Causal Structural Learning via Local Graphs.

Area of Science:

• Causal Inference
• Machine Learning
• Network Analysis

Background:

• Learning causal structures is challenging in high-dimensional data with unmeasured confounders and selection bias.
• Existing methods struggle with real-world networks containing hub nodes.

Purpose of the Study:

• To develop a new causal structure learning algorithm, local Fast Causal Inference (lFCI).
• To address limitations of standard algorithms in sparse, high-dimensional settings with latent and selection variables.

Main Methods:

• Proposed a novel local notion of sparsity tailored for complex network structures.
• Developed the local FCI (lFCI) algorithm, a variant of the Fast Causal Inference algorithm.
• Introduced an assumption for local determination of conditional dependencies.

Main Results:

• lFCI demonstrates consistency under the new sparsity condition and local dependency assumption.
• The algorithm offers reduced computational and sample complexity compared to standard FCI.
• lFCI achieves state-of-the-art performance on large random networks, especially those with hub nodes.

Conclusions:

• lFCI provides an effective solution for causal discovery in sparse, high-dimensional settings with unmeasured confounders and selection bias.
• The algorithm's ability to handle hub nodes makes it suitable for real-world network analysis.