Analyzing incomplete sequences using gaussian hidden Markov models

This paper studies the methods of incomplete sequence analysis using Gaussian hidden Markov models (HMMs). We present Marginalization algorithm, which can be applied both for training HMM on incomplete sequences and for recognition of incomplete sequences using HMMs. In addition, we present a modification of Viterbi algorithm that can be used for decoding and imputation of incomplete sequences using HMM. Both presented algorithms significantly outperform the standard methods of incomplete sequence analysis, namely: elimination of missing observations in sequences followed by “gluing” of the remaining subsequences into one sequence and imputation of missing observations with the mean of the neighboring observations.