Jointly Learning Consistent Causal Abstractions Over Multiple Interventional Distributions


In this presentation we review the definition of abstraction between structural causal models and we frame the problem of learning a mapping between them. We discuss the challenges of learning a causal abstraction that minimizes the abstraction error in terms of interventional consistency. We then suggest an approach based on a relaxation and parametrization of the problem, leading to a solution based on differentiable programming. The solution approach is evaluated both on synthetic and real-world data.

Slides here