Fairness-aware Edit of Thresholds in a Learned Decision Tree Using a Mixed Integer Programming Formulation

抄録

Fairness in machine learning is an emerging topic in recent years. In this paper, we propose a post-processing method for editing a given decision tree to be fair according to a specified discrimination criterion by modifying its branching thresholds in internal nodes. We propose a mixed integer linear programming (MIP) formulation for the problem, which can deal with several other constraints flexibly and can be solved efficiently by any existing solver. By experiments, we confirm the effectiveness of our approach by comparing existing post-processing methods.