PBGAN: Path Based Graph Attention Network for Heterophily

Typical Graph Neural Networks (GNNs) have achieved noticeable performance in node classification tasks with the strategy of neighbor aggregation, which benefits from the homophily assumption of graphs (nodes with the same class tend to be connected). However, the performance of many models degrades tensely when data tend to be heterophilic. Therefore, we need to optimize the neighbor aggregation strategy to deal with these two different types of graphs. In this paper, we propose a Path Based Graph Attention Network (PBGAN) to address this issue. It utilizes structure and path information in the high-order neighborhood and employs Recurrent Neural Network (RNN) to learn attention coefficients according to the shortest paths between node pairs. This learning method makes the attention coefficients better reflect the relational importance between nodes, which is practical for processing heterophilic graphs. Empirical results demonstrate that the PBGAN achieves comparable performance in both homophilic and heterophilic datasets, proving the effectiveness of the proposed attention learning method.