Causal Abstraction: Definition, Measures and Learning
Date:
In this seminar we will provide an overview of the fast-developing field of causal abstraction. After a quick review of the formalism of structural causal models (SCMs), we will introduce the notion of causal abstraction; we will review the main frameworks proposed in the literature for expressing causal abstraction between SCMs, highlighting their strengths and limitations. We will then focus on the problem of measuring the consistency of SCMs at different levels of abstractions, and show that different measures may captures different aspects of causal abstraction. Finally, we will present the problem of learning causal abstractions between two SCMs and offer an overview of recent works tackling this challenge. We will conclude with a discussion of some open problems and future directions of research.