Patent: Patent Application Network
The Patent dataset is a heterogeneous graph constructed from the United States Patent and Trademark Office. Our HTAG is constructed using source data provided by the Harvard USPTO Patent Dataset [1], covering the period from 2010 to 2018. After processing, it contains three types of entities—patents (2,762,187 nodes), inventors (2,873,311 nodes), and examiners (10,641 nodes)—as well as two types of directed relations connecting three types of entities—an inventor “invents” a patent, and an examiner “examines” a patent. Patents are associated with their titles and abstracts, and all the other types of entities are textless. We concatenate the patent title and abstract and pass it to a MiniLM sentence encoder, generating a 384-dimensional feature vector for each patent node. Additionally, each patent is assigned a primary International Patent Classification (IPC) code, and the patent application date is included as metadata.
Prediction task. The task is to predict the main IPC codes of patent applications at the class level. There are 120 IPC codes at the class level in our dataset. We also use both Micro-F1 and Macro-F1 scores as classification metrics.
Dataset splitting. The splitting strategy is the time-based split. Specifically, we train on patents published up to 2014, validate on patents published in 2015, and test on patents published since 2016.
Dataset statistics.
| # Nodes | # Edges | # Classes | # Splits |
|---|---|---|---|
| 5,646,139 | 8,833,738 | 120 | Train: 1,705,155 |
| Patent: 2,762,187 | Patent-Inventor: 6,071,551 | Valid: 374,275 | |
| Inventor: 2,873,311 | Patent-Examiner: 2,762,187 | Test: 682,757 | |
| Examiner: 10,641 |
[1] Suzgun, M., Melas-Kyriazi, L., Sarkar, S., Kominers, S. D., & Shieber, S. (2023). The harvard uspto patent dataset: A large-scale, well-structured, and multi-purpose corpus of patent applications. Advances in neural information processing systems, 36, 57908-57946.
Dataset Construction
- Download patent data (2010-2018) from Harvard USPTO Patent Dataset to
./raw/ - Run
extract_from_hupd.py - Run
graph_builder.ipynb