SUPExplainer: Subgraph-Level Explanations for Graph Neural Networks via Reinforcement Learning Mechanism
Xinlian Yi, Yushan Zhang, Xinghui Li · 2022
Graph neural networks(GNNs) have attracted extensive attention due to their effective applications in various graph tasks[1] [2]. However, GNNs regarded as black-boxes generally meet the challenges of explanation and confidence. To address the above challenge, we propose SUPExplainer, a novel Subgraph-level explanation approach, which can provide a reliable explanation for GNN model. Given an original graph and a trained GNN, SUPExplainer efficiently select the significant subgraphs via Reinforcement Learning(RL) Mechanism to represent the critical structural information of the graph. The model is trained to maximize cumulative rewards through proximal policy optimization(PPO) and acts in an explanation-based environment. However, the system can only receive global rewards, which can not attribute cooperative contributions to different subgraphs. To solve such a problem, we further adopt approximate Shapley value which enhances the sample efficiency and reduces computations. Furthermore, experimental results indicate that SUPExplainer can identify important graph structures in explanation more intuitively, and is quantitatively superior to some baseline methods in reasonable calculation level.