Structure Learning Under Missing Data.

Area of Science:

• Causal inference
• Machine learning
• Statistics

Background:

• Causal discovery aims to infer graphical causal models from data.
• Missing data is a pervasive challenge in real-world causal discovery problems.
• Standard algorithms applied to incomplete data can lead to biased inferences due to selection bias.

Purpose of the Study:

• To develop robust methods for causal discovery in the presence of missing data.
• To adapt existing structure learning algorithms to handle systematically missing entries.
• To address scenarios with known and unknown missing data models.

Main Methods:

• Developed an algorithm for causal discovery with known missing data models.
• Proposed approaches for causal discovery when the missingness model is unknown.
• Validated methods through simulations comparing performance against standard algorithms.

Main Results:

• The proposed adjustments effectively account for missing data in causal discovery.
• The methods demonstrated superior performance compared to standard structure learning algorithms.
• Successful causal graph inference was achieved across different missing data scenarios.

Conclusions:

• Adjusted structure learning algorithms provide accurate causal graph inference with missing data.
• The developed methods offer reliable solutions for causal discovery problems with incomplete datasets.
• This work advances the field of causal inference by addressing a critical practical limitation.