Scalable Classification of Univariate and Multivariate Time Series

Abstract

Time Series Classification (TSC) is important in many applications including IoT, medical, stock market analysis, economic forecasting, process and quality control and Big Data systems. The problem of time series classification has been studied separately for univariate (UTS) and multivariate (MTS) time series using different datasets and techniques. In this paper, we propose a unique, off-the-shelf approach to classifying time series that improves the current state-of-the-art accuracy for UTS and its generalization to MTS. Our technique maps each time series to a Gramian Angular Difference Field (GADF), interprets that as an image and uses Google’s pre-trained Convolutional Neural Network (trained on Inception v3) to map the GADF images into a 2048-dimensional vector space. Then, a multilayer perceptron (MLP) with three hidden layers, and a softmax activation function at the output is used to achieve the final classification. Our method yields competitive results for training and prediction in the UTS case while delivering superior results for MTS datasets. Unlike many published results, our technique is robust in the presence of variable length time series with missing data points, and scales well with the size of dataset.