Open AccessTypically, time series classifiers require signal pre-processing (filtering signals from noise and artifact removal, etc.), enhancement of signal features (amplitude, frequency, spectrum, etc.), classification of signal features in space using the classical techniques and classification algorithms of multivariate data. We consider a method of classifying time series, which does not require enhancement of the signal features. The method uses the shapelets of time series (time series shapelets) i.e. small fragments of this series, which reflect properties of one of its classes most of all.
Despite the significant number of publications on the theory and shapelet applications for classification of time series, the task to evaluate the effectiveness of this technique remains relevant. An objective of this publication is to study the effectiveness of a number of modifications of the original shapelet method as applied to the multivariate series classification that is a littlestudied problem. The paper presents the problem statement of multivariate time series classification using the shapelets and describes the shapelet–based basic method of binary classification, as well as various generalizations and proposed modification of the method. It also offers the software that implements a modified method and results of computational experiments confirming the effectiveness of the algorithmic and software solutions.
The paper shows that the modified method and the software to use it allow us to reach the classification accuracy of about 85%, at best. The shapelet search time increases in proportion to input data dimension.