Structural and statistical uncertainty in observational causal machine learning at scale

Thesis

Causal machine learning (Causal ML) tackles various tasks, including causal-effect inference, causal reasoning, and causal structure discovery. This thesis explores uncertainty for Causal ML methods that scale to large datasets and complex, high-dimensional input/output modalities, such as images, text, time series, and videos. Scalability is essential for efficiently processing vast amounts of information and predicting complex relationships.

As we scale and achieve greater modeling flexibility, communicating the unknown becomes increasingly important. We examine two primary types of uncertainty: statistical and structural. Statistical uncertainty arises when fitting machine learning models to finite datasets. Addressing this uncertainty allows predicting a range of plausible causal effects that shrink with more training examples, facilitating better-informed decision-making and indicating areas needing improved understanding. Structural uncertainty arises from imprecise knowledge of the causal structure and generally requires further assumptions about the data-generating process or interaction with the world.

In this thesis, we develop scalable Causal ML methods that navigate statistical and structural uncertainty effectively. We demonstrate the importance of considering scalability and uncertainty in Causal ML algorithm design and application, enhancing decision making and knowledge acquisition. Our contributions aim to advance the Causal machine learning field and provide a foundation for future research.

Authors: Jesson, A

• DOI: 10.5287/ora-2n7a8v44x
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: observational data; deep learning; causal inference; uncertainty quantification
• Subjects: Machine learning