Research map / Graph learning / GNN architectures
Generating Diagnostic and Actionable Explanations for Fair 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 Geb, a multi-objective evolutionary framework that finds compact biased subgraphs as diagnostic explanations and counterfactual debiased subgraphs (structure and features) as actionable explanations for fair GNNs, using an influenced-nodes gradient approximation to avoid retraining.
Problem
Explaining which subgraph patterns cause biased behavior in GNNs and what actionable edits could rectify that bias, while keeping explanations compact and classification accuracy intact.
Approach
Formulates fairness explanation generation as multi-objective combinatorial optimization (compactness, bias attribution, accuracy) solved by a dedicated evolutionary algorithm, with an influenced-nodes-based gradient approximation that updates the GNN in one gradient step instead of retraining, plus a two-stage diagnosis-then-repair pipeline.
Key results
- Geb variants reduce ΔSP by roughly 3% to 9% relative to vanilla GNN counterparts while keeping AUC comparable (e.g., Geb-GCN NBA ΔSP 0.5 vs GCN 7.9).
- The influenced-nodes gradient approximation cuts running time by 10% to 93% without significantly affecting AUC or fairness.
- Geb yields more compact explanations than GEAR and NIFTY, and a theoretical fairness risk bound links debiasing subgraphs to lower statistical parity risk.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Geb to FairGAT | FairGAT · NBA | ∆SP: 0.7±0.3 -> 0.5±0.3 | Helped |
| Apply Geb to FairGAT | FairGAT · NBA | AUC: 77.5±0.7 -> 74.9±0.7 | Hurt |
| Apply Geb to FairGAT | FairGAT · Pokec-z | ∆SP: 0.5±0.3 -> 0.6±0.3 | No clear effect |
| Apply Geb to FairGAT | FairGAT · Pokec-z | ∆EO: 0.8±0.3 -> 0.6±0.3 | Helped |
| Apply Geb to FairGAT | FairGAT · Pokec-n | ∆SP: 0.6±0.3 -> 0.2±0.1 | Helped |
| Apply Geb to FairGAT | FairGAT · Pokec-n | AUC: 74.9±0.4 -> 74.8±0.2 | No clear effect |
| Remove the identified biased subgraphs without performing stage two (ablation) | Geb-GCN-S1 and Geb-GAT-S1 · NBA, Pokec-z, and Pokec-n | ∆SP | Helped |
| Remove the identified biased subgraphs without performing stage two (ablation) | Geb-GCN-S1 and Geb-GAT-S1 · NBA, Pokec-z, and Pokec-n | AUC | Hurt |
| Replace influenced-nodes gradient approximation with a few gradient-descent update steps (ablation) | Geb · NBA, Pokec-z, and Pokec-n | classification and fairness performance | No clear effect |
| Train the model from scratch instead of using influenced-nodes gradient approximation (ablation) | Geb · NBA, Pokec-z, and Pokec-n | classification and fairness performance | No clear effect |
| Remove counterfactual features (ablation) | Geb · NBA, Pokec-z, and Pokec-n | fairness | Hurt |
Surprising findings
- Randomly flipping positive edges to negative does not reliably improve fairness while preserving accuracy, supporting heuristic search over random edits.
Limitations
- Geb is slower than other baselines, though framed as a one-time cost
- Some Geb variants show notable ΔSP or AUC trade-offs on Pokec datasets (e.g., Geb-GCN Pokec-n ΔSP 6.3)
- Gradient approximation relies on a local-dependence assumption that may not hold for distant influence
Experimental setup
Datasets: NBA Pokec-z Pokec-n
Benchmarks: NBA Pokec-z Pokec-n
Metrics: AUC statistical parity (ΔSP) equal opportunity (ΔEO) compactness running time
Baselines: ALFR ALFR-e Debias Debias-e FCGE FairGCN FairGAT NT-FairGNN GE-REFEREE GEAR NIFTY
Builds on: NIFTY GEAR FairGCN REFEREE Koh and Liang influence functions
Closest papers by meaning
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.88)
- Generating In-Distribution Proxy Graphs for Explaining Graph Neural Networks (ICML 2024 · similarity 0.87)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.87)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.87)
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.86)
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (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.