2020/09/09 by Kirtan Padh, Padh, Kirtan, Diego Antognini +7 · 2 citations
Computer Science · Social Sciences · #Artificial Intelligence (cs.AI) #Ethics and Social Impacts of AI #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Gender Politics and Representation #Information Retrieval (cs.IR) #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Qualitative Comparative Analysis Research
paper · pdf · doi:10.48550/arxiv.2009.04441
openalex publication_date 2020/09/09 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
The goal of fairness in classification is to learn a classifier that does not\ndiscriminate against groups of individuals based on sensitive attributes, such\nas race and gender. One approach to designing fair algorithms is to use\nrelaxations of fairness notions as regularization terms or in a constrained\noptimization problem. We observe that the hyperbolic tangent function can\napproximate the indicator function. We leverage this property to define a\ndifferentiable relaxation that approximates fairness notions provably better\nthan existing relaxations. In addition, we propose a model-agnostic\nmulti-objective architecture that can simultaneously optimize for multiple\nfairness notions and multiple sensitive attributes and supports all statistical\nparity-based notions of fairness. We use our relaxation with the\nmulti-objective architecture to learn fair classifiers. Experiments on public\ndatasets show that our method suffers a significantly lower loss of accuracy\nthan current debiasing algorithms relative to the unconstrained model.\n