A Philosophical Theory of Fairness for Prediction-Based Decisions

Abstract

The computer science community has produced a flurry of research on defining fairness for prediction-based decisions. This philosophical contribution takes this literature seriously as it analyzes some of the most widely discussed, and employed fairness definitions from the ML literature. It defends a theory of fairness for prediction-based decisions and uses it to interpret the so-called "Trolley Problem of Machine Learning" – the choice between predictive parity and equalised odds as fairness metrics. Like other contributions to the machine learning literature, our theory takes as its point of departure the Rawlsian theory of Equality of Opportunity. Unlike previous approaches with a similar inspiration, it redefines the main concepts of that theory by introducing moral concepts at a higher level of abstraction. This theory aims to be neutral with respect to substantive moral theories about the justification of inequality. Its main virtue is to make it transparent how, given a moral view concerning what justifies inequality, what does not justify inequality, and what is luck, one may assess the decisions taken by a rule that has certain statistical properties, which can be determined, for example, by considering its confusion table.

Keywords: fairness, machine learning, algorithmic fairness, theories of justice, John Rawls, equality of opportunity, fairness dilemma