CAREER: Robust, Interpretable, and Fair Allocation of Scarce Resources in Socially Sensitive Settings

这项学院早期职业发展计划(CALEAR)赠款将通过改善分配稀缺资源以满足基本需求的关键公共部门系统，为促进国家繁荣和经济福利做出贡献。这些系统在复杂、不确定、依赖时间的环境中运行，作为公共部门服务，它们必须是透明的，满足潜在冲突的利益相关者目标，并被构建为在部署时在各种环境中按预期运行。为了解决这些问题，该项目构建了一个计算高效的框架，以设计对不确定性健壮、可解释和公平的策略。这些系统将学习并正确地平衡利益相关者的价值判断，并考虑到潜在的偏见、激励和差异。指导这项研究的核心用例将集中在将稀缺的住房资源分配给那些经历无家可归的人。该项目将通过与洛杉矶无家可归者服务局和无家可归专家的合作来促进。将研究与教育相结合的计划包括设计一门新的“社会影响分析”课程和一个在线实验平台，以教育学生和普通公众在社会敏感环境中分配资源。外展活动将侧重于促进STEM领域的多样性、公平性和包容性，包括与好莱坞STEM学院的长期合作伙伴关系和与Code.org非营利组织的新合作。这项研究将推进数据驱动的稳健优化模型，以很好地处理信息的不完备性和非平稳性，并得出可提供概率性能保证的易处理的模型。该项目将提供一种学习和聚合不完整和相互冲突的利益相关者偏好的方法，并提供新的公平的机器学习(ML)算法。这项研究将开发一种新的稳健排队理论框架，该框架利用学习的偏好、结果预测和观察数据来设计政策，确保在开放世界部署时按计划工作。最后，该项目将提供工具，帮助委员会评估政策，预测其后果，并了解公平、效率和可解释性之间的权衡。该框架在几个方面为健壮的优化文献做出了贡献。为决策相关信息发现、指数型多偶然性、目标非线性的多阶段稳健优化问题提供了一种建模和求解方案。这是第一次对稳健公平的ML的模型和方法进行研究。最后，在鲁棒排队理论的框架下，提供了一种从噪声观测数据中执行反事实策略评估和优化的方法。这项研究为人工智能和市场营销提供了新的基于优化的偏好诱导和聚合技术，为ML提供了通用稳健和公平的工具，并为排队理论提供了基于因果推理的系统评估和设计方法。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。

This Faculty Early Career Development Program (CAREER) grant will contribute to the advancement of national prosperity and economic welfare by improving critical public sector systems that allocate scarce resources to satisfy basic needs. These systems operate in complex, uncertain, time-dependent environments, and as public sector services, they must be transparent, satisfy potentially conflicting stakeholder objectives, and be constructed to perform as intended in a variety of environments when deployed. To address these problems, this project constructs a computationally efficient framework to design policies that are robust to uncertainty, interpretable, and fair. The systems will learn and correctly balance stakeholder value judgements and account for underlying biases, incentives, and disparities. The central use case that will guide the research will focus on allocating scarce housing resources to those experiencing homelessness. The project will be facilitated by a collaboration with the Los Angeles Homeless Services Authority and with homelessness experts. The plan to integrate research and education includes the design of a new course on “Analytics for Social Impact” and of an online experimental platform to educate students and the general public about resource allocation in socially sensitive settings. Outreach activities will be focused on the promotion of diversity, equity, and inclusion in STEM fields and include a long-term partnership with the STEM Academy of Hollywood and a new collaboration with the Code.org non-profit.This research will advance data-driven robust optimization models that cope well with information incompleteness and non-stationarity, and derive tractable models that offer probabilistic performance guarantees. This project will provide a methodology for learning and aggregating incomplete and conflicting stakeholder preferences and offer new fair machine learning (ML) algorithms. The research will develop a novel robust queuing theory framework that leverages learned preferences, outcome predictions, and observational data to design policies guaranteed to work as planned when deployed in the open world. Finally, the project will provide tools that help committees evaluate policies, anticipate their consequences, and understand the trade-offs between fairness, efficiency, and interpretability. The framework contributes to the robust optimization literature in several regards. It provides a modeling and solution scheme for multi-stage robust optimization problems with decision-dependent information discovery, exponentially many contingencies, and non-linear objective. It is the first study on models and methods for robust and fair ML. Finally, it provides a method for performing counterfactual policy evaluation and optimization from noisy observational data in the robust queuing theory framework. The research contributes to AI and marketing with new optimization-based techniques for preference elicitation and aggregation, to ML with general purpose robust and fair tools, and to queuing theory with causal inference-based approaches for system evaluation and design.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.