Learning method to generate shapelets from time-series dataset

In time-series data mining, shapelet-based methods have been widely studied. A shapelet is a subsequence of time-series data and corresponds to a representative partial pattern of time-series datasets in a given class. The learning time-series shapelets (LTS) is a supervised learning method and achieves high accuracy time-series classification by learning shapelets and classifiers simultaneously. However, it needs to note that a shapelet generated by the LTS method does not represent a partial pattern of time-series data in a given class, but rather corresponds to a boundary that distinguishes classes effectively. To generate shapelets that represent time-series data belonging to a single class, it is necessary to develop an alternative method to the LTS method. In this paper, we proposed a learning method that is inspired by LTS method for generating shapelets from time-series data of a single class. To properly evaluate whether the proposed method can generate diverse shapelets that fit partial patterns of time-series data, we applied the proposed method to 18 open time-series datasets for training or testing. Our experiments demonstrated that the proposed method was able to generate shapelets from training time-series data belonging to each class; the shapes of generated shapelets were diverse and fitted partial patterns of testing time-series data.