PCMHS-Based Algorithm for Bayesian Networks Online Structure Learning

Given Bayesian Networks online structure learning problem, the paper presents an algorithm based on importance sampling and Parallel Crossover Metropolis-Hasting Sampler for evaluating online samples and network structure learning. The algorithm firstly selects the best samples for online structure learning using importance sampling method, and adjusts them according to the existed reliable network structure. Then on the basis of mutual information among nodes of the network, it initializes several parallel Markov Chains converging to Boltzmann distribution. At last new reliable network structure is formed by evaluating the learned structures in the process of iteration. The experimental result on standard data set shows that the algorithm can achieve online structure adjustment, and meanwhile has a high convergence speed, integration and learning accuracy.