Automated Causal Discovery with Observational Data via Directed Graphical Models - New Theory and Methods

在许多科学领域，包括生物学、心理学、神经科学、气候科学、机器人和量子力学，确定因果关系是至关重要的。虽然确定因果关系的黄金标准仍然是受控实验，但这可能是昂贵的、不道德的，在许多情况下甚至是不可能的。因此，从被动观察的数据(而不是实验数据)中确定因果关系通常是可取的，有时也是唯一的选择。在这个项目中，PI将开发一系列的因果发现方法，这些方法在理论上是可靠的，在实际应用中有助于从观测数据中识别因果关系。将开发与拟议方法配套的高效开放源码软件，该项目还为研究生提供研究培训机会。所提出的方法将基于有向图形模型(DGM)。尽管DGM在各学科中很受欢迎，但由于几个突出的挑战，使用DGM从观测数据中建立因果关系在理论上和方法上都仍然困难。首先，由于马尔可夫等价类的原因，DGM通常是不可识别的，在马尔可夫等价类中，所有DGM编码相同的条件独立性集合，因此在没有进一步假设的情况下彼此不能区分。其次，DGM类在边缘化条件下不是封闭的，因此结构学习可能会被不可测量的混杂因素所误导。第三，现有的绝大多数方法依赖于数据生成机制上相对较强的分布假设，当假设被严重违反时，这可能会导致显著的估计偏差。该项目旨在通过开发新的非IID数据的DGM并在存在混杂因素和模型错误指定的情况下建立其因果可识别性理论来应对这三个挑战。作为验证，所提出的方法将用于从基因组数据集中对基因调控网络进行逆向工程。结果将通过研讨会、出版物和新的研究生课程传播。这一奖项反映了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。

Establishing causality is crucial in many fields of science including biology, psychology, neuroscience, climate science, robotics, and quantum mechanics. While the gold standard for establishing causality remains controlled experimentation, it can be expensive, unethical, and even impossible in many cases. Therefore, establishing causality from passively observed data (as opposed to experimental data) is often desirable and, sometimes, the only option. In this project, the PI will develop a series of causal discovery methods that are theoretically sound and practically useful for identifying causality with observational data. Efficient open-source software accompanying the proposed methods will be developed and the project also provides research training opportunities for graduate students. The proposed methods will be based on directed graphical models (DGMs). Despite the popularity of DGMs across disciplines, using DGMs to establish causality from observational data remains difficult, both theoretically and methodologically, due to several prominent challenges. First, DGMs are generally non-identifiable due to Markov equivalence class in which all DGMs encode the same set of conditional independencies and hence are not distinguishable from each other without further assumptions. Second, the class of DGMs is not closed under marginalization and therefore the structure learning can be misled by unmeasured confounders. Third, the vast majority of existing methods rely on relatively strong distributional assumptions on the data generating mechanism, which can cause significant estimation biases when the assumptions are seriously violated. This project aims to address these three challenges by developing new DGMs for non-iid data and establishing their causal identifiability theories in the presence of confounders and model misspecification. As validation, the proposed methods will be used to reverse engineer gene regulatory networks from genomic datasets. Results will be disseminated through workshops, publications, and new graduate courses.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.