A Proposal of a Fair Voting Ensemble Classifier Using Multi-objective Optimization

Current fair machine learning techniques attempt to maintain the trained pipeline accuracy while improving the fairness of its predictions through a myriad of mathematical definitions. Often, proposed methods introduce fairness as a constraint instead of an objective or focus on a single definition of fairness and accuracy, overlooking the differences between privileged and unprivileged groups. Therefore, this paper proposes a multi-objective optimization design where accuracy and fairness are considered as objectives and treated in a constrained multi-objective problem. The optimization result is a set of weights assigned to a soft-voting ensemble with pre-trained classifiers. These weights differ between privileged and unprivileged groups, thus maximizing at the same time fairness and accuracy. The proposed method is tested against popular, state-of-the-art implementations in datasets commonly used in literature. Results show that the proposed method is consistently ranked in the first positions across all objectives – both in fairness and accuracy.