenario
for ξ = 0 (Fig. 8), which represents for example a scenario
where the immunizing information is about vaccination
campaigns (individuals who have been administered vac-
cination do not lose immunity). For every combination of
parameters we have absence of endemic state even with
the highest considered value of r0. The two parameters
that represent how individuals are likely to get involved
both in the immunization and in the information spread-
ing (ω and ψ) seem to have the same impact on the delay
of the infection.
The value ξ = 0.5 (Fig. 9) describes the scenario when
the information is about a good practice (e.g., boiling
water, using mosquito nets, etc.), which loses its effec-
0.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψfractionofinfectedpopulation(λ=0.5,γ=0.5)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψfractionofinfectedpopulation(λ=0.8,γ=0.4)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψfractionofinfectedpopulation(λ=0.8,γ=0.2)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψfractionofinfectedpopulation(λ=1.0,γ=0.1)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψtimetoreachthestationarystate(λ=0.5,γ=0.5)0306090120150180210240270>3000.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψtimetoreachthestationarystate(λ=0.8,γ=0.4)0306090120150180210240270>3000.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψtimetoreachthestationarystate(λ=0.8,γ=0.2)0306090120150180210240270>3000.00.20.40.60.81.0ξ0.00.20.40.60.81.0ω=ψtimetoreachthestationarystate(λ=1.0,γ=0.1)0306090120150180210240270>300 8
Figure 8: Fraction of infected population at the stationary state (top row) and time required to reach the stationary state
(bottom row), for different combinations of λ
γ (from left to right 1, 2, 4, 10, respectively) and ξ = 0. White spaces show that
no stationary state is reached during the period of observation.
Figure 9: Fraction of infected population at the stationary state (top row) and time required to reach the stationary state
(bottom row), for different combinations of λ
γ (from left to right 1, 2, 4, 10, respectively) and ξ = 0.5. White spaces show that
no stationary state is reached during the period of observation.
tiveness or it is stopped being used by a person with rate
ξ. For this case we can notice that the fraction of in-
fected population is independent from ψ, as rows in the
density plot are of the same color. This suggests that,
for this scenario, the rate at which people lose immunity
does not affect the size of the endemic state.
V. CONCLUSIONS
In this paper we have presented a model that describes
the spreading of disease in a population where individuals
move between geographic areas, extracted from cellular
network records. We have showed the evolution of the
disease and we have evaluated two types of countermea-
sures, namely the quarantine of central geographic areas
and a collaborative "viral" information campaign among
0.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.5,γ=0.5)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.8,γ=0.4)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.8,γ=0.2)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=1.0,γ=0.1)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.5,γ=0.5)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.8,γ=0.4)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.8,γ=0.2)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=1.0,γ=0.1)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.5,γ=0.5)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.8,γ=0.4)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=0.8,γ=0.2)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωfractionofinfectedpopulation(λ=1.0,γ=0.1)0.00.10.20.30.40.50.60.70.80.91.00.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.5,γ=0.5)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.8,γ=0.4)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=0.8,γ=0.2)0306090120150180210240270>3000.00.20.40.60.81.0ψ0.00.20.40.60.81.0ωtimetoreachthestationarystate(λ=1.0,γ=0.1)0306090120150180210240270>300 9
the population, by inferring the underlying social struc-
ture from the call records.
Our future research agenda includes the investigation
of analytical aspects of the model, such as the derivation
of the critical reproductive ratio R0, i.e., the value that
corresponds to the transition between an endemic and an
endemic-free infection. Currently, the model is based on
the assumption of a static mobility matrix: our goal is
to refine the model by introducing time-dependent ma-
trices, also exploring the application of the recent the-
oretical results related to temporal networks. We also
plan to refine the model introducing specific contact rates
for each metapopulation, potentially based on more fine-
grained information about the number of encounters and
the number of calls of each individual, if available. Fi-
nally, we plan to explore hybrid countermeasures, such as
concurrent partial restrictions of mobility and targeted
information campaigns.
Acknowledgments
The authors thank Charlotte Sophie Mayer for use-
ful and fruitful discussions. This work was supported
through the EPSRC Grant "The Uncertainty of Iden-
tity: Linking Spatiotemporal Information Between Vir-
tual and Real Worlds" (EP/J005266/1).
[1] C.G. Helman et al., Culture, health and illness. (Arnold,
Hodder Headline Group, London, United Kingdom,
2001), No. Ed. 4.
[2] D. Lazer, A.S. Pentland, L. Adamic, S. Aral, A.L.
Barabasi, D. Brewer, N. Christakis, N. Contractor,
J. Fowler, M. Gutmann, et al., Life in the network: the
coming age of computational social science Science 323,
721 (2009).
[3] A. T. Campbell, S. B. Eisenman, N. D. Lane, E. Miluzzo,
R. Peterson, H. Lu, X. Zheng, M. Musolesi, K. Fodor,
and G.-S. Ahn, The Rise of People-Centric Sensing IEEE
Internet Computing Special Issue on Mesh Networks
(2008).
[4] 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 net-
works Proceedings of the National Academy of Sciences
104, 7332 (2007).
[5] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi, Un-
derstanding individual human mobility patterns Nature
453, 779 (2008).
[6] N. Eagle and A. Pentland, Eigenbehaviors: Identifying
structure in routine Behavioral Ecology and Sociobiology
63, 1057 (2009).
[7] J. Manyika, M. Chui, B. Brown, J. Bughin, R. Dobbs,
C. Roxburgh, and A. H. Byers, Big data: The next fron-
tier for innovation, competition, and productivity McK-
insey Global Institute (2011).
[8] United Nations Statistics Division, Standard Country
and Area Codes for Statistical Use.
[9] The health of the people the African regional health report
(World Health Organization, Regional Office for Africa,
Brazzaville, Republic of Congo, 2013).
[10] V. Colizza, A. Barrat, M. Barthelemy, A.-J. Valleron,
and A. Vespignani, Modeling the Worldwide Spread of
Pandemic Influenza: Baseline Case and Containment
Interventions PLoS Med 4, e13 (2007).
[11] J. M. Epstein, D. M. Goedecke, F. Yu, R. J. Morris, D. K.
Wagener, and G. V. Bobashev, in Controlling Pandemic
Flu: The Value of International Air Travel Restrictions,
PLoS ONE 2, e401 (2007).
[12] S. Meloni, N. Perra, A. Arenas, S. G´omez, Y. Moreno,
and A. Vespignani, Modeling human mobility responses to
the large-scale spreading of infectious diseases Scientific
Reports 1, (2011).
[13] S. Funk, M. Salath´e, and V. A. A. Jansen, Modelling the
influence of human behaviour on the spread of infectious
diseases: a review Journal of The Royal Society Interface
7, 1257 (2010).
[14] V.D. Blondel, M. Esch, C. Chan, F. Clerot, P. Deville,
E. Huens, F. Morlot, Z. Smoreda, and C. Ziemlicki, Data
for Development: the D4D Challenge on Mobile Phone
Data arXiv preprint arXiv:1210.0137 (2012).
[15] J. R. Norris, Markov Chains (Cambridge University
Press, Cambridge, United Kingdom, 1998).
[16] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and
E. Lefebvre, Fast unfolding of communities in large net-
works Journal of Statistical Mechanics: Theory and Ex-
periment 2008, P10008 (2008).
[17] M. J. Keeling and P. Rohani, Modeling infectious diseases
in humans and animals (Princeton University Press,
Princeton, NJ, 2011).
[18] J. Tang, S. Scellato, M. Musolesi, C. Mascolo, and
V. Latora, Small-world Behavior in Time-varying Graphs
Physical Review E 81, (2010), 055101(R).
[19] P. Holme and J. Saramaki, Temporal networks Physics
Reports 519, (2012).
[20] CIA, The World Factbook, 2012.
|
1309.3908 | 1 | 1309 | 2013-09-16T11:35:19 | Exploring Image Virality in Google Plus | [
"cs.SI",
"cs.CY",
"cs.MM",
"physics.soc-ph"
] | Reactions to posts in an online social network show different dynamics depending on several textual features of the corresponding content. Do similar dynamics exist when images are posted? Exploiting a novel dataset of posts, gathered from the most popular Google+ users, we try to give an answer to such a question. We describe several virality phenomena that emerge when taking into account visual characteristics of images (such as orientation, mean saturation, etc.). We also provide hypotheses and potential explanations for the dynamics behind them, and include cases for which common-sense expectations do not hold true in our experiments. | cs.SI | cs | Exploring Image Virality in Google Plus
Marco Guerini
Trento RISE
Trento - Italy
Jacopo Staiano
University of Trento
Trento - Italy
Email: [email protected]
Email: [email protected]
Davide Albanese
Fondazione Edmund Mach, CRI-CBC
San Michele allAdige (TN) - Italy
Email: [email protected]
3
1
0
2
p
e
S
6
1
]
I
S
.
s
c
[
1
v
8
0
9
3
.
9
0
3
1
:
v
i
X
r
a
Abstract—Reactions to posts in an online social network show
different dynamics depending on several textual features of the
corresponding content. Do similar dynamics exist when images
are posted? Exploiting a novel dataset of posts, gathered from
the most popular Google+ users, we try to give an answer to such
a question. We describe several virality phenomena that emerge
when taking into account visual characteristics of images (such as
orientation, mean saturation, etc.). We also provide hypotheses
and potential explanations for the dynamics behind them, and
include cases for which common-sense expectations do not hold
true in our experiments.
I.
INTRODUCTION
How do things become ‘viral’ on the Internet? And what
exactly do we mean by ‘influence’? Since marketing and
industry people want their messages to spread in the most
effective and efficient way possible,
these questions have
received a great deal of attention, particularly in recent years,
as we have seen a dramatic growth of social networking on
the Web. Generally speaking, virality refers to the tendency of
a content either to spread quickly within a community or to
receive a great deal of attention by it. In studying the spreading
process we will focus on the content and its characteristics,
rather than on the structure of the network through which
the information is moving. In particular, we will investigate
the relationships between visual characteristics – of images
enclosed in Google+ posts – and virality phenomena. We will
use three virality metrics: plusoners, replies and resharers.
This exploratory work stems from the use people make of
social networking websites such as Google+, Facebook and
similar: we hypothesized that perceptual characteristics of an
image could indeed affect the virality of the post embedding it,
and that – for example – cartoons, panorama or self-portraits
picture affect users’ reactions in different ways. The aim of
this paper is to investigate whether signs of such “common-
sense” intuition emerge from large-scale data made available
on popular social networking websites like Google+ and, in
such case, to open discussion on the associated phenomena.
The paper is structured as follows: first, we review previous
works addressing the topic of virality in social networks,
and particularly some focusing on content impact. Then, after
describing the dataset collected and used for this work, we
proceed with the study of virality of Google+ posts and the
characteristics of their content. We also discuss the behavior of
virality indexes in terms of their alternative use, arguing that
plusones and comments fulfill a similar purpose of followers’
“appreciation” while reshares have a different role of “self-
representation”. Finally, we investigate possible interactions
between image characteristics and users’ typology, in order to
understand to what extent results are generalizable or typical
of a community, gathered around a common interest.
II. RELATED WORKS
Several researchers studied information flow, community
building and similar processes using Social Networking sites
as a reference [1], [2], [3], [4]. However, the great majority
concentrates on network-related features without taking into
account the actual content spreading within the network [5].
A hybrid approach focusing on both product characteristics
and network related features is presented in [6]: the authors
study the effect of passive-broadcast and active-personalized
notifications embedded in an application to foster word of
mouth.
Recently, the correlation between content characteristics
and virality has begun to be investigated, especially with
regard to textual content; in [7], for example, features derived
from sentiment analysis of comments are used to predict the
popularity of stories. The work presented in [8] uses New York
Times’ articles to examine the relationship between emotions
evoked by the content and virality, using semi-automated
sentiment analysis to quantify the affectivity and emotionality
of each article. Results suggest a strong relationship between
affect and virality; still,
the virality metric considered is
interesting but very limited: it only consists of how many
people emailed the article. The relevant work in [9] measures
a different form of content spreading by analyzing which are
the features of a movie quote that make it “memorable” online.
Another approach to content virality, somehow complementary
to the previous one, is presented in [10], trying to understand
which modification dynamics make a meme spread from
one person to another (while movie quotes spread remaining
exactly the same).
More recently, some works tried to investigate how dif-
ferent textual contents give rise to different reactions in the
audience: the work presented in [11] correlates several viral
phenomena with the wording of a post, while [12] show that
specific content features variations (like the readability level
of an abstract) differentiate among virality level of downloads,
bookmarking, and citations. Still, to our knowledge, no attempt
has been made yet to investigate the relation between visual
content characteristics and virality.
III. DATA DESCRIPTION
the Google+ API1, we
Using
public
posts
from the
979
top
followed
users
harvested
the
in
1https://developers.google.com/+/api/
Google+ (plus.google.com),
the
socialstatistics.com website on March 2nd 20122.
The time span for the harvesting is one year, from June 28th
2011 (Google+ date of launch) to June 29th 2012.
reported
by
as
We decided to focus on the most popular users for several
reasons: (i) the dataset is uniform from the point of view of
sample role, i.e. VIPs, (ii) the behavior of the followers is
consistent – e.g. no friendship dynamics – and (iii) extraneous
effects due to followers network is minimized, since top
followed users’ network is vast enough to grant that, if a
content is viral, a certain amount of reactions will be obtained.
We defined 3 subsets of our dataset, comprising respec-
tively: (i) posts containing a static image, (ii) posts containing
an animated image (usually, gif), (iii) posts without attach-
ments (text-only). All other posts (containing as attachment
videos, photo albums, links to external sources) were dis-
carded. Statistics for our dataset are reported in Table I. For
each post, we considered three virality metrics3:
• Plusoners: the number of people who +1’d;
• Replies: the number of comments;
• Resharers: the number of people who reshared.
TABLE I.
AN OVERVIEW OF THE GOOGLE+ DATASET.
Global
actors
posts
published interval
979
289434
6/28/11–6/29/12
Posts with static images
actors
posts
min/max/median posts per actor
min/max/median plusoners per post
min/max/median replies per posta
min/max/median resharers per post
950
173860
1/3685/ 65.5
0/9703/33.0
0/571/12.0
0/6564/4.0
actors
posts:
min/max/median posts per actor
min/max/median plusoners per post
min/max/median replies per post
min/max/median resharers per post
Posts with animated images
344
12577
1/2262/3.0
0/5145/17.0
0/500/7.0
0/6778/10.0
Posts without attachments, text-only
actors
posts
min/max/median posts per actor
min/max/median plusoners per post
min/max/median replies per post
min/max/median resharers per post
aReplies count is cut around 500 by the API service.
939
102997
1/1744/41.0
0/20299 /16.0
0/538/17.0
0/13566/1.0
In Figures 1 and 2 we display the evolution over time of
the network underlying our dataset (using a week as temporal
unit), and of the reactions to posts given by users, respectively.
We notice that:
1) the average number of reactions per user shows quite
different trends depending on the metric considered: while
replies tend not to be affected by the growth of the
network, reshares and, to a lesser degree plusones, show
an ever-growing trend.
2) The temporal plot of the average number of followers per
user (Figure 1) in our dataset (in Google+ terminology,
2The dataset presented and used in this work will be made available to the
community for research purposes.
3Since the API provide only an aggregate number, we cannot make any
temporal analysis of how reactions to a post were accumulated over time.
the number of people who circled them) shows a gra-
dient increase around weeks 28/29. Interestingly, this is
reflected in the plot of reactions over time (Figure 2): the
gradient increases around the same weeks, for reshares
and plusones; these effects are most probably due to
Google+ transitioning from beta to public in late Septem-
ber 2011 (a similar phenomenon is reported also in [13]).
3) Finally, the orders of magnitude of such growths are very
different: we notice that while reactions increase of a
factor of 7 over the time period we took into account,
the total number of followers increased of a factor of 25.
600000
500000
400000
300000
200000
100000
s
r
e
w
o
l
l
o
f
f
o
r
e
b
m
u
n
.
g
v
A
0
0
10
20
30
Week
40
50
60
Fig. 1. Average number of followers per user, at 1-week temporal granularity.
Plusoners
Resharers
Replies
250
200
150
100
50
s
n
o
i
t
c
a
e
r
f
o
r
e
b
m
u
n
.
g
v
A
0
0
10
20
30
Week
40
50
60
Fig. 2. Average number of reactions per user, at 1-week temporal granularity.
This value represents the average number of reactions elicited by each user’s
posts over 1-week time-slices.
The relative amount of followers’ reactions does not sig-
nificantly increase as the network grows4. As detailed in the
next section, our analyses are based on comparing probability
distributions: e.g. we evaluate if grayscale images have a
significantly higher or smaller probability of reaching a certain
virality score than colored ones. In the following analyses,
for the sake of clarity, our discussion will not
take into
account the normalization factor (i.e. the size of the audience
when a content is posted). Indeed, we have run the same
analyses normalizing the virality indexes of a given post
against its potential audience: i) the effects are still visible,
ii) the effects are consistent both in significance and sign
with the not-normalized distributions, but iii) differences have
lower magnitude (explained by the fact that virality indexes
should be normalized using the actual audience – e.g. the
4It has been noted how (see, for instance, http://on.wsj.com/zjRr06), espe-
cially in the time frame we consider, users’ activity did not increase much in
front of the exploding network size.
followers exposed to the content). Thus, since we are interested
in comparing the virality of different image categories and
our preliminary experiments showed that by normalizing the
indexes their comparisons, their sign, and the derived inter-
pretations still hold, we choose to report the non-normalized
version of the results that are more intuitively readable.
In the following sections, after the analyses of text-only
posts and of posts containing an animated image, we will
consider the subset of static images as the reference dataset.
Exemplar pictures taken from the dataset are shown in Figure
3, depicting some image categories that we will take into
account in the following sections.
Fig. 3. Exemplar pictures from the dataset.
IV. DATA ANALYSIS
Virality metrics in our dataset follow a power-law-like
distribution thickening toward low virality score. In order to
evaluate the “virality power” of the features taken into account,
we compare the virality indexes in terms of empirical Comple-
mentary Cumulative Distribution Functions (CCDFs). These
functions are commonly used to analyse online social networks
in terms of growth in size and activity (see for example
[14], [15], or the discussion presented in [16]) and also for
measuring content diffusion, e.g. the number of retweets of a
given content [17]. Basically, these functions account for the
probability p that a virality index will be greater than n and
are defined as follows:
F (n) =
number of posts with virality index > n
total number of posts
(1)
For example,
the probability of having a post with
more than 75 plusoners is indicated with Fplus(75) =
P(#plusoners > 75). In the following sections we use CCDFs
to understand the relation between image characteristics and
post virality; in order to assess whether the CCDFs of the
several types of posts we take into account show signifi-
cant differences, we will use the Kolmogorov-Smirnov (K–
S) goodness-of-fit test, which specifically targets cumulative
distribution functions.
A. Image vs. text-only
First of all, we aim to understand what is the impact
of “adding an image to a post” in Google+. Some studies
[18] already show that posts containing an image are much
more viral
than simple plain-text posts, and that various
characteristics of image based banners affect viewer’s recall
and clicks [19]. This finding can be explained in light of a
“rapid cognition” model [20], [21]. In this model, the user
has to decide in a limited amount of time, and within a
vast information flow of posts, whether to take an action on
a particular post (e.g. to reply, reshare, give it a plusone).
Thus, pictures, and the characteristics thereof analyzed in the
following sections, might play a role of paramount importance
in her decision-making process as she exploits visual cues
that grab her attention. In some respects, the rapid cognition
model is reminiscent of the mechanisms by which humans
routinely make judgments about strangers’ personality and
behavior from very short behavioral sequences and non-verbal
cues [22], [23].
In order to investigate the general impact of images we
compared posts containing a picture with posts containing
only text. While our findings overall coincide with [18],
some interesting phenomena emerged. First, we see that the
probability for a post with an image to have a high number of
resharers is almost three times greater ( Fresh(10) = 0.28 vs.
0.10, K–S test p < 0.001), see Figure 4.c. Still, the CCDFs
for the other virality indexes show different trends:
• Posts containing images have lower probability of being
viral when it comes to number of comments ( Frepl(50)
= 0.33 vs. 0.22, K–S test p < 0.001), see Figure 4.b.
This can be explained by the fact that text-only posts
elicit more “linguistic-elaboration” than images (we also
expect that the average length of comments is higher for
text-only posts but we do not investigate this issue here).
• Also, if we focus on simple appreciation (plusoners in
Figure 4.a), results are very intriguing: while up to about
75 plusoners the probability of having posts containing
images is higher, after this threshold the situation cap-
sizes. This finding can be of support to the hypothesis
that, while it is easier to impress with images in the
information flow — as argued with the aforementioned
“rapid cognition” model — high quality textual content
can impress more.
B. Static vs. Animated
Animated images add a further dimension to pictures
expressivity. Having been around since the beginning of the
Internet (the gif format was introduced in late 80’s), animated
images have had alternate fortune, especially after the wide
spread of services like youtube and the availability of broad-
band. Nonetheless, they are still extensively used to produce
simple animations and short clips. Noticeably, the value of
simple and short animations has been acknowledged by Twitter
with the recently released Vine service.
)
n
(
F
)
n
(
F
)
n
(
F
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
a
with image
without attachments
0
50
100
150
200
250
300
350
400
number of plusoners (n)
b
0
50
100
150
200
250
300
350
400
number of replies (n)
c
0
10
20
40
30
70
number of resharers (n)
50
60
80
90 100
)
n
(
F
)
n
(
F
)
n
(
F
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
a
static image
animated image
0
50
100
150
200
250
300
350
400
number of plusoners (n)
b
0
50
100
150
200
250
300
350
400
number of replies (n)
c
0
10
20
40
30
70
number of resharers (n)
50
60
80
90 100
Fig. 4. Virality CCDFs for posts with image vs. text-only posts.
Fig. 5. Virality CCDFs for static vs. animated images.
Whether a post contains a static or animated image has a
strong discriminative impact on all virality indexes, see Figure
5. With respect to plusoners and replies, static images tend to
show higher CCDFs (respectively two and three times more,
Fplus(75) = 0.30 vs. 0.17, Frepl(50) = 0.22 vs. 0.08, K–S test
p < 0.001), while on resharers the opposite holds.
The fact
that Fresh(n) is two times higher for posts
containing animated images ( Fresh(10) = 0.48 vs. 0.27, K–
S test p < .001) can be potentially explained by the fact that
animated images are usually built to convey a small “memetic”
clip - i.e. funny, cute or quirky situations as suggested in [24].
In order to verify this hypothesis we have annotated a small
random subsample of 200 images. 81% of these animated
images were found to be “memetic” (two annotators were
used, positive example if the image score 1 at least on one of
the aforementioned dimensions, annotator agreement is very
high — Cohen’s kappa 0.78). These findings indicate that
animated images are mainly a vehicle for amusement, at least
on Google+.
C. Image Orientation
We then focused on the question whether image orientation
(landscape, portrait and squared) has any impact on virality
indexes. We included squared images in our analysis since they
are typical of popular services a la Instagram. These services
enable users to apply digital filters to the pictures they take and
confine photos to a squared shape, similar to Kodak Instamatic
and Polaroids, providing a so-called “vintage effect”.
We have annotated a small random subsample of 200 im-
ages. 55% of these images were found to be “Instagrammed”
(two annotators were used, positive example if the image
)
n
(
F
)
n
(
F
)
n
(
F
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
a
vertical
square
horizontal
0
50
100
150
200
250
300
350
400
number of plusoners (n)
b
0
50
100
150
200
250
300
350
400
number of replies (n)
c
0
10
20
40
30
70
number of resharers (n)
50
60
80
90 100
Fig. 6. Virality CCDFs for image orientation.
is clearly recognized as modified with a filter; annotator
agreement is high – Cohen’s kappa 0.68). Note that, if we
include also black and white squared pictures without any
other particular filter applied (b/w is one of the ”basic” filter
provided by Instagram) the amount of Instagrammed pictures
rises to 65%. Obviously, the ratio of pictures modified with this
and similar services could be higher; here, we rather wanted to
identify those pictures that were clearly recognized as seeking
for the aforementioned “vintage effect”.
While the orientation seems not to have strong impact
on resharers, with a mild prevalence of horizontal pictures
(see Figure 6.c), plusoners and replies tend to well discrimi-
nate among various image orientations. In particular, portrait
images show higher probability of being viral than squared
images than, in turn, landscapes (see Figure 6.a and 6.b).
Furthermore, CCDFs indicate that vertical images tend to
be more viral than horizontal ones ( Fplus(75) = 0.38 vs. 0.26,
Frepl(50) = 0.38 vs. 0.17, K–S test p < 0.001). Hence, while
squared images place themselves in the middle in any metric,
landscape images have lower viral probability for plusones and
replies but slightly higher probability for reshares.
This can be partially explained by the fact that we are
analyzing “celebrities” posts. If the vertically-orientated image
contains the portrait of a celebrity this is more likely to be
appreciated rather that reshared, since the act of resharing can
also be seen as a form of “self-representation” of the follower
(we will analyze the impact of pictures containing faces in the
following section). The opposite holds for landscapes, i.e. they
are more likely to be reshared and used for self-representation.
D. Images containing one face
In traditional mono-directional media (e.g. tv, billboards,
etc.) a widely used promotion strategy is the use of testimo-
nials, especially celebrities endorsing a product. Is the same
strategy applicable to Social Media? Understanding the effect
of posting images with faces by most popular Google+ users
(and hypothesizing that those are their faces) is a first step in
the direction of finding an answer.
We computed how many faces are found in the images,
along with the ratio of the area that include faces and the
whole image area, using the Viola-Jones [25] face detection
algorithm. We considered images containing one face vs.
images containing no faces. We did not consider the surface of
image occupied by the face (i.e. if it is a close-up portrait, or
just a small face within a bigger picture). The discriminative
effect of containing a face on virality is statistically significant
but small. Still, the pictures containing faces tend to have mild
effect on resharers (slightly higher replies and plusoners but
lower resharers as compared to images with no faces).
In order to verify the hypothesis mentioned earlier, i.e.
that self-portraits tend to be reshared less, we also focused
on a subsample of images containing faces that cover at least
10% of the image surface (about 6400 instances). In this
case, the differences among indexes polarize a little more
(higher plusoners and comments, lower resharers), as we were
expecting. Unfortunately, images with even higher face/surface
ratio are too few to further verify the hypotheses.
E. Grayscale vs. Colored
The impact and meaning of black-and-white (i.e. grayscale)
photographic images has been studied from different perspec-
tives (e.g. semiotics and psychology) and with reference to
different fields (from documentary to arts and advertising).
)
n
(
F
)
n
(
F
)
n
(
F
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
a
mean bright. ≤ 0.85
mean bright. > 0.85
0
50
100
150
200
250
300
350
400
number of plusoners (n)
b
0
50
100
150
200
250
300
350
400
number of replies (n)
c
0
10
20
40
60
30
70
number of resharers (n)
50
80
90 100
Fig. 7. Virality CCDFs for image Brightness.
Rudolf Arnheim, for example, argues that color produces
essentially emotional experience, whereas shape corresponds
to intellectual pleasure [26]. Hence, black-and-white photog-
raphy, because of its absence of expressive colors, focuses on
shapes that require intellectual reflection and brings to explore
aesthetic possibilities. We want to understand if such functions
and effects can be spotted in our virality indexes.
In order to have a “perceptual” grayscale (some images
may contain highly desaturated colors and so perceived as
shades of gray) we dichotomized the dataset according to the
mean-saturation index of the images, using a very conservative
threshold of 0.05 (on a 0-1 scale).
As can be seen in Figure 8.a and 8.b, colored images
(with saturation higher than 0.05) have a higher probability of
collecting more plusoners and replies as compared to images
with lower saturation (grayscale). In particular the probability
functions for replies is more than two times higher ( Frepl(50)
values are 0.26 vs. 0.10, K–S test p < 0.001). Instead, image
saturation has no relevant impact on resharers.
F. Very Bright Images
After converting each image in our dataset to the HSB
color space, we extracted its mean Saturation and Brightness.
the HSB (Hue/Saturation/Brightness) color
More in detail,
space describes each pixel
in an image as a point on a
cylinder: the Hue dimension representing its color within the
set of primary-secondary ones, while Saturation and Brightness
describe respectively how close to the pure color (i.e. its Hue),
and how bright it is.We split the dataset according to images
mean brightness using a threshold of 0.85 (in a scale included
between 0 and 1). Usually images with such an high mean
brightness tend to be cartoon-like images rather than pictures.
)
n
(
F
)
n
(
F
)
n
(
F
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
1.0
0.8
0.6
0.4
0.2
0.0
a
mean sat. ≤ 0.05
mean sat. > 0.05
0
50
100
150
200
250
300
350
400
number of plusoners (n)
b
0
50
100
150
200
250
300
350
400
number of replies (n)
c
0
10
20
40
30
70
number of resharers (n)
50
60
80
90 100
Fig. 8. Virality CCDFs for Grayscale vs. Colored images.
Previous research
[27] has shown that pixel brightness is
expected to be higher in cartoon-like (or significantly “pho-
toshopped’) than in natural images.
Image brightness level has a strong impact on plusoners and
replies, and a milder one on resharers. Brighter images have a
lower probability of being viral on the first two indexes (Figure
7.a and 7.b) and a higher probability on the latter (Figure
7.c). In particular, lower brightness images have a plusone
and reshare probability almost two times higher ( Fplus(75) =
0.31 vs. 0.18, Frepl(50) = 0.23 vs. 0.12, K–S test p < 0.001),
while for resharers it is 27% higher in favor of high brightness
images ( Fresh(10) = 0.33 vs. 0.26). Surprisingly, analyzing
a small random subsample of 200 very bright images, we
found that while 88% of these images contained some text,
as we would have expected, only 13% were cartoon/comics
and only 13% contained the real picture of an object as
subject, even if highly ”photoshopped”. Above all, only a
small amount of these images (21%) was considered funny or
memetic5. The great majority comprised pictures containing
infographics, screenshots of software programs, screenshots
of social-networks posts and similar. In this respect we are
analyzing a content that is meant to be mainly informative,
and is somehow complementary to the content of animated
pictures (mainly intended for amusement, see IV-B).
G. Vertical and Horizontal edges
Finally, we want to report on an explorative investigation
we made. We focused on the impact of edges intensity on posts
virality. The intensity of vertical/horizontal/diagonal edges was
5Two annotators were used, four binary categories were provided (contain-
text/comics/real-picture-obj/funny). The overall inter-annotator agreement on
these categories is high, Cohen’s kappa 0.74.
computed using Gaussian filters, based on code used in [28]
in the context of real-time visual concept classification. The
probability density of the average edges intensity follows a
gaussian-like distribution, with mean of about 0.08 (both for
horizontal and vertical edges). We divided images into two
groups: those having an average edge intensity below the
sample mean, and those having an average edge intensity
above the mean. Results showed that images with horizontal
edge intensity below the sample mean are far more viral
on the plusoners and replies indexes, while vertical are less
discriminative. Results for horizontal hedges are as follows:
Fplus(75) = 0.36 vs. 0.22, Frepl(50) = 0.27 vs. 0.14, Fresh(10)
= 0.25 vs. 0.29, K–S test p < 0.001. While these results do
not have an intuitive explanation, they clearly show that there
is room for further investigating the impact of edges.
H. Virality Indexes Correlation
From the analyses above, virality indexes seem to “move
together” (in particular plusoners and replies) while resharers
appear to indicate a different phenomenon. We hypothesize
that plusoners and replies can be considered as a form of
endorsement, while reshares are a form of self-representation.
This explains why, for example, pictures containing faces
are endorsed but not used for self-representation by VIPs’
followers. On the contrary, animated images that usually
contain funny material are more likely to provoke reshares for
followers’ self-representation. In fact, people usually tend to
represent themselves with positive feelings rather than negative
ones (especially popular users, see [29]), and positive moods
appear to be associated with social interactions [30], [31].
TABLE II.
VIRALITY INDEXES CORRELATION ON THE DATASETS
Pearson
MIC
Static images
plusoners vs. replies
plusoners vs. resharers
replies vs. resharers
0.723
0.550
0.220
Animated Images
plusoners vs. replies
plusoners vs. resharers
replies vs. resharers
Text Only
plusoners vs. replies
plusoners vs. resharers
replies vs. resharers
0.702
0.787
0.554
0.802
0.285
0.172
0.433
0.217
0.126
0.304
0.396
0.205
0.529
0.273
0.185
This is supported also by the correlation analysis of the
three virality indexes, reported in Table II, made on the
various datasets we exploited. In this analysis we used both
the Pearson coefficient and the recent Maximal Information
Coefficient (MIC), considering plusoners ≤ 1200, replies ≤
400 e resharers ≤ 400. MIC is a measure of dependence
introduced in [32] and it is part of the Maximal Information-
based Nonparametric Exploration (MINE) family of statistics.
MIC is able to capture variable relationships of different
nature, penalizing similar levels of noise in the same way. In
this study we use the Python package minepy [33].
In particular, from Table II we see that: plusones and replies
always have a high correlation while replies and resharers
always correlate low. Plusoners and reshares, that have a mild
correlation in most cases, correlate highly when it comes to
funny pictures, i.e. animated ones. This can be explained by
a specific “procedural” effect: the follower expresses his/her
appreciation for the funny picture and, after that, he/she
reshares the content. Since resharing implies also writing a
comment in the new post, the reply is likely not to be added
to the original VIP’s post.
In Table III we sum up the main findings of the paper, com-
paring the various CCDFs: animated images and infographics
have much higher probability of being reshared, while colored
images or images containing faces have higher probability
of being appreciated or commented. Finally, black and white
pictures (grayscale) turn out to be the least “viral” on Google+.
TABLE III.
SUMMARY OF MAIN FINDINGS OF THE ANALYSIS.
very bright
grayscale
color
animated
one-face > 10% area
Fplus(75)
0.18
0.21
0.31
0.17
0.35
Frepl(50)
0.12
0.11
0.24
0.08
0.30
Fresh(10)
0.33
0.28
0.27
0.48
0.23
V. USER ANALYSIS
Finally, we investigate if there is any relevant interaction
between images characteristics and VIP’s typology. In Table IV
we report demographic details6 on the Google+ dataset, as
provided by the users in their profile pages.
TABLE IV.
USER DEMOGRAPHICS IN THE GOOGLE+ DATASET.
User-category
Technology
Photography
Music
Writing
Actor
Entrepreneur
Sport
Artist
TV
Company
Website
Politician
No Category
Organization
Not Available
Other
Total
35 (19%)
41 (24%)
96 (59%)
26 (21%)
21 (36%)
12 (29%)
Female (%) Male (%)
110 (61%)
130 (76%)
48 (29%)
76 (63%)
34 (59%)
29 (71%)
22 (55%)
21 (60%)
11 (33%)
-
11 (31%)
8 (24%)
-
-
-
6 (43%)
-
-
-
-
19 (86%)
8 (57%)
-
-
1 (33%)
257 (27%)
2 (67%)
510 (54%)
Neutral (%)
36 (20%)
1 (1%)
19 (12%)
19 (16%)
3 (5%)
-
18 (45%)
3 (9%)
14 (42%)
28 (100%)
23 (100%)
3 (14%)
-
9 (100%)
7 (100%)
-
Total (%)
181 (19%)
172 (18%)
163 (17%)
121 (13%)
58 (6%)
41 (4%)
40 (4%)
35 (4%)
33 (3%)
28 (3%)
23 (2%)
22 (2%)
14 (1%)
9 (1%)
7 (1%)
3 (0%)
183 (19%)
950 (100%)
In order to investigate possible user category effects in our
dataset — that is, if our analyses are also influenced by the
type of user posting images rather than by the actual content
solely, we evaluated the entropy for each image category over
the 16 user categories (as defined in Table IV). In Table V
we report the contingency table of image-category entropy
distributions over user-categories. Looking at the Kullback-
Leibler (KL) divergence of specific image categories with
respect to the reference distribution (i.e., taken as the total
number of images posted by each user-category), we observe
very few but interesting effects due to specific user-categories.
In particular, while all the KL divergences are very small,
two of them (for Grayscale and High Brightness, reported in
6No Category denotes users that do not provide any personal information
and for which it was not possible to trace back their category; Not Available
denotes seven accounts that were no more publicly accessible when we
gathered demographic info; Other denotes very rare and unusual category
definitions. The Neutral gender refers to pages afferent to ”non-humans” like
products, brands, websites, firms, etc.
Bold) are an order of magnitude greater than other classes.
Interestingly the divergence is explained mainly by the distri-
bution gap in only two User’s categories. For High Brightness
the gap is mainly given by Technology user category that dou-
bles its probability distribution (from 22% to 40%) and Music
and Photography that reduce their probability distribution to
one third. This divergence from the reference distribution
is consistent with the analysis of the content we made in
section IV-F: these images where mainly infographics and
screenshots of software programs and social networks (so
mainly connected to technology). For Grayscale the gap is
mainly given by Photography users category that rises by 50%
its probability distribution and Music, that reduces it to one
third. This gap is consistent with the idea, expressed in Section
IV-E, that black-and-white photography is a particular form of
art expressivity mainly used by professionals.
VI. CONCLUSIONS
We have presented a study, based on a novel dataset of
Google+ posts, showing that perceptual characteristics of an
image can strongly affect the virality of the post embedding it.
Considering various kinds of images (e.g. cartoons, panorama
or self-portraits) and related features (e.g. orientation, anima-
tions) we saw that users’ reactions are affected in different
ways. We provided a series of analyses to explain the underly-
ing phenomena, using three virality metrics (namely plusoners,
replies and resharers). Results suggest
that plusoners and
replies “move together” while reshares indicate a distinct users’
reaction. In particular, funny and informative images have
much higher probability of being reshared but are associated
to different
image features (animation and high-brightness
respectively), while colored images or images containing faces
have higher probability of being appreciated and commented.
Future work will dig deeper into the assessment of relations
between visual content and virality indexes, adopting multi-
variate analysis that includes user’s categories (e.g. which is the
viral effect of b/w pictures taken by professional photographer
as compared to those taken by non professional users). We
will also extend our experimental setup in the following
ways: (a) taking into account compositional features of the
images, i.e. resembling concepts such as the well-known ”rule
of thirds”; (b) extracting and exploiting descriptors such as
color histograms, oriented-edges histograms; (c) building upon
the vast
literature available in the context of scene/object
recognition, dividing our dataset into specific categories in
order to analyse relations between categories, such as natural
images or sport images, and their virality.
ACKNOWLEDGMENT
The work of J. Staiano and M. Guerini has been partially
supported by the FIRB project S-PATTERNS and the Trento
RISE PerTe project, respectively.
REFERENCES
[1] K. Lerman and R. Ghosh, “Information contagion: an empirical study of
the spread of news on digg and twitter social networks,” in Proceedings
of ICWSM-10, Mar 2010.
[2] S. Jamali and H. Rangwala, “Digging digg : Comment mining, pop-
ularity prediction, and social network analysis,” in Proceedings of
International Conference on Web Information Systems and Mining,
2009.
TABLE V.
CONTINGENCY TABLE OF IMAGE-CATEGORY DISTRIBUTIONS OVER USER-CATEGORIES.
Grayscale
Colored
High Brightness
Low Brightness
Containing Face
Containing No Face
Squared
Vertical
Horizontal
7%
4%
5%
0%
8%
3%
0%
0%
0%
31%
0%
0%
27%
1%
1%
11%
0.173
6%
6%
6%
1%
7%
16%
0%
0%
0%
19%
0%
3%
22%
2%
2%
10%
0.002
9%
5%
7%
1%
6%
3%
0%
0%
0%
9%
0%
1%
40%
1%
2%
17%
0.259
6%
5%
6%
1%
7%
16%
0%
0%
0%
22%
0%
3%
20%
2%
2%
10%
0.003
5%
8%
6%
1%
7%
19%
0%
0%
0%
15%
0%
4%
19%
3%
2%
11%
0.027
7%
5%
6%
1%
7%
12%
0%
0%
0%
23%
0%
2%
24%
1%
2%
10%
0.006
4%
5%
5%
1%
8%
15%
0%
0%
2%
23%
0%
2%
16%
5%
1%
11%
0.047
5%
6%
7%
1%
5%
29%
0%
0%
0%
14%
0%
2%
18%
1%
1%
10%
0.076
7%
5%
6%
1%
8%
8%
0%
0%
0%
23%
0%
3%
25%
2%
2%
11%
0.029
Total
6%
5%
6%
1%
7%
14%
0%
0%
0%
20%
0%
2%
22%
2%
2%
11%
User-category
No Category
Actor
Artist
Company
Entrepreneur
Music
Not Available
Organization
Other
Photography
Politician
Sport
Technology
TV
Website
Writing
KL-divergence
[3] E. Khabiri, C.-F. Hsu, and J. Caverlee, “Analyzing and predicting
community preference of socially generated metadata: A case study
on comments in the digg community,” in Proceedings of ICWSM-09,
2009.
J. Z. Aaditeshwar Seth and R. Cohen, “A multi-disciplinary approach
for recommending weblog messages,” in The AAAI 2008 Workshop on
Enhanced Messaging, 2008.
[4]
[5] K. Lerman and A. Galstyan, “Analysis of social voting patterns on
digg,” in Proceedings of the first workshop on Online social networks,
ser. WOSP ’08. New York, NY, USA: ACM, 2008, pp. 7–12.
[Online]. Available: http://doi.acm.org/10.1145/1397735.1397738
[6] 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.
[7] S. Jamali, “Comment mining, popularity prediction, and social network
analysis,” Master’s thesis, George Mason University, Fairfax, VA, 2009.
J. A. Berger and K. L. Milkman, “Social Transmission, Emotion, and the
Virality of Online Content,” Social Science Research Network Working
Paper Series, December 2009.
[8]
[9] C. Danescu-Niculescu-Mizil, J. Cheng, J. Kleinberg, and L. Lee, “You
had me at hello: How phrasing affects memorability,” in Proceedings
of the ACL, 2012.
[10] M. Simmons, L. A. Adamic, and E. Adar, “Memes online: Extracted,
subtracted, injected, and recollected,” Proceedings of ICWSM-11, 2011.
[11] M. Guerini, C. Strapparava, and G. Ozbal, “Exploring text virality in
social networks,” in Proceedings of ICWSM-11, Barcelona, Spain, July
2011.
[12] M. Guerini, A. Pepe, and B. Lepri, “Do linguistic style and readability
of scientific abstracts affect their virality,” Proceedings of ICWSM-12,
2012.
[13] D. Schioberg, S. Schmid, F. Schneider, S. Uhlig, H. Schioberg,
and A. Feldmann, “Tracing the birth of an osn: social graph
and profile
the 3rd
Annual ACM Web Science Conference, ser. WebSci
’12. New
York, NY, USA: ACM, 2012, pp. 265–274.
[Online]. Available:
http://doi.acm.org/10.1145/2380718.2380753
in Proceedings of
in google+,”
analysis
[15]
[14] Y.-Y. Ahn, S. Han, H. Kwak, S. Moon, and H. Jeong, “Analysis of
topological characteristics of huge online social networking services,”
in Proceedings of the 16th international conference on World Wide Web.
ACM, 2007, pp. 835–844.
J. Jiang, C. Wilson, X. Wang, P. Huang, W. Sha, Y. Dai, and B. Y.
Zhao, “Understanding latent interactions in online social networks,”
in Proceedings of the 10th ACM SIGCOMM conference on Internet
measurement. ACM, 2010, pp. 369–382.
J. Leskovec, Dynamics of large networks. ProQuest, 2008.
[16]
[17] 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. ACM, 2010, pp. 591–600.
[18] TrackSocial, “Optimizing Facebook Engagement, whitepaper,” 2012.
[19] H. Li and J. L. Bukovac, “Cognitive impact of banner ad characteristics:
An experimental study,” Journalism & Mass Communication Quarterly,
vol. 76, no. 2, pp. 341–353, 1999.
[20] N. Ambady and R. Rosenthal, “Thin slices of expressive behavior as
predictors of interpersonal consequences: A meta-analysis.” Psycholog-
ical bulletin, vol. 111, no. 2, p. 256, 1992.
[21] D. Kenny, Interpersonal perception: A social relations analysis. The
Guilford Press, 1994.
[22] B. Lepri, R. Subramanian, K. Kalimeri, J. Staiano, F. Pianesi,
and N. Sebe, “Employing social gaze and speaking activity for
automatic determination of the extraversion trait,” in International
Conference on Multimodal Interfaces and the Workshop on Machine
Learning for Multimodal Interaction, ser. ICMI-MLMI ’10. New
York, NY, USA: ACM, 2010, pp. 7:1–7:8.
[Online]. Available:
http://doi.acm.org/10.1145/1891903.1891913
J. R. Curhan and A. Pentland, “Thin slices of negotiation: predicting
outcomes from conversational dynamics within the first 5 minutes.”
Journal of Applied Psychology, vol. 92, no. 3, p. 802, 2007.
[23]
[24] C. Dufour, “An investigation into the use of viral marketing for the
companies and the key success factors of a good viral campaign,” Ph.D.
dissertation, Dublin Business School, 2011.
[25] P. Viola and M. J. Jones, “Robust real-time face detection,” Int. J.
Comput. Vision, vol. 57, no. 2, pp. 137–154, May 2004. [Online].
Available: http://dx.doi.org/10.1023/B:VISI.0000013087.49260.fb
I.
and H. Rohrig,
Ianeva, A. P. de Vries,
[26] R. Arnheim, Art and visual perception. Stockholms Universitet, 1987.
“Detecting
[27] T.
cartoons: a case study in automatic video-genre classification,”
in Proceedings of the 2003 International Conference on Multimedia
and Expo - Volume 2, ser. ICME ’03. Washington, DC, USA:
IEEE Computer Society, 2003, pp. 449–452.
[Online]. Available:
http://dl.acm.org/citation.cfm?id=1170745.1171531
J. Uijlings, A. Smeulders, and R. Scha, “Real-time visual concept
classification,” Multimedia, IEEE Transactions on, vol. 12, no. 7, pp.
665 –681, nov. 2010.
[28]
[29] D. Quercia, J. Ellis, L. Capra, and J. Crowcroft, “In the mood for being
[30]
influential on twitter,” Proceedings of IEEE SocialCom’11, 2011.
J. R. Vittengl and C. S. Holt, “A time-series diary study of mood and
social interaction,” Motivation and Emotion, vol. 22, no. 3, pp. 255–275,
1998.
[31] M. De Choudhury, S. Counts, and M. Gamon, “Not all moods are
created equal! exploring human emotional states in social media,” in
Proceedings of ICWSM-12, 2012.
[32] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman,
G. McVean, P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C.
Sabeti, “Detecting novel associations in large data sets,” science, vol.
334, no. 6062, pp. 1518–1524, 2011.
[33] D. Albanese, M. Filosi, R. Visintainer, S. Riccadonna, G. Jurman, and
C. Furlanello, “minerva and minepy: a c engine for the mine suite and
its r, python and matlab wrappers,” Bioinformatics, vol. 29, no. 3, pp.
407–408, 2013.
|
1102.0257 | 1 | 1102 | 2011-02-01T19:44:09 | Emergence through Selection: The Evolution of a Scientific Challenge | [
"cs.SI",
"cs.AI",
"math.DS",
"physics.soc-ph"
] | One of the most interesting scientific challenges nowadays deals with the analysis and the understanding of complex networks' dynamics and how their processes lead to emergence according to the interactions among their components. In this paper we approach the definition of new methodologies for the visualization and the exploration of the dynamics at play in real dynamic social networks. We present a recently introduced formalism called TVG (for time-varying graphs), which was initially developed to model and analyze highly-dynamic and infrastructure-less communication networks such as mobile ad-hoc networks, wireless sensor networks, or vehicular networks. We discuss its applicability to complex networks in general, and social networks in particular, by showing how it enables the specification and analysis of complex dynamic phenomena in terms of temporal interactions, and allows to easily switch the perspective between local and global dynamics. As an example, we chose the case of scientific communities by analyzing portion of the ArXiv repository (ten years of publications in physics) focusing on the social determinants (e.g. goals and potential interactions among individuals) behind the emergence and the resilience of scientific communities. We consider that scientific communities are at the same time communities of practice (through co-authorship) and that they exist also as representations in the scientists' mind, since references to other scientists' works is not merely an objective link to a relevant work, but it reveals social objects that one manipulates, select and refers to. In the paper we show the emergence/selection of a community as a goal-driven preferential attachment toward a set of authors among which there are some key scientists (Nobel prizes). | cs.SI | cs |
Emergence through Selection: The Evolution of
a Scientific Challenge
Walter Quattrociocchi1 and Frederic Amblard2
1 University of Siena Pian Dei Mantellini 44, 53100, Siena, Italy
Tel: +39 0577 233 710
[email protected]
2 IRIT - Universit Toulouse 1 Capitole, 2, rue du Doyen Gabriel Marty
31042 Toulouse Cedex 9
Tel: +33 (0) 561 128 795
[email protected]
Summary. One of the most interesting scientific challenges nowadays deals with
the analysis and the understanding of complex networks' dynamics and how their
processes lead to emergence according to the interactions among their components.
In this paper we approach the definition of new methodologies for the visualiza-
tion and the exploration of the dynamics at play in real dynamic social networks. We
present a recently introduced formalism called TVG (for time-varying graphs), which
was initially developed to model and analyze highly-dynamic and infrastructure-less
communication networks. As an application context, we chose the case of scientific
communities by analyzing a portion of the ArXiv repository (ten years of publica-
tions in physics). The analysis presented in the paper passes through different data
transformations aimed at providing different perspectives on the scientific commu-
nity and its evolutions.
On a first level we discuss the dataset by means of both a static and temporal
analysis of citations and co-authorships networks. Afterward, as we consider that
scientific communities are at the same time communities of practice (through co-
authorship) and that a citation represents a deliberative selection pointing out the
relevance of a work in its scientific domain, we introduce a new transformation aimed
at capturing the interdependencies between collaborations' patterns and citations'
effects and how they make evolve a goal oriented systems as Science.
Finally, we show how through the TVG formalism and derived indicators, it is
possible to capture the interactions patterns behind the emergence (selection) of
a sub-community among others, as a goal-driven preferential attachment toward a
set of authors among which there are some key scientists (Nobel prizes) acting as
attractors on the community.
Key words: social networks analysis, evolving structures, scientific commu-
nities, emergence, time-varying graphs, temporal metrics, social selection.
2
Walter Quattrociocchi and Frederic Amblard
1 Introduction
One of the most interesting scientific challenges nowadays deals with the
analysis and the understanding of social networks' dynamics and how their
processes lead to emergence according to the interactions at play among
their components. The research efforts in this area strive to understand
what are the driving forces behind the evolution of social networks and
how they are articulated together with social dynamics, e.g., opinion dy-
namics, the epidemic or innovation diffusion, the teams formation and so
on ([7, 22, 15, 3, 28, 9, 31, 5, 30]). In this paper we approach the definition
of new methodologies for the visualization and the exploration of the dynam-
ics within real dynamic social networks. As an example, we chose the case
of scientific communities by analyzing a portion of the ArXiv repository (ten
years of publications in physics) focusing on the social determinants (e.g. goals
and potential interactions among individuals) behind the emergence and the
resilience of scientific communities. In particular, the analysis addresses the
co-existence of co-authorships' and citations' behaviors of scientists by focus-
ing on the most proficient and cited authors interactions' patterns and, in
turn, on how they are affected by the selection process of citations. Such a
social selection a) produces self-organization because it is played by a group of
individuals which act, compete and collaborate in a common environment in
order to advance Science and b) determines the success (emergence) of both
topics and scientists working on them.
On the one hand, the studies on scientific network dynamics deal with the
understanding of the factors that play a significant role in their evolution, not
all of them being neither objective nor rational e.g., the existence of a star
system [38], [24], [25], [1] the blind imitation concerning the citations [20], the
reputation and community affiliation bias [8]. On the other hand, having some
elements to understand such dynamics could enable for a better detection of
the hot topics and of the vivid subfields and how the scientific production is
advanced with respect to selection process inside the community itself. Among
the available data to analyze such a system, a subset of the publications in
a given field is the most frequently used such as in [29], [23], [26], and [33].
The scientific publications correspond to the production of such a system and
clearly identify who are the producers (the authors), which institution they
belong to (the affiliation), which funded project they are working on (the ac-
knowledgement) and what are the related publications (the citations), having
most of the time a public access to these data explain also a part of its frequent
use in the analyses of the scientific field. Classical analyses concern either the
co-authorships network ([1, 24]) or the citation network ([11, 34]), more rarely
the institutional network ([28]). Moreover, such networks are often considered
as static and their structure is rarely analyzed overtime (an exception is the
one performed by [33] on Physical Review).
The illustrative analysis presented in the paper passes through different
data transformations aimed at providing different perspectives on the scien-
Emergence through Selection: The Evolution of a Scientific Challenge
3
tific network and its evolutions. On a first level we discuss the dataset by
means of both static and temporal analysis of citations and co-authorships
networks. A second level of analysis consists in transforming the data in order
to explicit the interdependencies between the co-authorships and citations by
analyzing the scientists' representations of the collaboration structure within
the scientific field. Such a representation is captured through the network of
cited collaborations ([32]), i.e. from a publication we have several references to
other papers, each one corresponds to a promotion of the scientists authoring
the work.
One of the problem when trying to characterize such a dynamic structure
is that classical indicators from either graph theory or social network analysis
cannot be applied directly. Therefore, we used an algebra, the Time-Varying
Graphs (TVG) ([4]) that enables to take into account the dynamical aspects
of networks and allows for the definition of temporal indicators ([35]) to char-
acterize patterns in evolving structures.
Through our approach, we capture the attractiveness played by famous
authors on co-authorship behaviors and on the sub-communities structural
evolution.
2 Context
In [24] the network of scientific collaborations, explored upon several databases,
shows a clustered and small world structure. Moreover, several differences be-
tween the collaborations' patterns of the different fields studied are captured.
Such differences have been deepened in [25] with respect to the number of pa-
pers produced by a given group of authors, the number of collaborations and
the topological distances between scientists. Peltomaki and Alava in [27] pro-
pose a new emulative model aimed at approximating the growth of scientific
networks, by incorporating bipartition and sub-linear preferential attachment.
A model for the self-assembly of creative teams based on three parameters (e.g.
team size, the rate of newcomers in the scientific production and the tendency
of authors to collaborate with the same group) has been outlined in [9]. Con-
nectivity patterns in a citations network have been studied in relation to the
development of the DNA theory [11]. The work of Klemm and Eguiluz ( [12])
observed that real networks (e.g. movie actors, co-authorship in science, and
word synonyms) growing patterns are characterized by a clustering trend that
reaches an asymptotic value larger than regular lattices of the same average
connectivity.
In the field of social network analysis several works have approached the
problem of temporal metrics [10, 14, 13]. Actually, the focus is on the definition
of instruments able to capture the intrinsic properties of complex systems'
evolution, that is, characterizing the interdependencies and the co-existence
between local behaviors (interactions) and their global effects (emergence)
[6, 21, 40, 7, 30]. The research approach to characterize the evolution patterns
4
Walter Quattrociocchi and Frederic Amblard
of social networks, at the very beginning was mainly based upon simulations,
while in the past few years, due to the large availability of real datasets,
either the methodology of analysis and the object of research have changed
([37, 19, 13, 5, 17]). In particular, in the latter paper Leskovec states as central
problem, for the social networks in general and for the scientific communities
networks analysis in particular, the definition of mathematical models able
to deal and to reproduce all the properties of dynamical real world networks
such as the shrinking diameter ([18]), or the "small world" effect [39]. Actually
instruments and paradigms affording this challenge are mainly based upon
stochastic definitions ([16]) or conceptualized as a sequence of snapshots of
the network at different times [36].
3 Preliminaries
3.1 Time-Varying Graphs
The time-varying graph (TVG) formalism, recently introduced in [4], is a
graph formalisms based on an interaction-centric point of view and offers
concise and elegant formulation of temporal concepts and properties [35].
Let us consider a set of entities V (or nodes), a set of relations E among
entities (edges), and an alphabet L labeling any property of a relation (label);
that is, E ⊆ V ×V ×L. The set E enables multiple relations between any given
pair of entities, as long as these relations have different properties, that is, for
any e1 = (x1, y1, λ1) ∈ E, e2 = (x2, y2, λ2) ∈ E, (x1 = x2 ∧ y1 = y2 ∧ λ1 =
λ2) =⇒ e1 = e2.
Relationships between entities are assumed to occur over a time span T ⊆
T, namely the lifetime of the system. The temporal domain T is assumed to
be N for discrete-time systems or R for continuous-time systems. The time-
varying graph structure is denoted by the set G = (V, E, T, ρ, ζ), where ρ :
E × T → {0, 1}, called presence function, indicates whether a given edge is
present at a given time, and ζ : E × T → T, called latency function, indicates
the time it takes to cross a given edge if starting at a given date. As in this
paper the focus is on the temporal and structural analysis of a social network,
we will deliberately omit the latency function and consider TVGs described
as G = (V, E, T, ρ).
TVGs as a sequence of footprints.
Given a TVG G = (V, E, T, ρ), one can define the footprint of this graph
from t1 to t2 as the static graph G[t1,t2) = (V, E[t1,t2)) such that ∀e ∈ E, e ∈
E[t1,t2) ⇐⇒ ∃t ∈ [t1, t2), ρ(e, t) = 1. In other words, the footprint aggregates
interactions over a given time window into static graphs. Let the lifetime
T of the time-varying graph be partitioned in consecutive sub-intervals τ =
[t0, t1), [t1, t2) . . . [ti, ti+1), . . .; where each [tk, tk+1) can be noted τk. We call
sequence of footprints of G according to τ the sequence SF(τ ) = Gτ0, Gτ1, . . ..
Emergence through Selection: The Evolution of a Scientific Challenge
5
Expressing other Temporal Concepts
A sequence of couples J = {(e1, t1), (e2, t2) . . . , (ek, tk)}, such that {e1, e2, ..., ek}
is a walk in G is a journey in G if and only if ∀i, 1 ≤ i < k, ρ(ei, ti) = 1 and
ti+1 ≥ ti. The departure(J ) and arrival(J ) are respectively the starting date
t1 and the last date tk of a journey J .
Journeys can be thought of as paths over time from a source to a des-
tination and therefore have both a topological and a temporal length. The
length of J is the number J = k of couples in J (i.e., the
topological
J =
number of hops);
arrival(J ) − departure(J ).
length is its end-to-end duration:
its temporal
4 Exploring the Dataset
4.1 The Dataset
As mentioned in the introduction, the scientific community analyzed in this
work has been extracted from the hep-th (High Energy Physics Theory) por-
tion of the arXiv website, an on-line repository available at http://arxiv.org/.
The dataset is composed by a collection of papers and therefore their
related citations over the period within January 1992 to May 2003. For each
paper the set of authors, the dates of the on-line publications on arXiv.org,
and the references are provided. There are 352 807 citations within the total
amount of 29 555 papers written by 59 439 authors. The broadness of the
time window covered allows us to explore the dataset in order to extract,
capture and characterize the evolution of the interactions patterns within the
community by means of different data transformations.
4.2 The Networks Description
From the dataset, we can easily derive two graphs. The first, namely the
co-authorships network, having authors as nodes and the undirected links
standing for the relation of co-authoring a paper. The second, the citations
network, where nodes are the papers and the links (directed) are the references
among papers. More formally, the derived graphs can be defined as:
•
•
the co-authorship network as Ga : (Va, E) where nodes in Va are the
authors and links e ∈ E connect nodes co-authoring a paper.
the citations network as Gc : (Vc, E) where the nodes in Vc are the
papers and each edge e ∈ E corresponds to a reference to another paper.
In Table 1 we provide measures about the citations and collaborations
networks.
The diameters - e.g., the longest shortest path between to pairs of nodes
(respectively authors and papers) - of both networks have high values. The
6
Walter Quattrociocchi and Frederic Amblard
Network Indicators
Network Diameter
Network Modularity
Ga
26
Gc
37
0,706 0.617
Network Average Clustering Coefficient 0.5006 0.156
Table 1. Co-authorships and Citations Graph Measures
modularity, measuring how a network can be partitioned into modules or
subparts, has high values on both graphs. Whereas the average clustering
coefficient of the collaborations graph is higher than in the citations graph.
The networks are composed by several connected islands with few inter-
connections within them, and the co-authorships network is more clustered of
the citations graph.
5 Temporalizing the Dataset
In this section we are going to explicit the temporal aspects, (i.e. the structural
evolution) of the citations and co-authorships networks. The transformation is
performed through the time-varying graphs formalism defined in the previous
section.
We derive two time-varying graphs: the temporal co-authorships network,
with undirected edges and authors as nodes where a link stands for the re-
lations of co-authoring a paper; and the temporal citations network having
papers as nodes and the links (directed) representing the citations from a pa-
per to another one. The temporal dimension of both networks is derived by
the paper's submission date. The temporal co-authorship network has edges
labeled with the date of submission, while the temporal citations network has
the nodes labeled with the publication date of papers citing other papers.
More formally, we can define
•
•
a : (V, E, T, ρ)
the temporal co-authorships network as a quadruplet Gt
where the nodes in v ∈ V are the authors and links e ∈ E connect a couple
of scientists co-authoring a paper. The temporal domain T = [ta, tb) of the
function ρ, is the lifetime of each node v that in this context is assumed
as ta to be the submission date of the paper and tb = ∞.
c : (V, E, T, ρ) where
the temporal citations network as a quadruplet Gt
the nodes in the set V are the papers and each edge e ∈ E corresponds to a
citation to another paper. As for the co-authorships network, the temporal
dimension T = [ta, tb) of the presence function ρ of Gt
c is defined within
the submission date of papers and ∞.
Emergence through Selection: The Evolution of a Scientific Challenge
7
5.1 Citations Network Evolution
In this section we show the evolution of the temporal citations network Gt
c by
using the sequence of footprints as defined in section 3.1. The values are com-
puted by aggregating the interactions occurring at each sub-interval SF (τ )
having τ fixed to one year. Figure 1(a) shows the evolution of the cluster-
ing coefficient the curve is characterized by a stable trend attesting on low
values. The density evolution, which is shown in Figure 1(b), presents the
same low and decreasing behavior, meaning that both the distances and in-
terconnections among nodes (citations within papers) are stable for all the
time windows observed. Also the modularity, shown in Figure 1(c), has a
decreasing but stable trend.
(a) Average Clustering Coefficient
(b) Density
(c) Modularity
Fig. 1. Citations Graph Evolution
5.2 Co-authorships Network Evolution
The temporal co-authorships graphs presents a different structural evolution
with respect to the temporal graph of citations analyzed in the previous sec-
tion. As before, here the values are computed by aggregating the interactions
8
Walter Quattrociocchi and Frederic Amblard
at each sub-interval SF (τ ) where τ is fixed to one year. The average cluster-
ing coefficient evolution in the time interval observed, that is shown in Figure
2(a), has an oscillating trend with higher values than the ones reached by
the temporal citations graph. In addition, Gt
a has a more modular and denser
structure, as shown in Figure 2(c) and Figure 2(b). The captured trends are
characterized by a decreasing (and not stable) trend.
(a) Average Clustering Coefficient
(b) Density
(c) Modularity
Fig. 2. Co-authorships Graph Evolution
6 Expliciting Interactions
In this section we provide an additional data transformation in order to cap-
ture more details about the evolution of our scientific network. The dynam-
ics in scientific communities are based upon competitions and collaborations
among authors and groups of scientists. In the analysis we want to capture a)
the resulting emerging effects caused by these two opposite motivations and b)
how they are expressed in terms of collaborations and citations patterns. The
dataset analyzed in the current paper presents two explicit interactions: the
Emergence through Selection: The Evolution of a Scientific Challenge
9
papers' co-authorships and the citations between papers. In addition, there
is an implicit level of interaction that depends upon the goals behind any re-
search paper: the quality and, at the same time, the necessity to be highly cited
(competition). Hence, often both the collaborations' and citations' strategies
are optimized in order to have the highest impact with respect to the problem
addressed and to collect the highest number of citations. How such processes
affect the scientific production and the scientific communities structural evo-
lution?
6.1 Deriving the Interactions Graph
We approach the data transformation in order to explicit the co-authorships
and cited collaboration patterns and their interdependencies. The dataset is
transformed in an undirected graph, that we call the interaction network,
having as nodes the authors, weighted links representing the co-authorship
on a paper, and when a paper is cited by another work, the links' weights,
connecting the authors of the referenced paper, are incremented.
More formally, the graph of the cited co-authorships is defined as quadru-
pled Gcc : (V, E, T, ρ) on a discrete time. The nodes v ∈ V are the authors,
the set of edges E represents the collaborations on a paper's production.
The nodes appear on the graph the first time a paper they wrote has been
published, and the interaction L is weighted with a variable wi, namely the
strength value of a collaboration, that is incremented at each citation received
by a paper produced by a given couple of nodes (u, v).
In the following section we analyze the behaviors and the interaction's
strategies within the most cited authors' network, such a graph, namely
Gi, is a subset of the global interaction network Gcc : (V, E, T, ρ).
In particular the nodes considered in the analysis are only the authors
having links' strength values wi ≥ 150, that is, all the groups having more
than 150 citations on a work. Such a network in its maximum expansion,
during the 10 years temporal window observed, is composed by 12 583 nodes
and 84 512 edges.
6.2 The Phase Transition
Figure 3(b) shows the density values for each element of the temporal sequence
of footprints SF (τ ) of the interactions network of the most proficient scientists
Gi. The time interval τ is fixed to one year.
The density trend starts with very low values and then an increase of the
graph's sparsity occurs during its evolution with a very low counter-trend
during the period between 1999 and 2000.
The growing rate of the modularity, computed on SF (τ ) is shown in 3(c).
It is characterized by an increasing rate until 1993, then it reaches its highest
values during the period between 1999 and 2000, but through a smoothed
rate. As far as we can see by the modularity evolution, the interconnections
10
Walter Quattrociocchi and Frederic Amblard
(a) Average Clustering Coefficient
(b) Density
(c) Modularity
Fig. 3. Collaborations Graph Evolution
among separated groups of authors starts in 1993, then their interconnection
continues, but with a more stable rate. Looking at the curve of the average
clustering coefficient shown in (Figure 3(a)), we can see a phase transition
occurring between 1999 and 2000 and separating a monotone trend from a
decreasing one.
We can interpret the modularity trend as showing that nodes during the
first phase are divided in several and separated groups, while after the phase
transition of the clustering coefficient, the connections among these groups
start to become denser causing a network structure with a smaller number
of larger communities (modules) - e.g. the network tends toward a structural
homogeneity.
6.3 Zooming on Interconnections
In order characterize the phenomena behind the phase transition outlined
in the previous section, in Figure 4 we show the trend of the average ratio
between nodes and edges in both the whole interaction network (in black)
and the network of the most proficient scientists (in red). As we can see, the
phase transition, evinced in the clustering coefficient evolution in Figure 3(a),
is not caused by an increase of the number of authors in the period between
1999 and 2000, neither it is a pattern related to the whole dataset.
Emergence through Selection: The Evolution of a Scientific Challenge
11
Fig. 4. The average number of edges per node for the entire network of the cited
co-authorships (in black) and for the network of the most proficient authors (in red)
In Table 6.3 we present the evolution of the average degree, the average
path length and of the degree power law within the temporal window observed.
As for the previous indicators these values are computed on the sequence of
footprints SF (τ ) with τ fixed to one year of Gi.
Year Average Degree Average Path Length Power Law
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
0,0095
0,0176
0,012
0,0135
0,132
0,0118
0,106
0,066
0,64
0,6
1,16
1,13
1,12
1,12
3,92
3,79
3,82
1
1
1
0
-1,386
-1,79
-2,16
-2,27
-2,5
-2,5
-5,08
-5,27
-5,25
Table 2. Other interaction network's measurements
In bold the values when the phase transition occurs. Neither the average
path length, indicating the average distances among nodes, the power law
degree, measuring how closely the degree distribution of a network follows
a power-law scale and the evolution of average degree, counting the average
number of connections at each node, are immune to the phase transition.
12
Walter Quattrociocchi and Frederic Amblard
7 Goals and Preferential Attachment
As the Time-Varying graphs is an interaction-centric formalism. In this section
we will show how such a modeling approach is compliant with one of the
widely diffused platforms for network analysis and how it is possible to show
the punctual evolution as in a movie of the temporal networks.
The most interesting emerging phenomenon from the previous section is
a phase transition in the evolution of the structure of the most proficient
authors' network occurring between 1999 and 2000. According to the tem-
poral analysis, such changes in the network are caused by a particular trend
regarding the interconnections among nodes (authors) of the most cited co-
authorships graph Gi.
In this section we outline the network evolution by showing the formation
of the biggest community at the beginning of the phase transition (1998) until
its maximum expansion (2002). In Figure 5 we provide a sequence of screen-
shots showing the nodes' aggregation patterns. The pictures are obtained
through the Gephi platform ([2]). At the beginning (Figure 5(a)) there are
several separated components, that start to connect (Figure 5(b)).
(a) Several separated connected
components
(b) that start to connect with each
other
Fig. 5. Connections within the islands
Notice that the edges are emphasized with respect to the strength value
wi counting the number of citations of each couple of nodes. The component
(group of authors) in the center is highly cited and it is playing as an attractor
on the neighboring nodes as it is shown in Figure 6(a) until the maximum level
of connections in the group is reached, as shown in Figure 6(b).
7.1 Zooming on the Attractors
In this section we provide a more detailed vision on such a process of aggre-
gation toward the attractors. Let starts by introducing Figure 7 showing the
Emergence through Selection: The Evolution of a Scientific Challenge
13
(a) The number
connections
within authors continues to increase.
of
(b) The maximum level of con-
nectivity is reached.
Fig. 6. The growing phase of interaction among authors
number of citations received at each semester by the most cited paper in our
dataset.
Fig. 7. the citations trend of the most cited paper for each semester
The citations rate has a strong increase after two semesters. The third
semester coincides with the interval (1999-2000) of the phase transition cap-
tured in the previous section.
Hence, in order to understand the effect of this paper, in the following we
will show a sequence of snapshots of the network structure in the neighbor of
the authors of the most cited paper when it appears in our database. Notice
that in the following pictures, the nodes' diameters are proportional to the
total amount of citations received by their papers.
14
Walter Quattrociocchi and Frederic Amblard
At the beginning there are only separated components as shown in Figure
8(a). Then a large node appears (Figure 8(b)), and near appears a node with
a smaller diameter but with a higher number of links. The biggest node is one
of the authors of the most cited paper and as we can see, the node has a very
low number of connections (collaborations) in that time interval.
(a) Before
(b) one of the authors (the biggest
node) of the most cited paper and a
smaller node with an higher degree
appear
Fig. 8. The appearence of one of the most cited authors
(a) The group of authors of the
most cited papers appears. The au-
thors are the two big nodes and the
smaller hub
(b) The portion of the graph be-
comes denser
Fig. 9. Densification through the hub node
In Figure 9(a) the fat node (a Nobel prize) and the hub node are connected,
they publish a paper together with another node with a large diameter. Several
islands start to link the clique formed and as we can see in Figure 9(b) the
process of diffusion continues by means of new hubs.
The sequence of snapshots shows the interactions patterns behind the for-
mulation of the "String Theory" and of its consequent developments.
Emergence through Selection: The Evolution of a Scientific Challenge
15
Authors in the same community start to migrate toward the island of
the authors of the most cited paper. The increasing community densification
causes the formation of a giant component around the group authoring the
most cited paper that in turns makes the network become denser and homo-
geneous as emerged in the analysis in the previous sections.
It is a goal-driven preferential attachment e.g., the mechanism used to
explain the power law degree distributions in social networks - due to the
number of citations (representing the emergence through selection) to a given
group. Authors tend to join highly cited groups to satisfy both the quality and
the possibility to be highly cited requirements. Moreover, considering that at
the beginning there are several separated groups, the phenomenon can be in-
terpreted as a three-fold process with a first phase as the exploration of ideas
by means of separated works afforded by separated groups, a second one when
a part of the ideas explored starts to be cited more than the others, and a
third one when authors tend to join groups that have produced highly cited
works. The process tripartition resembles the phases of the natural selection,
e.g. the exploration, the selection and migration. In this context such a (social)
selection a) produces self-organization because it is played by a group of in-
dividuals which act, compete and collaborate in order to advance science and
b) determines the success (emergence) of a topic and of the scientists working
on it.
7.2 Characterizing the Community Evolution
Table 7.2 summarizes the network evolution by means of a) basic indicators,
e.g. the number of nodes, the number of edges and the community's diameter)
and b) aggregated indicators, e.g. the cyclomatic number, the alpha, beta,
and gamma index. The cyclomatic number counts the number of cycles on the
graph, its magnitude characterizes the development of the nodes' accessibility.
The alpha index is the ratio between the number of cycles in the graph and
their possible maximum value. The range of the alpha index spread within
0 to 1, that are from no cycles to a completely interconnected network. The
beta index, is a simple measure of connectivity. It relates the total number
of edges to the total number of nodes. The higher the value, the greater the
connectivity is. The gamma index measures the ratio between the number
of edges on the network and the maximum number of possible edges among
nodes. The gamma index spreads within 0 and 100, respectively indicating
the minimum and the maximum number of edges between nodes.
As we can see from the evolution of these parameters, the aggregation
pattern among separated components is evident for each one of the metric
proposed. In terms of nodes that join the community and their mutual con-
nections, the diameter over time passes through a phase of expansion and
then tends to stabilize.
16
Walter Quattrociocchi and Frederic Amblard
Measures April 00 October 00 April 01 October 01 April 02 October 02 April 03
Vertices:
Edges:
67
106
70
110
72
114
65
99
10
35
66
100
10
35
0,017
1,52
52,38
0,016
1,51
52,08
8
40
0,018
1,58
54,3
8
41
0,017
1,57
53,92
8
43
0,017
1,58
54,28
23
29
6
17
0,73
1,69
61,9
51
75
10
25
0,02
1,47
51,02
Diameter:
Cyclomatic:
Alpha:
Beta:
Gamma:
Table 3. network measurement of the biggest community
8 Conclusions
In this paper we characterize the evolution of a scientific community extracted
by the ArXiv's hep-th (High Energy Physics Theory) repository. The analysis
starts with a static vision on the dataset by showing the structure of the cita-
tions and co-authorships graphs derived by the dataset. Then by adding the
temporal dimension on both networks we characterize the structural changes
of the co-authorships and citations graphs. The temporal dimension and the
metrics used for the analysis were formalized using Time-Varying Graphs
(TVG), a mathematical framework designed to represent the interactions and
their evolution in dynamically changing environments.
Since we are interested in the relationships between collaborations and
citations behaviors of scientists, we focus on the network of most cited authors
and on its structural evolution where several interesting aspects emerge. The
network evolves toward a denser structure, a phase transition occurs in the
1999-2000 time interval causing the homogenization of communities.
Through our approach, we capture the role played by famous authors
on co-authorship behaviors. They act as attractors on the community. The
driving force is a sort of preferential attachment driven by the number of
citations received by a given group, that in terms of the goal of any scientific
community indicates a strategy oriented to the community belonging.
Furthermore, the evolution of the network from a sparse and modular
structure to a denser and homogeneous one, can be interpreted as a three-fold
process reflecting the natural selection. The first phase is the exploration of
ideas by means of separated works, once some ideas start to be cited (selected)
more than others, then authors tend to join groups that have produced highly
cited works. The selection is performed by individuals in a goal oriented envi-
ronment and such a (social) selection produces self-organization because it is
played by a group of individuals which act, compete and collaborate in order
to advance Science. In fact, the driving force is an emergent effect of the inter-
dependencies between citations and the goal of the scientific production since
the social selection determines the emergence of a topic and of the scientists
working on it by determining the so called preferential attachment toward
groups and topics having high potential of citations. Finally, we show that
Emergence through Selection: The Evolution of a Scientific Challenge
17
the migration of authors toward the most cited authors (attractors) expresses
through a hub node, - e.g. a node with few citations and several co-authorships.
In the next future we are going to outline the behavior of the most profi-
cient scientist in terms of their aggregation patterns, and on how their works
are diffused within the community, that is, characterizing the reasons behind
the selection process causing the network structural evolution. Such aspects
will be addressed both with new analyses on different datasets and by means
multi-agent simulations. The former stream will be devoted to the definition
of new patterns, the latter will be used for the understanding of how changing
some parameters of the network influences the evolution, and consequently
the quality, of the scientific production.
9 Acknowledgments
This work was partially supported by the Future and Emerging Technolo-
gies programme FP7-COSI-ICT of the European Commission through project
QLectives (grant no.: 231200).
References
1. Jeong H. Neda Z. Ravasz E. Schubert A. Barabasi, A.L. and T. Vicsek. Evolution
of the social network of scientific collaborations. Physica A, vol.311, pages 590 --
614, 2002.
2. M. Bastian, S. Heymann, and M. Jacomy. Gephi: An open source software for
In International AAAI Conference on
exploring and manipulating networks.
Weblogs and Social Media, 2009.
3. C.T Butts and K.M. Carley. Structural change and homeostasis in organizations:
A decision-theoretic approach, 2002.
4. A. Casteigts, P. Flocchini, W. Quattrociocchi, and N. Santoro. Time-varying
graphs and dynamic networks. Technical Report University of Carleton, Canada,
2010.
5. C. Castellano, S. Fortunato, and V. Loreto. Statistical physics of social dynam-
ics. 2007.
6. Jorn Davidsen, Holger Ebel, and Stefan Bornholdt. Emergence of a small world
from local interactions: Modeling acquaintance networks. Phys. Rev. Lett.,
88(12):128701, Mar 2002.
7. G. Deffuant, D. Neau, F. Amblard, and Gerard Weisbuch. Mixing beliefs among
interacting agents. Advances in Complex Systems, 3:87 -- 98, 2001.
8. N. Gilbert. Referencing as persuasion. Social Studies of Science, vol.7, pages
113 -- 22, 1977.
9. R. Guimera, B. Uzzi, J. Spiro, and L.A. Amaral. Team Assembly Mechanisms
Determine Collaboration Network Structure and Team Performance. Science,
308(5722):697 -- 702, 2005.
10. P. Holme. Network reachability of real-world contact sequences. Physical Review
E, 71(4):46119, 2005.
18
Walter Quattrociocchi and Frederic Amblard
11. Norman P. Hummon and Patrick Doreian. Connectivity in a citation network:
The development of dna theory, 1989.
12. Konstantin Klemm and V´ıctor M. Egu´ıluz. Highly clustered scale-free networks.
Physical Review E, 65(3):036123+, Feb 2002.
13. G. Kossinets, J. Kleinberg, and D. Watts. The structure of information pathways
in a social communication network. In Proc. of the 14th ACM SIGKDD Intl.
Conf. on Knowledge Discovery and Data Mining (KDD 2008), pages 435 -- 443,
2008.
14. V. Kostakos. Temporal graphs. Physica A: Statistical Mechanics and its Appli-
cations, 388(6):1007 -- 1023, 2009.
15. M. Lelarge. Diffusion of innovations on random networks: Understanding the
chasm, 2009.
16. Jure Leskovec, Deepayan Chakrabarti, Jon Kleinberg, and Christos Faloutsos.
17. Jure Leskovec, Deepayan Chakrabarti, Jon M. Kleinberg, Christos Faloutsos,
and Zoubin Ghahramani. Kronecker graphs: An approach to modeling networks.
Journal of Machine Learning Research, 11:985 -- 1042, 2010.
18. Jure Leskovec, Jon Kleinberg, and Christos Faloutsos.
19. Jure Leskovec, Jon M. Kleinberg, and Christos Faloutsos. Graph evolution:
Densification and shrinking diameters. TKDD, 1(1), 2007.
20. M.H. MacRoberts and B.R. MacRoberts. Problems of citation analysis. Scien-
tometrics, vol.36, no.3, pages 435 -- 444, 1996.
21. M. Mataric. Designing emergent behaviors: From local interactions to collective
intelligence. In In Proceedings of the International Conference on Simulation of
Adaptive Behavior: From Animals to Animats, volume 2, pages 432 -- 441, 1992.
22. C. Moore and M.E.J. Newman. Epidemics and percolation in small-world net-
works. Phys. Rev. E, 61:5678 -- 5682, 2000.
23. M. E. J. Newman. Clustering and preferential attachment in growing networks.
In Physical Review E, vol.64, 2001.
24. M. E. J. Newman. The structure of scientific collaboration networks. 98(2):404 --
409, January 2001.
25. M. E. J. Newman. Coauthorship networks and patterns of scientific collabo-
ration. In Proceedings of the National Academy of Sciences, pages 5200 -- 5205,
2004.
26. M. E. J. Newman. Who is the best connected scientist? a study of scientific
coauthorship networks. In Complex Networks, lecture notes in Physics, 2004.
27. Matti Peltomaki and Mikko Alava. Correlations in bipartite collaboration net-
works. J.STAT.MECH., page P01010, 2006.
28. W.W Powell, D.R. White, and K.W. Koput. Network dynamics and field evolu-
tion: The growth of interorganizational collaboration in the life sciences. Amer-
ican Journal of Sociology, vol.110, no.4, pages 1132 -- 1205, 2005.
29. D.J. De Solla Price. Networks of scientific papers. Science, vol.149, no.3683,
pages 510 -- 515, 1965.
30. W. Quattrociocchi, R. Conte, and E. Lodi. Simulating opinion dynamics in
heterogeneous communication systems. ECCS 2010 - Lisbon Portugal, 2010.
31. W. Quattrociocchi, M. Paolucci, and R. Conte. On the effects of informational
cheating on social evaluations: image and reputation through gossip. Interna-
tional Journal of Knowledge and Learning, 5(5/6):457 -- 471, 2009.
32. Walter Quattrociocchi and Frederic Amblard. Selection in scientific networks.
arxiv:1012.4396, 2010.
Emergence through Selection: The Evolution of a Scientific Challenge
19
33. F .Radicchi, S. Fortunato, B. Markiness, and A. Vespignani. Diffusion of scien-
tific credits and the ranking of scientists. Physical Review E, vol.80, 2009.
34. S. Redner. Citation statistics from 110 years of physical review. Physical Review,
Physics Today, vol.58, pages 49 -- 54, 2005.
35. N. Santoro, W. Quattrociocchi, P. Flocchini, A. Casteigts, and F. Amblard.
Time varying graphs and social network analysis: Temporal indicators and met-
rics. Technical Report University of Carleton, Canada, 2010.
36. J. Tang, S. Scellato, M. Musolesi, C. Mascolo, and V. Latora. Small-world
behavior in time-varying graphs. Arxiv preprint arXiv:0909.1712, 2009.
37. Carla Taramasco, Jean-Philippe Cointet, and Camille Roth. Academic team
formation as evolving hypergraphs. Scientometrics, April 2010.
38. C.S Wagner and K. Leydesdorff. Network structure, self-organization, and the
growth of international collaboration in science. Research Policy vol 34 n10,
pages 1608 -- 1618, 2005.
39. Duncan J. Watts. Networks, dynamics and the small world phenomenon. AJS,
1999.
40. David R. Woolley. PLATO: The Emergence of Online Community, 1994.
|
1905.10348 | 1 | 1905 | 2019-04-20T15:55:13 | Predicting Brazilian court decisions | [
"cs.SI",
"cs.CL"
] | Predicting case outcomes is useful but still an extremely hard task for attorneys and other Law professionals. It is not easy to search case information to extract valuable information as this requires dealing with huge data sets and their complexity. For instance, the complexity of Brazil legal system along with the high litigation rates makes this problem even harder. This paper introduces an approach for predicting Brazilian court decisions which is also able to predict whether the decision will be unanimous. We developed a working prototype which performs 79% of accuracy (F1-score) on a data set composed of 4,043 cases from a Brazilian court. To our knowledge, this is the first study to forecast judge decisions in Brazil. | cs.SI | cs | Predicting Brazilian court decisions
Andr´e Lage-Freitas, H´ector Allende-Cid, Orivaldo Santana, and L´ıvia de Oliveira-Lage
[email protected], [email protected], [email protected], [email protected]
1
Abstract -- Predicting case outcomes is useful but still an
extremely hard task for attorneys and other Law professionals.
It is not easy to search case information to extract valuable
information as this requires dealing with huge data sets and
their complexity. For instance, the complexity of Brazil legal
system along with the high litigation rates makes this problem
even harder.
This paper introduces an approach for predicting Brazilian
court decisions which is also able to predict whether the decision
will be unanimous. We developed a working prototype which
performs 79% of accuracy (F1-score) on a data set composed of
4, 043 cases from a Brazilian court. To our knowledge, this is the
first study to forecast judge decisions in Brazil.
Index Terms -- jurimetrics, legal outcome forecast, predictive
algorithms
I. INTRODUCTION
Since Code of Hammurabi1, we have been trying to improve
legal certainty in human relationships by making public the
law and the rulings of courts. In addition to publicizing the
laws, legal systems usually provide further support to legal
certainty through judicial decisions. These decisions are useful
not only for judging specific situations, but also to influence
society behavior by exposing the legal consequences of our
actions. Thereby, predicting legal decisions is fundamental to
understand the legal consequences of our actions as well as
for supporting law professionals to improve the quality of their
work.
In Brazil for example, lower court judges decisions might be
appealed to Brazilian courts (Tribiunais de Justic¸a) to be re-
viewed by second instance court judges. In an appellate court,
judges decide together upon a case and their decisions are
compiled in Agreement reports named Ac´ordaos. Similar to
lower court decisions, Ac´ordaos include Report, Fundamentos,
Votes2, and further metadata such as judgment date, attorneys,
judges, etc. These Agreements documents are very useful for
understanding jurisprudence thus guiding lawyers and court
members about the decisions. For instance, attorneys often
use these documents to prepare cases while judges should take
9
1
0
2
r
p
A
0
2
]
I
S
.
s
c
[
1
v
8
4
3
0
1
.
5
0
9
1
:
v
i
X
r
a
Andr´e Lage-Freitas is with Universidade Federal de Alagoas.
H´ector Allende-Cid is with Pontificia Universidad Catlica de Valparaso.
Orivaldo Santana is with Universidade Federal do Rio Grande do Norte.
L´ıvia de Oliveira-Lage is a Prosecutor at States Attorneys Office at
Procuradoria Geral do Estado de Alagoas.
April 20th, 2019. Authorship: A.L.F. proposed the research problem,
supervised this work, and executed the experiments. H.A.C. proposed and
developed NLP and machine learning algorithms. O.S. verified the method-
ology. L.O.L. supervised and reviewed the Law aspects of this work. All
authors discussed the results and contributed to the final manuscript. The
authors declare that have no competing interests and that received no funding
for this study.
1https://en.wikipedia.org/wiki/Code of Hammurabi
2C.f. Brazilian Law: Art. 489, Lei 13105/15.
them into account -- or even use them as guidelines -- for next
decisions.
In order to understand Ac´ordao decisions, one has to read
the subject at the summary, read the decision Report, how each
judge voted in this case (Votes), and the final decision which
can be unanimous or not. Moreover, each Ac´ordao might have
more than one decision -- regarding one or more appealed case
claims -- which can increase the Ac´ordao complexity. This
problem becomes harder as there usually are hundreds -- and
sometimes thousands -- of Ac´ordaos related to the case on
which a Law professional is working.
A very common and extremely important task for Law
professionals is to speculate how a specific court would decide
given the ideas and the facts which compose the case. For
example, this is useful for preparing and tuning a case to
have a favourable decision. Hence, attorneys can rely on
substantial assumptions on how judges will decide based on
their arguments. Although this information can be found in
public Ac´ordaos, the myriads of available documents make
this task very complex and error prone, even for experiment
lawyers.
In addition to Brazil, several other legal systems in the world
share the very same problem of predicting legal decisions.
The challenge is hence generalized as how to predict legal
decisions with a satisfactory level of accuracy to support the
work of attorneys, judges, and other professionals such as
counters and real state offices. By satisfactory, we mean that
the quality of the prediction in terms of accuracy should be
better -- or even higher -- than Law experts.
it still
Nevertheless,
is very hard to perform any legal
decision prediction with satisfactory accuracy, even though
computers have been used for such challenge for decades [7].
For instance, Ashley and Bruninghaus [2] propose a method
for classifying and predicting cases which is able to meet
91.8% of accuracy, however the evaluation takes only into
account a small data set (146 cases). Katz et al. [6] uses
historical data for predicting USA Supreme Court decisions
by classifying decisions in two and three classes and by
presenting judge profiles. That approach reaches 70.2% of
accuracy for predicting case decisions and is assessed by on
a data set with 28, 000 cases. Also using data from the USA
Supreme Court, Ruger et al. [11] exposes how the prediction
of Law experts performs in comparison to a trained statistical
model for different Law fields by using less than 200 cases.
In [1], Aletras et al. uses Support-vector Machine (SVM) for
predicting if cases violate Articles 3, 6, and 8 of European
Convention on Human Rights of Human Rights. Their results
acheived 78% of accuracy performed on 584 European Court
cases separated by subjects.
Other related work takes advantage of machine learning
techniques to support further legal tasks. In [8], the authors
propose a framework for automatically judging legal decisions
by using Attention neural network models. They applied the
approach for divorce decisions in China. In [13], Shulayeva
et al. separates legal principles from case facts on legal docu-
ments by using a Naive Bayseian Multimodal classifier. In [4],
the authors proposes to use transfer learning to recognize
the same words which have different meanings in different
contexts, i. e., name-entity linking task. In [3], the authors uses
Bayesian networks to classify legal decisions from a Brazilian
Labor court and conclude that both employees and employers
are roughly successful in their litigation. Last, Ruhl et al. [12]
overviews some perspectives on how complex systems are
useful for supporting policy-makers on legal-related topics
such as appellate jurisprudence and tax policy analysis.
We are also motivated by recent results that show that
intelligent systems can perform better than Law experts3. Our
hypothesis is that by taking advantage of Natural Language
Processing (NLP) and Machine Learning techniques it
is
able to build a system that meets high quality legal decision
predictions. Different from the closest related works of this
paper [1], [2], [6] which address United States and European
courts, we propose an approach for legal decision prediction
for Brazilian courts which also predicts whether the court
decision will be unanimous. Moreover, in contrast to [1], [2],
we trained a model at thousand-scale data set with 4, 043
cases. Moreover, in contrast to [1], our approach does not
rely on a binary classification problem -- since it uses three
possible prediction results -- nor require that case data set
should be separated by specific Law articles, hence being a
more generic approach.
The reminder of this paper is structured as follows. In
Section II, we present details on the aforementioned prob-
lem such as the case study and the methodology employed.
Section III exposes the results while Section IV concludes our
investigation and proposes future directions on this subject.
II. MATERIAL AND METHODS
The research question which guides our study is how to
predict legal decisions with a satisfactory level of accuracy
for Brazilian courts by including the prediction of the court
unanimous behavior. Next sections provide further informa-
tion about our assumptions and the proposed methodology.
A. A generic approach
We focus on Brazilian courts as Brazil legal system is not
trivial. We believe that if we are able to solve this problem
for such complex legal system, our approach would also fit
other simpler or it could be straightforwardly adapted for more
complex legal systems which share similarities. Nevertheless,
it
is worth to state that other legal systems also rely on
similar documents. For instance, in Indiana Court of Appeal
(United States), the Appellate Court decisions are composed
by a group of three judges whose decisions (opinions) are
divided in Case Summary, Facts, Procedural History, and the
3https://www.bbc.com/news/technology-41829534
2
court conclusion at the end of that document. In France, the
Appellate Court (Cour d'appel) also renders decisions coming
from the agreement of three judges. That decision is called an
Arret whose structure is also composed of legal basis for the
appeal, case history, and the final decision.
Further, we share the same assumption of Aletras et. al [1]:
"there is enough similarity between (at least) certain chunks of
the text of published judgments and applications lodged with
the Court and/or briefs submitted by parties with respect to
pending cases".
B. Decision labels and data set
Regarding the flow process of a Brazilian appeal, when
lawyers lodge appeals at a court it is analyzed by a group
composed of three judges to check whether the appeal is able
to be judged by the court. If the appeal does not meet the
formal requirements, the appeal is identified as not cognized
(recurso nao conhecido) hence not
judged by the court.
Otherwise, the appeal is therefore judged and classified in
various categories. We therefore assumed that court decisions
can be classified by using the following labels:
• not-cognized, when the appeal was not accepted to
be judged by the court;
• yes, for full favourable decisions;
• partial, for partially favourable decisions;
• no, when the appeal was denied;
• prejudicada,
to mean that
the case could not be
judged for any impediment such as the appealer died or
gave up on the case for instance;
• administrative, when the decision refers to a court
administrative subject as conflict of competence between
lower court judges.
In addition to the decision labels, an orthogonal concern of
Brazilian court decisions -- as well as for other legal systems
-- refers to its unanimity aspect, being labeled as:
• unanimity which means that the decision was unani-
mous among the three judges that voted in the case; and
• not-unanimity by meaning that one of the judges
disagreed on the decision.
With respect to the data set, we relied on 4, 762 decisions
(Ac´ordaos) from a State higher court (appellate court), the
Tribunal de Justic¸a de Alagoas. From this data set, we re-
moved the decisions that had repeated descriptions to not bias
the sample thus resulting 4, 332 examples. Repeated decision
descriptions occur owing to very similar cases which share
the same description. Moreover, for the sake of predictability,
we removed all the decisions classified as prejudicada,
not-cognized and admnistrative as these labels refer
to very peculiar situations which are not useful for prediction
purposes addressed by this paper. Finally, the total amount of
examples were 4, 043 cases.
C. Methodology
Figure 1 depicts an overview of our methodology. From
the legal decision data set, we extracted and separated the
texts which hold information about the case description, the
decision, and their unanimity aspect. This Natural Language
Pre-processing task includes removing stop-words and word
suffixes for improving the capacity of word representation.
Then, we took advantage of Term Frequency-inverse Docu-
ment Frequency (tdf-df) statistics to increase the importance
of relevant words while decreasing the importance of general
repetitive words not relevant to the addressed problem. As
follows, we used texts which refers to decisions and unanimity
to classify them to one of the possible labels (c.f. Section II-B).
As a result, we built a structured training data set depicted by
Table I.
Next, we used 80% of the data set to train a Machine
Learning model which was then assessed by using the latter
20% of the data set (c.f. Figure 1). To train a model means to
automatically find out which parameters are the most suitable
for predicting decisions based on the training data set. Because
we address decision and unanimous predictions, it requires to
train two models to address both predictions. Once trained, the
models can be used to predict decision and unanimity given
a case description. Last, to evaluate our approach, we used
the F1-score metrics and performed 5-fold cross-validation to
improve the practicability of our approach. Results are thus
exposed as success accuracy rate in percentage.
3
label. The accuracy of case outcome prediction in this situation
was 74.07% (σ2 = 0.00029) for the F1-score metrics. This
result is very interesting as it does not validate our previous
assumption that the not regularly distributed data set would
strongly bias the model.
With respect to predicting the unanimous behaviour of the
Tribunal de Justic¸a de Alagoas Brazilian court, our approach
scored 98.46% (σ2 = 0.000031) for the F1-score metrics.
This assessment was performed in a data set with 2, 274
cases. From the 4, 332 data set -- which had no repeated
decision descriptions -- , we removed the samples that either
our classifier did not managed to label or the decision itself
did not had any information about unanimity. The resulting
data set had 2, 289 samples. As follows, we removed from
this data set the decision whose labels were prejudicada,
not-cognized and admnistrative -- as these labels
are not relevant for the predictive addressed problem -- re-
sulting in a data set with 2, 274 examples. The distribution of
unanimity and not-unanimity labels are depicted by
Table IV.
The very-high unanimous predictive accuracy of 98.46% is
explained by the fact that most of decisions are unanimous,
therefore the model was biased to this label. We indeed
expected that most decisions were unanimous since this is well
known by law experts. However, the great difference between
unanimous and non unanimous decision is a surprising result.
In order to understand how our approach would perform when
predicting unanimity by using a more uniformly distributed
data set, we therefore performed another evaluation by ran-
domly removing decisions whose label was unanimity to
meet the same number of not-unanimity decisions. The
resulting data set had 90 examples, half of them labeled as
unanimity and the other half not-unanimity. With this
configuration, our prototype reached 76.94% (σ2 = 0.015)
F1-score accuracy.
Fig. 1. Methodology. Specific data from case outcomes are extracted then
classified to different labels. A machine learning model is hence trained and
assessed by using 5-fold cross-validation.
Furthermore, in order to assess the exposed methodology,
we developed a working prototype in Python. We used the
NLTK framework [9] for Natural Language Processing in such
a way that our prototype is easily configurable for various lan-
guages in addition to Portuguese. The prototype also provides
a graphical user interface which can be accessible from any
Web browser.
III. RESULTS
Our approach was able to score 78,99% F1-score (σ2 =
0.000017) when predicting legal decision for the Tribunal
de Justic¸a de Alagoas Brazilian court by using 4, 043 judge
decisions. The number of samples for each label is exposed
in Table II.
In order to analyze our prediction over a more uniformly
distributed data set, we randomly removed 1, 549 no-labeled
decisions to have the same number of partial-labeled
decisions. Table III depicts the distributions of each decision
IV. CONCLUSION
This paper proposes a methodology for predicting Brazilian
court legal decisions which is able to reach 79% of accuracy
when employed for a Brazilian court data set with 4, 043
cases. Our approach is able to predict case outcomes by using
three different labels: yes, no, partial. Moreover, the proposed
method also predicts whether the decision will be unanimous,
which fits not only Brazil legal system, but also several others
whose decisions are judged by more than one judge. The
unanimity prediction performance of our approach is 77% of
accuracy. To our knowledge, this is the first study to predict
Brazil legal decisions.
Moreover, our approach is easy to use as it only requires that
users provide the description of their litigation and the output
will be one of the aforementioned case outcome predictive
label along with its predictive unanimity label. These infor-
mation are very useful for attorneys, judges, and other Law
professionals as they provide practical support for their work.
Moreover, our contribution also includes a working prototype
which can be configured to further languages as well as for
different data sets.
Data
Sample 1
Sample 2
Sample 3
Decision description
Direito Processual Civil...
Apelac¸ao criminal...
Apelac¸ao C´ıvel em Ac¸ao Ordin´aria...
Decision
Recurso conhecido e provido
Recurso conhecido e parcialmente provido
Recurso conhecido e nao provido
Unanimity
Unanimidade
Decisao unnime
Decisao unanime
Label
yes
patial
no
Unanimity label
unanimity
unanimity
unanimity
4
TRAINING DATA SET INCLUDES DECISION TEXTS AND LABELS WHICH WERE CLASSIFIED ACCORDING TO RESPECTIVE DECISION TEXTS. E.G., IN
SAMPLE 1, provido WAS CLASSIFIED AS A FAVORABLE ( Y E S) DECISION AND Unanimidade WAS CLASSIFIED AS U N A N I M I T Y.
TABLE I
[6] D. M. Katz, M. J. Bommarito, and J. Blackman. A general approach
for predicting the behavior of the Supreme Court of the United States.
PLOS ONE, 12(4):e0174698, apr 2017.
[7] L. Loevinger. Jurimetrics: The Methodology of Legal Inquiry. Law and
Contemporary Problems, 28(1):5, 1963.
[8] S. Long, C. Tu, Z. Liu, and M. Sun. Automatic Judgment Prediction via
Legal Reading Comprehension. Technical report, Tsinghua University,
Beijing, sep 2018.
[9] E. Loper and S. Bird. NLTK. In Proceedings of the ACL-02 Workshop
on Effective tools and methodologies for teaching natural language
processing and computational linguistics -, volume 1, pages 63 -- 70,
Morristown, NJ, USA, 2002. Association for Computational Linguistics.
[10] P. H. Luz de Araujo, T. E. de Campos, R. R. R. de Oliveira, M. Stauffer,
S. Couto, and P. Bermejo. LeNER-Br: A Dataset for Named Entity
Recognition in Brazilian Legal Text. In Proceedings of the International
Conference on the Computational Processing of Portuguese, pages 313 --
323, Canela, 2018.
[11] T. W. Ruger, P. T. Kim, A. D. Martin, and K. M. Quinn. The Supreme
Court Forecasting Project: Legal and Political Science Approaches to
Predicting Supreme Court Decisionmaking. Columbia Law Review,
104(4):1150, may 2004.
[12] J. B. Ruhl, D. M. Katz, and M. J. Bommarito. Harnessing legal
complexity. Science, 355(6332):1377 -- 1378, mar 2017.
[13] O. Shulayeva, A. Siddharthan, and A. Wyner. Recognizing cited facts
and principles in legal judgements. Artificial Intelligence and Law,
25(1):107 -- 126, mar 2017.
Labels
N. of decisions
partial
866
yes
762
no
2,415
TABLE II
DISTRIBUTION OF DECISIONS ACCORDING TO THEIR LABELS.
Labels
N. of decisions
no
866
partial
866
yes
762
TABLE III
DISTRIBUTION OF DECISIONS ACCORDING TO THEIR LABELS WHEN
RANDOMLY REMOVING N O-LABELED DECISION SAMPLES TO CREATE A
REGULAR DISTRIBUTED DATA SET.
Labels
N. of decisions
not-unanimity
45
unanimity
2,229
DISTRIBUTION OF DECISIONS ACCORDING TO JUDGE UNANIMOUS
TABLE IV
BEHAVIOR.
Although we believe that our contribution is quite satisfac-
tory given the accuracy rate aforementioned, future investiga-
tions might consider comparing our results with Law experts,
as performed in [11] and by current Lawtech products such
as Case Crunch and LawGeex4. Other future work includes
to investigate whether taking advantage of existent Named-
entity recognition data sets for Brazilian law documents [10]
improve the prediction quality. Furthermore, the assessment
of the proposed method can be performed on larger and/or
different data sets, such as the European Court of Human
Rights for instance. Ultimately, despite the various directions
one might take to leverage our work, we believe that Mireille
Hildebrandt's viewpoint on "agnostic machine learning" and
its consequences to the Rule of Law [5] should be taken into
account when designing and using a legal predictive system.
REFERENCES
[1] N. Aletras, D. Tsarapatsanis, D. Preoiuc-Pietro, and V. Lampos. Predict-
ing judicial decisions of the European Court of Human Rights: a Natural
Language Processing perspective. PeerJ Computer Science, 2:e93, oct
2016.
[2] K. D. Ashley and S. Bruninghaus. Automatically classifying case texts
and predicting outcomes. Artificial Intelligence and Law, 17(2):125 --
165, jun 2009.
[3] R. Barros, F. Lorenzi, and L. K. Wives. Recent Trends and Future
Technology in Applied Intelligence, volume 10868 of Lecture Notes in
Computer Science. Springer International Publishing, Cham, 2018.
[4] A. Elnaggar, R. Otto, and F. Matthes. Named-Entity Linking Using
Deep Learning For Legal Documents: A Transfer Learning Approach.
Technical report, Technische Universitat, Munchen, 2018.
[5] M. Hildebrandt. Algorithmic regulation and the rule of law. Philosoph-
ical Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences, 376(2128):20170355, sep 2018.
4https://www.artificiallawyer.com/2018/02/26/lawgeex-hits-94-accuracy-in-
nda-review-vs-85-for-human-lawyers/
|
1811.08366 | 1 | 1811 | 2018-11-20T17:01:16 | Temporal Graph Offset Reconstruction: Towards Temporally Robust Graph Representation Learning | [
"cs.SI",
"cs.LG"
] | Graphs are a commonly used construct for representing relationships between elements in complex high dimensional datasets. Many real-world phenomenon are dynamic in nature, meaning that any graph used to represent them is inherently temporal. However, many of the machine learning models designed to capture knowledge about the structure of these graphs ignore this rich temporal information when creating representations of the graph. This results in models which do not perform well when used to make predictions about the future state of the graph -- especially when the delta between time stamps is not small. In this work, we explore a novel training procedure and an associated unsupervised model which creates graph representations optimised to predict the future state of the graph. We make use of graph convolutional neural networks to encode the graph into a latent representation, which we then use to train our temporal offset reconstruction method, inspired by auto-encoders, to predict a later time point -- multiple time steps into the future. Using our method, we demonstrate superior performance for the task of future link prediction compared with none-temporal state-of-the-art baselines. We show our approach to be capable of outperforming non-temporal baselines by 38% on a real world dataset. | cs.SI | cs |
Temporal Graph Offset Reconstruction: Towards Temporally Robust Graph
Representation Learning
Stephen Bonner †, John Brennan †, Ibad Kureshi ¶, Georgios Theodoropoulos §,
Andrew Stephen McGough ‡ and Boguslaw Obara †
†Department of Computer Science, Durham University, Durham, UK
{s.a.r.bonner, j.d.brennan, ibad.kureshi, boguslaw.obara}@durham.ac.uk
‡ School of Computing, Newcastle University, Newcastle, UK [email protected]
§ School of Computer Science and Engineering, SUSTech, Shenzhen, China [email protected]
¶ InlecomSystems, Brussels, Belgium [email protected]
Abstract -- Graphs are a commonly used construct for rep-
resenting relationships between elements in complex high di-
mensional datasets. Many real-world phenomenon are dynamic
in nature, meaning that any graph used to represent them is
inherently temporal. However, many of the machine learning
models designed to capture knowledge about the structure
of these graphs ignore this rich temporal information when
creating representations of the graph. This results in models
which do not perform well when used to make predictions
about the future state of the graph -- especially when the delta
between time stamps is not small. In this work, we explore
a novel training procedure and an associated unsupervised
model which creates graph representations optimised to predict
the future state of the graph. We make use of graph convo-
lutional neural networks to encode the graph into a latent
representation, which we then use to train our temporal offset
reconstruction method, inspired by auto-encoders, to predict
a later time point -- multiple time steps into the future. Using
our method, we demonstrate superior performance for the task
of future link prediction compared with none-temporal state-
of-the-art baselines. We show our approach to be capable of
outperforming non-temporal baselines by 38% on a real world
dataset.
Keywords-graph representation learning; dynamic graphs;
I. INTRODUCTION
Graphs have become a common metaphor in numerous
scientific domains for representing high-dimensional com-
plex data which is loosely topologically related, common
examples being social media networks or citation networks.
As such they provide a useful abstraction for how data is
related. Graphs allow for complex analysis to be performed
such as identifying the missing link within a graph (a
person whom you might know or that paper you must
read), however, to date almost all of the prediction work
which has been performed on graphs has been focused
on analysis in the topological domain as opposed to the
temporal domain. This is strange as almost all graphs change
with time (making new fiends or publishing new papers).
We formally define a graph G = (V, E) as a finite set of
vertices V , with a corresponding set of of edges E. Elements
of E are unordered tuples {u, v} where u, v ∈ V . Elements
in V and E may have a label or some associated features,
although these are not needed for this work. In order to
perform analysis on graphs we need a mechanism which
converts the formal graph representation into a format which
is amenable for machine learning -- graph representation
learning, sometimes referred to as graph embeddings.
Recently, the field of graph representation learning has
received a lot of interest as a way of analysing high-
dimensional graphs via the use of machine learning. Graph
representation learning, comprises a set of techniques which
learn latent representations of a graph, which can then be
used as input to machine learning models for downstream
prediction tasks [1]. The majority of graph representation
learning techniques have focused upon learning vertex em-
beddings [2] and reconstructing missing edges [1]. As such,
the goal of graph representation learning is to learn some
function f : V → Rd which maps from the set of vertices V
to a set of embeddings of the vertices, where d is the required
dimensionality. Resulting in f being a matrix of dimensions
V by d, i.e. an embedding of size d for each vertex in
the graph. However, the majority of graph representation
learning approaches to date ignore the temporal aspect of
dynamic graphs, resulting in models which have no concept
of how the graph will change in the future.
There are many cases in the real-world where a machine
learning model will be trained using historical data and
then used for inference at a later point in time. A primary
example of this being the recommender systems literature,
where graphs can be used to model the relationship between
users and items [3]. Here a model is trained to recommend
items to users based on historically collected data, and
then used to make predictions about newly arriving data.
However, the underlying graph structure is dynamic and
could undergo large changes between time points, leading
to the existing model suffering from degraded predictive
performance, forcing the model to be retrained.
This paper introduces a new training procedure and as-
sociated model which can be used to create vertex level
representations which encode information about how the
graph will evolve into the future,
leading to a model
which is more robust to temporal change. To achieve this,
we introduce the temporal offset reconstruction method to
create graph representations which are explicitly designed
to predict the next time point for a dynamic graph. We
show that this offset method results in vertex representations
which perform better when used to make predictions on later
time points. Further, we make use of graph convolutional
neural networks [4], combined with our temporal offset
reconstruction method to show that graph convolutions can
be used to capture the dynamics of an evolving graph dataset.
To the best of our knowledge, this is the first time this has
been done. In summary, we make the following contributions
during this paper:
• Temporal Offset Reconstruction - We introduce the
temporal offset reconstruction method to create vertex
representations which contain temporal, in addition to
structural, information about the vertex. Achieved by
explicitly training a model to predict the future rather
than the current state of the graph.
• Graph Convolution-based Model - We explore the
use of unsupervised graph convolutional networks to
capture graph dynamics.
• Detailed Experimentation - We perform detailed ex-
perimentation on the task of future link prediction
using benchmark graph datasets, containing synthetic
and empirical evolutionary graphs.
To aid the reproducibility of our work, we open-source
all of our PyTorch [5] based source-code1 and experimen-
tation scripts, as well as presenting results upon benchmark
datasets.
The remainder of this paper is structured as follows: In
Section II we review the relevant literature, Section III we
discuss our methodology and introduce our temporal offset
approach, Section IV details the experimental setup, Section
V presents the results and Section VI draws conclusions.
II. RELATED WORKS
A. Graph Representation Learning
Traditionally graph representations were created via tech-
niques based on matrix factorization, where a mapping to
a lower dimensional space is found such that pair-wise
relationships in the original graph are preserved. Examples
of such approaches include Laplican eigenmaps [6], Graph
Factorization [7], GraGrep [8] and HOPE [9]. More recently,
models originally designed for Natural Language Processing
(NLP) have been adapted to create graph embeddings. Such
approaches use random walks to create 'sentences' which
can be used as input to language inspired models such as
Word2Vec [10]. NLP inspired graph embedding approaches
include DeepWalk [11], Node2Vec [1] and Hyperbolic em-
beddings [12].
1https://github.com/sbonner0/temporal-offset-reconstruction
Recently graph specific neural network based models
have been created which are inspired by Convolutional
Neural Networks (CNNs) from the field of computer vision.
Such approaches attempt to create a differential model for
learning directly from graph structures. Many Graph CNN
approaches operate in the spectral domain of the graph, using
eigenvectors derived from the Laplacian matrix of a graph
[4]. Early approaches to define convolution operators on
graphs often had large memory and computation complexi-
ties and were thus unsuited to many real world graphs [13].
Later more efficient spectral methods were proposed
which reduced the complexity of the filtering operations
whilst still operating on the entire adjacency matrix [4],
[14]. The Graph Convolutional Network (GCN) approach
has proven to be particularly effective [4]. The GCN ap-
proaches uses a layer-wise propagation rule to aggregate
information from a vertices 1-hop neighbourhood to create
it's representation, this layer-wise rule can be stacked k
times to aggregate information from k-hops away from a
given vertex. The requirement to have the whole adjacency
matrix available in memory means that the GCN approach
struggles to scale to massive graphs. To tackle this problem,
Graph-Sage [15] learns to aggregate features from a fixed
size sample of a vertices neighbourhood and as such can
be applied to new vertices which have joined the graph.
However, the approach mandates that all vertices in the
graph have features available and the performance can vary
depending upon the neighbourhood sampling strategy [16].
Thus far, all the graph specific models discussed have
been supervised approaches, requiring the graph's to have
labels. There have been a modest selection of unsuper-
vised graph specific neural models, many are based on
auto-encoders - a type of neural network whose task is
to reconstruct the input data after being projected into a
lower-dimension [17]. Structural Deep Network Embedding
(SDNE) uses an auto-encoder to reconstruct each row in a
graph's adjacency matrix [18]. A more recent auto-encoder
employs a generative model to adversarially regularise the
embeddings to help improve performance [19]. Work has
also explored the use of GCNs as the basis of a convolutional
auto-encoder model [20], producing state-of-the-art results
for link-prediction in citation graphs.
B. Temporal Embeddings
1) Natural Language Temporal Embeddings: Some ap-
proaches from temporal language based models are created
to model how word use evolves over time. For example,
work has been performed to use cosine similarity to au-
tomatically measure how a word changes, relative to it's
neighbours, over time and identify anomalies [21]. To over-
come this non-convex problem, work has been performed in
creating diachronic word embedding by aligning different
embedding snapshots using orthogonal Procrustes, making
the learning not end to end [22]. In later work, a dynamic
Word2Vec model for word embedding is created which
attempts to solve the non-convex problem common to em-
beddings via Bayesian variational black-box inference [23].
The approach creates embeddings which change smoothly
over time and are better able to predict the change in context
for a given word than previous methods.
2) Graph Specific Temporal Embeddings: To date, there
have been few attempts to consider the temporal change
of a graph when creating its embedding. However the
existing approaches can broadly be split into two categories:
Temporal Walk and Adjacency Matrix Factorisation based.
Nguyen et al. [26], propose a model
to incorporate
temporal information when creating graph embeddings via
random walks by capturing individual changes (edge ad-
dition/deletion for example) within a graph. The authors
propose a temporal random walk to create the input data,
however their approach creates more complex and rich
Temporal Walk-Based Approaches -
Many of the temporal graph embedding approaches which
exist are based on data created via temporal walks, which are
random walks over dynamic graphs. Perhaps the first such
approach is that of STWalk [24]. In this work, the authors
aim to learn node trajectories via the use of random walks
which learn representations that considers all the previous
time-steps of a temporal graph. In the best performing
approach presented, the authors learn two representations
for a given vertex simultaneously which are concatenated
to create the final embedding. The first representation is a
normal DeepWalk embedding designed to capture the spatial
information for a vertex. Wile the second representation is
learned across a specially constructed graph structure, where
each vertex is connected to it's 1-hop neighbourhood from
each previous time-step. However the approach is not end-
to-end and requires the user to manually chose how many
time steps to consider.
Yu et al. [25], propose NetWalk, a vertex-level dynamic
graph embedding model using random walks designed to
facilitate anomaly detection in streaming graphs. The ap-
proach captures a collection of short random walks from
the graph which are then passed into an auto-encoder based
model to create the vertex representations. In addition to
the usual reconstruction based loss term, an additional term
is added to minimise the pair-wise distance between the
representations of vertices occurring within the same walk.
To apply this approach to the domain of streaming graphs,
where changes to graph are being made online, the approach
maintains for each vertex a list of neighbouring vertices
which is updated as the graph changes. If changes in a
vertices neighbour list occur, new random walks will be gen-
erated and the representations updated. The final anomaly
detection is performed via a dynamic clustering model on
the vertex representations. However, the created embeddings
are not capable of capturing temporal dynamics or reliably
predicting the future state of a graph.
temporal walks via a biasing process. The approach can be
used to add temporal information into any embedding model
which relies on random walks as input data, however the
paper explicitly explores a model based on the Skip-Gram
architecture and shows the predictive performance increases
over non-temporal baselines.
Adjacency Matrix Factorisation Approaches -
Goyal et al. [27], propose a model for creating dynamic
graph embeddings, entitled DynGEM. In this approach
they extend the auto-encoder graph embedding model of
SDNE [18] to consider dynamic graphs. To do this, they
use a method similar to Net2net [28], which is designed to
transfer the learned knowledge from one neural network to
a second model. This technique allows them to add more
neurons to the auto-encoder, appropriate to the increasing
graph size, via a heuristic approach entitled PropSize. The
use of the Net2net technique means that the model can be ex-
panded while ensuring the learned function is approximately
preserved. The process for training the model is as follows,
at the fist time snapshot of the graph, a complete graph auto-
encoder is trained as described in the SDNE paper [18]. For
the next time step of the model, a new auto encoder is trained
reusing the weights from the previous step, with any new
neurons being added according to the PropSize heuristic.
However the approach does not explicitly predict the future
state of the graph, rather, it transfers knowledge from the
previous timestamp to help the current auto-encoder better
reconstruct the current time-step.
Additionally there have been attempts to incorporate tem-
poral aspects into GCNs [29], [30], however both approaches
focus upon supervised learning and do not explicitly use the
models to predict the future graph state.
III. METHOD
A. Motivation
Many empirical graph datasets comprise a stream of con-
tinually evolving changes, as connections within the graph
are formed and broken. One can consider the graph then as
a series of snapshots, with each snapshot contains the graph
as it is at that point in time. More formally, we can redefine
a graph G to be a temporal graph G(cid:48) = {G0, G1, ..., Gt},
where each graph snapshot Gi (I ∈ [0, t]) contains a
corresponding vertex set Vi and edge set Ei.
In many real-world use cases of machine learning, a
model is trained on historical data and then used to make
predictions about new events at a future point
in time.
An example of where this practice is common is in the
recommender systems industry where recent state-of-the-art
systems, for recommending items to users, are based on
graph convolutions [3], [31]. However, to date, the majority
of models for creating graph representations do not consider
how the graph evolves over time. This could potentially re-
sult in models which have good initial predictive capability,
but whose performance will degrade as the graph continues
to change over time.
B. Temporal Graph Offset Reconstruction
We propose to tackle the challenging problem of creating
temporal robust graph embeddings by training a model to
explicitly recreate a future time step of the graph. We name
this approach temporal offset reconstruction as the input and
target graph are offset by at least one time point. More
concretely, a graph Gi is used as input to model θ(Gi)
which learns a representation for each vertex in Gi such that
it's output can accurately predict the graph Gi+δ. Ideally,
we want to create a model θ(Gi) which can perform this
temporal offset reconstruction using the graphs Gi and Gi+δ
alone, Gi+δ = θ(Gi), requiring no pre-processing steps
which could affect the models performance (e.g. random
walk procedures), no pre-computed vertex features and no
labels required or used.
In the remainder of this section, we will detail the graph
convolutions used to create the vertex representations, the
models we explore to perform the temporal offset recon-
struction and the training procedure.
C. Graph Convolutions
To perform the graph encoding to create the initial vertex
representations, we propose to utilise the spectral graph
convolution networks introduced by Kipf [4]. GCNs are fully
differential end-to-end models for learning from graphs,
meaning that no pre-processing step is required.
One can consider a GCN to be a differentiable func-
tion for aggregating information from a vertices immediate
neighbourhood [15], [16]. The GCN framework introduced
by Kipf produces a representation Z for each vertex in a
graph, provided in the form of it's adjacency matrix2 A
with a set of vertex level features X, taking the form of
Z = GCN (A, X). When no vertex features are present the
identity matrix I of A can be used such that X = I. As with
any deep neural network, a GCN can contain many layers
which aggregate the data. The operation performed at each
layer by the GCN is as follows [4]:
GCN (l)(H (l), A) = σ( AH (l)W (l)) ,
(1)
where l is the current layer number, W (l) is the weight
matrix of that layer, H are the features computed at the
previous layer or is equal to X at l = 0, and σ is a non-linear
activation function like the rectified linear unit (ReLU). To
help with feature scaling, A is the matrix A normalised by
it's degree matrix D and it's identity matrix I such that
A = (D− 1
2 (A + I)D− 1
2 ) [4].
Essentially, we can consider this GCN function to be ag-
gregating a weighted average of the neighbourhood features
for each vertex in the graph. Stacking multiple GCN layers
2The adjacency matrix A = (ai,j ) for a graph is symmetric matrix of
size V by V , where (ai,j ) is 1 if an edge is present and 0 otherwise.
has the effect of increasing the number of hops from which
a vertex-level representation can aggregate information - for
example, a three layer GCN will aggregate information from
three-hops within the graph to create each representation.
The original methods presented in the literature required
GCN based models to be trained via supervised learning,
where the final vertex representation is tuned via provided la-
bels for a specific task -- classification as a common example
[4], [15]. This is a key difference between GCNs and other
graph embedding approaches, as these commonly require
no labels and thus are applicable on a broader selection of
graphs. Recently, extensions to the GCN framework have
been made which allows for convolutional auto-encoders
for graph datasets [20]. Auto-encoders are a type of un-
supervised neural network model which attempt to compress
input data to a low-dimensional space, and then reconstruct
the original data directly from the learned representation. We
will explore similar concepts but using them for temporal
reconstruction in the following section.
D. Model Overview
For creating our temporally offset graph embeddings, we
will explore the use of both non-probabilistic and variational
encoder models. These are related to the convolutional graph
auto-encoders of Kipf [20], however, we are exploring the
creation of two models explicitly to reconstruct a future state
of the graph, rather than just to capture the current graph.
Below we detail the specifics of both the non-probabilistic
(TO-GAE) and variational (TO-GVAE) models used for
temporal offset reconstruction.
TO-GAE: TO-GAE is the non-probabilistic interpretation
of the temporally offset graph auto-encoder concept, where
the goal is to learn a low-dimensional representation of Ai
from Gi, via an encoding from a GCN Zi = GCN (Ai, Xi),
such that it can be used to predict accurately the structure of
some future time step δ of the graph via a product between
Zi and it's transpose passed through a logistic sigmoid unit
σ:
Ai+δ = σ(ZiZ T
i ).
(2)
For all the work presented in the paper, the GCN model
used to learn Zi is a two layer model.
TO-GVAE: TO-GVAE is a variational interpretation of
the temporally offset graph auto-encoder concept. Again the
goal is to learn a vertex level representation for future graph
reconstruction but using ideas from Bayesian learning [32].
This variational method differs from the non-probabilistic
version outlined above as instead of directly learning the
mapping Z, we instead learn a distribution from which Z
is sampled. Using a variational approach to create the latent
space has been shown to create more robust and meaningful
embeddings, resulting in better performing models [20],
[32].
V (cid:89)
As TO-GVAE is a Bayesian style model, we must define
a model with which to perform inference. This again makes
use of the GCN layers outlined in Section III-C to learn
a mean µ and a variance γ vector used to parametrise the
Gaussian distribution N from which Zi is finally sampled:
q(ZiXi, Ai) =
N (zvµv, diag(γ2
v ))
(3)
v=1
where q is our approximation of the of the true distribution
we are interesting in capturing -- p(Ai+δZi). Here both
µ and γ are separate individual GCN layers GCN (µ) and
GCN (γ) respectively. These layers however take their input
from a common first layer GCN (0).
Once the inference model specified in Equation 3 has
created the various required parameters, a generative model
is created to predict the next time step in the graph. The
generative model we use is again based on the inner-product
between the latent representations, given as follows:
p(Ai+δZi) =
p(Axy = 1σ(zxzT
y )),
(4)
V (cid:89)
V (cid:89)
x=1
y=1
(cid:104)
(cid:105)−
where Axy are the elements from Ai and z are the rows for
each vertex taken from Zi.
To train the model, common for variational methods [20],
[32], we directly optimise the lower bound L with regards
to the model parameters:
L = Eq(ZiXi,Ai)
logp(Ai+δZi)
KL(q(ZiAi, Xi)p(Zi)),
(5)
where KL is the Kullback-Leibler distance between p and q.
For this model we use a Gaussian prior as the distribution
for p(Zi).
E. Model Parameters and Training Procedure
After initial experimentation we found two layers of con-
volution to give the optimal performance for both models,
with the first layer comprising 32 filters, whilst the second
having 16 filters. For training both the models we use full-
batch gradient descent via the ADAM algorithm with a
learning rate of 0.001 for a total of 50 epochs. We also
found the use of an additional term in the loss functions
which penalises the model parameters for getting too large
via L2 to help model performance. All of our models, as
well as the comparative baselines, have been implemented
in the PyTorch library [5].
IV. EXPERIMENTAL SETUP
A. Evaluation Overview and Methodology
As the primary goal of our approach is to create vertex
representations which are better able to encode information
about how the graph will evolve into the future, we will be
using the important task of link prediction as our primary
task. Formally the task of link prediction in the context of
machine learning can be defined as follows: given a subset
of edges Etrain ⊂ E from graph G = (V, E), learn a
model which can accurately predict the remaining edges
Etest = E − Etrain [33]. Many recent methods attempt
to solve this problem via vertex embedding similarity -- i.e.
vertices with more similar embeddings, according to some
metric, are more likely to be connected via an edge [1], [11],
[20].
During the evaluation, we will be investigating two ways
a model trained on temporal graph data could be used for
inference:
• Evolution Pattern Prediction -- For this inference ap-
proach, the original input graph G0 is kept constant,
whilst the rest of the time series G1, ..., Gt is used as
the targets for prediction to measure the models ability
to predict the future graph changes.
• Future Link Prediction -- For this inference approach,
the trained model is kept constant whilst each graph
snapshot Gi is passed in. The model is then evaluated
by making predictions on hold out edges from the same
snapshot to test how well the model can predict edges
in future snapshots of the graph.
Graph edges are predicted as follows: Given the learned
vertex embeddings, the adjacency matrix is reconstructed
via a dot-product of the embedding matrix A(cid:48) = σ(ZZ T).
This reconstructed adjacency matrix is compared with the
true graph to assess how well the embedding is able to
reconstruct the future graph. For the future link prediction
task, edges from the graph are randomly removed before the
graph is fed into the model. The embeddings are then used
to predict the hold-out set of edges.
B. Performance Metrics
As one can consider the task of link prediction to be
that of a binary classification problem (an edge can only
be present or not), we make use of two standard binary
classification metrics:
• Area Under the Receiver Operating Characteristic
Curve (AUC) -- The ratio between the True Positive
Rate (TPR) and False Positive Rate (FPR) measured at
various classification thresholds.
• Mean Average Precision (AP) -- The mean average pre-
cision across the set of test edges: P recision(AP ) =
T P +F P , where T P denotes the number of true posi-
tives the model predicts, and F P denotes the number
of false positives.
T P
In both cases large values are better.
C. Datasets
We make use of the following empirical graph datasets
presented in Table I when performing our experimental
(a) AUC score
(b) AUC score on new edges
(c) AP score
(d) AP score on new edges
Figure 1: AUC and AP scores on the Cora dataset evolved via the configuration method with a 25% chance of edges being
rewired per time step. Values are presented for the whole graph and only on new edges which have been altered since the
graph used for training.
(a) AUC score
(b) AUC score on new edges
(c) AP score
(d) AP score on new edges
Figure 2: AUC and AP scores on the Citeseer dataset evolved via the configuration method with a 25% chance of edges
being rewired per time step. Values are presented for the whole graph and only on new edges which have been altered since
the graph used for training.
Table I: Empirical Graph Datasets
Dataset
V
E
Temporal
Reference
cit-HepPh
cora
citeseer
34,546
2,708
3,327
421,578
5,429
3,327
Snapshots
Synthetic
Synthetic
[34]
[20]
[20]
evaluation. For the two datasets which contain no inherent
temporal information, we generate a synthetic evolutionary
trajectory for the graph using one of the rewire processes
detailed in section IV-D. For the empirical cit-HepPh dataset,
we create six snapshots based on a linear partitioning of the
graph's timeline.
D. Random Rewire Process
In order to have large amounts of temporal graph data
which has been evolved via a controllable process, we make
use of the random rewire methodology [35], [36]. The rewire
process alters a given source graph's degree distribution by
randomly altering the source and target of a set number of
edges. During this re-wire process, it is not guaranteed that
the source or target of the edge will be altered, indeed it is
not always possible due to the graphs topology. Also, the
rewiring process does not change the total number of edges
or vertices within the graph. We employ two types of random
rewire in this work:
• Erdos - The edges are rewired such that the resulting
topology of the graph begins to resemble a Erdos-R´enyi
graph, were edges are uniformly distributed between
vertices.
• Configuration - The edges are rewired in such a way
that each vertex approximately preserves it's associated
number of edges, creating graphs with a similar degree
distribution to the original.
E. Baseline Approaches
Graph convolutional auto-encoders have been shown to
have state-of-the-art performance in the link-prediction task
for high-dimensional graphs [20]. As such, we compare
our approach to both non-probabilistic and variational ver-
sions of the convolutional graph auto-encoder, these will be
denoted as GAE and GVAE in the results section. These
approaches are trained as originally detailed by the authors
[20], meaning that temporal information is not utilised.
V. RESULTS
This section will present the results of the experimental
evaluation. As outlined in Section IV-A, we are testing two
01234Temporal Graph Snapshot ID0.550.600.650.700.750.800.85AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.650.70AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.600.650.700.750.800.85AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.400.450.500.550.600.65AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.550.600.650.700.750.800.85AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.550.600.650.700.75AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.650.700.750.800.85AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.40.50.60.7AP ScoreGAEGVAETO_GAETO_GVAE (a) AUC score
(b) AUC score on new edges
(c) AP score
(d) AP score on new edges
Figure 3: AUC and AP scores on the Cora dataset evolved via the configuration method with a 50% chance of edges being
rewired per time step. Values are presented for the whole graph and only on new edges which have been altered since the
graph used for training.
(a) AUC score
(b) AUC score on new edges
(c) AP score
(d) AP score on new edges
Figure 4: AUC and AP scores on the Citeseer dataset evolved via the configuration method with a 50% chance of edges
being rewired per time step. Values are presented for the whole graph and only on new edges which have been altered since
the graph used for training.
ways to evaluate the models: evolution pattern prediction and
future link prediction. We present results on the datasets in-
troduced in Section IV-C, which contain both simulated and
empirical evolving graph datasets. All the results presented
are the mean, with the standard deviation, of ten repeats of
the evaluation procedure, using a random train/test split for
each repetition.
A. Simulated Graph Evolution
We first present results using time series simulated via the
random rewire processes introduced in Section IV-D on both
the cora and citeseer datasets. For this experiment, we train
using the original unaltered graph to reconstruct the next
graph in the time series. We then measure the performance
of the resulting embeddings at predicting edges in future
graph snapshots using the two inference approaches outlined
in Section IV-A.
1) Evolution Pattern Prediction: Figure 1 shows how
well the models perform upon the Cora dataset with smaller
possible perturbations introduced by the rewire process. The
figure shows how both the temporally offset methods we
introduce generally demonstrate greater performance than
the baseline approaches. When considering the AUC score
on new edges, we can see a large increase in performance.
Our methods also show performance above the baselines
when reconstructing the full graph, demonstrating that the
temporal offset process does not harm the abilty to predict
edges which have not changed.
Figure 2 contains the results for the Citeseer dataset,
again using the smaller rewire probability. This figure con-
tinues the trends established in the previous figure, with
the temporally offset methods beating the baseline methods.
However, this time it is the variational approach which often
demonstrates the greater performance, especially when only
new edges are considered.
The next two sets of figures show results for when a large
possible perturbation is made in-between each graph time
step. Figure 3 demonstrates that, as was expected, when even
large steps are made between graphs, the gap between our
approaches and the baselines also increases. The temporally
offset methods show a clear increase in performance, even
when using the model to make predictions about graphs
later in the time series, which will have a quite different
topological structure to the graph used to train the model.
Figure 4 shows the results for the Citeseer dataset us-
ing the higher level of possible perturbation. The figure
continues the trend of the previous results by showing the
temporally offset models to be better at predicting future
01234Temporal Graph Snapshot ID0.500.550.600.650.700.750.80AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.650.700.75AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.550.600.650.700.750.80AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.650.70AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.550.600.650.700.750.80AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.650.700.75AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.600.650.700.750.80AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.450.500.550.600.650.700.75AP ScoreGAEGVAETO_GAETO_GVAE (a) AUC score on Cora
(b) AP score on Cora
(c) AUC score on Citeseer
(d) AP score on Citeseer
Figure 5: AUC and AP sscoes for the future link prediction task on both the Cora and Citeseer datasets evolved using the
Erdos rewire method with E/2 edges having the chance of being rewired. The results presented are scores for predicting
only new edges which have appeared after the original graph used for training the model.
(a) AUC score on Cora
(b) AP score on Cora
(c) AUC score on Citeseer
(d) AP score on Citeseer
Figure 6: AUC and AP scores for the future link prediction task on both the Cora and Citeseer datasets evolved using the
Erdos rewire method with the complete set of E having the chance of being rewired. The results presented are scores for
predicting only new edges which have appeared after the original graph used for training the model.
changes in the graph, even when considering both the
unchanged and new edges.
2) Future Link Prediction: For assessing the ability of the
various models to continue to make accurate link predictions
as the graph undergoes heavy topological change, we make
use of the Erdos-based random rewire method. Figure 5
shows the results for only new edges on both the cora and
citeseer datasets when 50% of the edges have the chance of
being rewired in-between each graph snapshot. The figure
shows that our temporally offset training method is more
robust to the Erdos rewired edges, as it displays a higher
level of predictive performance particularly when regarding
the AUC metric. Interesting, of the two temporally offset
models, it is the non-probabilistic approach which displays
greater performance across both datasets.
Figure 6 highlights the performance on new edges when
all the edges have the chance to be rewired between graph
snapshots. The results show that the temporal offset ap-
proaches are generally more robust to the large change in
graph topology between graph snapshots. With both demon-
strating greater performance, especially at earlier points in
the time series. Continuing the trend established in the
previous experiment, the non-probabilistic temporally offset
model out performs the variational approach. We hypoth-
esise that as the variational approach is a more complex
model, it is over fitting more strongly to the non-rewired
original graph edges, making it less able to learn the Erdos
pattern of rewiring.
B. Empirical Time-Series
1) Evolution Pattern Prediction: Table II displays the
results for the task of evolution pattern prediction for all
models for the cit-HepPh dataset. The table shows that both
the temporally offset methods significantly outperform the
baseline approaches on both whole graph and new edges
metrics at
this particular task. The gap in performance
between the temporally offset and normal approaches for
this empirical dataset is larger than on the previous synthetic
results, indicating that our approach is much better able to
learn the temporal dynamics of real datasets. We can also
see that the variational temporally offset approach is often
the best performing of the two approaches, particularly at
later time-points.
2) Future Link Prediction: Table III highlights the results
for all models at the task of future link prediction on the cit-
HepPh dataset. We can see that compared with the previous
task, all models are more closely matched. The results
show all approaches to have a good predictive performance,
even at later time points, however, the temporally offset ap-
proaches still outperform the baselines. Here it is interesting
to note that the non-probabilistic model almost always out
performs the variational one.
01234Temporal Graph Snapshot ID0.500.550.600.650.70AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.350.400.450.50AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.450.500.550.600.650.700.75AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.350.400.450.50AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.65AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.4250.4500.4750.5000.525AP ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.500.550.600.650.70AUC ScoreGAEGVAETO_GAETO_GVAE01234Temporal Graph Snapshot ID0.450.500.550.60AP ScoreGAEGVAETO_GAETO_GVAE Model
Metric
G1
G2
G3
G4
G5
GVAE
GAE
TO-GVAE
TO-GAE
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
0.699(±0.0133)
0.8023(±0.0089)
0.5358(±0.0103)
0.5875(±0.0071)
0.653(±0.0159)
0.7817(±0.0103)
0.4665(±0.0172)
0.5541(±0.0131)
0.9943(±0.0004)
0.9925(±0.0011)
0.995(±0.001)
0.989(±0.003)
0.9944(±0.0012)
0.9915(±0.0028)
0.9955(±0.0009)
0.9882(±0.0043)
0.6327(±0.0036)
0.7459(±0.0056)
0.513(±0.0036)
0.5698(±0.0055)
0.5939(±0.0062)
0.7293(±0.0069)
0.4512(±0.0033)
0.5416(±0.0055)
0.873(±0.0022)
0.9197(±0.0015)
0.8203(±0.0042)
0.8615(±0.0043)
0.8702(±0.0029)
0.9176(±0.0022)
0.8159(±0.0029)
0.8588(±0.0025)
0.5913(±0.0022)
0.7036(±0.001)
0.5012(±0.0098)
0.5663(±0.0067)
0.5546(±0.0034)
0.6875(±0.0023)
0.4526(±0.0041)
0.5434(±0.0031)
0.7728(±0.0031)
0.8515(±0.0027)
0.7076(±0.0016)
0.776(±0.0023)
0.7629(±0.0062)
0.8461(±0.002)
0.6972(±0.0087)
0.7711(±0.0035)
0.5821(±0.0049)
0.6853(±0.0035)
0.5034(±0.0067)
0.5598(±0.0079)
0.5343(±0.0043)
0.6643(±0.006)
0.4436(±0.0031)
0.5286(±0.0077)
0.726(±0.0084)
0.8158(±0.0056)
0.6641(±0.0089)
0.7396(±0.0055)
0.711(±0.0095)
0.8077(±0.0062)
0.6449(±0.0084)
0.7285(±0.005)
0.5771(±0.0122)
0.685(±0.0095)
0.4979(±0.0112)
0.552(±0.0111)
0.5343(±0.0043)
0.6643(±0.006)
0.4436(±0.0031)
0.5286(±0.0077)
0.7281(±0.0045)
0.8177(±0.0034)
0.6591(±0.008)
0.7367(±0.0044)
0.711(±0.0095)
0.8077(±0.0062)
0.6449(±0.0084)
0.7285(±0.005)
Table II: Evolution pattern prediction results presented as mean values with standard deviation for both the whole graph and
new edges on the cit-HepPh dataset across all models trained using G0. A bold value indicates the highest score for that
metric for the given graph snapshot.
Model
Metric
G1
G2
G3
G4
G5
GVAE
GAE
TO-GVAE
TO-GAE
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
AUC
AP
NE-AUC
NE-AP
0.9702(±0.0009)
0.9759(±0.0006)
0.9586(±0.002)
0.9521(±0.0016)
0.9885(±0.0006)
0.9902(±0.0004)
0.9837(±0.0008)
0.9803(±0.0008)
0.9957(±0.0004)
0.9944(±0.0007)
0.9964(±0.0003)
0.9921(±0.0008)
0.9961(±0.0001)
0.9943(±0.0003)
0.997(±0.0001)
0.9922(±0.0006)
0.9336(±0.0026)
0.9461(±0.0021)
0.9149(±0.0005)
0.9103(±0.0012)
0.9794(±0.0018)
0.9815(±0.0014)
0.9732(±0.0022)
0.9685(±0.002)
0.9871(±0.0006)
0.9869(±0.0005)
0.9841(±0.0005)
0.979(±0.0006)
0.99(±0.0001)
0.9887(±0.0001)
0.9882(±0.0002)
0.9825(±0.0)
0.8796(±0.0018)
0.9016(±0.0012)
0.8552(±0.0029)
0.8581(±0.0028)
0.9545(±0.0011)
0.9616(±0.0011)
0.9447(±0.0015)
0.9442(±0.002)
0.9651(±0.0014)
0.9693(±0.0009)
0.9578(±0.0022)
0.9549(±0.0021)
0.9751(±0.0009)
0.9757(±0.0008)
0.9701(±0.0011)
0.9646(±0.0014)
0.8503(±0.0018)
0.8762(±0.0013)
0.8232(±0.002)
0.8289(±0.0023)
0.9412(±0.0007)
0.9518(±0.0007)
0.9303(±0.0007)
0.9332(±0.0012)
0.9534(±0.0009)
0.9596(±0.0008)
0.9439(±0.0011)
0.9427(±0.0017)
0.9645(±0.0003)
0.967(±0.0008)
0.9579(±0.0003)
0.9536(±0.0013)
0.8499(±0.0009)
0.8761(±0.0007)
0.8231(±0.0039)
0.8289(±0.0034)
0.9412(±0.0007)
0.9518(±0.0007)
0.9303(±0.0007)
0.9332(±0.0012)
0.9524(±0.001)
0.9591(±0.0011)
0.9427(±0.0003)
0.9419(±0.0015)
0.9645(±0.0003)
0.967(±0.0008)
0.9579(±0.0003)
0.9536(±0.0013)
Table III: Future link prediction results presented as mean values with standard deviation for both the whole graph and new
edges on the cit-HepPh dataset across all models trained using G0. A bold value indicates the highest score for that metric
for the given graph snapshot.
VI. CONCLUSION
Whilst a lot of focus has recently been placed on finding
methods for learning graph representations which are accu-
rately able to make predictions about the current state of
the graph, few works have investigated how these models
perform for temporal graphs. This paper has explored the
use of graph specific neural networks to learn vertex level
representations which are accurately able to make predic-
tions about future time points of evolving graphs. This is
achieved by explicitly training the models to predict future
time steps via the temporal offset reconstruction method. To
assess our approach, we perform extensive experimentation
on both real-world and synthetic evolving graph datasets.
There is large scope for future continuation of this re-
search. We plan to investigate if the use of topological vertex
properties can be used as input to the model to improve
performance, instead of relying upon the identity matrix of
the graph. We also plan to investigate if an inductive style
model [15] could be used to replace the GCN as this would
allow us to vary the number of vertices between time steps
and allow for batch-based training.
ACKNOWLEDGEMENTS
We gratefully acknowledge the support of NVIDIA Cor-
poration with the donation of the GPU used for this re-
search. Additionally we thank the Engineering and Physical
Sciences Research Council UK (EPSRC) for funding.
REFERENCES
[1] A. Grover and J. Leskovec, "node2vec : scalable feature learn-
ing for networks," International Conference on Knowledge
Discovery and Data Mining, 2016.
[2] P. Goyal and E. Ferrara, "Graph embedding techniques,
applications, and performance: a survey," arXiv preprint
arXiv:1705.02801, 2017.
[3] R. v. d. Berg, T. N. Kipf, and M. Welling, "Graph convolu-
tional matrix completion," arXiv preprint arXiv:1706.02263,
2017.
[4] T. N. Kipf and M. Welling, "Semi-supervised classifica-
tion with graph convolutional networks," arXiv preprint
arXiv:1609.02907, 2016.
[5] A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang,
Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer,
"Automatic differentiation in pytorch," 2017.
[6] M. Belkin and P. Niyogi, "Laplacian eigenmaps and spectral
techniques for embedding and clustering," Advances in neural
information processing systems, pp. 585 -- 591, 2002.
[7] A. Ahmed, N. Shervashidze, S. Narayanamurthy, V. Josi-
fovski, and A. J. Smola, "Distributed large-scale natural graph
factorization," International conference on World Wide Web,
pp. 37 -- 48, 2013.
[8] S. Cao, W. Lu, and Q. Xu, "Grarep: Learning graph represen-
tations with global structural information," ACM International
on Conference on Information and Knowledge Management,
pp. 891 -- 900, 2015.
[9] M. Ou, P. Cui, J. Pei, Z. Zhang, and W. Zhu, "Asymmetric
transitivity preserving graph embedding," International Con-
ference on Knowledge Discovery and Data Mining, pp. 1105 --
1114, 2016.
[10] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and
J. Dean, "Distributed representations of words and phrases
and their compositionality," in Advances in neural information
processing systems, 2013, pp. 3111 -- 3119.
[11] B. Perozzi, R. Al-Rfou, and S. Skiena, "DeepWalk: online
learning of social representations," International Conference
on Knowledge Discovery and Data Mining, 2014.
[12] B. Chamberlain, M. D. Clough, and James, "Neural embed-
dings of graphs in hyperbolic space," KDD Workshop on
Mining and Learning with Graphs (MLG), 2017.
[13] J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun, "Spectral
networks and locally connected networks on graphs," arXiv
preprint arXiv:1312.6203, 2013.
[14] M. Defferrard, X. Bresson, and P. Vandergheynst, "Convolu-
tional neural networks on graphs with fast localized spectral
filtering," in Advances in Neural Information Processing
Systems, 2016, pp. 3844 -- 3852.
[15] W. Hamilton, Z. Ying, and J. Leskovec, "Inductive repre-
sentation learning on large graphs," in Advances in Neural
Information Processing Systems, 2017, pp. 1024 -- 1034.
[16] J. Chen, T. Ma, and C. Xiao, "Fastgcn: fast learning with
graph convolutional networks via importance sampling,"
arXiv preprint arXiv:1801.10247, 2018.
[17] P. Baldi, "Autoencoders, unsupervised learning, and deep
architectures," in Proceedings of ICML workshop on Unsu-
pervised and Transfer Learning, 2012, pp. 37 -- 49.
[18] D. Wang, P. Cui, and W. Zhu, "Structural deep network
embedding," in International Conference on Knowledge Dis-
covery and Data Mining. ACM, 2016, pp. 1225 -- 1234.
[19] W. Yu, C. Zheng, W. Cheng, C. C. Aggarwal, D. Song,
B. Zong, H. Chen, and W. Wang, "Learning deep network
representations with adversarially regularized autoencoders,"
in International Conference on Knowledge Discovery & Data
Mining. ACM, 2018, pp. 2663 -- 2671.
[20] T. N. Kipf and M. Welling, "Variational graph auto-encoders,"
arXiv preprint arXiv:1611.07308, 2016.
[21] Y. Kim, Y.-I. Chiu, K. Hanaki, D. Hegde, and S. Petrov,
language
"Temporal analysis of language through neural
models," arXiv preprint arXiv:1405.3515, 2014.
[22] W. L. Hamilton, J. Leskovec, and D. Jurafsky, "Diachronic
word embeddings reveal statistical laws of semantic change,"
arXiv preprint arXiv:1605.09096, 2016.
[23] R. Bamler and S. Mandt, "Dynamic word embeddings," arXiv
preprint arXiv:1702.08359, 2017.
[24] S. Pandhre, H. Mittal, M. Gupta, and V. N. Balasubrama-
nian, "Stwalk: learning trajectory representations in temporal
graphs," in Proceedings of the ACM India Joint International
Conference on Data Science and Management of Data.
ACM, 2018, pp. 210 -- 219.
[25] W. Yu, W. Cheng, C. C. Aggarwal, K. Zhang, H. Chen, and
W. Wang, "Netwalk: A flexible deep embedding approach
for anomaly detection in dynamic networks," in Iternational
Conference on Knowledge Discovery & Data Mining. ACM,
2018, pp. 2672 -- 2681.
[26] G. H. Nguyen, J. B. Lee, R. A. Rossi, N. K. Ahmed,
E. Koh, and S. Kim, "Continuous-time dynamic network
embeddings," in 3rd International Workshop on Learning
Representations for Big Networks (WWW BigNet), 2018.
[27] P. Goyal, N. Kamra, X. He, and Y. Liu, "Dyngem: Deep
embedding method for dynamic graphs," arXiv preprint
arXiv:1805.11273, 2018.
[28] T. Chen, I. Goodfellow, and J. Shlens, "Net2net: Accel-
erating learning via knowledge transfer," arXiv preprint
arXiv:1511.05641, 2015.
[29] F. Manessi, A. Rozza, and M. Manzo, "Dynamic graph
convolutional networks," arXiv preprint arXiv:1704.06199,
2017.
[30] Y. Seo, M. Defferrard, P. Vandergheynst, and X. Bresson,
"Structured sequence modeling with graph convolutional re-
current networks," arXiv preprint arXiv:1612.07659, 2016.
[31] R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamil-
ton, and J. Leskovec, "Graph convolutional neural net-
works for web-scale recommender systems," arXiv preprint
arXiv:1806.01973, 2018.
[32] D. P. Kingma and M. Welling, "Auto-encoding variational
bayes," arXiv preprint arXiv:1312.6114, 2013.
[33] V. Mart´ınez, F. Berzal, and J.-C. Cubero, "A survey of link
prediction in complex networks," ACM Computing Surveys,
vol. 49, no. 4, p. 69, 2017.
[34] J. Leskovec and A. Krevl, "SNAP Datasets: Stanford large
network dataset collection," http://snap.stanford.edu/data, Jun.
2014.
[35] S. Bonner, J. Brennan, G. Theodoropoulos, I. Kureshi, and
A. S. McGough, "Deep topology classification: A new ap-
proach for massive graph classification," in International
Conference on Big Data.
J. Brennan,
IEEE, 2016, pp. 3290 -- 3297.
I. Kureshi, M. Stephen, and
G. Theodoropoulos, "Efficient comparison of massive graphs
through the use of graph fingerprints," in KDD Workshop on
Mining and Learning with Graphs (MLG), 2016.
[36] S. Bonner,
|
Subsets and Splits