FAI: Using Machine Learning to Address Structural Bias in Personnel Selection

今天，美国的人事选拔从业人员主要受到两种知识流的指导：1)与就业机会有关的法律方面的发展；2)在社会、行为和经济科学方面的发现的积累，这些发现指导了人事选拔中公认的专业实践。最近关于机器学习公平性的文献提供了第三种知识流，从业者在设计人员选择系统时可以很容易地利用这一知识流，但是机器学习文献和两种传统知识流之间缺乏整合，留下了相当大的差距，阻碍了它们的有效整合。本研究项目的重点是弥合这一差距，将机器学习建立为人力资源管理人员选择系统设计的第三个支柱。项目的结果告知政策制定者和技术开发者人员选择公平性的重要因素。它还通过更好地将人员选择的实证研究与公平的机器学习算法的技术设计联系起来，促进了关于在人力资源管理中使用机器学习的讨论。该项目的研究植根于它整合的多学科知识的实质性机构，以便在当前的法律结构中公平地选择人员。具体而言，该项目开发了一个理论框架，展示了在现行就业机会法下，人员选择系统的不同设计特征（从预测器选择到分期设计）如何影响和塑造帕累托前沿（在选择有效性和公平性之间的权衡方面）。理论框架的发现说明了人员选择系统的设计特征与内部使用的机器学习算法之间保持一致的重要性。因此，该项目的一个关键组成部分是一系列研究任务，将理论发展、算法设计、系统实现和实证研究结合起来，将机器学习技术适当地置于当前的法律和工业环境中，以进行人员选择。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。

Today, personnel selection practitioners in the United States are primarily guided by two streams of knowledge: 1) the development on the legal front pertaining to employment opportunities, and 2) the accumulation of findings in social, behavioral, and economic sciences that guide the accepted professional practices in personnel selection. The recent literature on fairness in machine learning offers a third stream of knowledge that practitioners can readily tap into when designing their personnel selection systems, yet a lack of integration between the machine learning literature and the two conventional streams of knowledge leaves a considerable gap preventing their effective integration. This research project focuses on bridging the gap to establish machine learning as the third pillar for the design of personnel selection systems in human resource management. The outcomes of the project inform policy makers and technology developers the factors important to the fairness of personnel selection. It also facilitates discussions about the use of machine learning in human resource management, by better connecting the empirical research of personnel selection with the technical design of fair machine learning algorithms.The research in the project is rooted in the substantive bodies of multidisciplinary knowledge it integrates to enable fair personnel selection in the current legal structure. Specifically, the project develops a theoretical framework demonstrating how different design characteristics of a personnel selection system, from predictor selection to staging designs, influence and shape the Pareto front (in terms of tradeoff between selection validity and fairness) achievable under the prevailing employment opportunity laws. The findings from the theoretical framework speak to the importance of alignment between the design characteristics of a personnel selection system and the machine learning algorithms used within. Consequently, a key component of the project is a series of research tasks that combine theory development, algorithmic design, system implementation, and empirical research to properly situate the machine learning techniques within the current legal and industrial environments for personnel selection.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.