Localized shapelets selection for interpretable time series classification

Shapelet-based methods have attracted increasing attention in time series research due to their good classification performance. Most existing approaches distinguish time series from different classes by evaluating the discriminative ability of all subsequeces according to the distance information, and the extracted shapelets might not always be real subsequeces of the original time series, which lead to expensive computation and poor interpretability. However, the information about where the shapelet is located in the time series is of importance to discriminate classes. In this paper, we propose a lo calized shapelets selection approach for interpretable time series classification. Specifically, a location measure and distance measure are defined to evaluate the discriminative ability of each shapelet candidate, and then the shapelet transformation process also integrates the location information of shapelets to provide a more interpretable insight in the classification result. Extensive experiments show that our proposed method is competitive on accuracy and efficiency compared with 18 baselines on UCR repository, and the location information effectively contributes to the improvement of shapelet interpretability.