Research map / Graph learning / GNN architectures
Chasing Fairness in Graphs: A GNN Architecture Perspective
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 Message Passing (FMP), a GNN aggregation scheme derived from a unified optimization framework that aggregates neighbor information and then debiases by pushing demographic group representation centers together, trained with vanilla cross-entropy and no data pre-processing.
Problem
GNN message passing amplifies societal bias relative to MLPs, and prior fairness methods for graphs rely on data pre-processing or fair training losses, while architecture-level fairness is less explored and lacks transparency about how sensitive attributes are used.
Approach
Derive a message-passing layer from a fairness-plus-smoothness optimization objective (Laplacian smoothing plus an l1 demographic-parity penalty), solved with Fenchel conjugates and a primal-dual predictor-corrector scheme, so each layer aggregates neighbors then applies an explicit sensitive-attribute-based debiasing step in probability space.
Key results
- On NBA, FMP reaches 73.33% accuracy with ΔDP 18.92%, the best accuracy and lowest DP among listed methods.
- On Pokec-z, FMP reduces demographic parity to 0.81% versus 3.25% for GCN and 1.28% for JKNet, with accuracy 70.50% comparable to GCN's 71.78%.
- FMP reduces demographic parity by 49.69% and 56.86% relative to the lowest-bias baseline on Pokec-z and Pokec-n, respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FMP instead of the strongest listed baseline | Node classification · Pokec-z | demographic parity ($\Delta_{DP}$): 1.28 $\pm$0.96 -> \textbf{0.81} $\pm$ 0.40 | Helped |
| Use FMP instead of the strongest listed baseline | Node classification · Pokec-z | equal opportunity ($\Delta_{EO}$): 1.79 $\pm$ 0.82 -> 1.73 $\pm$ 1.03 | Helped |
| Use FMP instead of GCN | Node classification · Pokec-z | accuracy: \textbf{71.78} $\pm$ 0.37 -> 70.50 $\pm$ 0.50 | Hurt |
| Use FMP instead of the strongest listed baseline | Node classification · Pokec-n | demographic parity ($\Delta_{DP}$): 1.47 $\pm$ 1.12 -> \textbf{0.66} $\pm$ 0.40 | Helped |
| Use FMP instead of the strongest listed baseline | Node classification · Pokec-n | equal opportunity ($\Delta_{EO}$): 3.03 $\pm$ 1.77 -> 1.47 $\pm$ 0.87 | Helped |
| Use FMP instead of GCN | Node classification · Pokec-n | accuracy: \textbf{73.09} $\pm$ 0.28 -> 72.16 $\pm$ 0.33 | Hurt |
| Use FMP instead of the strongest listed baseline | Node classification · NBA | demographic parity ($\Delta_{DP}$): 20.12 $\pm$ 16.18 -> \textbf{18.92} $\pm$ 2.28 | Helped |
| Use FMP instead of GAT | Node classification · NBA | equal opportunity ($\Delta_{EO}$): 13.00 $\pm$ 13.37 -> \textbf{13.33} $\pm$ 5.89 | Mixed |
| Use FMP instead of ML1 | Node classification · NBA | accuracy: 72.70 $\pm$ 1.19 -> \textbf{73.33} $\pm$ 1.85 | Helped |
| Use FMP instead of adversarial debiasing or demographic regularization | GNN and MLP backbones · Pokec-z, Pokec-n, and NBA | demographic parity-accuracy trade-off | Helped |
Surprising findings
- Many standard GNNs (GAT, GCN, SGC, APPNP) show higher demographic parity bias than MLP on all three datasets.
- On NBA, aggregation can hurt accuracy when neighbors rarely share labels, so GNNs underperform MLP there.
Limitations
- Fairness evaluated only for binary sensitive attributes and node classification; the paper restricts to vanilla cross-entropy and demographic parity-style objectives.
- Evaluated on only three small or medium datasets (NBA, Pokec-z, Pokec-n) with 5 runs; baseline hyperparameter tuning is limited.
- Accuracy is sometimes lower than the best baseline (e.g., Pokec-z and Pokec-n), trading off some performance for fairness.
Experimental setup
Datasets: Pokec-z Pokec-n NBA
Benchmarks: Pokec-z Pokec-n NBA
Metrics: accuracy demographic parity (ΔDP) equal opportunity (ΔEO)
Baselines: MLP GCN GAT SGC APPNP JKNet ML1 adversarial debiasing adding regularization Fair Mixup
Builds on: GCN GAT APPNP PAPC (proximal alternating predictor-corrector) Fair Mixup
Closest papers by meaning
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.93)
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.91)
- Adversarial Attacks on Fairness of Graph Neural Networks (ICLR 2024 · similarity 0.91)
- FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.90)
- Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.89)
- DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning (ICML 2025 · similarity 0.89)
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.