Traditional algorithmic fairness notions rely on label feedback, which can only be elicited from expert critics. However, in most practical applications, several non-expert stakeholders also play a major role in the system and can have distinctive opinions about the decision making philosophy. For example, in kidney placement programs, transplant surgeons are very wary about accepting kidney offers for black patients due to genetic reasons. However, non-expert stakeholders in kidney placement programs (e.g. patients, donors and their family members) may misinterpret such decisions from the perspective of social discrimination. This paper evaluates group fairness notions from the viewpoint of non-expert stakeholders, who can only provide binary \emph{agreement/disagreement feedback} regarding the decision in context. Specifically, two types of group fairness notions have been identified: (i) \emph{definite notions} (e.g. calibration), which can be evaluated exactly using disagreement feedback, and (ii) \emph{indefinite notions} (e.g. equal opportunity) which suffer from uncertainty due to lack of label feedback. In the case of indefinite notions, bounds are presented based on disagreement rates, and an estimate is constructed based on established bounds. The efficacy of all our findings are validated empirically on real human feedback dataset.
翻译:通过从非专家利益相关者征询异议反馈来实现包容性公平评估
翻译后的摘要:
传统的算法公平概念依赖于标签反馈,只能从专家评论家那里引出。然而,在大多数实际应用中,还有一些非专家利益相关者也在系统中发挥重要作用,并且可能对决策思路有不同的意见。例如,在肾脏移植项目中,移植外科医生非常谨慎地接受黑人病人的肾脏提供,因为基因原因。但是,肾脏放置计划中的非专家利益相关者(例如患者、捐献者及其家人)可能会从社会歧视的角度误解这样的决策。本文从非专家利益相关者的角度评估组公平概念,他们只能对上下文中的决策提供二元“赞成/反对的反馈”。具体地,已经确定了两种类型的组公平概念:(i)“明确的概念”(例如校准),可以使用异议反馈进行准确评估,并且(ii)“不确定的概念”(例如机会均等性),由于缺乏标签反馈而受到不确定性的影响。在不确定概念的情况下,根据异议率提供边界,并根据已建立的边界构建估计值。我们所有发现的有效性都通过真实人类反馈数据得到经验证实。