Profile-matrix-based shapelet discovery for time series binary classification

Abstract

Due to the characteristics of interpretability and high accuracy, shapelet-based time series classification (S-TSC) has attracted considerable attention on data mining community over the past decades. An effective way for the S-TSC is shapelet transformation which produces a transformed dataset by the best k shapelets, where each of the k features in the dataset represent the distance between a time series and a shapelet. It can not only optimize the shapelet selection process, but also the transformed dataset can be combined with any traditional classifiers. However, the method requires evaluating a large amount of shapelets that will consume huge time. In this paper, we propose a matrix profile based method for shapelet discovery to handle time series binary classification. Firstly, this method calculates the distance from time series in one class to time series in another class. Secondly, it selects the three pairs of time series with the farthest distance, and finds the critical region corresponding to each set by the Matrix Profile, and extracts the shapelets from them. By this way, the shapelet candidates for evaluation can be significantly reduced. Experiments show that the method can improve the discovery speed of shapelets to a certain extent while maintaining high accuracy.