Mosaic: An Accurate and Efficient Kernel-Based Multivariate Time Series Classifier

Multivariate time series classification has received significantly less attention than its univariate counterpart. Many current state-of-the-art multivariate time series classification methods are naive generalizations of their univariate versions without explicitly considering inter-dimensional correlations. This paper presents Mosaic, a kernel-based multivariate time series classification method. In contrast to many existing approaches, Mosaic specifically focuses on extracting inter-dimensional correlations through transforming the entire multivariate time series instances into multi-dimensional latent space, novel pooling operations, and cross-validation-based feature ensembling. When evaluated on the University of East Anglia (UEA) multivariate time series classification datasets, Mosaic achieves the best overall accuracy compared to existing kernel methods and performs competitively with HIVE-COTE 2 (HC2), currently the best-performing time series classifier while being at least one order of magnitude faster than HC2.