Research map / Graph learning / Graph self-supervised learning
FairWire: Fair Graph Generation
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.
Analyzes structural bias in link prediction with multi-valued sensitive attributes, proposes a batch-wise intra/inter-edge ratio fairness regularizer, and embeds it in a diffusion-based graph generator (FairWire) to produce fairer synthetic graphs.
Problem
Structural bias in graph topology leads to disparate link prediction outcomes across sensitive groups, and graph generative models (diffusion) amplify this bias; existing fairness methods assume binary sensitive attributes and fair graph generation work requires class labels.
Approach
Theoretical bound on statistical parity for link prediction under multi-valued sensitive attributes yields a regularizer that pushes the intra-/inter-edge ratio per sensitive group toward its group-size-determined optimum, applied batch-wise; FairWire adds this regularizer to a discrete diffusion graph generator that also generates synthetic sensitive attributes, so fairness can be enforced withou
Key results
- Link prediction: L_FairWire reduces Cora Δ_SP from 27.01% (GNN) to 4.76% with AUC 92.18% vs 94.43%.
- Graph generation: FairWire reduces Cora Δ_SP from 35.72% (GraphMaker synthetic) to 12.91% with AUC 86.49% vs 87.29%.
- Diffusion-generated synthetic graphs amplify structural bias relative to originals on all four datasets (e.g., Cora Δ_SP 27.71% -> 35.72%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the FairWire regularizer to GNN link prediction training | GCN on Cora · Cora link prediction | ΔSP: 27.01 ± 1.38 -> 4.76 ± 0.24 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Cora · Cora link prediction | ΔEO: 9.11 ± 1.43 -> 2.05 ± 0.37 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Cora · Cora link prediction | AUC: 94.43 ± 0.74 -> 92.18 ± 1.03 | Hurt |
| Add the FairWire regularizer to GNN link prediction training | GCN on Citeseer · Citeseer link prediction | ΔSP: 27.40 ± 1.24 -> 8.62 ± 0.80 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Citeseer · Citeseer link prediction | ΔEO: 7.37 ± 1.33 -> 1.29 ± 0.68 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Citeseer · Citeseer link prediction | AUC: 96.16 ± 0.28 -> 96.00 ± 0.23 | No clear effect |
| Add the FairWire regularizer to GNN link prediction training | GCN on Amazon Photo · Amazon Photo link prediction | ΔSP: 32.65 ± 0.95 -> 27.75 ± 0.52 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Amazon Photo · Amazon Photo link prediction | ΔEO: 8.01 ± 0.52 -> 7.11 ± 0.41 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Amazon Computer · Amazon Computer link prediction | ΔSP: 23.70 ± 0.79 -> 20.44 ± 0.44 | Helped |
| Add the FairWire regularizer to GNN link prediction training | GCN on Amazon Computer · Amazon Computer link prediction | ΔEO: 5.51 ± 0.79 -> 4.24 ± 0.51 | Helped |
| Use FairWire instead of fairness-agnostic GraphMaker | Diffusion graph generation on Cora · Cora graph generation | ΔSP: 35.72 ± 1.74 -> 12.91 ± 6.35 | Helped |
| Use FairWire instead of fairness-agnostic GraphMaker | Diffusion graph generation on Cora · Cora graph generation | ΔEO: 13.27 ± 0.81 -> 4.31 ± 3.59 | Helped |
Surprising findings
- Synthetic graphs from the diffusion generator exhibit higher Δ_SP and Δ_EO than the original graphs on all datasets, i.e., generative modeling amplifies structural bias.
- Some FairWire settings with large lambda cause substantial utility drops (e.g., Citeseer AUC 83.47% at lambda=1 in generation).
Limitations
- FairAdj could not be run on Amazon Photo/Computer due to memory limits.
- Fairness results in generation have high variance on some datasets (e.g., Cora Δ_SP ±6.35).
- Analysis relies on assumptions A1-A3 (e.g., intra-edges dominate inter-edges, non-extreme group sizes) that may not hold generally.
Experimental setup
Datasets: Cora Citeseer Amazon Photo Amazon Computer
Benchmarks: Cora Citeseer Amazon Photo Amazon Computer
Metrics: AUC statistical parity (Δ_SP) equal opportunity (Δ_EO) 1-Wasserstein degree distance 1-Wasserstein clustering coefficient distance
Baselines: GNN Adversarial regularization FairDrop FairAdj GraphMaker
Builds on: structured denoising diffusion for discrete data (D3PM-style) GraphMaker FairAdj FairDrop
Closest papers by meaning
- FDGen: A Fairness-Aware Graph Generation Model (ICML 2025 · similarity 0.88)
- Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction (ICML 2024 · similarity 0.86)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.86)
- FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.86)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.86)
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 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.