Explainable Fairness and Propensity Score Matching

ABSTRACT Fairness is a key requirement for artificial intelligence applications. The assessment of fairness is typically based on group‐based measures, such as statistical parity, which compares the machine learning output for the different population groups of a protected variable. Although intuitive and simple, statistical parity may be affected by the presence of control variables, correlated with the protected variable. To remove this effect, we propose to employ Shapley values, which measure the additional difference in output specifically due to the protected variable. To remove the possible impact of correlations on Shapley values, we compare them across different subgroups of the most correlated control variables, checking for the presence of Simpson's paradox, for which a fair model may become unfair when conditioning on a control variable. We also show how to mitigate unfairness by means of a propensity score matching that can improve statistical parity, building a training sample that matches similar individuals in different protected groups. We apply our proposal to a real‐world database containing 157,269 personal lending decisions and show that both logistic regression and random forest models are fair when all loan applications are considered, but become unfair for high loan amounts requested. We show how propensity score matching can mitigate this bias.