Similarity-based approaches represent a promising direction for time seriesanalysis. However, many such methods rely on parameter tuning, and some haveshortcomings if the time series are multivariate (MTS), due to dependenciesbetween attributes, or the time series contain missing data. In this paper, weaddress these challenges within the powerful context of kernel methods byproposing the robust \emph{time series cluster kernel} (TCK). The approachtaken leverages the missing data handling properties of Gaussian mixture models(GMM) augmented with informative prior distributions. An ensemble learningapproach is exploited to ensure robustness to parameters by combining theclustering results of many GMM to form the final kernel. We evaluate the TCK on synthetic and real data and compare to otherstate-of-the-art techniques. The experimental results demonstrate that the TCKis robust to parameter choices, provides competitive results for MTS withoutmissing data and outstanding results for missing data.
展开▼
