Gerrymandering Individual Fairness

TL;DR

This paper investigates the vulnerability of individual fairness to gerrymandering in score prediction models, revealing its weaknesses and proposing potential improvements for fairness measures.

Contribution

It demonstrates that gerrymandering individual fairness is feasible and discusses how to strengthen fairness notions to address identified weaknesses.

Findings

Gerrymandering individual fairness in score prediction is possible.

Individual fairness can be a weak fairness measure depending on feature space and metric.

Proposes ideas to improve the robustness of fairness notions.

Abstract

Individual fairness, proposed by Dwork et al., is a fairness measure that is supposed to prevent the unfair treatment of individuals on the subgroup level, and to overcome the problem that group fairness measures are susceptible to manipulation, or gerrymandering. The goal of the present paper is to explore the extent to which it is possible to gerrymander individual fairness itself. It will be proved that gerrymandering individual fairness in the context of predicting scores is possible. It will also be argued that individual fairness provides a very weak notion of fairness for some choices of feature space and metric. Finally, it will be discussed how the general idea of individual fairness may be preserved by formulating a notion of fairness that allows us to overcome some of the problems with individual fairness identified here and elsewhere.