Reliable Early Classification on Multivariate Time Series with Numerical and Categorical Attributes

Abstract

Early classification on multivariate time series has recently emerged as a novel and important topic in data mining fields with wide applications such as early detection of diseases in healthcare domains. Most of the existing studies on this topic focused only on univariate time series, while some very recent works exploring multivariate time series considered only numerical attributes and are not applicable to multivariate time series containing both of numerical and categorical attributes. In this paper, we present a novel methodology named REACT (Reliable EArly ClassificaTion), which is the first work addressing the issue of constructing an effective classifier on multivariate time series with numerical and categorical attributes in serial manner so as to guarantee stability of accuracy compared to the classifiers using full-length time series. Furthermore, we also employ the GPU parallel computing technique to develop an extended mechanism for building the early classifier efficiently. Experimental results on real datasets show that REACT significantly outperforms the state-of-the-art method in terms of accuracy and earliness, and the GPU implementation is verified to substantially enhance the efficiency by several orders of magnitudes.