Recently a parametric family of fairness metrics to quantify algorithmic fairness has been proposed based on generalized entropy which have been originally used in economics and public welfare. Since these metrics have several advantages such as quantifying unfairness at the individual-level and group-level, and unfold trade-off between the individual fairness and group-level fairness, algorithmic fairness requirement may be given in terms of generalized entropy for a fair classification problem. We consider a fair empirical risk minimization with a fairness constraint specified by generalized entropy. We theoretically investigate if the fair empirical fair classification problem is learnable and how to find an approximate optimal classifier of it.
翻译:最近,基于经济和公共福利最初使用的普遍的昆虫,提出了一套用来量化算法公平性的量化公平指标。由于这些衡量标准有若干优点,如在个人和群体一级量化不公平,在个人公平与群体公平之间展开权衡,因此,在公平分类问题上,算法公平要求可以是通用的昆虫。我们考虑公平的经验风险最小化,而普遍昆虫则规定了公平性限制。我们理论上调查公平的经验公平分类问题是否可学,以及如何找到一个大致最佳分类方法。