Research map / Graph learning / Graph self-supervised learning
Graph Segmentation and Contrastive Enhanced Explainer for Graph Neural Networks
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 GSCExplainer, a model-agnostic GNN explainer that segments graphs into explanatory and redundant subgraphs and uses edge-perturbation contrastive learning plus an effectiveness loss to improve explanation fidelity.
Problem
Model-agnostic GNN explainers rely on the GNN's own predictions as supervision because ground-truth explanatory subgraphs are scarce, which can bias explanations away from the model's true decision basis.
Approach
Partition the input graph by edge importance into explanatory and redundant subgraphs, then build positive and negative pairs by edge-perturbing one subgraph while keeping the other fixed, and apply a contrastive loss alongside a prediction-validity loss so explanatory subgraphs are learned without ground-truth labels.
Key results
- On MUTAG at edge ratio 0.5, GSCExplainer reaches 75.1% explanation accuracy vs 71.4% for OrphicX, the best baseline.
- On NCI1 at edge ratio 0.9, GSCExplainer reaches 84.5% vs 85.4% for OrphicX, so it does not win at every sparsity level.
- On Tree-Cycles GSCExplainer achieves 100% accuracy at all sparsity levels, matching or exceeding OrphicX.
- Removing graph contrastive learning drops accuracy by 6.4 points on MUTAG and 9.1 points on NCI1 (average).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add graph contrastive learning (GCL) to GSCExplainer (ablation) | GSCExplainer · MUTAG | prediction accuracy: drops of 6.4 percentage points | Helped |
| Add graph contrastive learning (GCL) to GSCExplainer (ablation) | GSCExplainer · NCI1 | prediction accuracy: drops of 9.1 percentage points; decline exceeding 10 percentage points | Helped |
| Add graph contrastive learning (GCL) to GSCExplainer (ablation) | GSCExplainer · BA-SHAPES, 5 or 6 explanatory-subgraph edges | prediction accuracy: 94.1 vs 63.2; 95.6 vs 63.2 | Helped |
| Add graph contrastive learning (GCL) to GSCExplainer (ablation) | GSCExplainer · TREE-CYCLES | prediction accuracy: 100 100 100 100 100 | No clear effect |
Surprising findings
- On NCI1 at edge ratio 0.9, GSCExplainer (84.5%) is slightly below OrphicX (85.4%).
Limitations
- Explanation accuracy is evaluated only against the GNN's own predictions, not against ground-truth explanations.
- Sensitive hyperparameters (edge perturbation ratio alpha, number of negatives m) require tuning.
- The paper does not state compute resources or release code.
Experimental setup
Datasets: MUTAG NCI1 BA-Shapes Tree-Cycles
Benchmarks: MUTAG NCI1 BA-Shapes Tree-Cycles
Metrics: explanation accuracy (acc_exp) sparsity
Baselines: OrphicX GNNExplainer PGExplainer Gem
Builds on: GNNExplainer PGExplainer Gem GraphCL
Closest papers by meaning
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.90)
- EiG-Search: Generating Edge-Induced Subgraphs for GNN Explanation in Linear Time (ICML 2024 · similarity 0.90)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
- Higher Order Structures for Graph Explanations (AAAI 2025 · similarity 0.89)
- Stratified GNN Explanations through Sufficient Expansion (AAAI 2024 · similarity 0.89)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · 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.