An Efficient Entropy-Based Causal Discovery Method for Linear Structural Equation Models With IID Noise Variables.

Area of Science:

• Causal inference
• Machine learning
• Information theory

Background:

• Discovering causal relationships from observational data is crucial.
• Existing methods like LiNGAM struggle with Gaussian noise and measurement errors.
• A robust method is needed for accurate causal discovery.

Purpose of the Study:

• To propose a novel entropy-based two-phase iterative algorithm.
• To address limitations of existing algorithms in handling Gaussian noise and measurement errors.
• To accurately identify unique causal structures from observational data.

Main Methods:

• An entropy-based two-phase iterative algorithm is developed.
• Phase one identifies exogenous variables using entropy for arbitrary distributions.
• Phase two revises exogenous variables to remove measurement error effects.

Main Results:

• The algorithm effectively identifies causal structures in data with arbitrary distributions and measurement errors.
• Experimental results on real-world data demonstrate superior effectiveness and stability.
• Application to mobile-base-station data confirms algorithm efficacy.

Conclusions:

• The proposed algorithm advances causal discovery by handling complex data conditions.
• It offers a robust solution for identifying causal relationships where other methods fail.
• The method shows significant promise for real-world applications with noisy data.