Research map / Learning theory and algorithms / Fairness
Towards Harmless Rawlsian Fairness Regardless of Demographic Prior
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 VFair, which minimizes the variance of training losses inside the set of near-optimal empirical losses via a dynamic gradient/loss-level update, achieving harmless Rawlsian fairness without demographic priors.
Problem
Group fairness when no demographic/sensitive attributes are available during training, aiming to improve Rawlsian (worst-group) fairness without hurting overall utility ("harmless" fairness), for both classification and regression.
Approach
Use the variance (standard deviation) of per-example training losses as a demographic-free fairness proxy, minimized as a secondary objective inside the optimal set of the mean loss; a dynamic coefficient lambda (set by gradient-conflict and non-negative-weight constraints) combines the primary and secondary gradients so the mean loss is never harmed.
Key results
- On regression benchmarks VFair reduces MUD on Law School from 7.33 (ERM) to 6.63 and VAR from 4.89 to 3.66 with non-significant utility change (12.95 vs 12.88 MSE).
- On COMPAS regression, VFair lowers MUD from 2.50 (ERM) to 0.93 and VAR from 3.23 to 0.47 while utility stays at 23.15 vs 23.08.
- On classification, fairness gains over ERM are mostly not statistically significant (e.g., UCI Adult accuracy 84.74 vs 84.67, p=0.08), with VAR the main exception; VFair gives notable gains only on CelebA (WU 91.09 vs 89.77).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VFair instead of ERM | regression models · Law School | MUD: 7.33{ (0.14)} -> 6.63{ (0.18)} | Helped |
| Use VFair instead of ERM | regression models · Law School | Utility (MSE): 12.88{ (0.12)} -> 12.95{ (0.11)} | No clear effect |
| Use VFair instead of ERM | regression models · COMPAS | MUD: 2.50{ (1.17)} -> 0.93{ (0.21)} | Helped |
| Use VFair instead of ERM | regression models · COMPAS | VAR: 3.23{ (0.8)} -> 0.47{ (0.07)} | Helped |
| Use VFair instead of ERM | regression models · C & C | MUD: +2.02 | No clear effect |
| Use VFair instead of ERM | regression models · AgeDB | MUD: +0.03 | No clear effect |
| Use VFair instead of ERM | F1-score classifiers · UCI Adult | MUD: 6.87 -> 5.82 | Helped |
| Use VFair instead of ERM | F1-score classifiers · CelebA | MUD: 19.39 -> 14.39 | Helped |
| Use VFair instead of ERM | regression models · AgeDB | Utility (MSE): +0.68 | Helped |
| Use both λ₁ and λ₂ rather than λ₁ alone (ablation) | VFair · COMPAS classification | Utility (accuracy): 66.45}{ (0.85) -> 66.80}{ (0.27) | Helped |
| Add λ₂ to λ₁ (ablation) | VFair · COMPAS classification | MUD: 6.60}{ (2.40) -> 6.25}{ (0.80) | Helped |
| Use λ₁ only, omitting λ₂ (ablation) | VFair · Law School classification | Utility (accuracy): 85.40}{ (0.30) -> 45.39}{ (28.53) | Hurt |
Surprising findings
- Regression tasks achieve significant fairness gains without utility loss, whereas classification tasks show limited fairness improvement under the same harmless setting because accuracy is quantized.
- Uniform (near-constant-prediction) classifiers can emerge when optimizing variance alone; the mean-loss-preserving lambda is needed to avoid collapse.
Limitations
- Computational cost is roughly twice that of ERM.
- Classification fairness gains are limited under accuracy-based utility due to quantized metrics.
- Only tested on i.i.d. tabular and image datasets; non-IID prediction not addressed.
Experimental setup
Datasets: UCI Adult Law School COMPAS CelebA Communities & Crime AgeDB
Benchmarks: UCI Adult Law School COMPAS CelebA Communities & Crime AgeDB
Metrics: Utility (accuracy / MSE) WU (worst-group utility) MUD (maximum utility disparity) TUD (total utility disparity) VAR (variance of prediction error) F1-score
Baselines: ERM DRO ARL FairRF MPFR BPF FKL
Builds on: Distributionally robust optimization (DRO) Adversarially reweighted learning (ARL) Worst-case fairness Variance-based generalization bounds (Bennett's inequality)
Closest papers by meaning
- Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.87)
- Achievable Fairness on Your Data With Utility Guarantees (NeurIPS 2024 · similarity 0.87)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.87)
- Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.87)
- On Optimal Steering to Achieve Exact Fairness (NeurIPS 2025 · similarity 0.87)
- Stay Fair! Ensuring Group Fairness in Diffusion Models Across Guidance Scales (NeurIPS 2026 · 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.