Research map / Graph learning / GNN architectures
Self-Interpretable Graph Learning with Sufficient and Necessary Explanations
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.
Proposes SUNNY-GNN, a self-interpretable graph learner that augments generated explanations by adding or removing edges and applies a contrastive loss with distance and confidence coefficients to make explanations both sufficient and necessary, improving node classification accuracy and fidelity.
Problem
Self-interpretable graph learning methods lose prediction accuracy relative to standard GNNs because their built-in explanations are not both sufficient and necessary, so biased explanations hurt GNN performance.
Approach
Augment each generated explanatory subgraph into positive samples (edges added) and negative samples (edges removed), then use a contrastive loss to pull positives closer and push negatives apart, with distance coefficients to make samples hard and confidence coefficients to filter unreliable ones.
Key results
- SUNNY-GNN improves average prediction accuracy over the base GNNs by 1.1% and outperforms the best self-interpretable baseline by 3.5% on average (up to 6.0%) across six homogeneous datasets.
- On explainability fidelity, SUNNY-GNN outperforms baselines by 13.1% on average (up to 33.5%), achieving fid- near 0 and high fid+ (e.g., 90.24% fid+ on Cora with GCN).
- On heterogeneous graphs, SUNNY-GNN with Simple-HGN reaches 63.10% on IMDB, 95.73% on DBLP and 92.10% on ACM, outperforming the GSAT variant by 0.8% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SUNNY-GNN explanations and contrastive training to GCN | GCN · Citeseer, Cora, Pubmed, Amazon, Coauthor-CS, Coauthor-Physics | accuracy: Average impro. (%): 3.6 ↑; 3.1 ↑; 3.4 ↑; 4.8 ↑; 3.2 ↑; 3.0 ↑ | Helped |
| add SUNNY-GNN explanations and contrastive training to GAT | GAT · Citeseer, Cora, Pubmed, Amazon, Coauthor-CS, Coauthor-Physics | accuracy: Average impro. (%): 3.2 ↑; 3.1 ↑; 2.3 ↑; 3.0 ↑; 3.4 ↑; 6.0 ↑ | Helped |
| add SUNNY-GNN to Simple-HGN | Simple-HGN · IMDB | accuracy: 62.24±0.7 -> 63.10±0.9 | Helped |
| add SUNNY-GNN to Simple-HGN | Simple-HGN · DBLP | accuracy: 95.70±0.8 -> 95.73±0.5 | No clear effect |
| add SUNNY-GNN to Simple-HGN | Simple-HGN · ACM | accuracy: 91.44±0.5 -> 92.10±0.6 | Helped |
| add contrastive loss Lcts (ablation) | SUNNY-GNN with GAT · Citeseer | accuracy: 69.01±1.2 -> 71.26±0.7 | Helped |
| add contrastive loss Lcts (ablation) | SUNNY-GNN with GAT · Citeseer | fid+: 76.50±1.4 -> 79.25±2.4 | Helped |
| add contrastive loss Lcts (ablation) | SUNNY-GNN with GAT · Citeseer | fid−: 1.78±0.4 -> 0.46±0.5 | Helped |
| add distance coefficients δ (ablation) | SUNNY-GNN with GAT · Citeseer | accuracy: 70.78±0.7 -> 71.26±0.7 | Helped |
| add distance coefficients δ (ablation) | SUNNY-GNN with GAT · Citeseer | fid+: 79.13±1.8 -> 79.25±2.4 | No clear effect |
| add distance coefficients δ (ablation) | SUNNY-GNN with GAT · Citeseer | fid−: 0.47±0.5 -> 0.46±0.5 | No clear effect |
| add confidence coefficients η (ablation) | SUNNY-GNN with GAT · Citeseer | accuracy: 70.48±0.8 -> 71.26±0.7 | Helped |
Surprising findings
- Existing self-interpretable graph learning methods can degrade GNN accuracy by up to 15.2% relative to GNNs trained without built-in explanations.
- Removing the contrastive loss causes a 3.4% drop in overall performance on Citeseer with GAT, indicating that the self-interpretation objective alone harms accuracy.
Limitations
- SE-GNN could not be run on Amazon and Coauthor-Physics due to memory costs above 90 GB.
- Training cost scales with the number of augmented samples and local computation graph size.
- Heterogeneous extension is described as preliminary and the authors leave more complex scenarios to future work.
Experimental setup
Datasets: Citeseer Cora Pubmed Amazon Photo Coauthor-CS Coauthor-Physics IMDB DBLP ACM
Benchmarks: Citeseer Cora Pubmed Amazon Photo Coauthor-CS Coauthor-Physics IMDB DBLP ACM
Metrics: classification accuracy negative fidelity (fid-) positive fidelity (fid+)
Baselines: GSAT CAL SE-GNN ProtGNN GNNExplainer PGExplainer ReFine Simple-HGN
Builds on: GSAT CAL SE-GNN ProtGNN Simple-HGN
Closest papers by meaning
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.93)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.92)
- Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs (ICLR 2025 · similarity 0.91)
- GraphTrail: Translating GNN Predictions into Human-Interpretable Logical Rules (NeurIPS 2024 · similarity 0.91)
- Redundancy Undermines the Trustworthiness of Self-Interpretable GNNs (ICML 2025 · similarity 0.91)
- Graph Segmentation and Contrastive Enhanced Explainer for Graph Neural Networks (AAAI 2025 · similarity 0.90)
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.