Research map / Graph learning / Graph self-supervised learning
FDGen: A Fairness-Aware Graph Generation Model
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 FDGen, a diffusion-based fair graph generation model that mitigates both structural and node-feature bias via a disentangled fairness regularizer, evaluated on four attributed graph datasets.
Problem
Graph generation models inherit societal biases; existing fair graph generation work addresses only structural bias and ignores feature bias in generated node attributes.
Approach
Disentangles node representations into channels, identifies the sensitive-related channel with a discriminator, and applies an MMD-based fairness regularizer plus inter/intra-group edge loss on the sensitive-irrelevant channels within a graph diffusion model (GraphARM-based), jointly reducing feature and structural bias.
Key results
- On Cora node classification with GCN, FDGen reaches 80.05% accuracy with delta DP 13.88% and delta EO 11.95%, the lowest fairness gaps among the compared methods (FairGen-GCN 14.16% and 13.35%).
- Ablations show FDGen-WD (regularizing full representations) lowers generation quality, while FDGen-WS (feature bias only) yields worse fairness than full FDGen.
- The paper reports FDGen achieves comparable or better generation quality than baselines while significantly improving graph fairness metrics on all four datasets (results shown only in figures).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FDGen-generated graphs instead of the original graph | GCN node classification · Cora | ∆DP (%): 27.01 ± 1.38 -> 13.88 ± 0.24 | Helped |
| Use FDGen-generated graphs instead of the original graph | GCN node classification · Cora | accuracy (%): 82.43 ± 0.34 -> 80.05 ± 1.03 | Hurt |
| Use FDGen instead of FairGen | GCN node classification · Cora | ∆EO (%): 13.35 ± 1.24 -> 11.95 ± 0.37 | Helped |
| Apply fair regularization directly to complete node representations rather than sensitive-irrelevant representations (FDGen-WD) (ablation) | FDGen · Graph generation | graph generation quality | Hurt |
| Mitigate only feature bias, leaving structural bias unmitigated (FDGen-WS) (ablation) | FDGen · Graph generation | fairness | Hurt |
| Mitigate only feature bias instead of both feature and structural bias (FDGen-WS) (ablation) | FDGen · Graph generation | graph generation quality | Helped |
Surprising findings
- FairWire-style and FairGen-style structural-only fairness methods leave substantial feature-level disparity, which FDGen's joint treatment reduces.
Limitations
- Evaluation limited to four datasets with binary sensitive attributes
- Generation quality and fairness results are reported mainly in figures without full numeric tables in the text
Experimental setup
Datasets: Cora Citeseer Photo Computer
Benchmarks: Cora Citeseer Photo Computer
Metrics: accuracy F1 delta DP delta EO MMD degree distribution MMD clustering coefficient fair degree distribution fair clustering coefficient
Baselines: GraphARM FairAdj FG2AN FairGen FairWire
Builds on: GraphARM FairWire FairGen distance covariance
Closest papers by meaning
- FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.88)
- FairWire: Fair Graph Generation (NeurIPS 2024 · similarity 0.88)
- Fairness-Aware Estimation of Graphical Models (NeurIPS 2024 · similarity 0.86)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.86)
- Intersectional Unfairness Discovery (ICML 2024 · similarity 0.86)
- Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction (ICML 2024 · similarity 0.85)
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.