On the Connection Between Counterfactual Fairness, Statistical Parity and Individual Fairness
Abstract: The relations among observational fairness notions (those defined based on data distributions) have been studied in the literature, yet the relations between counterfactual fairness and observational fairness notions remain less explored. In this paper, we study the relations between counterfactual fairness and two kinds of observational fairness, statistical parity and individual fairness. In particular, we are interested in understanding whether a predictor trained using counterfactually fair representations (Zuo et al., 2023) can satisfy individual fairness and statistical parity. We show that, for a certain type of causal model called the Gaussian Causal Model (GCM), counterfactual fairness can imply both statistical parity and individual fairness. We also identify another class of causal models under which counterfactual fairness implies statistical parity. Experiments on both synthetic and real-world data demonstrate that counterfactually fair representation can enhance fairness in machine learning models without compromising performance, outperforming methods designed for observational fairness.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Niki_Kilbertus1
Submission Number: 3716
Loading