MEGA: Explaining Graph Neural Networks with Network Motifs

Feng Ding, Naiwen Luo, Shuo Yu, Tingting Wang, Feng Xia

Abstract: Graph Neural Networks (GNNs) are powerful tools for graph representation. However, GNNs have remained black boxes, leading to the lack of explainability. As a consequence, the application of GNNs has been severely limited. Existing methods focus on generating an explanation with important nodes and edges but pay less attention to high-order structures (e.g., network motif), which are more intuitive and important for graph data. The explanations generated by the explainer will thus ignore the high-order information contained in multi-node neighbours, resulting in a decrease in the human comprehensibility of the explanation subgraph. In this paper, we propose a motif-aware GNNs explainer (MEGA), wherein a motif-aware subgraph generation module and a counterfactual optimization layer are employed. MEGA can provide high-quality counterfactual explanations for GNNs while focusing on high-order features of graph data. We justify the effectiveness of the proposed MEGA on both synthetic and real-world datasets. Experimental results show that MEGA outperforms state-of-the-art baselines while keeping explanations at a smaller level.

External IDs:dblp:conf/ijcnn/DingLYWX23