Fair Decision Making via Automated Repair of Decision Trees

Abstract

Data-driven decision-making allows more resource allocation tasks to be done by programs. Unfortunately, real-life training datasets may capture human biases, and the learned models can be unfair. To resolve this, one could either train a new, fair model from scratch or repair an existing unfair model. The former approach is liable for unbounded semantic difference, hence is unsuitable for social or legislative decisions. Meanwhile, the scalability of state-of-the-art model repair techniques is unsatisfactory. In this paper, we aim to automatically repair unfair decision models by converting any decision tree or random forest into a fair one with respect to a specific dataset and sensitive attributes. We built the FairRepair tool, inspired by automated program repair techniques for traditional programs. It uses a MaxSMT solver to decide which paths in the decision tree could be flipped or refined, with both fairness and semantic difference as hard constraints. Our approach is sound and complete, and the output repair always satisfies the desired fairness and semantic difference requirements. FairRepair is able to repair an unfair decision tree on the well-known COMPAS dataset [2] in 1 minute on average, achieving 90.3% fairness and only 2.3% semantic difference. We compared FairRepair with 4 state-of-the-art fairness learning algorithms [10, 13, 16, 18]. While achieving similar fairness by training new models, they incur 8.9% to 13.5% semantic difference. These results show that FairRepair is capable of repairing an unfair model while maintaining the accuracy and incurring small semantic difference.