Algorithm Design: A Fairness-Accuracy Frontier

Abstract

Algorithm designers increasingly care not only about accuracy but also about fairness across predefined groups. We study the trade-off between these objectives and characterize it by a fairness-accuracy frontier: the set of outcomes that cannot be simultaneously improved in both dimensions. The shape of this frontier is governed by a simple property of the inputs, which we call group skew . In particular, reducing accuracy for both groups to increase fairness is justified if and only if inputs are group skewed. We also study an information design problem in which a designer regulates inputs but another agent chooses the algorithm. We show that, when inputs are not group skewed, banning group identity or other informative inputs is strictly suboptimal.