Abstract

Much of the recent literature in machine learning (ML) fairness has focused on statistical group-based notions of fairness, where the goal is to achieve or equalize some model performance metric across protected groups. While this framework has received widespread mathematical attention, this talk will discuss several of its practical and philosophical limitations. First, there is a practical issue of enforcing group-based fairness constraints when the data on protected groups is incomplete or noisy. Second, even with perfect data, we discuss rule-based notions of fair treatment that group-based fairness notions still cannot philosophically capture. Finally, even the most heavily fairness-constrained ML model might still fall short in satisfying societal needs due to choices in problem formulation and downstream interventions. Thus, we argue that the typical view of the ML life cycle in ML research needs to be expanded to capture a full spectrum of societal impacts.