Improved Local Search with Momentum for Bayesian Networks Structure Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Statistics

Background:

• Bayesian Networks structure learning (BNSL) is computationally challenging.
• Exact search methods for BNSL are accurate but resource-intensive.
• Local search methods for BNSL can handle large networks but often get stuck in local optima.

Purpose of the Study:

• To develop novel operators for perturbing Bayesian Network structures.
• To create a framework that uses these operators to escape local optima during structure learning.
• To improve the accuracy and efficiency of BNSL algorithms.

Main Methods:

• Proposed two novel operators and one derived operator to perturb network structures while maintaining acyclicity.
• Designed a framework incorporating an influential perturbation factor using these operators.
• Integrated the perturbation factor to escape local optima in the search for optimal Bayesian Network structures.

Main Results:

• The proposed algorithm demonstrates competitive performance against state-of-the-art constraint-based methods.
• The algorithm achieves equivalent or superior solutions compared to state-of-the-art exact search and hybrid methods.
• Experimental results validate the effectiveness of the perturbation-based framework in BNSL.

Conclusions:

• The novel operators and framework effectively address the local optimum problem in BNSL.
• The proposed approach offers a practical and efficient solution for learning complex Bayesian Network structures.
• This work contributes to advancing the field of Bayesian Networks structure learning.