Fairer Machine Learning Through Multi-objective Evolutionary Learning

Qingquan Zhang, Jialin Liu, Zeqi Zhang, Junyi Wen, Bifei Mao, Xin Yao

Abstract: Dilemma between model accuracy and fairness in machine learning models has been shown theoretically and empirically. So far, dozens of fairness measures have been proposed, among which incompatibility and complementarity exist. However, no fairness measure has been universally accepted as the single fairest measure. No one has considered multiple fairness measures simultaneously. In this paper, we propose a multi-objective evolutionary learning framework for mitigating unfairness caused by considering a single measure only, in which a multi-objective evolutionary algorithm is used during training to balance accuracy and multiple fairness measures simultaneously. In our case study, besides the model accuracy, two fairness measures that are conflicting to each other are selected. Empirical results show that our proposed multi-objective evolutionary learning framework is able to find Pareto-front models efficiently and provide fairer machine learning models that consider multiple fairness measures.