Learning-based shapelets discovery by feature selection for time series classification

Shapelet-based methods have attracted widespread attention over the past decade in the time series classification for their benefits of high classification accuracy and good interpretability. The primary challenge of shapelet-based methods is to find discriminative shapelets that best distinguish different classes. Although the existing shapelet-based methods have achieved encouraging results, the number of obtained shapelets is still so large that the shapelets are not discriminative enough for the classification, and some of these shapelets are of less interpretability to reveal the most important patterns of different classes. In this paper, we propose a novel learning-based shapelets discovery method by feature selection (LSDF) for time series classification, of which the significance is transforming the shapelets discovery task into an optimization problem and efficiently learning interpretable shapelets. The value at each time point is regarded as a feature and Marginal Fisher Analysis (MFA) is combined with fused lasso to obtain the discriminative features (shapelets). The experimental results on real-world datasets from University of California, Riverside (UCR) repository show that LSDF outperforms the state-of-the-art shapelet-based and non-shapelet-based methods in classification accuracy and running time. Finally, the interpretability of shapelets learned by our methods is shown from different perspectives.