acceptodds
Under review as a conference paper at ICLR 2027

RAWLS–SHAPLEY FAIRNESS: A NASH-BARGAINING APPROACH TO EQUITABLE MACHINE LEARNING

Abstract

We propose Rawls–Shapley (RS) learning, a machine-learning fairness objective grounded in Yaari’s (1981) formal result that applying Rawls’ Difference Principle through Shapley’s utility-transfer framework uniquely recovers the Nash bargaining solution. Translating this result into a training objective, we define each group’s utility gain and aggregate these gains through a Nash-product formulation. This gives rise to the Rawls–Shapley objective, whose optimization automatically places greater emphasis on groups with lower utility gains, thereby incorporating Rawlsian priority into the learning process in a principled and differentiable manner. We instantiate RS in two settings: RS-Known, which uses observed group labels, and RS-Proxy, which trains a supervised proxy predictor on group labels to derive soft group memberships (a “limited demographics” regime, distinct from the fully label-free setting). We evaluate both variants on Adult income and COMPAS recidivism benchmarks alongside empirical risk minimization (ERM), Adversarially Reweighted Learning (ARL), Logistic Regression, and XGBoost, reporting accuracy, AUC, and equalized-odds gap across protected groups. Empirically, RS-Proxy achieves the lowest equalized-odds gap among the RS and non-parity neural objectives on Adult (0.145), while Equalized-Odds Regularization (EqOdds-Reg) achieves the best parity across both datasets. The two RS variants achieve broadly similar worst-group accuracy. On COMPAS, where group labels are noisy and less predictable, all methods converge to similar performance, consistent with our graceful-degradation analysis.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Rawls–Shapley (RS) learning, Adversarially Reweighted Learning (ARL), Nash-welfare solution, Shapley’s utility-transfer

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_CCn5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Rawls-Shapley learning, a group-fairness objective motivated by the connection between distributive justice and nash bargaining. It maximizes the product of group utility gains, where utility is the average probability assigned to the correct label. the resulting gradients assign larger coefficents to groups with lower utility gains. The paper includes two implementations: RS-Known using observed demographic groups and RS-proxy for supervised soft memberships. The paper comes with theoretical guarantees and a set of experiments on classic benchmark fairness datasets.

Critical strengths

Strength #1: The objective is simple, interpretable, intuitive and principled: it provides a way to give greater marginal weight to groups with lower utility.

Strength #2: The paper gives a careful analysis, showing that Nash welfare does not guarantee the best worst-group outcome, and that its theoretical optimality guarantees do not directly apply to the solutions found by neural network training.

Critical weaknesses

Weakness #1: The experiments lack a matched baseline that maximizes the arithmetic mean of the same soft group utilities, so they do not isolate the benefit of Nash aggregation from the change in training loss. Both RS variants also report lower mean worst-group accuracy and AUC than ERM on both datasets, without uncertainty estimates to assess these differences. Some additional baselines/ablations I'd suggest are group-balanced soft utility, min-max soft utility, and just overall soft utility without group-weightings.

Weakness #2: Incremental contribution relative to existing wealfare + bargaining approaches. Relevant work (not cited in the paper) include [1] which studies welfare aggregation over groups, including geometric-mean welfare and [2] which optimizes a Nash product of group improvements in classification trees.

[1] Cousins, Cyrus. "An axiomatic theory of provably-fair welfare-centric machine learning." Advances in Neural Information Processing Systems 34 (2021): 16610-16621. [2] Liapis, Georgios I., and Lazaros G. Papageorgiou. "Optimisation-based classification tree: a game theoretic approach to group fairness." International Conference on Optimization and Learning. Cham: Springer Nature Switzerland, 2024.

Questions for authors
  1. Evidence for the Nash aggregation: how much does it outperform simpler baselines mentioned above?
  2. What is the novelty compared to the two related works mentioned above?
Reviewer_EvLHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Rawls–Shapley (RS) learning, a fairness-aware approach that aims to improve performance for disadvantaged groups without sacrificing the performance of other groups. Inspired by Nash bargaining theory, the method maximizes the product of group utilities, automatically assigning greater importance to groups with lower performance.

The authors introduce two variants: RS-Known, which uses observed sensitive-group labels, and RS-Proxy, which estimates group membership when demographic information is limited. They provide theoretical characterizations under convexity assumptions and evaluate the methods on Adult and COMPAS against standard and fairness-aware baselines.

The results are mixed. RS-Proxy improves the equalized-odds gap over standard training on Adult, but the gains do not consistently extend to COMPAS, and direct equalized-odds regularization generally achieves better parity.

Critical strengths
  1. Intuitive and well-motivated fairness objective. The Nash-welfare formulation provides a smooth alternative to strict worst-group optimization, prioritizing disadvantaged groups while allowing all groups to contribute to training. This is a meaningful perspective on fairness beyond enforcing equal error rates.
  2. Consideration of limited demographic information. The RS-Proxy variant addresses the practical challenge of incomplete access to sensitive attributes. Its behavior when group membership is difficult to predict is also characterized theoretically.
Critical weaknesses
  1. Limited methodological and theoretical novelty. The central objective is an application of the established Nash-welfare formulation, and the inverse-utility weighting follows directly from differentiating the logarithmic objective. The main theoretical characterization relies on existing results and convexity assumptions that generally do not hold for neural networks. It is unclear what fundamentally new learning guarantees the method provides.

  2. Insufficient empirical evidence of improvement. The evaluation is limited to two standard datasets, and neither proposed variant consistently outperforms existing methods. On Adult, RS-Known actually increases the equalized-odds gap from 0.1685 to 0.1979, while RS-Proxy reduces it to 0.1448. Direct equalized-odds regularization achieves a substantially lower gap of 0.1037. On COMPAS, the improvements are similarly limited. Moreover, the proposed objective does not directly enforce equalized odds, making the relationship between its theoretical motivation and empirical fairness metrics somewhat unclear.

Questions for authors

See weaknesses.

Reviewer_g2t6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper uses the Nash bargaining solution as the training objective, which places more weight on worse-off groups without putting all the weight on the single worst-off group. The authors propose two variants: RS-Known, which uses observed group labels, and RS-Proxy, which uses a classifier to produce soft group memberships. They report experiments on two datasets.

Critical strengths
  • Bringing the Nash bargaining solution into Rawlsian fairness for ML training is a novel perspective.
Critical weaknesses
  • The theoretical contribution is limited: the results are either already known or follow from simple derivations.
  • The experimental evaluation is limited. Only two datasets are used, there are no ablation studies, the method does not consistently outperform ERM, and important baselines are missing.
Questions for authors
  • The paper does not compare with baselines that target the same problem [1,2]. How does the proposed method perform against them?
  • In the actual experiments, what are the group weights 1/(G_s + \epsilon)? If the weights are close across groups, the proposed method may behave like group-balanced ERM. Could the authors include this baseline for comparison?
  • Why is the soft-correctness score used to measure utility? Could the authors discuss or compare alternative utility choices?

[1] Sagawa, Shiori, et al. "Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization." arXiv preprint arXiv:1911.08731 (2019).

[2] Martinez, Natalia, Martin Bertran, and Guillermo Sapiro. "Minimax pareto fairness: A multi objective perspective." International conference on machine learning. PMLR, 2020.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.