CAREER: Advancing Equity in Selection Problems Through Bias-Aware Optimization

这项教师早期职业发展计划（Career）资助将通过开发系统的方法来减少由于隐性偏见导致的劳动力不平等，为促进国家健康、繁荣和福利做出贡献。广泛使用的自动化申请人筛选技术带来了系统性筛选STARS（通过其他途径培训的熟练工人）的高风险，而对不同背景的候选人的选择施加配额可能违反反歧视法。该奖项支持开发基本方法，以透明地处理候选人评估数据中的背景偏见，而无需诉诸配额或公平约束（由于法律要求）。这项跨学科研究将为从业者和政策制定者提供工具，以了解系统中的低效率，从而导致招聘和大学招生政策的协同设计。附带的教育计划旨在为高中生开发STEAM （STEM+艺术）研讨会，通过“招聘经理”模拟游戏、针对政策和法律专业人士的研讨会、道德或课程的设计，以及对专注于STEM少数族裔的学生的持续指导，让他们了解数据中的偏见。本研究将通过使用反事实和因果分析来构建候选人评估数据的基数和序数变异性集，分别产生双线性优化问题和序数组合优化问题，从而开发基于其背景的数据变异性模型的基本方法。该研究将开发新的技术来解决这些具有挑战性的问题类，使用参数优化和顺序理论作为开始，并找到易于处理的解决方案。这项工作将为我们如何处理上下文数据提供一个根本性的转变，同时推进有序、鲁棒、参数和一般离散优化的理论。这项工作将解决与公平-效率权衡相关的重要政策设计问题，例如，当数据与上下文相关时，偏见感知技术对公平、多样性和公平性的影响，在构建可变性集时改变“风险”参数对候选人选择的影响，并强调利用有限资源减少可变性的方法。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。

This Faculty Early Career Development Program (CAREER) grant will contribute to the advancement of national health, prosperity, and welfare by developing systematic approaches to reducing workforce inequality due to implicit bias. Widely used automated applicant screening technologies bring high risk of systematically screening-out STARS (skilled workers trained by alternate routes), while imposing quotas on the selection of candidates from different backgrounds can violate anti-discrimination laws. This award supports the development of fundamental methodologies for transparently handling contextual biases in candidate evaluation data without resorting to quotas or fairness constraints (due to legal requirements). This interdisciplinary research will provide tools for practitioners and policymakers to understand the inefficiencies in the system, leading to a synergistic design of policies for hiring and college admissions. The accompanying educational plan aims to develop STEAM (STEM+art) workshops for high school students to understand biases in data through a “hiring manager” simulation game, a workshop geared towards policy and law professionals, the design of courses in Ethical OR, and the continued mentorship of students with a focus on STEM minorities.This research will develop fundamental methodologies to model variability in data due to its context, by using counterfactual and causal analysis to construct cardinal and ordinal variability sets for candidates’ evaluation data, yielding bilinear optimization problems and ordinal combinatorial optimization respectively. The research will develop new techniques to address these challenging problem classes using parametric optimization and order theory as a start and find tractable solutions. This work will provide a fundamental shift in how we process contextual data, while advancing the theories of ordinal, robust, parametric, and general discrete optimization. The work will address important policy-design questions related to equity-efficiency trade-offs, e.g., the impact of bias-aware techniques on equity, diversity, and fairness when data is contextual, the impact of changing "risk" parameters in the construction of variability sets on candidate selection and highlight ways to use limited resources to reduce variability.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.