Introduction to Causal Discovery
Abstract
Causal discovery is a subfield of causal inference that focuses on uncovering causal relationships from observational data. In this talk, we focus on two main families of causal discovery algorithms designed for observational settings: constraint-based algorithms and noise-based algorithms. Constraint-based algorithms rely on conditional independence tests to recover causal relationships, while noise-based algorithms exploit asymmetries in the noise distribution.We present two foundational constraint-based algorithms—the PC algorithm and the FCI algorithm—as well as the LiNGAM algorithm, a representative of noise-based algorithms. We discuss the intuition and assumptions behind the algorithms, but without providing rigorous proofs. We discuss the incorporation of background knowledge and some other practical aspects of causal discovery.
https://ckassaad.github.io/