FAI: Building a Fair Recommender System for Foster Care Services within the Constraints of a Sociotechnical System

美国有超过40万名年轻人在寄养家庭，每年有超过2万人在没有与家人团聚的情况下离开寄养家庭。对于这些年轻人来说，结果很糟糕。到19岁时，只有59%的人完成了高中学业，20%的人无家可归，27%的男性被监禁。这些结果对青年和社会产生了重大和代价高昂的影响。幸运的是，寄养中的青年有可能有资格接受职业培训等服务，这可能会增加他们获得积极生活结果的机会。然而，个案工作者通常只能在青少年遇到相关需要或危机后，才能确定他们是否需要这些服务。为了解决这一问题，该项目将制定一种算法，帮助个案工作者在危机发生之前确定需要服务的青年，并以公平和公正的方式分配这些服务。关键是，算法的每一步都将得到寄养青年和社会工作者的投入。所开发的方法将为长期资金不足和人手不足的寄养机构释放资源，增加决策程序的透明度和问责性，并提供一个关键工具，在危机发生之前识别需要服务的青年。在开发服务建议算法时，将解决三个基本研究目标。首先，将开发方法来帮助领域专家识别和缓解训练数据中难以发现的社会偏见。第二，将制定方法，以确定在服务分配决定方面的多个公平角度(例如，寄养青年和个案工作者)。最后，将制定一种方法来建议寄养青年的服务分配战略，方法是根据多个关于公平的观点，平衡确保公平分配的目标和增加所有青年取得积极生活成果的几率的目标。将对所有方法进行广泛评估，并与寄养服务中的多个利益攸关方紧密结合。我们还将提供证据，证明我们的方法对其他领域的普遍性，以及它们的效率和准确性的可证明界限。利益相关者整合包括与当地寄养机构合作，每年为10,000多名青年提供服务，以及由具有寄养经验的青年组成的青年咨询委员会，他们将在模型开发中发挥重要作用。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

There are more than 400,000 youth in foster care in the US, and each year more than 20,000 age out of foster care without reunifying with their families. The outcomes for these youth are poor. By age 19, only 59% finish high school, 20% have been homeless, and 27% of males have been incarcerated. These outcomes have significant and costly impacts for the youth and for society. Fortunately, youth in foster care are potentially eligible to receive services, such as vocational training, that may improve their chances at positive life outcomes. However, caseworkers are typically only able to identify youth for these services after the youth experiences a relevant need or crisis. To address this problem, this project will develop an algorithm to assist caseworkers in identifying youth in need of services before crises occur, and to allocate those services in a fair and just way. Critically, the algorithm will be informed at every step by inputs from foster youth and caseworkers. The methods developed will free resources for foster care agencies that are chronically underfunded and understaffed, increase procedural transparency and accountability in decision-making, and provide a critical tool to identify youth who need services before crises occur.In developing the service recommendation algorithm, three fundamental research objectives will be addressed. First, methods will be developed to help domain experts identify and mitigate hard-to-find social biases in training data. Second, methods will be developed to identify multiple perspectives of fairness (e.g. from foster youth and case workers) with respect to service allocation decisions. Finally, a method will be developed to recommend service allocation strategies for foster youth in ways that balance the goal of ensuring fair distribution, according to multiple perspectives on fairness, and the goal of increasing the odds of positive life outcomes of all youth. All approaches will be evaluated extensively, with tight integration with multiple stakeholders in foster care. We will also provide evidence of the generality of our methods to other domains and provable bounds on their efficiency and accuracy. Stakeholder integration includes a partnership with a local foster care agency servicing over 10,000 youth per year, and a Youth Advisory Council of youth with experience in foster care who will play an important role in model development.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.