Edge Partition Modulated Graph Convolutional Networks

Graph convolutional networks (GCNs), which propagate the node features through the edges and learn how to transform the aggregated features under label supervision, have achieved great success in supervised feature extraction for both graph-level and node-level classification tasks. However, GCNs typically treat the graph adjacency matrix as given and ignore how the edges could be formed by different latent inter-node relations. In this paper, we introduce a relational graph generative process to model how the observed edges are generated by aggregating the node interactions over multiple overlapping node communities, each of which represents a particular type of relation that contributes to the edges via a logical OR mechanism. Based on this relational generative model, we partition each edge into the summation of multiple relation-specific weighted edges, and use the weighted edges in each community to define a relation-specific GCN. We introduce a variational inference framework to jointly learn how to partition the edges into different communities and combine relation-specific GCNs for the end classification tasks. Extensive evaluations on real-world datasets have demonstrated the working mechanisms of the edge partition modulated GCNs and their efficacy in learning both node and graph-level representations.