Community-influencing path explanation for link prediction in heterogeneous graph neural network

Abstract

Most existing research on the interpretability of Graph Neural Networks (GNNs) for Link Prediction (LP) focuses on homogeneous graphs, with relatively few studies on heterogeneous graphs. Community is a crucial structure of a graph and can often improve LP performance. However, existing GNN explanation methods for heterogeneous LP rarely consider the impact of communities, leading to generated explanations that do not align with human understanding. To fill this gap, we consider community influence in GNN explanation for heterogeneous LP. We first demonstrate the effectiveness of communities in GNN explanations for heterogeneous LP through a preliminary analysis. Under this premise, we propose CI-Path, a Community-Influencing Path explanation for heterogeneous GNN-based LP that considers the influence of communities throughout the entire learning process. Specifically, we conduct degree centrality pruning and employ a community detection algorithm for data preprocessing. Then we propose a community-influencing objective, comprising community-influencing prediction loss and community-influencing path loss. Finally, we identify the reasonable explanatory paths that are the shortest with the minimum sum of node degrees and the fewest number of communities crossed. Extensive experiments on five heterogeneous datasets demonstrate the superior performance of CI-Path compared to baselines. Our code is available at https://github.com/wenyhsmile/CI-Path.