Research map / Graph learning / GNN architectures
Train Once and Explain Everywhere: Pre-training Interpretable Graph Neural Networks
Jun Yin, Chaozhuo Li, Hao Yan, Jianxun Lian, Senzhang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add explainer pre-training on synthetic PT-Motifs (ablation) | π-GNN · Mutag | explanation ROC-AUC: 95.29 -> 99.81 | Improved |
| add explainer pre-training on synthetic PT-Motifs (ablation) | π-GNN · BA-2Motifs | explanation ROC-AUC: 93.19 -> 99.33 | Improved |
| add explainer pre-training on synthetic PT-Motifs (ablation) | π-GNN · Spurious-Motif b=0.9 | explanation ROC-AUC: 89.64 -> 96.39 | Improved |
| add explainer pre-training on synthetic PT-Motifs (ablation) | π-GNN · Graph-SST2 | accuracy: 83.48 -> 88.05 | Improved |
| add explainer pre-training on synthetic PT-Motifs (ablation) | π-GNN · Spurious-Motif b=0.7 | accuracy: 71.02 -> 77.52 | Improved |
| add structural pattern learning module (vs N-thread GNN encoders) (ablation) | π-GNN · Mutag interpretation / Graph-SST2 prediction | ROC-AUC / accuracy: 3.85% / 1.08% average decrease when removed | Improved |
| add hypergraph refining module (ablation) | π-GNN · Mutag interpretation / Graph-SST2 prediction | ROC-AUC / accuracy: 6.56% / 1.84% average decrease when removed | Improved |
| keep both modules (vs removing both, π-GNN-ALL) (ablation) | π-GNN · Mutag interpretation / Graph-SST2 prediction | ROC-AUC / accuracy: 7.56% / 2.16% average decrease when removed | Improved |
| shrink pre-training set from 80,000 to 10,000 graphs (PT-Motifs-S) (ablation) | π-GNN · Mutag | explanation ROC-AUC: 99.81 -> 96.28 | Hurt |
| shrink pre-training set from 80,000 to 50,000 graphs (PT-Motifs-M) (ablation) | π-GNN · Graph-SST2 | accuracy: 88.05 -> 87.59 | No clear effect |
| pre-train on small PT-Motifs-S instead of no pre-training (ablation) | π-GNN · Ogbg-Molhiv | AUC: 79.71 -> 79.82 | No clear effect |
| pre-train on heavily imbalanced dataset (b = 0.7, i about 9.3) (ablation) | π-GNN · Mutag, Molhiv, Graph-SST2 | interpretation/prediction | Hurt |
Closest papers by meaning
- How Interpretable Are Interpretable Graph Neural Networks? (ICML 2024 · similarity 0.91)
- From GNNs to Trees: Multi-Granular Interpretability for Graph Neural Networks (ICLR 2025 · similarity 0.91)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.90)
- The Intelligible and Effective Graph Neural Additive Network (NeurIPS 2024 · similarity 0.89)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.89)
- GraphTrail: Translating GNN Predictions into Human-Interpretable Logical Rules (NeurIPS 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.
