Achieving Universal Fairness in Machine Learning: A Multi-objective Optimization Perspective

As automatic decision-making systems advance rapidly, ensuring fairness has become an indispensable requirement in machine learning. While numerous fairness-aware learning algorithms have emerged in recent years, most primarily emphasize promoting a single fairness definition centered around a single sensitive attribute. These approaches fail to meet the real-world demand to satisfy multiple fairness definitions across various sensitive attributes simultaneously. To this end, we introduce the concept of Universal Fairness, which involves achieving multiple definitions of fairness for multiple sensitive attributes simultaneously. Due to conflicting objectives in the optimization process, we propose a multi-objective optimization framework, UFair, designed to attain Pareto optimality among different fairness and utility objectives. Theoretically, we demonstrate that UFair can converge to the optimal weights at rate \(\mathcal {O}(\frac{1}{\sqrt{T}})\) . Empirically, extensive experiments conducted on three real-world datasets validate that our UFair efficiently optimizes different fairness constraints alongside utility goals simultaneously.