FAI: Fair AI in Public Policy - Achieving Fair Societal Outcomes in ML Applications to Education, Criminal Justice, and Health and Human Services

该项目通过提高对如何以公平和负责任的方式将机器学习方法应用于高风险、真实世界环境的理解，促进了机器学习(ML)服务于社会公益的潜力。政府机构和非营利组织使用ML工具为后续决策提供信息。然而，越来越多的学者、记者和政策制定者对ML技术在不同政策领域(包括儿童福利、卫生和刑事司法)的社会福利和负担分配中发挥的突出(和日益增长的)作用表示担忧。这些决定中的许多都对他们的受试者的生活产生了长期的影响。如果应用不当，它们可能会伤害本已脆弱和历史上处于不利地位的社区。这些担忧引发了越来越多的研究努力，旨在了解差距并开发旨在最大限度地减少或缓解差距的工具。到目前为止，这些努力对现实世界应用的影响有限，过于狭隘地关注抽象的技术概念和计算方法，而忽视了这些方法影响的决策和社会结果。这样的努力通常也无法将工作置于现实世界的背景下，也无法从受ML辅助决策影响最大的社区获得投入。这个项目寻求与政府机构和非营利组织密切合作，填补当前研究和实践中的这些空白。这个项目借鉴了计算机科学、统计学和公共政策的学科观点。它的第一个目标是探索政策目标和ML公式之间的映射。这一目标侧重于必须参考哪些事实才能对公平做出一致的决定，并将对公平的评估与接受决定的人的短期和长期社会结果联系起来。这项工作提供了与合作伙伴、政策制定者和受影响社区接触的实用方法，以将期望的公平目标转化为可计算的可处理措施。它的第二个目标是调查整个ML决策支持渠道的公平性，从政策目标到数据，再到模型再到干预。它探讨了数据收集、推算、模型选择和评估的不同方法如何影响结果工具的公平性。该项目的第三个目标是在政策领域模拟ML辅助决策的长期社会结果，最终指导设计促进公平的方法的扎根方法。该项目的总体目标是弥合公平ML中的积极研究和政策领域中的应用之间的分歧。它通过创新的教学和培训活动，扩大代表不足的群体对研究和技术设计的参与，增进公众、从业者和立法者之间的科学和技术理解，并与合作伙伴机构产生直接的积极影响，做到了这一点。该奖项反映了NSF的法定使命，并通过使用基金会的智力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 project advances the potential for Machine Learning (ML) to serve the social good by improving understanding of how to apply ML methods to high-stakes, real-world settings in fair and responsible ways. Government agencies and nonprofits use ML tools to inform consequential decisions. However, a growing number of academics, journalists, and policy-makers have expressed apprehension regarding the prominent (and growing) role that ML technology plays in the allocation of social benefits and burdens across diverse policy areas, including child welfare, health, and criminal justice. Many of these decisions impart long-lasting effects on the lives of their subjects. When applied inappropriately, they can harm already vulnerable and historically-disadvantaged communities. These concerns have given rise to a growing number of research efforts aimed at understanding disparities and developing tools that aim to minimize or mitigate them. To date, these efforts have been limited in their impact on real-world applications by focusing too narrowly on abstract technical concepts and computational methods at the expense of addressing the decisions and societal outcomes these methods affect. Such efforts also commonly fail to situate the work in real-world contexts or to draw input from the communities most affected by ML-assisted decision-making. This project seeks to fill these gaps in current research and practice in close partnership with government agencies and nonprofits.This project draws upon disciplinary perspectives from computer science, statistics, and public policy. Its first aim explores the mapping between policy goals and ML formulations. This aim focuses on what facts must be consulted to make coherent determinations about fairness, and anchors those assessments of fairness to near- and long-term societal outcomes for people subject to decisions. This work offers practical ways to engage with partners, policymakers, and affected communities to translate desired fairness goals into computationally tractable measures. Itssecond aim investigates fairness through the entire ML decision-support pipeline, from policy goals to data to models to interventions. It explores how different approaches to data collection, imputation, model selection, and evaluation impact the fairness of resulting tools. The project’s third aim is concerned with modeling the long-term societal outcomes of ML-assisted decision-making in policy domains, ultimately to guide a grounded approach to designing fairness-promoting methods. The project’s over-arching objective is to bridge the divide between active research in fair ML and applications in policy domains. It does that through innovative teaching and training activities, broadening the participation of under-represented groups in research and technology design, enhancing scientific and technological understanding among the public, practitioners, and legislators, and delivering a direct positive impact with partner agencies.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.