LiNGAM-SF: Causal Structural Learning Method With Linear Non-Gaussian Acyclic Models for Streaming Features.

Area of Science:

• Causal inference
• Machine learning
• Time series analysis

Background:

• Causal structure learning for streaming features (CSLSFs) is challenged by limited precision and inability to detect latent confounders.
• Existing score-based methods often struggle with accuracy in dynamic data streams.

Purpose of the Study:

• To propose a novel causal structure learning method for streaming features using linear non-Gaussian acyclic models (LiNGAM-SFs).
• To address limitations in precision and latent confounder detection in current CSLSF methods.

Main Methods:

• Utilizing the causal identifiability of data within a LiNGAM framework for online learning.
• Employing the classical SF algorithm for causal skeleton learning and proving its properties.
• Introducing the identify causal directions in the presence of latent variables (ICDPLV) subalgorithm for direction identification.
• Implementing the detecting latent confounder (DLC) subalgorithm for global latent confounder detection.

Main Results:

• The proposed LiNGAM-SF method achieves at least an 11% average increase in precision compared to state-of-the-art methods.
• Experimental results confirm the method's effectiveness in detecting latent confounders.
• The ICDPLV subalgorithm successfully distinguishes between possible causal structures and identifies directions.

Conclusions:

• LiNGAM-SFs offer a significant advancement in online causal structure learning for streaming data.
• The method effectively addresses precision limitations and latent confounder detection challenges.
• This work represents the first application of LiNGAMs for online causal structure learning.