Research map / Learning theory and algorithms / Fairness
Fairness-Aware Meta-Learning via Nash Bargaining
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 a two-stage fairness-aware meta-learning method that first uses a Nash Bargaining Solution over group hypergradients to resolve hypergradient conflicts and reach the Pareto front, then optimizes a specific fairness objective.
Problem
Naive one-stage meta-learning for group fairness (LtR, FORML, Meta-gDRO) produces hypergradients that conflict with subgroup update directions, causing unstable convergence and compromised performance and fairness.
Approach
Frame hypergradient aggregation as a cooperative bargaining game among groups, derive the Nash Bargaining Solution for gradient aggregation without linear independence assumptions (G^T G alpha = 1/alpha), and use it in an early stage before switching to the fairness objective.
Key results
- On Bank Telemarketing, FORML two-stage raised Overall AUC from 0.706 to 0.779 and reduced Max-gAUCD from 0.039 to 0.029.
- Claimed up to 10% overall performance improvement and 67% disparity drop across fairness settings.
- LtR's hypergradient alignment rate improved from 60% to 80% with bargaining; FORML went from 0 to 76.9% on Bank Telemarketing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add two-stage Nash bargaining before LtR fairness optimization | LtR meta-learning · Adult Income (sex) | Overall AUC: 0.830 -> 0.830 (+.000) | No clear effect |
| Add two-stage Nash bargaining before LtR fairness optimization | LtR meta-learning · Adult Income (race) | Overall AUC: 0.803 -> 0.805 (+.002) | Helped |
| Add two-stage Nash bargaining before FORML fairness optimization | FORML meta-learning · Bank Telemarketing (age) | Max-gAUCD: 0.039 -> 0.029 (-.010) | Helped |
| Add two-stage Nash bargaining before FORML fairness optimization | FORML meta-learning · Credit Default (sex) | Overall AUC: 0.554 -> 0.611 (+.057) | Helped |
| Add two-stage Nash bargaining before Meta-gDRO fairness optimization | Meta-gDRO meta-learning · Communities and Crime (blackgt6pct) | Worst-gAUC: 0.629 -> 0.629 (+.000) | No clear effect |
| Add two-stage Nash bargaining before FORML fairness optimization | FORML meta-learning · Titanic Survival (sex) | Max-gAUCD: 0.033 -> 0.011 (-.022) | Helped |
| Add two-stage Nash bargaining before Meta-gDRO fairness optimization | Meta-gDRO meta-learning · Student Performance (sex) | Worst-gAUC: 0.894 -> 0.900 (+.006) | Helped |
| Use two-stage Nash bargaining before FORML fairness optimization | FORML meta-learning · Bank Telemarketing (age) | Overall AUC: 0.706 -> 0.779 (+.073) | Helped |
Surprising findings
- Most epochs of one-stage methods had less than 3% of aggregated directions aligned with every subgroup's objective.
- On Credit Default and Communities and Crime, the method gave little or negative improvement, attributed to noisy validation sets and low bargaining success rates.
Limitations
- Gains are limited or negative on Credit Default and Communities and Crime due to noisy validation sets and low bargaining feasibility.
- NBS derived under disagreement point D=0 and local linear approximation; general D not studied.
- Theoretical explanation for Stage 2 staying near the Pareto front is left open.
Experimental setup
Datasets: Adult Income Credit Default Bank Telemarketing Communities and Crime Titanic Survival Student Performance
Benchmarks: Adult Income Credit Default Bank Telemarketing Communities and Crime Titanic Survival Student Performance
Metrics: Overall AUC Max-gAUCD Worst-gAUC
Baselines: Baseline (no meta-learning) DRO LtR FORML Meta-gDRO
Builds on: learning to reweight Meta-gDRO FORML nash bargaining solution for multi-task learning
Closest papers by meaning
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.86)
- FFB: A Fair Fairness Benchmark for In-Processing Group Fairness Methods (ICLR 2024 · similarity 0.86)
- FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.86)
- Achievable Fairness on Your Data With Utility Guarantees (NeurIPS 2024 · similarity 0.86)
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.86)
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 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.