Improved Parameter Uniform Priors in Bayesian Network Structure Learning

Abstract

Bayesian Dirichlet equivalent uniform score (BDeu) is often used in Bayesian structure learning. But it does not work well when data size is sparse because the equivalence of the prior parameter distribution isn’t suit for the specific data set. To break the rules of uniform and equivalent, the paper proposes the Bayesian Dirichlet Sparse score (BDs) which change distribution of prior parameter through the all zero items in the sparse data. The circulation principle of information entropy and simulations are used to explain the reason why BDs is better than BDeu when data size is sparse. In the experiments, we also verify the stability of BDs when hyperparameters change.