An effective EM algorithm for mixtures of Gaussian processes via the MCMC sampling and approximation

摘要

The Mixture of Gaussian Processes (MGP) is a powerful statistical model for characterizing multimodal data, but its conventional Expectation-Maximization (EM) algorithm (Dempster et al., 1977) is computationally intractable because of its time complexity. To solve this problem, some approximation techniques have been proposed in the conventional EM algorithm. However, these approximate EM algorithms are ineffective or limited in some situations. To implement the EM algorithm more effectively, we approximate the EM algorithm with simulated samples of latent variable via the Monte Carlo Markov Chain (MCMC) sampling, and design an MCMC EM algorithm. Experiments on both synthetic and real-world data sets demonstrate that our MCMC EM algorithm is more effective than the state-of-the-art EM algorithms on classification and prediction problems. (C) 2018 Elsevier B.V. All rights reserved.