Research map / Graph learning / Graph self-supervised learning
Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness
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 Fair Graph U-net, a graph U-net variant whose pooling, subgraph identification, and edge augmentation modules enforce group fairness while bounding individual unfairness, evaluated on German, Credit, and Bail node classification datasets.
Problem
Graph pooling in GNNs can amplify subgroup disparities and structural bias (intra-group edge over-linking) and typically ignores individual fairness, especially when group fairness constraints are enforced.
Approach
Combines fair top-k pooling with equal per-group selection, a probability distribution over valid rankings (inspired by Rawlsian difference principle and distributive justice) that spreads individual losses uniformly, fair edge augmentation with an intra/inter-group loss term, and fairness-aware PageRank-based subgraph identification within a Graph U-net.
Key results
- On German, Fair Graph U-net reports SPD 0.05, EOD 0.04, and MIU 0.05 versus 0.06/0.04/0.18 for the best fairness baseline FairAGG.
- On Credit, Fair Graph U-net achieves accuracy 0.75 and SPD 0.06 with MIU 0.06, outperforming FairGNN's SPD of 0.12.
- On Bail, Fair Graph U-net reaches SPD 0.04, EOD 0.02, and MIU 0.03, the lowest fairness gaps among methods reported, with accuracy 0.81 matching the top predictive results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Fair Graph U-net instead of FairAGG | Fair Graph U-net · German | MIU: 0.18 -> 0.05 | Helped |
| Use Fair Graph U-net instead of FairAGG | Fair Graph U-net · Credit | Accuracy: 0.65 -> 0.75 | Helped |
| Use Fair Graph U-net instead of FairAGG | Fair Graph U-net · Bail | EOD: 0.04 -> 0.02 | Helped |
| Remove the individual fairness component (ablation) | Fair Graph U-net · German, Credit, and Bail | MIU: 0.05 -> 0.11; 0.06 -> 0.13; 0.03 -> 0.06 | Hurt |
| Remove the individual fairness component (ablation) | Fair Graph U-net · German, Credit, and Bail | Accuracy: 0.72 -> 0.69; 0.74 -> 0.72; 0.81 -> 0.78 | Hurt |
| Remove the fair adaptive subgraph identification module and use fixed 2-hop subgraphs (ablation) | Fair Graph U-net · German, Credit, and Bail | fairness and performance | Hurt |
| Remove fair edge augmentation and connect nodes within a 2-hop radius without fairness constraints (ablation) | Fair Graph U-net · German, Credit, and Bail | fairness and performance | Hurt |
Surprising findings
- Graph U-net on Credit reached accuracy 0.76 yet had high SPD (0.18), showing predictive gains coexist with unfairness.
- The group-fairness-only variant (Fair Graph U-net-G) reduced both individual fairness (MIU) and overall accuracy, indicating that group-only constraints can hurt individual outcomes.
Limitations
- Fairness evaluated only on three small tabular-style graph datasets with a binary sensitive attribute.
- Predictive performance is comparable rather than uniformly best; some baselines have higher accuracy or F1 on certain datasets.
- Hyperparameter sensitivity: increasing beta and gamma degrades performance metrics.
Experimental setup
Datasets: German Credit Bail
Benchmarks: German Credit Bail
Metrics: Accuracy F1-score AUC SPD EOD MIU
Baselines: GCN Graph U-net SAGPool FairGNN Graphair FairAGG
Builds on: Graph U-net SAGPool FairGNN PageRank
Closest papers by meaning
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.91)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.90)
- DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning (ICML 2025 · similarity 0.89)
- FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.88)
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.88)
- Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.88)
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.