Group Fairness Under Distribution Shifts: Analysis and Robust Post-Processing

Ruicheng Xian, Han Zhao

Keywords: group fairness, classification, distribution shift, robustness

TL;DR: We analyze fair classifiers under distribution shift and propose a robust post-processing algorithm.

Abstract: Group fairness, as a statistical notion, is sensitive to distribution shifts, which may invalidate the fairness guarantees of classifiers trained with non-robust algorithms. In this work, we analyze randomized fair classifiers and derive upper bounds on fairness violation and excess risk under distribution shift, decomposed into covariate shift, and concept shift—changes in the distribution of group labels (and other variables considered by the fairness criterion) conditioned on the input. Our bounds are general and apply to both multi-class and attribute-blind settings; notably, we show that attribute-blind classifiers incur an additional dependency on the fairness tolerance in their excess risk, suggesting the robustness benefits of attribute awareness. Next, we propose a robust post-processing algorithm that learns fair classifiers with respect to an uncertainty set constructed by modeling the potential covariate and concept shifts, aligning with the decomposition in our analysis. We evaluate our algorithm under geographic shifts in the ACSIncome dataset, demonstrating improved fairness on unseen regions, with additional evaluations performed under noisy group labels and worst-case covariate shifts.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 5446