Toward a better trade-off between performance and fairness with kernel-based distribution matching

As recent literature has demonstrated how classifiers often carry unintended\nbiases toward some subgroups, deploying machine learned models to users demands\ncareful consideration of the social consequences. How should we address this\nproblem in a real-world system? How should we balance core performance and\nfairness metrics? In this paper, we introduce a MinDiff framework for\nregularizing classifiers toward different fairness metrics and analyze a\ntechnique with kernel-based statistical dependency tests. We run a thorough\nstudy on an academic dataset to compare the Pareto frontier achieved by\ndifferent regularization approaches, and apply our kernel-based method to two\nlarge-scale industrial systems demonstrating real-world improvements.\n