Why Self-Inconsistency Arises in GNN Explanations and How to Exploit It

Researchers identify why Graph Neural Network explanations produce inconsistent results when re-applied to their own outputs, attributing this to context perturbation during re-explanation. They propose Self-Denoising, a training-free post-processing method that improves explanation quality with minimal computational overhead.

Graph Neural Networks (GNNs) have become critical tools for decision-making in finance, chemistry, and social networks, yet their interpretability remains problematic. Self-Interpretable GNNs attempt to solve this by generating human-readable explanations for their predictions. The discovery that these explanations contradict themselves when reapplied—a phenomenon called self-inconsistency—undermines trust in GNN-based systems, particularly in high-stakes domains like credit assessment or fraud detection where explainability is legally mandated.

This research addresses a fundamental gap in understanding GNN behavior. By isolating re-explanation-induced context perturbation as the root cause, the authors provide a mechanistic explanation for why certain edges in explanatory graphs prove unreliable. The latent signal assignment hypothesis further clarifies how regularization techniques inadvertently create sensitivity to input perturbations. This theoretical groundwork is essential for building trustworthy AI systems.

The proposed Self-Denoising method offers practical value by requiring no model retraining—only a single additional forward pass. This positions it as an immediate solution for practitioners deploying existing GNN models. The 4-6% computational overhead is negligible compared to improved explanation reliability. For developers and researchers, this work provides both diagnostic tools and remediation strategies. Institutions implementing GNNs for regulatory compliance benefit most, as inconsistent explanations create audit risks.

Future work should explore whether self-inconsistency affects downstream applications like link prediction or node classification differently. Testing SD across more specialized domains—molecular graphs, knowledge graphs, social networks—would validate generalizability. Understanding whether self-inconsistency correlates with actual prediction errors remains an open question with significant implications for GNN deployment.

• →Self-inconsistency in GNN explanations stems from context perturbation when models re-analyze their own explanatory subgraphs.
• →Self-Denoising calibrates GNN explanations with a single forward pass and adds only 4-6% computational overhead.
• →Not all edges in explanatory graphs respond equally to perturbation; latent signal assignment determines sensitivity patterns.
• →Conciseness regularization, commonly used to improve explainability, can inadvertently increase sensitivity to context changes.
• →The method is model-agnostic and training-free, enabling immediate adoption across existing GNN frameworks without retraining.