Learning Representations of Partial Subgraphs by Subgraph InfoMax

Dongkwan Kim, Jiho Jin, Jaimeen Ahn, Alice Oh

Keywords: Graph Neural Network, Subgraph, Mutual Information Maximization

Abstract: Subgraphs are important substructures of graphs, but learning their representations has not been studied well. Particularly, when we have partial subgraphs, existing node- or subgraph-level message-passing is likely to produce suboptimal representations. In this paper, we propose Intra- and Inter-Subgraph InfoMax, a model that learns subgraph representations under incomplete observation. Our model employs subgraph summaries at two different levels while maximizing the mutual information between the subgraph summaries and the node representations. By doing so, we reconstruct the representation of the underlying subgraph and improve its expressiveness from different angles of the local-global structure. We conduct experiments on three real-world datasets under training and evaluation protocols designed for this problem. Experimental results show that our model outperforms baselines in all settings.

One-sentence Summary: We propose Intra- and Inter-Subgraph InfoMax, a model that learns a subgraph representation under partial observation.