Research and Implement of Structure Learning Algorithm for Hybrid Bayesian Networks

Traditional Bayesian networks structural learning usually needs domain experts providing some priori information to reduce the search space of network structures, so the accuracy of attained result depends on the experts’ comprehension to dataset to some extent. To overcome this drawback, the paper proposes a novel hybrid three-phase algorithm HBN, which firstly using the concept of pseudo-BN. In VO learning phase, we draw lessons from the index information gain in Algorithm ID3 assisting to sort variables. In constructing pseudo-BN phase, the paper designs a novel scoring function and improves the algorithm finding an approximate minimal d-separating set. In closing operation phase, we eliminate the fake of pseudo-BN. Finally, lots of experiments show algorithm HBN is relatively adaptive to construct medium and small sized network structure and possesses some merits such as fleet construction and satisfactory classificatory accuracy compared with TAN and NBC, though the run time needed is longer than other two models.