Towards explaining graph neural networks via preserving prediction ranking and structural dependency

Graph Neural Networks (GNNs) have demonstrated their efficacy in representing graph-structured data, but their lack of explainability hinders their applicability to critical tasks. Existing GNNs explainers fail to consider the prediction ranking consistency between the original graph and the explanation, which is critical for preserving the fidelity of the explainer. Moreover, the structural dependency in the graph, reflecting the distinctive learning schema of the model, is ignored in current GNN explainers. To this end, we propose the NeuralSort based Plackett-Luce model to guide the parameter learning of the explainer via a differentiable ranking loss to ensure the explainer's fidelity to the GNNs. Additionally, a graph transformation schema explicitly modeling the edge dependency is proposed for constructing the mask generator. By integrating the aforementioned strategies, we propose a novel framework for explaining GNNs in a faithful manner. Through comprehensive experiments both for node classification and graph classification on BA-Shapes, BA-Community, Graph-Twitter, and Graph-SST5 datasets, the proposed framework achieves 149.67%, 51.43%, 40.747%, and 28.87% improvements compared with the state-of-the-art explainers in terms of fidelity to the GNNs. Data and code are available at https://github.com/ymzhang0103/RDPExplainer.