Efficient class-specific shapelets learning for interpretable time series classification

Abstract

In the last decade, time series classification approaches based on time-independent shapelets, have received considerable attention due to their high prediction accuracy and intuitive interpretability . However, most existing shapelets discovery approaches find shapelets by evaluating the discriminatory power of all subsequences of the series, which is computationally expensive even with certain speed-up techniques. Even though some shapelet learning approaches learn the near-to-optimal shapelets from the training series rather than searching from numerous segments, they still have significant drawbacks in their performance regarding the accuracy, efficiency, and interpretability due to the numerous class-shared shapelets with fixed lengths. Thus, we propose a new shapelet learning approach that can learn as few as possible class-specific and variable-length shapelets. Extensive experiments demonstrate that our proposed method is competitive about classification accuracy over 18 baselines on 25 datasets, outperforms 2 orders of magnitude about efficiency, and is more interpretable than existing classifiers .