A Hybrid Stochastic Sampling Algorithm for Bayesian Network Structure Learning
Chunlin Hu · Jisuanji gongcheng · 2014
According to slow convergence speed of stochastic sampling algorithm, based on uniform sampler and independent sampler, by improving convergence speed from the initial sample, sampling method and proposal distribution, a hybrid markov chain Monte Carlo sampling algorithm(HSMHS) is put forward in this paper. Based on mutual information between network nodes, it generates initial samples of network structure. In iteration sampling phase, according to certain probability distribution, it randomly selects uniform sampler or independent sampler, and computs proposal distribution of independent sampler based on the current samples to improve the mixing of sampling process. It can be proved that sampling process of HSMHS converges to the posterior probability of network structure, and the algorithm has a good learning accuracy. Experimental results on standard data set also verify that both learning efficiency and precision of HSMHS outperform classical algorithms MHS, PopMCMC and Order-MCMC.