AutoShapelet: Reconstructable Time Series Shapelets

Abstract

Time series shapelets is a snippets of time series that can distinguish one class from others. In the last decade, many researches show that time series shapelets is not only one of the most promising classification techniques, but also a desirable solution because it is simply an explainable result to the experts. However, Two main drawbacks of time series shapelets discovery are speed and the appearance of the candidates and its representative, i.e. the time series shapelets itself. In this paper, we do not improve the running time of discovering the time series shapelets, but we propose a new method to learn the shape of time series shapelets, instead of picking one from candidates. The number of candidates can be vary from ten thousands to millions subsequences or even more depended on the length of the candidates. In this paper, autoencoder technique is applied to reduce the complexity of candidates from the higher-dimensional space to the much smaller-dimensional space, to highlight the potential candidates as the representatives, to learn the shapes of those candidates instead of the individual one, and to reconstruct the more-smooth time series shapelets. Our time series shapelets, named autoshaplets, is not fit to the exact value of the training data anymore, which normally is noisy according to the real observation. The experiment results demonstrate that the new generated shapelets can achieve higher accuracy compared to the exact shapelets, and it is less sensitive to the training data.