Dimension Selection Strategies for Multivariate Time Series Classification with HIVE-COTEv2.0

Abstract

Multivariate time series classification (MTSC) is an area of machine learning that deals with predicting a discrete target variable from multidimensional time dependent data. The possible high dimensionality of multivariate time series can affect the training time and possibly accuracy of complex classifiers, which often scale poorly in dimensions. We explore dimension filtering algorithms for high dimensional MTSC used in conjunction with the state of the art MTSC algorithm, HIVE-COTEv2.0. We apply and adapt recently proposed selection algorithms and propose new methods based on the ROCKET classifier built on single dimensions. We find that, for high dimensional MTSC problems, the best approach can on average filter between 50 % and 60 % of dimensions without significant loss of accuracy, reducing train time by a similar proportion.