PPGCN: A Message Selection Based Approach for Graph Classification

Recently, the Graph Convolutional Networks (GCNs) have achieved state-of-the-art performance in many graph data related tasks. However, traditional GCNs may generate redundant information in the message passing phase. In order to solve this problem, we propose a novel graph convolution named Push-and-Pull Convolution (PPC), which follows the message passing framework. On the one hand, for each star-shaped subgraph, PPC uses a node pair based message generation function to calculate the message pushed by each local node to the central node. On the other hand, in the message aggregation substep, each central node pulls valuable information from the messages pushed by its local nodes based on a gate network with pre-perceiving function. Based on the PPC, a new network named Push-and-Pull Graph Convolutional Network (PPGCN) is proposed for graph classification. PPGCN stacks multiple PPC layers to extend the receptive field of each node, then applies a global pooling layer to get the graph embedding based on the concatenation of all PPC layers’ outputs. The new network is permutation invariant and can be trained end-to-end. We evaluate the performance of PPGCN in 6 graph classification datasets. Compared with state-of-the-art baselines, PPGCN achieves the top-1 accuracy on 4 of 6 datasets.