Restricted Bayesian Network Structure Learning

Learning the structure of a Bayesian network from data is a difficult problem, as its associated search space is superexponentially large. As a consequence, researchers have studied learning Bayesian networks with a fixed structure, notably naive Bayesian networks and tree-augmented Bayesian networks, which involves no search at all. There is substantial evidence in the literature that the performance of such restricted networks can be surprisingly good. In this paper, we propose a restricted, polynomial time structure learning algorithm that is not as restrictive as both other approaches, and allows researchers to determine the right balance between classification performance and quality of the underlying probability distribution. The results obtained with this algorithm allow drawing some conclusions with regard to Bayesian-network structure learning in general.