Reproduce “Conditions and Assumptions for Constraint-based Causal Structure Learning” on OpenResearch

About this paper

Kayvan Sadeghi, Terry Soo

This research provides a theoretical framework for constraint-based causal structure learning that relaxes the restrictive faithfulness assumption. It introduces alternative, potentially testable conditions such as ordered stability and path-stability, under which algorithms can accurately discover the Markov equivalence class of the true causal graph, with specific applications to Structural Causal Models.

Key insights

• The study shows that faithfulness is not a necessary condition for correct causal discovery; a combination of weaker, more specific properties like ordered stability and path-stability is sufficient.
• Many of the newly identified conditions for accurate causal discovery, especially for SCMs, are based on intrinsic properties of the probability distribution or SCM functions, making them, in principle, empirically testable.
• The global Markov property holds for general SCMs, even in the presence of latent variables, providing a robust theoretical basis for SCM-based causal inference.

Summary by alphaXiv.