Comparative evaluation of score criteria for dynamic Bayesian Network structure learning.

Area of Science:

• Computer Science
• Statistics
• Artificial Intelligence

Background:

• Dynamic Bayesian Networks (DBNs) are key probabilistic models for temporal processes.
• Structure learning in DBNs influences model performance.
• Score-based and hybrid DBN learning methods rely on specific score criteria.

Purpose of the Study:

• To investigate the impact of diverse score criteria on Dynamic Bayesian Network structure learning.
• To compare the performance of commonly used scores against Akaike-based and Bayesian-based information criteria.

Main Methods:

• Adapted Akaike-based criteria: AIC, CAIC, KIC, AIC4.
• Adapted Bayesian-based criteria: BIC, BICadj, HBIC, BICQ.
• Evaluated these criteria within score-based and hybrid DBN structure learning frameworks.

Main Results:

• Performance variations were observed across different score criteria.
• Specific criteria demonstrated differential effects on DBN structure learning outcomes.
• The study provides empirical evidence on criterion impact.

Conclusions:

• The choice of score criterion significantly affects Dynamic Bayesian Network structure learning.
• Akaike-based and Bayesian-based information criteria offer viable alternatives for DBN structure learning.
• Further analysis is warranted to optimize criterion selection for specific temporal modeling tasks.