Adapting Fairness-Intervention Algorithms to Missing Data

Abstract

Missing values in real-world data pose a significant and unique challenge to algorithmic fairness. Different demographic groups may be unequally affected by missing data, and standard procedures for handling missing values such as impute-then-classify can exacerbate discrimination. In this thesis, we analyze how missing values in data affect algorithmic fairness. We discuss the fairness risks posed by disparate missingness and the limitations of impute-then-classify. We then prove a theoretical result that characterizes the reduction in fairness-accuracy potential when training a model on imputed data. We present scalable adaptive algorithms for fair classification with missing values. These algorithms can be combined with any preexisting fairness-intervention algorithm to handle all possible missing patterns while avoiding the disadvantages of impute-then-classify. Numerical experiments with an array of state-of-the-art fairness interventions demonstrate that our adaptive algorithms consistently achieve higher fairness and accuracy than impute-then-classify across many different datasets, particularly when the missingness pattern of the data conveys significant information about the predictive label.