MSAX: Multivariate Symbolic Aggregate Approximation for Time Series Classification

Abstract

Time Series (TS) analysis is a central research topic in areas such as finance, bioinformatics, and weather forecasting, where the goal is to extract knowledge through data mining techniques. Symbolic aggregate approximation (SAX) is a state-of-the-art method that performs discretization and dimensionality reduction for univariate TS, which are key steps for TS representation and analysis. In this work, we propose MSAX, an extension of this algorithm to multivariate TS that takes into account the covariance structure of the data. The method is tested in several datasets, including the Pen Digits, Character Trajectories, and twelve benchmark files. Depending on the experiment, MSAX exhibits comparable performance with state-of-the-art methods in terms of classification accuracy. Although not superior to 1-nearest neighbor (1-NN) and dynamic time warping (DTW), it has interesting characteristics for some classes, and thus enriches the set of methods to analyze multivariate TS.

Supported by national funds through Fundação para a Ciência e a Tecnologia (FCT) through projects UIDB/50021/2020 (INESC-ID), UIDB/50022/2020 (LAETA, IDMEC), UIDB/50008/2020 (IT), PREDICT (PTDC/CCI-CIF/29877/2017), and MATISSE (DSAIPA/DS/0026/2019).