Research map / Learning theory and algorithms / Fairness

A Comprehensive View of Fairness through Distributional Stability

NeurIPS 2026 · Learning theory and algorithms › Fairness · approach: Fair allocation and selection · 0 citations (OpenAlex)

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.

Frames fairness as distributional stability under perturbations of group composition, shows DP/EO and individual fairness arise as special cases, and proposes STABLE, a convex-mixture-of-reweighted-predictors SOCP with generalization bounds.

Problem

Standard group and individual fairness notions are evaluated at a fixed data distribution, so fairness degrades under deployment-time shifts in demographic composition, and guarantees are not uniform across deployment contexts.

Approach

Defines fairness as stability of prediction-rate functionals under perturbations that resample protected-group proportions; DP and EO correspond to stability under global and within-positive-class resampling, with the unfairness gap equal to the Lipschitz constant of the functional in the perturbation parameter. Controls both the mean level and variance of the rate across a distribution of perturb

Key results

• On Adult under EO, STABLE achieves unfairness 0.013 and deploy gap 0.013 with accuracy 0.844, the best among compared methods.
• On COMPAS under DP, STABLE reaches unfairness 0.012 and deploy gap 0.006 with accuracy 0.680, matching ERM accuracy while reducing unfairness from 0.132.
• On Adult under DP, STABLE attains the lowest unfairness (0.011) but accuracy drops to 0.753 versus 0.827 for ERM+DP.

Surprising findings

• On Adult under DP, the stability-based method loses about 7 accuracy points relative to ERM+DP despite lowest unfairness.
• Static DP/EO-constrained baselines can still drift below target when deployed in high-proportion states, showing controlling static unfairness does not preclude deployment bias.

Limitations

• Addresses a linearized convex surrogate of the risk rather than the original problem, introducing a linearization gap that can reduce accuracy.
• Restricted to binary classification with binary sensitive attributes in the main text; equalized odds only in appendix.
• Individual fairness case is not turned into a practical method; the methodology is focused on group fairness.

Experimental setup

Datasets: Adult Income COMPAS ACS Income

Benchmarks: Adult Income COMPAS ACS Income

Metrics: unfairness gap deployment gap accuracy prediction rate under demographic shift

Baselines: ERM+DP ERM+EO PostProc Shifty Group DRO

Builds on: Demographic parity Equal opportunity Individual fairness (Lipschitz) Group DRO Shifty PostProc

Closest papers by meaning

• Addressing Polarization and Unfairness in Performative Prediction (AAAI 2026 · similarity 0.89)
• Fairness Transferability Subject to Bounded Distribution Shift (NeurIPS 2022 · similarity 0.88)
• Stability and Multigroup Fairness in Ranking with Uncertain Predictions (ICML 2024 · similarity 0.87)
• Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · similarity 0.87)
• Geometry of Relaxed Fair Regression: A Unified Framework for Aware and Unaware Settings (NeurIPS 2026 · similarity 0.87)
• Fair Participation via Sequential Policies (AAAI 2024 · similarity 0.86)