Regularization of Hidden Markov Models Embedded into Reproducing Kernel Hilbert Space

摘要

Hidden Markov models (HMMs) are well-known probabilistic graphical models for time series of discrete, partially observable stochastic processes. In this paper, we discuss an approach to extend the application of HMMs to non-Gaussian continuous distributions by embedding the belief about the state into a reproducing kernel Hilbert space (RKHS), and reduce tendency to overfitting and computational complexity of algorithm by means of various regularization techniques, specifically, Nystr?m subsampling. We investigate, theoretically and empirically, regularization and approximation bounds, the effectiveness of kernel samples as landmarks in the Nystr?m method for low-rank approximations of kernel matrices. Furthermore, we discuss applications of the method to real-world problems, comparing the approach to several state-of-the-art algorithms.