Research map / Graph learning / GNN architectures
Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations 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 DAB-GNN, which uses three GNN-based disentanglers to isolate and amplify attribute, structure, and potential biases, then debiases with a bias contrast optimizer and a Wasserstein-based fairness harmonizer for fair node classification.
Problem
Fairness in GNN node embeddings is compromised by attribute bias, structure bias, and potential bias arising from message passing, which existing fairness-aware GNNs treat jointly in a single entangled embedding.
Approach
Disentangle node embeddings into three bias-specific components (attribute via attribute-kNN graph message passing, structure via random features on the input graph, potential via residual of an MLP over concatenated embeddings), then debias each with a Wasserstein fairness regularizer and enforce separation between components with a Frobenius-norm contrast loss.
Key results
- On Recidivism, DAB-GNN reaches AUC 93.41 and SP 0.73, versus 86.84 AUC and 8.13 SP for L3-Vanilla, per the paper's reported improvement.
- DAB-GNN achieves the best SP and EO on all five datasets in the main table, e.g., SP 0.44 and EO 0.45 on Credit versus 0.40 SP for FairVGNN, which has the lowest SP there.
- Ablation: removing the fairness harmonizer loss (w/o L_fh) increases SP on NBA from 1.12 to 9.50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAB-GNN instead of the strongest ACC baseline, PFR-AX | NBA node classification · NBA | ACC: 70.63 -> 71.39 | Helped |
| Use DAB-GNN instead of the strongest ACC baseline, CAF | Recidivism node classification · Recidivism | ACC: 86.79 -> 89.99 | Helped |
| Use DAB-GNN instead of the strongest ACC baseline, FairVGNN | Credit node classification · Credit | ACC: 77.92 -> 78.19 | Helped |
| Use DAB-GNN instead of the strongest ACC baseline, FairVGNN | Pokec_n node classification · Pokec_n | ACC: 69.51 -> 67.18 | Hurt |
| Use DAB-GNN instead of the strongest ACC baseline, NIFTY | Pokec_z node classification · Pokec_z | ACC: 65.71 -> 68.56 | Helped |
| Add AbDisen (ablation) | DAB-GNN · NBA | SP: 6.08 -> 1.12 | Helped |
| Add SbDisen (ablation) | DAB-GNN · NBA | SP: 5.97 -> 1.12 | Helped |
| Add PbDisen (ablation) | DAB-GNN · NBA | SP: 6.52 -> 1.12 | Helped |
| Add the fairness harmonizer (ablation) | DAB-GNN · NBA | SP: 9.50 -> 1.12 | Helped |
| Add the bias contrast optimizer (ablation) | DAB-GNN · NBA | SP: 6.76 -> 1.12 | Helped |
| Add AbDisen (ablation) | DAB-GNN · Recidivism | ACC: 88.82 -> 89.99 | Helped |
| Add SbDisen (ablation) | DAB-GNN · Recidivism | ACC: 91.59 -> 89.99 | Hurt |
Surprising findings
- On the Credit dataset, FairVGNN reports lower SP (0.40) than DAB-GNN (0.44), so DAB-GNN is not best on every fairness metric.
- Removing individual disentanglers or regularizers sometimes improves accuracy, so the most impactful component for accuracy varies across datasets.
Limitations
- Potential bias is defined loosely and its nature is acknowledged as uncertain by the authors.
- Several competitors run out of memory on the larger Pokec datasets and on Credit, limiting full comparisons.
- Results are reported only for node classification in the main text; link prediction is mentioned but not evaluated.
Experimental setup
Datasets: NBA Recidivism Credit Pokec_n Pokec_z
Benchmarks: NBA Recidivism Credit Pokec_n Pokec_z
Metrics: ACC AUC F1 Statistical Parity Equality of Opportunity
Baselines: L1-Vanilla GCN L3-Vanilla GCN FairGNN NIFTY EDITS FairVGNN CAF GEAR BIND PFR-AX PostProcess FairSIN
Builds on: FairGNN EDITS GCN FairSIN
Closest papers by meaning
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.89)
- DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning (ICML 2025 · similarity 0.89)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.88)
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.88)
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.87)
- Adversarial Attacks on Fairness of Graph Neural Networks (ICLR 2024 · 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.