GraphExplainer: A LIME-Based Graph Neural Network Explanation Framework
Huihui Helen Wang, Yulin Liu, Hao Liu, Biyin Zhang · 2024
Graph Neural Networks (GNNs) have been widely applied due to their powerful data processing capabilities in many fields, including in computer vision, social networks, molecular chemistry and the financial field. However, similar to deep learning, GNNs also suffer from complexity and opacity in information processing and decision-making, significantly hindering their application in critical domains where erroneous predictions can lead to severe consequences. Given the exquisitely intricate nonlinear transformations that occur during iterations, elucidating the effectiveness of GNN models becomes an arduous undertaking. In the paper, we propose GraphExplainer, a versatile framework for explaining GNN models. By identifying the most critical tight subgraphs for node to be explained, we obtain a local neighborhood space within which a nonlinear interpretable model is locally learned. More profoundly, in the endeavor to explain a node, we first determine the neighborhood space through its tight subgraphs, generate a nonlinear interpretable model, and then employ HSIC Lasso to calculate the K most representative features as explanations. Experiments demonstrate that GraphExplainer provides explanations with higher accuracy.