A Comparative Evaluation of Post-hoc Explainability Methods for Graph Neural Networks in Node Classification

Abstract

Graph neural networks achieve strong predictive accuracy on node classification tasks while offering little insight into which structural elements drive a given prediction. A growing family of post-hoc explainers addresses this gap by attributing importance to edges or nodes in a trained model's computational neighborhood, yet the methods are rarely compared under a shared protocol. This paper evaluates five representative post-hoc explainers, spanning perturbation-based masking, parameterized inductive masking, gradient-based attribution, and Shapley-value sampling, on two fixed backbone architectures across seven node classification benchmarks: four synthetic graphs with planted ground-truth motifs and three citation networks. Explanations are compared on fidelity, sparsity, ground-truth consistency, and stability under small feature perturbations. Gradient-based and Shapley-based attributions produce the sparsest and most stable explanations, the perturbation-based GNNExplainer achieves the highest fidelity on the graph convolutional backbone, and no single method dominates once ground-truth consistency is considered, with agreement between explanation edges and planted motifs remaining close to chance across all five methods. A backbone-sensitivity analysis further shows that switching from a graph convolutional network to a graph attention network changes the relative ranking of methods, indicating that explainer quality is not a fixed property of the explainer alone. The results are reported with explicit attention to the gap between what a model was trained to predict from and what an explanation recovers, and the paper closes by outlining the scope conditions under which each method family is preferable.