Early Classification of Multivariate Time Series on Distributed and In-Memory Platforms

Abstract

With the popularity of Internet of Things (IOT) applications, various kinds of active sensors are deployed and multivariate time series datasets are generated rapidly. Early classification of multivariate time series is an emerging topic in data mining due to the wide applications in many domains. The unique part of early classification lies in that it uses only earlier part of time series data to reach classification results with the same accuracy as by methods using complete time series information. Although a number of relevant studies have been presented recently, most of them didn't consider the issues of data scale and execution efficiency simultaneously. The main research issue of this paper falls in how to mine interpretable patterns from multivariate time series data, with which effective classification models can be constructed to ensure the accuracy as well as earliness. To take into account the issues of data scale and execution efficiency simultaneously, we explore distributed in-memory computing techniques and multivariate shapelets-based approaches to construct a Spark-based inmemory mining framework to parallelize the feature extraction process. We implement a framework with Multivariate Shapelets Detection (MSD) method as a based example. Through empirical evaluation on various kinds of sensory datasets, the scalability of the framework is evaluated in terms of efficiency while ensuring the same accuracy and reliability in early classification of multivariate time series. This work is the first one to realize multivariate time series early classification on Spark distributed in-memory computing platform, which can serve as a good base for an extension to other time series classification methods based on shapelet feature extraction.