An improved greedy equivalent search method based on relative entropy.

Area of Science:

• Computational statistics
• Machine learning
• Bioinformatics

Background:

• Greedy Equivalence Search (GES) is a widely used Bayesian network structure learning algorithm.
• GES faces computational challenges when starting from an empty graph.
• Initializing GES with a graph reflecting strong dependencies can improve performance.

Purpose of the Study:

• To propose a novel three-phase algorithm for generating an effective initial graph for GES.
• To enhance the efficiency and accuracy of Bayesian network structure learning.
• To evaluate the proposed method against existing algorithms and real-world data.

Main Methods:

• A three-phase approach is proposed to construct an initial graph.
• Relative entropy is utilized to measure variable relationships.
• The generated graph is transformed into the equivalence class space (E-space) for GES optimization.

Main Results:

• The proposed algorithm significantly improves the efficiency of GES.
• The accuracy of Bayesian network structure learning is enhanced by the new method.
• Comparative analyses demonstrate superior performance against state-of-the-art methods.

Conclusions:

• The developed three-phase algorithm provides a superior initialization strategy for GES.
• This approach offers a more efficient and accurate method for Bayesian network structure learning.
• The algorithm shows promise for applications including analysis of the COVID-19 pandemic data.