Minimax AUC Fairness: Efficient Algorithm with Provable Convergence

The use of machine learning models in consequential decision making often exacerbates societal inequity, particular yielding disparate impact on members marginalized groups defined by race and gender. area under the ROC curve (AUC) is widely used to evaluate performance a scoring function learning, but studied algorithmic fairness less than other metrics. Due pairwise nature AUC, defining an AUC-based group metric pairwise-dependent may involve both intra-group inter-group AUCs. Importantly, considering only one category AUCs not sufficient mitigate unfairness AUC optimization. In this paper, we propose minimax bias mitigation framework that incorporates while maintaining utility. Based Rawlsian framework, design efficient stochastic optimization algorithm prove its convergence minimum group-level AUC. We conduct numerical experiments synthetic real-world datasets validate effectiveness proposed algorithm.