Fairer Machine Learning Through Multi-objective Evolutionary Learning

Towards Fairness-Aware Multi-Objective Optimization

“…Recently, multi-objective evolutionary learning strategy is adopted to handle such multi-or many-objective optimization problem with respect to the model accuracy and different fairness metrics [22], [130]. Adversarial learning is also used to mitigate unfairness by maximizing the prediction ability and minimizing the adversary's ability in predicting protected attributes [104].…”

“…In other words, preferences and fairness are not necessarily consistent with each other. In a narrow sense, preferences are commonly articulated as threshold to determine the tolerance of unfairness [131], [133], [142], [143], or the trade-off preference between the fairness and model accuracy [22], [130]. That is, preferences in ML are meant for assisting in finding acceptable fair outcomes for DMs.…”

“…In essence, all machine learning problems are optimization problems [193] where the optimization objectives are typically designed with respect to the learning model, such as the maximization of model performance (e.g., accuracy), maximization of diversity [194], minimization of model complexity [195], maximization of communication efficiency in federated learning [187], [188], maximization of robustness [196], [197], among many others [193], [198]. Specifically, fairness metrics are regarded as extra objectives in the optimization to ensure the fairness of the learning model [22], [29], [30]. Different from conventional learning methods to solve one of the objectives or a scalar function of multiple objectives, the Paretobased multi-objective machine learning [195], [199], [200] simultaneously optimizing multiple objectives has received increased interest recently.…”

“…When fairness is taken into account, an alarming issue of ML is the dilemma between model accuracy and fairness, since the fairness may impact the accuracy [21], [22]. In addition, some fairness measures are conflicting with each other.…”

“…1, an example of the PMO is presented, where solutions (a, b and c) are the best desired solutions in the corresponding ROIs under different reference vectors. Therefore, when the objectives (f 1 and f 2 ) are set to the utility and fairness of ML, then the weight vectors can be extracted according to the relative importance between the objectives, and the solutions are the best trade-off under different weights [22], [29], [30]. In a joint optimization between fairness and utilities of ML, Zhang et al [29] argue that ensuring fairness in decision making is not just a matter of creating new algorithms that are "fairer" but it is an iterative and interactive process that involves multiple stakeholders.…”