Research map / Graph learning / GNN architectures
Redundancy Undermines the Trustworthiness of Self-Interpretable GNNs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that self-interpretable GNN explanations are inconsistent across random seeds because of redundant edges caused by loose conciseness constraints, and proposes Explanation Ensemble (averaging explanations from independently trained models) as a tuning-free mitigation.
Problem
Self-interpretable GNNs trained with different random seeds produce inconsistent and sometimes inaccurate explanations, which undermines user trust in high-stakes applications.
Approach
Identifies redundancy from weak conciseness constraints (global hyperparameters) as the root cause of explanation inconsistency and inaccuracy, and proposes Explanation Ensemble, which averages edge importance scores over independently trained models so redundant edges with high variance get lower average weights.
Key results
- Explanation Ensemble improves SHD and AUC in all 48 evaluated cases (GIN backbone, four self-interpretable frameworks, four datasets) and ACC in 46 of 48.
- SWA improves explanation consistency in only 7 of 16 tests, so training instability does not fully explain inconsistency.
- On BA-2MOTIFS, EE raises AUC for Type I attention-based models by up to 7.40 points (e.g., BENZENE 83.50 to 88.61 with EA; EE SHD reductions up to 9.10 on MUTAGENICITY).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| aggregate explanations from independently trained models with Explanation Ensemble | self-interpretable GNNs · BA-2MOTIFS, 3MR, BENZENE, MUTAGENICITY | SHD | Helped |
| aggregate explanations from independently trained models with Explanation Ensemble | self-interpretable GNNs · BA-2MOTIFS, 3MR, BENZENE, MUTAGENICITY | AUC | Helped |
| aggregate explanations from independently trained models with Explanation Ensemble | self-interpretable GNNs · BA-2MOTIFS, 3MR, BENZENE, MUTAGENICITY | ACC | Helped |
| apply Stochastic Weight Averaging | self-interpretable GNNs · four evaluated datasets | SHD | Mixed |
| apply Stochastic Weight Averaging | self-interpretable GNNs · four evaluated datasets | AUC | Mixed |
| add embedding alignment loss | self-interpretable GNNs · BA-2MOTIFS | AUC: up to 3% | Hurt |
| add contrastive loss | self-interpretable GNNs · BA-2MOTIFS | explanation accuracy | Helped |
| add contrastive loss | self-interpretable GNNs · BENZENE | performance across all metrics | Hurt |
| increase ensemble size from 1 to 2 models (ablation) | Explanation Ensemble · BA-2MOTIFS, 3MR, BENZENE, MUTAGENICITY | SHD and AUC | Helped |
Surprising findings
- Explanation inconsistency persists even on synthetic datasets without spurious correlations, contrary to the spurious-correlation explanation from post-hoc work.
- Embedding alignment, which helps post-hoc explainers, fails to improve self-interpretable GNN explanations because the explainer aligns to a co-evolving target.
Limitations
- EE requires training multiple models, increasing training and deployment cost linearly with n.
- EE is incompatible with faithfulness metrics, and its effectiveness diminishes when irrelevant edges consistently receive high weights.
- Redundancy is shown to be hard to eliminate; EE is a reactive mitigation rather than a fix to the root cause.
Experimental setup
Datasets: BA-2MOTIFS 3MR BENZENE MUTAGENICITY
Benchmarks: BA-2MOTIFS 3MR BENZENE MUTAGENICITY
Metrics: SHD ROC-AUC accuracy
Baselines: SWA embedding alignment (EA) GIN without mitigation
Builds on: GIN GNN attention (GAT) CAL causal attention learning DIR/invariant rationale contrastive explanation regularization (Deng & Shen 2024) stochastic weight averaging
Closest papers by meaning
- Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs (ICLR 2025 · similarity 0.91)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.91)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.90)
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.90)
- Stratified GNN Explanations through Sufficient Expansion (AAAI 2024 · similarity 0.89)
- GNNX-BENCH: Unravelling the Utility of Perturbation-based GNN Explainers through In-depth Benchmarking (ICLR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.