Unified node, edge and motif learning networks for graphs

Abstract

Driven by the success of deep learning in recent years, many Graph Neural Networks have been proposed to address various tasks in graph learning. However, most of them are suffering from limitations on detecting and representing subgraph structures/information. First, most motif-based methods are relying on a predefined motif vocabulary or human designed algorithms to find motifs, suggesting they have not leveraged the power of deep networks to automatically learn underlying topological structures. Second, they mainly concentrate on node information and ignore the important roles of edge features and local structures. This paper presents a novel neural network framework (Node, Edge and MOtif Learning Network - NemoNet), integrating node and edge features in addition to the capacity of motif discovery and graph structure learning, eliminating these limitations. To address the diverse changes in subgraph topology, we convert graphs into rich structure-driven representation tensors before training the network. NemoNet is parameterized by motif filters that are initialized and updated during training. It is also designed to accept node and edge features as input and learn high-level embeddings from all information sources (node, edge and motif). Experiments on molecular datasets show NemoNet can learn graph structures, leverage these information sources effectively and achieve the highest accuracy (88.89% and 60.75%) in MUTAG (MUTAGenesis dataset) and PTC (Predictive Toxicology Challenge dataset), and a competitive performance against existing Graph Neural Networks in the other datasets.