XGExplainer: Robust Evaluation-based Explanation for Graph Neural Networks

Graph Neural Networks (GNNs) have emerged as a powerful tool for machine learning on graph datasets. Although GNNs can achieve high accuracy on several tasks, the explainability of the predictions remains a challenge. Existing works in GNN explainability aim to extract the key features contributing to the prediction made by a pre-trained model. For instance, perturbation-based methods focus on evaluating the potential explanatory subgraphs using the pre-trained model itself as an evaluator to determine whether the subgraphs capture the informative features. However, we show that this approach can fail to recognize informative subgraphs that become out-of-distribution relative to the training data. To address this limitation, we propose XGExplainer, a method designed to enhance the robustness of perturbation-based explainers. It achieves this by training a specialized GNN model, i.e., a robust evaluator model that aims at estimating the true graph distribution from randomized subgraphs of the input graph. Our method is geared towards enhancing the generalizability of existing explainability techniques by decoupling the pre-trained model from the evaluator, whose primary role is to gauge the informativeness of potential explanatory subgraphs. Our experiments show that XGExplainer consistently improves the performance of local and global explainer techniques and outperforms state-of-the-art methods on all datasets for node and graph classification tasks.