SEAL (learning from Subgraphs, Embeddings, and Attributes for Link prediction)

Link prediction is a key problem for network-structured data. Link prediction heuristics use some score functions, such as common neighbors and Katz index, to measure the likelihood of links. They have obtained wide practical uses due to their simplicity, interpretability, and (often) scalability. However, heuristic methods have strong assumptions on when two nodes are likely to have a link, which limits their effectiveness in networks where these assumptions fail. In this regard, a more reasonable way should be learning suitable "heuristics" from networks instead of using predefined ones.

We propose SEAL, a novel link prediction framework which automatically learn heuristics that suit the specific networks. SEAL extracts a local subgraph around each target link as its data representation, and learns a function mapping the subgraph patterns to link existence based on Graph Neural Networks (GNN). SEAL is also able to learn from node embeddings and node attributes as the additional information to the subgraph patterns. SEAL (learning from Subgraphs, Embeddings, and Attributes for Link prediction) has superior performance than existing link prediction methods, such as heuristics, latent feature methods, and embedding methods.

Code Download

For any questions or bug reports, please contact Yixin Chen and Muhan Zhang.

Reference

M. Zhang and Y. Chen, Link Prediction Based on Graph Neural Networks, Advances in Neural Information Processing Systems (NIPS-18). Spotlight presentation. (PDF)

Acknowledgement

Muhan and Yixin are supported by NSF grants CCF-1215302, IIS-1343896, DBI-1356669, CNS-1320921, and a Microsoft Research New Faculty Fellowship.