A New Metric for Classification of Multivariate Time Series

Abstract

Multivariate time series are an important kind of data collected in many domains, such as multimedia, biology and so on. We focus on discrimination metric for time series data; especially classify the multivariate time series as stationary or non-stationary. In this paper we present a new metric, the nonlinear trend of the cross-correlation matrix, for classification of multivariate time series, which could well depict the stationarity of multivariate time series. The proposed approach has been tested using two datasets, one natural and one synthetic, and is shown to our metric is more efficient than the benchmark metric in all cases. We take K-means clustering in the experiment.