An Enhanced Time Series Classification Using Linear-Regression Based Shape Descriptor

Abstract

Time series classification (TSC) has been used extensively with a wide range in many fields of real-world applications. Nearest Neighbor (NN) with Dynamic Time Warping (DTW) is regarded as one of the most effective and popular method used for a time series classification problem. Being a non-linear-alignment distance measure, DTW is known to function harmoniously with NN classifier as an effective tool in matching two time series sequences. However, despite the fact that DTW is specifically designed to discover an optimal alignment, it may not be able to achieve sensible local alignment. In this work, the shapeDTW framework was adopted in order to classify time series data based on the state-of-the-art HOG1D shape descriptor. Point-wise local structures were utilized in the alignment process; similarly-shaped structures are matched based on their levels of similarity. A generic alignment framework was provided by shapeDTW, which allows users to design the shape descriptors based on the data characteristics and domains. An enhanced shape descriptor representation called HOG1D-L was proposed based on two key concepts of the state-of-the-art HOG1D descriptor and linear regression. The proposed work on 84 UCR time series datasets was extensively tested and the results demonstrate that our approach can maintain or achieve better classification accuracy in most of the datasets.