Research map / Graph learning / Graph self-supervised learning
Robust Stochastic Graph Generator for Counterfactual 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 RSGG-CE, a residual-GAN graph autoencoder generator that samples counterfactual graphs from its latent space using partial-order edge sampling, evaluated against CF², CLEAR, MEG and G-CounteRGAN on Tree-Cycles and ASD.
Problem
Graph counterfactual explanations for graph classification: generating valid, plausible, minimally changed counterfactual graphs with low cost, without retraining per input and without heuristic perturbation rules.
Approach
A GCN discriminator and a graph-autoencoder generator (Residual GAN) trained only on the explainee class produce residual edge weights added to the input adjacency; at inference, edges are sampled in a partial order (existing edges first, then non-existing) and the oracle is queried only on the non-existing group until a valid counterfactual is found.
Key results
- On Tree-Cycles, RSGG-CE reaches Correctness 0.885 vs 0.530 for the next best (MEG), and GED 11.000 vs 27.564 for CF^2.
- On ASD, RSGG-CE reports Correctness 0.603 (about 8.84% gain over the second-best CLEAR at 0.554) and GED 234.853 vs 655.661 for CF^2.
- RSGG-CE runtime is 0.083 s on Tree-Cycles versus 4.811 s for CF^2 (the next fastest) and 80.000 s on ASD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RSGG-CE instead of MEG | Graph counterfactual explanations · Tree-Cycles (TC) | Correctness: 0.530 -> 0.885 | Helped |
| Use RSGG-CE instead of CLEAR | Graph counterfactual explanations · Autism Spectrum Disorder (ASD) | Correctness: 0.554 -> 0.603 | Helped |
| Increase the number of nodes per cycle (ablation) | RSGG-CE · Tree-Cycles (TC) | Correctness | Helped |
| Increase the number of cycles per graph (ablation) | RSGG-CE · Tree-Cycles (TC) | GED | Hurt |
| Increase the number of nodes per graph (ablation) | RSGG-CE · Tree-Cycles (TC) | Correctness | No clear effect |
| Increase the number of dataset instances (ablation) | RSGG-CE · Tree-Cycles (TC) | GED | No clear effect |
| Increase the number of dataset instances (ablation) | RSGG-CE · Tree-Cycles (TC) | Correctness | Helped |
Surprising findings
- Factual-based CF^2 has zero oracle calls but poor correctness, showing oracle-call count alone is misleading.
- G-CounteRGAN has the worst GED and tends to densely connect graphs by adding non-existing edges.
Limitations
- Node classification adaptation is only sketched and not evaluated.
- Evaluated only on two datasets (one synthetic, one small real dataset).
- Partial-order sampling is described as a modular example; alternative orderings are not explored.
Experimental setup
Datasets: Tree-Cycles ASD
Benchmarks: Tree-Cycles ASD
Metrics: Runtime GED Oracle Calls Correctness Sparsity Fidelity Oracle Accuracy
Baselines: MEG CF^2 CLEAR G-CounteRGAN
Builds on: Residual GAN CF^2 CLEAR G-CounteRGAN GRETEL
Closest papers by meaning
- Structure Your Data: Towards Semantic Graph Counterfactuals (ICML 2024 · similarity 0.88)
- LeapFactual: Reliable Visual Counterfactual Explanation Using Conditional Flow Matching (NeurIPS 2025 · similarity 0.88)
- Counterfactual Explanations for Misclassified Images: How Human and Machine Explanations Differ (Abstract Reprint) (AAAI 2024 · similarity 0.88)
- Generating Likely Counterfactuals Using Sum-Product Networks (ICLR 2025 · similarity 0.88)
- Graph Inverse Style Transfer for Counterfactual Explainability (ICML 2025 · similarity 0.87)
- V-CECE: Visual Counterfactual Explanations via Conceptual Edits (NeurIPS 2025 · similarity 0.87)
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.