An EM-MCMC algorithm for Bayesian structure learning

Structure of the network plays a central role in applications of Bayesian network. However, Bayesian network structure learning is a hard problem to solve. We propose an method, EM-MCMC algorithm for Bayesian network structure learning in which expectation maximization(EM) algorithm is applied to Bayesian parameter learning and Markov chain Monte Carlo(MCMC) method is used to sample for Bayesian structure. Differ from previous ones, we integrate accepting rejection sampling to MCMC acceptance function. We compare our method with EM-EM Bayesian structure learning that uses EM algorithm in parameter learning and structure learning, experimental results demonstrate the effectiveness and feasibility of our algorithm.