A Reductions Approach to Fair Classification

FATML’17 |

Published by Association for Computing Machinery

We present a systematic approach for achieving fairness in a binary classification setting. While we focus on two well-known quantitative definitions of fairness, our approach encompasses many other previously studied definitions as special cases. The key idea is to reduce fair classification to a sequence of cost-sensitive classification problems, whose solutions yield a randomized classifier with the lowest (empirical) error subject to the desired constraints. We introduce two reductions that work for any representation of the cost-sensitive classifier and compare favorably to prior baselines on a variety of data sets, while overcoming several of their disadvantages.

Publication Downloads

Fairlearn

May 16, 2018

A Python package that implements a variety of algorithms that mitigate unfairness in supervised machine learning.