Causal Structure Learning

Causal structure learning seeks graphs whose arrows represent direct cause-and-effect relationships, helping explain systems and predict what may change under interventions. Methods include conditional-independence tests such as PC and FCI, score-based and Bayesian graph search, functional models such as LiNGAM and additive noise models, and continuous optimization methods such as NOTEARS; interventions and temporal structure can further constrain the search. Work increasingly addresses time-varying systems, latent confounders, high-dimensional data, and ways to incorporate expert or semantic knowledge. A central challenge is that observational data often leave multiple graphs indistinguishable, while stronger identifying assumptions, hidden variables, noisy data, and scalability complicate reliable recovery and evaluation.