Contextual DAGs at AISTATS 2024

Ryan Thompson presented our paper on Contextual Directed Acyclic Graphs at AISTATS 2024.

In this paper we estimate directed acyclic graphs (DAGs) from observational data where the DAG structure can be different for distinct individuals based on contextual features. The significance of DAG estimation in causal discovery is well known and documented but here we are taking it one step further by allowing each individual to have a (potentially) different DAG. This setting has many applications (e.g., in personalized medicine) and, indeed, I think it is more realistic than estimating a DAG for the entire population.

The paper has some neat mathematical tricks :sparkles: that allowed us to develop a fast and scalable algorithm, given the complexity of having neural networks to map contextual features into a DAG. We also provide a convergence guarantee and some cool experiments, including the analysis of drug consumption.