A new representation method for multivariate time series classification problem using interval means and polar histogram densities

Abstract

Multivariate time series (MTS) classification is an instance of common time series data mining tasks and is ubiquitously found in many domains such as medicine, finance or human-computer interaction. Traditionally, the research community has approached the problem by extending the well-established methods available in the univariate time series (UTS) classification literature. In this work, a new feature based method is developed specifically for MTS classification and aims to capture not only features of individual univariate series—as the extension methods do— but also the interaction between them. The method utilizes simple interval statistics as the base feature and polar histogram densities to represent 2-way interactions. The feature vectors are processed with a random forest classifier for its ability to handle high dimensionality. The results are reported for benchmark datasets of various types and from a range of domains. The method provides a satisfyingly accurate and scalable solution to the problem. The 2-way interaction information significantly increases the accuracy in most of the cases while the extraction phase of this information dominates the computation time. The method is comparable to the state-of-the-art methods in the literature even though there is a significant room for improvement.