Reproduce “Local Causal Structure Learning in the Presence of Latent Variables and Selection Bias” on OpenResearch

About this paper

Zheng Li, Hao Zhang, Ruxin Wang, Ruichu Cai, Kun Zhang, Feng Xie

LoCaLS, an algorithm developed by researchers at the Shenzhen Institutes of Advanced Technology and collaborators, learns direct causes and effects of a target variable from observational data under both latent variables and selection bias. It demonstrates superior efficiency by requiring significantly fewer conditional independence tests than global methods, while maintaining comparable or improved target-specific accuracy on synthetic and real-world gene expression data.

Key insights

• Development of a novel definition for the Markov blanket that extends to scenarios involving both latent variables and selection bias, proving its minimality and conditional independence properties.
• Formal theoretical guarantees demonstrating that specific causal information (adjacencies and certain collider types) about a target variable, when learned locally, remains sound and consistent with the global causal structure.
• The proposed LoCaLS algorithm achieves both soundness and completeness for identifying identifiable direct causes and effects of a target variable under the challenging combined presence of latent variables and selection bias.

Summary by alphaXiv.