文章基本信息

标题：Are Parity-Based Notions of AI Fairness Desirable?
本地全文：下载
作者：James Foulds ; Shimei Pan
期刊名称：Bulletin of the Technical Committee on Data Engineering
出版年度：2020
卷号：43
期号：4
页码：51-73
出版社：IEEE Computer Society
摘要：It is now well understood that artificial intelligence and machine learning systems can potentially exhibit discriminatory behavior. A variety of AI fairness definitions have been proposed which aim to quantify and mitigate bias and fairness issues in these systems. Many of these AI fairness metrics aim to enforce parity in the behavior of an AI system between different demographic groups, yet parity-based metrics are often criticized for a variety of reasons spanning the philosophical to the practical. The question remains: are parity-based metrics valid measures of AI fairness which help to ensure desirable behavior, and if so, when should they be used? We aim to shed light on this question by considering the arguments both for and against parity-based fairness definitions. Based on the discussion we argue that parity-based fairness metrics are reasonable measures of fairness which are beneficial to maintain in at least some contexts, and we provide a set of guidelines on their use.