Title: Understanding the Origins of Self-Inconsistency in GNN Explanations and Leveraging It

Abstract: It has recently been noted that explanations generated by Self-Interpretable Graph Neural Networks (SI-GNNs) often exhibit self-inconsistency; specifically, reapplying the model to the subset of the graph it previously identified for explanation can yield a different result. Despite this observation, the underlying mechanisms driving such self-inconsistency have remained largely unclear. This study first isolates re-explanation-induced context perturbation as the immediate cause of these score variations. We subsequently propose a latent signal assignment hypothesis to elucidate why only specific edges respond to this perturbation, while also examining the influence of conciseness regularization on this assignment process. Since edges that lack consistency fail to offer reliable evidence for the model’s predictions, we introduce Self-Denoising (SD). This is a model-agnostic, training-free post-processing technique that refines explanations using merely one additional forward pass. Our empirical evaluations, conducted across various SI-GNN frameworks, backbone architectures, and standard benchmark datasets, validate our hypothesis. The results demonstrate that SD reliably enhances the quality of explanations, incurring only a marginal computational cost of approximately 4--6% in practical applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC