Dynamic Programming BN Structure Learning Algorithm Integrating Double Constraints under Small Sample Condition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Dynamic programming (DP) based Bayesian Network (BN) structure learning can achieve global optimal solutions.
• However, DP-based BN learning often yields inaccurate structures with small or incomplete sample data.
• Existing methods struggle with the reliability of learned BN structures under limited sample conditions.

Purpose of the Study:

• To propose a novel dynamic programming Bayesian Network structure learning algorithm incorporating double constraints for small sample scenarios.
• To enhance the accuracy and reliability of learned BN structures by addressing limitations of small sample sizes.
• To investigate the impact of integrating prior knowledge on the performance of BN structure learning.

Main Methods:

• Developed a dynamic programming BN structure learning algorithm with edge and path constraints.
• The double constraints were used to restrict the DP planning process and reduce the search space.
• Incorporated double constraints to guide the selection of optimal parent nodes, ensuring alignment with prior knowledge.

Main Results:

• The proposed double-constrained algorithm effectively limits the DP planning space and parent node selection.
• Simulation results demonstrate the algorithm's effectiveness in small sample conditions.
• Integrating prior knowledge significantly improved both the efficiency and accuracy of BN structure learning compared to non-integrating methods.

Conclusions:

• The double-constrained dynamic programming approach offers a robust solution for Bayesian Network structure learning with limited data.
• Prior knowledge integration is crucial for enhancing the performance of BN structure learning algorithms.
• The proposed method provides a more accurate and efficient way to learn BN structures, especially when dealing with small sample sizes.