Beyond Homophily: Neighborhood Distribution-guided Graph Convolutional Networks

Recently, Graph Convolutional Networks (GCNs) have achieved powerful success in various tasks related to graph data. It is usually believed that typical GCNs and their variants are constrained by the implicit homophily assumption, and fail to generalize to heterophilic scenario where most nodes have neighbors from different classes. While many efforts have been put into handling heterophilic graphs, most of them overlook the impact of neighborhood distribution of nodes. In this paper, we first conduct experimental investigation on both homophilic and heterophilic graphs, and surprisingly find that the neighborhood distribution of nodes with the same class tends to be similar. Based on this observation, we propose a novel neighborhood distribution-guided graph convolutional network, which can adaptively combine lower-order and higher-order neighborhood distributions into the graph convolutional process. To further enhance the model performance, we introduce a feature contrastive loss to optimize node representation by implicitly utilizing feature information. Experiments on seven real-world datasets demonstrate that our new approach exhibits superior performance compared to state-of-the-art methods with both homophily and heterophily.