Kernel similarity for classification and clustering of multi-variate time series with missing values.