A Bayesian Approach for Learning Bayesian Network Structures

We introduce a Bayesian approach method based on the Gibbs sampler for learning the Bayesian Network structure. For this, the existence and the direction of the edges are specified by a set of parameters. We use the non-informative discrete uniform prior to these parameters. In the Gibbs sampling, we sample from the full conditional distribution of these parameters, then a set of DAGs is obtained. For achieving a single graph that represents the best graph fitted on data, Monte Carlo Bayesian estimation of the probability of being the edge between nodes is calculated. The results on the benchmark Bayesian networks show that our method has higher accuracy compared to the state-of-the-art algorithms.