On Values in Fairness Optimization with Machine Learning

Abstract Statistical criteria of fairness, although controversial, bring attention to the multiobjective nature of many predictive modeling problems. In this article, I consider how epistemic and nonepistemic values help to justify the design of machine learning algorithms that optimize for more than one normative goal. I focus on a major design choice between biased search strategies that directly incorporate priorities for various objectives into an optimization procedure and unbiased search strategies that do not. I argue that both reliably generate Pareto optimal solutions such that various other values are relevant to making a rational choice between them.