Research on Bayesian Network Structure Score Function

Abstract

In order to induct a Bayesian network from data, the different structure score functions were proposed and applied to the structure learning of Bayesian network. This paper researches the inter-node correlation strength of network and presents a new score function, namely Bayesian conditional probability statistic (BCPS) score function. BCPS score function rigorous formula was given. The BCPS, BIC and BDe score function and K2 algorithm was applied to structure learning of Asia network and Alarm network. The results show that BCPS score function can estimate the fitting degree of data and structure rightly. The stability of K2-BCPS structure learning algorithm was better than K2-BCPS structure learning algorithm, and the Accuracy is superior to K2-BDe structure learning algorithm.