New NICE Method Addresses Scale Drift in Graph Neural Network Explanations
Explanations for Graph Neural Networks (GNNs) typically utilize a Perturb-Query approach, where input alterations lead to insights about the significance of graph components based on changes in predictions. However, these perturbations can result in significant distribution shifts, undermining the dependability of the explanations derived from predictions. Previous solutions have concentrated on refining the perturbed graphs or stabilizing model outputs. The latest study reassesses the perturbation process itself. The researchers reveal that Element-wise Masking (EM) diminishes edge-induced messages to near zero, causing a deterministic scale contraction across message-passing layers, a phenomenon termed Scale Drift. To address this, they propose Noise Corruption (NC), which introduces matched-norm noise to each message, maintaining the original scale characteristics while providing meaningful explanations. This research, titled 'NICE: Scale-Stable Perturbations for Graph Neural Network Explanations via Noise Corruption,' can be found on arXiv under identifier 2608.16038.
Key facts
- Post-hoc GNN explainers typically follow a Perturb-Query paradigm.
- Perturbations can cause substantial distribution shift, affecting explanation reliability.
- Element-wise Masking (EM) suppresses edge-induced messages toward zero.
- EM causes deterministic scale contraction across message-passing layers, termed Scale Drift.
- Scale Drift conflates information corruption with deviations in propagation scale.
- The paper proposes Noise Corruption (NC) as a scale-stable alternative.
- Noise Corruption perturbs messages using matched-norm noise.
- The paper is available on arXiv under identifier 2608.16038.
Entities
Institutions
- arXiv