Research map / Federated learning and privacy / Privacy attacks and secure learning
No Prejudice! Fair Federated Graph Neural Networks for Personalized Recommendation
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 F²PGNN, a fair federated personalized GAT-based recommender that regulates per-user learning rates using privacy-protected group statistics and adds local differential privacy to gradients and group statistics.
Problem
Group unfairness in graph-based recommender systems trained in federated settings, where computing fairness requires sensitive user attributes and the full interaction graph at the server, which conflicts with privacy.
Approach
Each client scales its local gradient by a fairness factor L computed from server-broadcast group statistics (P, Q), so better-performing groups learn more slowly; the group statistics and model updates are protected with local differential privacy, and a privacy-preserving inductive graph expansion adds higher-order neighbors via encrypted item IDs.
Key results
- F2PGNN improves fairness over F2MF by 47.65% (ML-100K G), 54.92% (ML-100K A), 84.22% (ML-1M G), 95.21% (ML-1M A) and 99.78% (Amazon-Movies A) at the threshold beta.
- Utility gain of 1.12% to 15.9% RMSE over datasets, except a 10.58% RMSE increase on Amazon-Movies (A).
- With LDP, worst RMSE increase w.r.t. no clipping is 1.88% (ML-100K G), 0.27% (ML-1M G), and 0.474% (Amazon-Movies A).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use F²PGNN instead of F2MF | Federated recommendation · ML-100K (G) | disparity: 47.65% | Helped |
| Use F²PGNN instead of F2MF | Federated recommendation · ML-100K (A) | disparity: 54.92% | Helped |
| Use F²PGNN instead of F2MF | Federated recommendation · ML-1M (G) | disparity: 84.22% | Helped |
| Use F²PGNN instead of F2MF | Federated recommendation · ML-1M (A) | disparity: 95.21% | Helped |
| Use F²PGNN instead of F2MF | Federated recommendation · Amazon-Movies (A) | disparity: 99.78% | Helped |
| Use F²PGNN instead of F2MF | Federated recommendation · Across datasets | RMSE: 1.12% ∼ 15.9%; 10.58% | Mixed |
| Increase fairness budget β (ablation) | F²PGNN · Across datasets | disparity: 9% ∼ 67% | Helped |
| Increase fairness budget β (ablation) | F²PGNN · ML-100K (A) | disparity | Mixed |
| Add LDP gradient clipping and noise (ablation) | F²PGNN · ML-100K (G) | RMSE: 1.88% | Hurt |
| Add LDP gradient clipping and noise (ablation) | F²PGNN · ML-1M (G) | RMSE: 0.27% | Hurt |
| Add LDP gradient clipping and noise (ablation) | F²PGNN · Amazon-Movies (A) | RMSE: 0.474% | Hurt |
| Increase LDP noise strength λ (ablation) | F²PGNN · Across datasets | RMSE | Hurt |
Surprising findings
- F2PGNN's behavior is unstable for ML-100K (Activity) beyond a certain fairness budget, and F2MF also shows instability above a threshold beta.
- On Amazon-Movies, F2PGNN has a higher RMSE than F2MF (about 1.18 vs 1.07), so F2MF is better in utility there.
Limitations
- Does not consider asynchronous updates or computational heterogeneity.
- Fairness and utility trade-off degrades on Amazon-Movies with a 10.58% RMSE increase, and performance is unstable on ML-100K (Activity) at high beta.
- Only one baseline (F2MF) is compared; evaluation is limited to binary sensitive attributes and synthetic activity attribute.
Experimental setup
Datasets: MovieLens-100K MovieLens-1M Amazon-Movies
Benchmarks: MovieLens-100K MovieLens-1M Amazon-Movies
Metrics: RMSE disparity (group unfairness)
Baselines: F2MF
Builds on: FedAvg FedPerGNN F2MF GAT
Closest papers by meaning
- FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference (NeurIPS 2024 · similarity 0.85)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.85)
- Federated Recommendation with Explicitly Encoding Item Bias (AAAI 2025 · similarity 0.85)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.85)
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.85)
- Interpolating Item and User Fairness in Multi-Sided Recommendations (NeurIPS 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.