Motif-Based Graph Representation Learning with Application to Chemical Molecules

This work considers the task of representation learning on attributed relational graph (ARG). Both nodes and edges in an ARG are associated with attributes/features allowing ARGs to encode rich structural information widely observed real applications. Existing neural networks offer limited ability capture complex interactions within local contexts, which hinders them from taking advantage expression power ARGs. We propose motif convolution module (MCM), a new motif-based technique better utilize information. The handle continuous edge node features is one MCM’s advantages over existing models. MCM builds vocabulary unsupervised way deploys novel operation extract context individual nodes, then used learn higher level representations via multilayer perceptron and/or message passing networks. When compared other approaches classifying synthetic graphs, our approach substantially at capturing context. also demonstrate performance explainability by applying it several molecular benchmarks.