Research map / Graph learning / Graph self-supervised learning
Graph Inverse Style Transfer for Counterfactual Explainability
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 GIST, a backtracking graph counterfactual explainer that overshoots the oracle's decision boundary and then learns to backtrack toward the input's spectral style via a transformer-based edge-probability generator, evaluated against forward-perturbation GCE baselines on 8 datasets.
Problem
Graph counterfactual explanations must flip a black-box GNN's prediction with minimal, structure-preserving changes, but existing forward perturbation methods often overshoot, discard structure, or need repeated oracle calls during training.
Approach
Reframes graph counterfactual generation as backtracking: overshoot the oracle boundary with a graph from another class, then learn a reverse transformer-based edge generator whose loss interpolates between spectral style alignment with the input (normalized Laplacian eigenvalues) and content preservation with the overshot graph.
Key results
- GIST achieves the highest validity on 4 of 8 datasets, e.g., 0.969 on AIDS and 0.791 on PROTEINS, and reports an average +7.6% validity gain over the second-best method.
- GIST reports fidelity of 0.957 on AIDS and 0.425 on PROTEINS, with the paper claiming an average +45.5% fidelity gain over the second-best method.
- Empirically, GIST's styled Laplacian closely matches the optimal convex combination of input and overshot spectra, with an average eigenvalue error of 2.051e-3 on sampled AIDS pairs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GIST instead of the second-best explainer | Graph counterfactual explainers · 8 graph classification benchmarks | validity: +7.6% | Helped |
| Use GIST instead of the second-best explainer | Graph counterfactual explainers · 8 graph classification benchmarks | fidelity: +45.5% | Helped |
| Set interpolation factor α = 0.5 (ablation) | GIST · Graph counterfactual generation | reconstruction error | Hurt |
Surprising findings
- CF-GNNExp. and GIST tie at 0.810 validity on BZR, and CF-GNNExp. outperforms GIST on MSRC21 validity tie aside, several baselines (CF^2, iRand) produce near-zero validity on several datasets.
- The paper notes that the backtracking method only guarantees Phi(G) != Phi(G*), not that G* keeps the overshot class, so class maintenance is not guaranteed in multiclass settings.
Limitations
- No guarantee that backtracked counterfactuals keep the class of the overshot graph in multiclass settings.
- Reduced performance at alpha=0.5 trade-off between content and style.
- Spectral alignment uses L1 over eigenvalues which needs padding for graphs of different sizes, and Wasserstein distance is left to future work.
Experimental setup
Datasets: AIDS BAShapes BBBP BZR COLORS-3 ENZYMES MSRC21 PROTEINS
Benchmarks: AIDS BAShapes BBBP BZR COLORS-3 ENZYMES MSRC21 PROTEINS
Metrics: validity fidelity GED sparsity runtime oracle calls
Baselines: iRand CF-GNNExp. CF^2 CLEAR RSGG-CE
Builds on: CycleGAN Gumbel-Softmax GRETEL
Closest papers by meaning
- Structure Your Data: Towards Semantic Graph Counterfactuals (ICML 2024 · similarity 0.88)
- COMRECGC: Global Graph Counterfactual Explainer through Common Recourse (ICML 2025 · similarity 0.88)
- Robust Stochastic Graph Generator for Counterfactual Explanations (AAAI 2024 · similarity 0.87)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.87)
- UNR-Explainer: Counterfactual Explanations for Unsupervised Node Representation Learning Models (ICLR 2024 · similarity 0.86)
- CoDy: Counterfactual Explainers for Dynamic Graphs (ICML 2025 · similarity 0.86)
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.