Improved Topology Features for Node Classification on Heterophilic Graphs

Yurui Lai, Taiyan Zhang, Rui Fan

Abstract: Graph neural networks (GNNs) are one of the most effective techniques for node classification tasks. However, standard GNNs strongly depend on the graph homophily assumption, and their accuracy can degrade substantially on heterophilic graphs. A number of recent works have found that using graph topology as an explicit feature can improve classification performance. However, we observe that the sparsity of graphs often limits the amount of information first order connectivity can provide. We propose an embedding method which uses higher-order connectivity information to further improve accuracy, while limiting the amount of extra computational overhead. We further observe that standard features-based GNNs and newer topology-based models each have their own strengths and weaknesses on the same graph, and we introduce a technique to combine information from both types of GNNs to achieve higher accuracy than either type alone. We conduct extensive experiments on graphs with a range of sizes and heterophily levels and show that our proposed GNN architecture achieves state of the art accuracy, especially on highly heterophilic graphs. We also conduct further experiments and ablations to validate the observations underlying our GNN’s design and analyze the importance of different components. Our code is available in https://github.com/lannester666/BoPE-GNN.