Blind pareto fairness and subgroup robustness

Jun 10, 2021·

Natalia Martinez

Martin Bertran

Afroditi Papadaki

Miguel Rodrigues

Guillermo Sapiro

· 1 min read

PDF Cite Code Video

Image credit: paper

Abstract

Much of the work in the field of group fairness addresses disparities between predefined groups based on protected features such as gender, age, and race, which need to be available at train, and often also at test, time. These approaches are static and retrospective, since algorithms designed to protect groups identified a priori cannot anticipate and protect the needs of different at-risk groups in the future. In this work we analyze the space of solutions for worst-case fairness beyond demographics, and propose Blind Pareto Fairness (BPF), a method that leverages no-regret dynamics to recover a fair minimax classifier that reduces worst-case risk of any potential subgroup of sufficient size, and guarantees that the remaining population receives the best possible level of service. BPF addresses fairness beyond demographics, that is, it does not rely on predefined notions of at-risk groups, neither at train nor at test time. Our experimental results show that the proposed framework improves worst-case risk in multiple standard datasets, while simultaneously providing better levels of service for the remaining population.

Type

Preprint

Publication

In Proceedings of the 38th International Conference on Machine Learning (ICML 2021)

In this work, we propose BPF, an algorithm that provably converges to a properly Pareto minimax solution, it requires minimal modifications to the standard learning pipeline of a standard model, and can scale easily to large datasets. Our results on a variety of standard fairness datasets show that this approach reduces worst-case risk as expected, and produces better models than competing methods for the low-risk population, thereby avoiding unnecessary harm. It identifies high risk samples and is easy to interpret since the user can control the optimal adversary through the use of a target worst partition size.

If a policymaker has a desired risk tradeoff instead of a target group size, we can search for the smallest partition size achieving this tradeoff using the proposed BPF; this now guarantees that the recovered model can satisfy this risk tradeoff for the worst possible partition up to a size, and for any smaller partition size there exists a partition such that this tradeoff is violated. Moreover, the tradeoffs between the worst and best group for a fixed group size can be controlled with the weight’s lower bound.

Future work includes incorporating additional domain-specific constraints on the worst partition and developing an algorithm that combines BPF with knowledge about somesubgroups that must be protected as well.

Last updated on Jun 10, 2021