Skip to Main Content
Hidden Markov models (HMMs) are widely employed in sequential data modeling both because they are capable of handling multivariate data of varying length, and because they capture the underlying hidden properties of time-series. Over the years, HMM-based clustering methods have been widely investigated and improved. However, their performance on noisy data and the effectiveness of similarity measure between sequences remain less explored. In this paper, we present a robust algorithm for sequential data clustering by combining spectral analysis with HMMs. We first derive Fisher kernels from continuous density HMMs for similarity matrix construction, and then apply spectral clustering algorithm to the mapped data. The eigenvector decomposition step in spectral analysis is critical for noise removal and dimensionality reduction. Experimental results on both synthetic and real-world data indicate that our proposed approach is more tolerant to noise and achieves improved accuracy compared to many state-of-the-art algorithms.