New Methods in High-Dimensional Causal Inference

随着深度学习算法在科学文献中的流行，以及谷歌、微软和Facebook等公司对其的使用和开发，它们在各种任务中的部署正在迅速加速。正因为如此，需要更好地理解深度学习的数学知识，更广泛地说，需要决策过程中的机器学习算法。需要的一项关键创新是如何将机器学习算法与因果建模过程相结合，这将帮助算法发展“推理能力”(例如，理解算法做出决策/预测的原因)。这个项目要解决的问题是如何对混杂因素进行建模，从而开发出具有理想抽样特性的因果效应估计器。此外，重要的是要有合理的程序，在计算上与观察值的数量成比例。在这个项目中，PI和团队将把他们的研究集中在两个领域。首先是理解深度学习算法的含义及其在流行的潜在结果模型的基本假设上的表现。在最近的工作中，PI发现了高斯过程分类算法、协变量重叠和因果效应估计量的正则性之间的基本紧张关系。这项研究的第一个目标将是看看深度学习算法是否也存在类似的现象。此外，还将探讨高斯过程的内插特性和基于深度学习的分类算法。该项目的第二部分将处理开发可扩展的因果效应估计算法。作为这项研究的一部分，将开发新的可计算扩展的因果效应估计算法。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。

With the popularity of deep learning algorithms in the scientific literature as well as by their use and development by companies such as Google, Microsoft and Facebook, their deployment in a variety of tasks is rapidly accelerating. Because of this, better mathematical understanding of deep learning, and more broadly, machine learning algorithms in decision making processes are needed. A key innovation that is needed is how to marry machine learning algorithms into causal modelling procedures, which will help algorithms to develop "reasoning capabilities" (e.g., understand why an algorithm is making the decision/prediction that it makes). The problem that will be addressed in this project is how to model confounders so as to develop causal effect estimators that have desirable sampling properties. In addition, it is important to have well-justified procedures that computationally scale with the number of observations. In this project, the PI and team will focus their research in two areas. The first will be to understand the implications of deep learning algorithms and their performance on foundational assumptions for the popular potential outcomes model. In recent work, the PI discovered a fundamental tension between Gaussian process classification algorithms, covariate overlap and regularity of causal effect estimators. The goal of the first aim of the research will be to see if a similar phenomenon holds for deep learning algorithms. In addition, the interpolation properties of Gaussian process and deep learning-based classification algorithms will be explored. The second part of the project will deal with developing scalable algorithms for causal effect estimation. New computationally scalable algorithms for causal effect estimation will be developed as part of this research.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.