Learnable Lengths and Shapes for Discriminative Time-series Subsequences

DOI

Bibliographic Information

Other Title
  • 長さと形を学習可能な判別波形パターン

Search this article

Description

In time-series classification, joint learning of discriminative subsequences called shapelets with the classifier has attracted much attention in recent years because of its high classification performance and interpretability. Previous research has proposed methods for learning only the shapes of shapelets through continuous optimization, but the shapelet lengths are fixed as hyperparameters and cannot be learned. In this study, we propose a method to jointly learn not only the shapes of shapelets but also their lengths through a continuous optimization problem. In particular, we theoretically show that the shapelet lengths can be learned to more clearly capture the differences between the classes while preserving interpretability. Experimental evaluations show improvements in AUC and total learning time on 71 UCR public time-series datasets, as well as the interpretability of the learned shapelets through case studies.

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top