CRII: SHF: Testing Fairness in Human Decisions with Algorithmic Bias

人类决策的不公平一直是我们社会中一个长期存在的问题，它不断威胁着历史上代表性不足的群体的平等权利。在决策过程中，机器学习软件从可能不公平的人类决策中学习，并做出可能不公平的预测，以指导未来的人类决策，这一问题已经加剧。一个这样的例子是人才招聘过程，人力资源人员筛选由机器学习软件排名的候选人简历列表，以决定谁值得面试。在这种情况下，人类决策的公平性尤其重要，因为(1)机器学习软件在接受不公平的人类决策训练时将继承偏见；(2)即使使用无偏见的机器学习软件，只要人力资源人员有偏见，最终的决策仍然可能是有偏见的。这个项目将从一个不同的方向来处理这个具有历史挑战性的问题。它学习用机器学习软件来模拟人类的决策，并利用该软件的算法偏差来检测人类决策中的偏差。这项工作可能会潜在地促进许多决策活动中的公平过程，例如人才招聘、信用卡审批和学校录取。与大多数关注提高算法公平性的现有研究不同，本研究旨在将从训练数据（人类决策作为因变量）中继承的算法偏差作为人类决策不公平性的指标。为此，该项目将使用一种新技术来测试人类决策的公平性。具体来说，机器学习模型在每个人口统计组中重新平衡阶级分布的被测人类决策上进行训练，然后用比较判断的回归测试套件测试学习模型。如果模型未能通过回归测试，那么它所训练的人类决策将被认为是不公平的。虽然很难直接测试人类是否有偏见，但使用比较判断来测试机器学习模型是否做出有偏见的预测更容易。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Unfairness in human decisions has been a long-standing issue in our society that constantly threatens the equality rights of historically underrepresented groups. This issue has been exacerbated in decision making processes where a machine learning software learns from potentially unfair human decisions and makes potentially unfair predictions to guide future human decisions. One such example is the talent hiring process where a human resource person screens a list of candidate resumes ranked by a machine learning software to decide who deserves an interview. Human decision fairness is especially important in this scenario since (1) the machine learning software will inherit the bias when trained on unfair human decisions; (2) even with unbiased machine learning software, the final decisions can still be biased as long as the human resource person is biased. This project will approach this historically challenging issue from a different direction. It learns to model the human decisions with a machine learning software and utilizes the algorithmic bias of that software to detect bias in the human decisions. This work could potentially advance the equity process in many decision making activities such as talent hiring, credit card approval, and school admission. In contrast to most of the existing research focusing on improving algorithmic fairness, this work aims to isolate the algorithmic bias inherited from the training data (with human decisions as dependent variables) as an indicator for human decision unfairness. To do this, the project will use a novel technique to test fairness in human decisions. Specifically, machine learning models are trained on the human decisions under test with re-balanced class distribution in each demographic group, then tests the learned model with a regression test suite of comparative judgements. If the model fails the regression test, the human decisions it was trained on will be considered as unfair. While it is difficult to directly test whether a human has bias, it is easier to test whether a machine learning model makes biased predictions using comparative judgements.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.