Perceptual and position-aware shapelet adaptation network for time series classification

Abstract

Shapelets are time series subsequences that effectively distinguish time series classes. Recently, time series classifiers based on shapelets have gained interest from the community thanks to their high accuracy and interpretable results. However, these shapelet-based methods still have some shortcoming in calculating sub-distance between shapelet and time series and the difficulty of training the minimum distance function. In this dissertation, we firstly propose a novel linear-complexity time series distance measure, Perceptually and Position-aware Important Subsequence Distance (PISD), that efficiently assesses the similarity of two time series based on a subsequence distance of their important subsequences. In contrast to well-known local shape-based measures, our method utilizes perceptually important points to automatically extract the important subsequences from time series, and leverages the position-aware corresponding subsequences instead of the whole time series to calculate subsequence distance. After that, we propose the novel shapelet-based classifier, called Shapelet Adaptation Network (SAN), which utilize PISD for extracting shapelet and transforming the time series to feature space. Our SAN also use the simple linear regression to learn directly from distance space instead of using soft-minimum function and transform the input time series to adapt with the chosen shapelets rather than optimizing the shapelet.