Self-Inconsistency in GNN Explanations and Self-Denoising Strategy

Researchers have discovered that the explanations generated by Self-Interpretable Graph Neural Networks (SI-GNNs) can exhibit self-inconsistency; when the model is applied again to its own explanatory subgraph, it produces varying explanations. This inconsistency is primarily due to context perturbation induced by re-explanation. A hypothesis regarding latent signal assignment clarifies why certain edges are more affected by this perturbation, while conciseness regularization plays a role in this assignment. The authors introduce Self-Denoising (SD), a training-free, model-agnostic post-processing approach that refines explanations with a single additional forward pass. Experiments conducted across various SI-GNN frameworks and backbone architectures validate its effectiveness.