Multi-cause Discrimination Analysis Using Potential Outcomes

Wen Huang, Yongkai Wu, Xintao Wu

Abstract: Discrimination analysis recently aroused wide attention in the fairness-aware learning field. Most existing causal modeling based fair learning research focuses on single cause effect of one protected attribute on decision. In this paper, we focus on discrimination discovery when multiple protected attributes and redlining attributes are present in addition to other covariates. We regard those protected and redlining attributes as multiple causes of the outcome variable. To deal with unobserved variables, especially hidden confounders, we adopt the potential outcome framework and leverage the state-of-the-art deconfounder algorithm to do causal inference under multiple causes. The deconfounder algorithm infers a latent variable as a substitute for unobserved confounders and then uses that substitute to perform causal inference. Our approach is more appropriate for discrimination discovery as it is able to relax the Markovian assumption and avoid the unidentifiability issue in structural causal modeling approaches. We conduct empirical evaluation on both synthetic data and real data. Empirical evaluation results demonstrate the effectiveness of our proposed approach.