强烈推荐3wolf · 2021-08
“Highly recommend it: Equality of Opportunity in Supervised Learning”
强烈推荐这篇论文
提出机会均等的公平性准则,并可事后调整任意预测器以满足它
论文针对指定敏感属性,提出监督学习中衡量歧视的准则,并在已知预测器、目标和群体成员信息的前提下,展示如何最优地调整任意已学得的预测器以消除按此定义的歧视。该框架还把分类不佳的代价从弱势群体转移给决策者,从而改善激励。作者认为这类准则是「无视的」,只依赖预测器、目标和受保护属性的联合统计,因此研究了此类度量能推断什么、不能推断什么,并以 FICO 信用分的案例说明。
讨论算法公平与贷款歧视时,评论者把它当作形式化公平定义的经典出处,用来说明对不同群体采用不同阈值,以及在最优算法上施加公平约束时会产生的权衡。
“Highly recommend it: Equality of Opportunity in Supervised Learning”
“Really make you see what may go wrong with bias in ML.”
“Google has actually released a paper on codifying what discrimination looks like for lending”
“discusses various fairness measures in the context of deciding loan applications based on credit scores.”
“once you have the most optimal algorithm. See for instance Equality of Opportunity in Supervised Learning.”