Fairer Machine Learning Through the Hybrid of Multi-objective Evolutionary Learning and Adversarial Learning

With growing concerns about the unwanted bias or discrimination in machine learning, a number of fairness-aware machine learning algorithms have been developed to mitigate bias in the prediction. Because the objectives of accuracy and fairness are antagonistic, it is hard to balance the trade-off between them. Recently, Multi-objective Evolutionary Learning (MOEL) framework has been proposed to train a set of Pareto models with the consideration of accuracy and fairness simultaneously. However, this framework prefers to train more accurate models rather than fairer models due to the lack of gradient in terms of fairness. In this paper, MOEL is enhanced through introducing Adversarial Learning (AL). The MOEL-AL framework aims to maximize a set of predictors' ability to predict true labels and minimize the ability of an adversarial network to predict the sensitive attributes from the predictors' output. Specifically, the adversarial network can be regarded as a proxy of the undifferentiable fairness metrics, so it is possible to propagate gradients in terms of both accuracy and fairness for the predictors during the back-propagation process. Besides, the adversarial strength for different predictors is adjusted dynamically according to their fairness metric. Compared with the state-of-the-art methods, experimental studies on seven well-known datasets show that our method can provide a set of fairer Pareto models with little drop on accuracy.