Causal Discovery in Linear Non-Gaussian Acyclic Model With Multiple Latent Confounders

Area of Science:

• Causal inference
• Statistical modeling
• Machine learning

Background:

• Causal discovery from observational data is crucial in scientific research.
• The linear non-Gaussian acyclic model (LiNGAM) is effective but struggles with multiple latent confounders.
• Existing methods face challenges in detecting latent confounders and uncovering complex causal relationships.

Purpose of the Study:

• To propose a hybrid causal discovery method for LiNGAM with multiple latent confounders (MLCLiNGAM).
• To address the challenges of detecting latent confounders and uncovering causal relations among observed and latent variables.
• To enhance causal discovery in complex observational datasets.

Main Methods:

• Utilizing a constraint-based method to learn the causal skeleton.
• Identifying causal directions through regression and independence tests on adjacent pairs.
• Detecting latent confounders using maximal clique patterns and reconstructing the causal structure.

Main Results:

• Theoretical results demonstrate the correctness and efficiency of the proposed MLCLiNGAM algorithms.
• Extensive experiments on synthetic and real data validate the method's effectiveness.
• The approach successfully addresses latent confounder detection and causal structure reconstruction.

Conclusions:

• The MLCLiNGAM method provides an effective solution for causal discovery in the presence of multiple latent confounders.
• The hybrid approach improves upon existing LiNGAM methods by integrating constraint-based and regression-based techniques.
• This work advances the field of causal inference from observational data, offering robust tools for scientific discovery.