Learning Decision Rules with Observational Data

对大规模和复杂数据的分析在社会中发挥着越来越重要的作用，机器学习方面的创新正在产生更强大的预测技术。然而，当我们使用这些数据来指导决策时，重要的是要认识到这些领域中的大多数数据集是观察性的，而不是随机的，并且需要仔细分析，以便得出关于部署潜在政策的因果关系的正确结论。该研究旨在开发数据驱动决策的新方法，这些方法可以利用机器学习的力量和表现力，同时严格建立在非随机数据因果推理的最佳实践之上。本项目围绕以下三个统计任务展开：(1)研究观察性研究中异质性治疗效果估计的问题，并开发一个可用于增强或神经网络等的灵活框架。所提出的方法的准确性取决于我们可以干预的因果信号的复杂性，而不是仅仅依赖于其他关联信号。(2)考虑福利最大化的结构化政策学习，并研究一种在非参数设置下后悔衰减为样本量的平方根反比的方法。(3)考虑从顺序随机数据中学习最优停止规则的问题，并提出一种新的鲁棒且计算可行的策略学习方法。所有这些结果背后的一个统一主题是，它们强调了如何使用半参数统计的经典思想来严格利用准确的机器学习预测器来解决决策问题。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。

The analysis of large-scale and complex data plays an increasingly central role in society, and innovations in machine learning are yielding ever more powerful predictive technologies. However, when we use such data to guide decision making, it is important to recognize that the majority of datasets in these domains are observational rather than randomized in nature, and require careful analysis in order to draw correct conclusions about the causal effect of deploying a potential policy. The research aims to develop new methods for data-driven decision making that can harness the power and expressiveness of machine learning, all while rigorously building on best practices for causal inference from non-randomized data.This project is centered around the following three statistical tasks: (1) Examine the problem of heterogeneous treatment effect estimation in observational studies, and develop a flexible framework that can be used with, e.g., boosting or neural networks. The accuracy of the proposed method depends on the complexity of the causal signal that we can intervene on, not on other merely associational signals. (2) Consider welfare maximizing structured policy learning, and study an approach whose regret decays as the inverse square root of the sample size in a non-parametric setting. (3) Consider the problem of learning optimal stopping rules from sequentially randomized data, and propose a new robust yet computationally feasible approach to policy learning in this setting. A unifying theme underlying all these results is that they highlight how classical ideas from semiparametric statistics can be used to rigorously leverage accurate machine learning predictors in decision-making problems.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.