Fast Search for Distance Dependent Chinese Restaurant Processes

Weiwei Feng, Peng Wang, Chuan Zhou, Li Guo, Peng Zhang

Abstract: The distance dependent Chinese Restaurant Processes(dd-CRP), a nonparametric Bayesian model, can model distance sensitive data. Existing inference algorithms for dd-CRP, such as Markov Chain Monte Carlo (MCMC) and variational algorithms, are inefficient and unable to handle massive online data, because posterior distributions of dd-CRP are not marginal invariant. To solve this problem, we present a fast inference algorithm for dd-CRP based on the A-star search. Experimental results show that the new search algorithm is faster than existing dd-CRP inference algorithms with comparable results.