The Measure and Mismeasure of Fairness: A Critical Review of Fair Machine Learning
share
In one broad class of supervised machine learning problems, researchers and practitioners seek to estimate the likelihood of a given event in a manner that is fair. For example, when estimating a loan applicant's risk of default, one might aim to create a statistical model that is free of race or gender bias. Over the last several years, three formal definitions of fairness have gained popularity in the computer science literature: (1) anti-classification, meaning that protected attributes---like race, gender, and their proxies---are not used to form predictions; (2) classification parity, meaning that common measures of predictive performance (e.g., false positive and false negative rates) are equal across groups defined by the protected attributes; and (3) calibration, meaning that conditional on predictions, outcomes are independent of protected attributes. Here we show that all three of these fairness definitions suffer from significant statistical limitations. In particular, requiring anti-classification or classification parity can, perversely, harm the historically disadvantaged groups they were designed to protect, while also violating common legal, economic, and social understandings of fairness; and calibration, while generally desirable, provides little guarantee that decisions are equitable. By highlighting these challenges in the foundation of fair machine learning, we hope to help researchers and practitioners productively advance the area.
READ FULL TEXT