Causal Discovery and Inference

In the realm of ever-evolving machine learning (ML), while it is essential to stay abreast of trends, our focus must transcend mere trend-chasing. As scientists, we are committed to addressing problems that possess the potential to positively transform our world—a notion we can term as “impact”-and for which we have a reasonable line of attack. We believe decision making (under constrained and uncertain environments, and where decisions can impact the world) is one of those problems.

The ML community have built incredibly powerful models for prediction. However, to use these models for truly impactful decision making they need to be lifted from learning predictive relationships in data to learning causal structures about the world. In this project we will develop new methods that leverage modern foundational machine learning (mostly based on probabilistic approaches) to address decision-making problems in a unifying framework. Our main line of attack can be viewed through the lenses of causal inference and estimation. In other words, ultimately, we care about answering “what if” questions with appropriate level of uncertainty or confidence, i.e., what will be the effect of changing a subset of variables in a complex, difficult to control, system? Examples of this abound and we are already working on practical applications where our developments will be critical, for example, (1) modelling well-being and (2) climate change.

Recent Developments

Directed Acyclic Graph (DAG) Estimation

• Contextual DAGs (Thompson et al., 2024)
• Bayesian DAG estimation via permutation-based distributions (Bonilla et al., 2024) and projection-induced distributions (Thompson et al., 2024)
• Optimal transport approachs to Parameter DAG estimation (Vo et al., 2024) and Structure DAG learning (Vo et al., 2024)
• Bayesian Granger Causality (Zhao & Bonilla, 2024)