Scoring Bayesian Networks of Mixed Variables.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automated Bayesian network learning is crucial for understanding complex systems.
• Existing methods often struggle with networks containing both continuous and discrete variables (mixed variables).
• Scalability is a key challenge in learning large, mixed-variable Bayesian networks.

Purpose of the Study:

• To develop two novel, scalable scoring functions for learning Bayesian networks with mixed variables.
• To provide efficient and flexible methods for automated Bayesian network structure discovery.

Main Methods:

• Introduced the Conditional Gaussian (CG) score for efficient handling of mixed variables.
• Developed the Mixed Variable Polynomial (MVP) score to model non-linear relationships.
• Integrated log-likelihood and degrees of freedom terms into a Bayesian Information Criterion (BIC) score.
• Proposed a structure prior for large network learning and a discrete case scoring simplification.
• Adapted scoring functions as conditional independence tests for constraint-based algorithms.

Main Results:

• The CG score offers high computational efficiency.
• The MVP score enables modeling of more complex, non-linear relationships.
• Empirical evaluations on simulated mixed-variable networks demonstrate the effectiveness of both methods.
• The proposed structure prior improves learning efficiency for large networks.

Conclusions:

• The CG and MVP scores provide effective and scalable solutions for learning Bayesian networks with mixed variables.
• These methods advance automated Bayesian network discovery, particularly for heterogeneous data.
• The scoring functions are adaptable to both search-and-score and constraint-based learning paradigms.