A two-phase filtering of discriminative shapelets learning for time series classification

Compared to the full-length methods for time series classification, shapelet-based methods acquire better interpretation, higher efficiency and precision since shapelets are discriminative features that well represent a time series. However, because of the large number of shapelets candidates, determining how to filter out shapelets with higher discriminability remains a challenge. In this paper, we propose a two-phase shapelets learning filtering framework for time series classification. Time series is first split into groups using the extreme key points, and local linear discriminant analysis with sparse group lasso regularizer is proposed to find projection vector. Then, a two-phase filtering framework is established to measure the sparsity of groups in order to quickly find the key group, where l2-norm is introduced in phase-1 and group sparsity degree is defined in phase-2 to filter sparse groups. Following that, only a few groups are used to extract shapelets and classify them, reducing the number of shapelets significantly. Finally, the group with the highest classification accuracy, i.e., the key group, is determined accurately. Extensive experiments on 28 time series datasets show that, when compared to other state-of-the-art shapelet-based classification methods, our proposed method achieves significant improvement and a competitive time cost.