Algorithms for Discovery of Multiple Markov Boundaries.

Area of Science:

• Machine Learning
• Causal Inference
• Statistical Modeling

Background:

• Markov boundary discovery is crucial for feature selection and understanding local causal structure.
• Existing algorithms primarily identify a single Markov boundary, which is insufficient for many real-world datasets.
• Multiple Markov boundaries are common and important in practical data analysis.

Purpose of the Study:

• To develop sound and efficient algorithms capable of discovering all Markov boundaries within a distribution.
• To address the limitations of current methods that fail to identify multiple Markov boundaries.

Main Methods:

• Introduction of the TIE* family of algorithms designed for comprehensive Markov boundary discovery.
• Extensive benchmarking study comparing TIE* against 26 state-of-the-art algorithms and variants.
• Evaluation across 15 diverse datasets from various application domains.

Main Results:

• The TIE* algorithm family demonstrates the capability to discover all Markov boundaries.
• Benchmarking confirms the broad applicability and efficiency of TIE* compared to existing methods.
• Successful application across a wide range of data sets highlights its practical utility.

Conclusions:

• TIE* provides a significant advancement in Markov boundary discovery, enabling the identification of all relevant boundaries.
• The developed algorithms offer a sound and efficient solution for complex data analysis scenarios.
• This work advances machine learning by providing tools for more complete feature selection and causal inference.