Improve Graph Convolutional Network Performance via Incorporating Message Filter

Graph Convolutional Network (GCN) has shown great potential in graph learning. However, since the message-passing mechanism treats all messages equally and updates representation of each node by transforming messages received from its neighbors and itself linearly, GCN lacks the capability to distinguish between good and bad messages for the task at hand and suffers from drawbacks such as over-smoothing. To overcome the limitation, we propose to incorporate message filter into GCN to improve message propagation and finally to enhance GCN performance. This is a flexible method, because for different tasks, different message filters can be designed from various viewpoints to achieve corresponding goals. Moreover, a message filter can be realized as a plug-and-play module and incorporated into GCN conveniently, without introducing new model parameters. We design a message filter for the task of node classification, which improves message aggregation of each node via strengthening the messages from the nodes with the same label and filtering out the messages from the nodes with different label. We theoretically prove the effectiveness of the message filter and evaluate the learned node representations quantitatively. Extensive experimental results demonstrate that incorporating the message filter into GCN makes a significant accuracy improvement in node classification.