Comparative evaluation of score criteria for dynamic Bayesian Network structure learning

Abstract

Dynamic Bayesian Networks (DBNs) are probabilistic models with a directional structure employed to model temporal processes. Three approaches to DBN structure learning are constraint-based, score-based, and hybrid. The score criterion determined in the score-based and hybrid approach has a certain effect on structure learning and this study aims to examine their performance by diversifying the score criteria used in DBN structure learning in addition to the scores commonly used in the literature. Thus, the Akaike-based information criteria as Akaike Information Criterion (AIC), Consistent AIC (CAIC), Kullback-Leibler Information Criterion (KIC), AIC4, and the Bayesian-based information criteria as Bayesian Information Criterion (BIC), Adjusted BIC (BICadj), Haughton BIC (HBIC), BICQ were adapted into the DBN structure learning. The obtained results were discussed.