DiCoLa: A Recursive Decomposition Framework for Causal Structure Learning with Latent Variables

Researchers have introduced DiCoLa, a novel recursive decomposition framework designed to enhance causal structure learning in the presence of latent variables. Traditional constraint-based causal discovery methods often struggle with high computational costs due to heavy reliance on conditional independence testing, particularly in high-dimensional settings. While existing divide-and-conquer strategies aim to mitigate this, they typically assume causal sufficiency, ignoring latent variables. This new study theoretically generalizes divide-and-conquer approaches to accommodate latent variables. The DiCoLa framework recursively breaks down global learning tasks into manageable subproblems and integrates their solutions through a principled reconstruction step to recover the global causal structure. The authors establish the soundness and completeness of the framework theoretically. Extensive experiments on synthetic data demonstrate significant improvements in computational efficiency across various causal discovery algorithms. Additionally, tests on real-world datasets confirm the practical effectiveness of the approach, offering a robust solution for complex causal inference challenges in machine learning and artificial intelligence.