Mining Markov Blankets Without Causal Sufficiency.

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Markov blankets (MBs) are crucial for causal discovery and Bayesian network (BN) structure learning.
• Existing MB discovery algorithms often assume causal sufficiency, ignoring potential latent common causes.
• Latent common causes are prevalent in real-world data, limiting the applicability of current methods.

Purpose of the Study:

• To develop an algorithm for discovering MBs that does not rely on the causal sufficiency assumption.
• To address the practical challenge of causal structure learning with unobserved confounding variables.
• To enable more robust causal discovery in complex, real-world datasets.

Main Methods:

• Utilized the maximal ancestral graph (MAG) model to represent latent common causes.
• Adapted the concept of MBs to settings without causal sufficiency.
• Proposed an effective and efficient algorithm for target MB discovery within an MAG framework.

Main Results:

• Developed a novel algorithm for Markov blanket discovery under the presence of latent common causes.
• Demonstrated the algorithm's effectiveness and efficiency through experiments.
• Validated the approach using both benchmark and real-world datasets.

Conclusions:

• The proposed algorithm successfully discovers Markov blankets without assuming causal sufficiency.
• This work advances causal structure learning by accommodating unobserved confounding.
• The findings are significant for applying Bayesian networks to real-world problems with complex causal relationships.