RSAST: Sampling Shapelets for Interpretable Time Series Classification
Nicolas Rojas Varela, Michael Franklin Mbouopda, Engelbert Mephu Nguifo · HAL (Le Centre pour la Communication Scientifique Directe) · 2023
Shapelet-based techniques are widely utilized in time series classification due to their combination of interpretability and accuracy. However, these approaches often face scalability challenges compared to other state-of-the-art (SOTA) techniques, primarily due to the large search space of subsequences in datasets with numerous or large instances. To address this problem, we propose RSAST, a method based on shapelet techniques, specifically SAST and STC. The method employs a stratified approach and specific statistical criteria to reduce the subsequence search space of these techniques. As a result, RSAST significantly decreases computation time while preserving classification performance and interpretability. We evaluated the scalability of RSAST, demonstrating the improvements in computation time compared to its baseline and other SOTA methods. Furthermore, experiments conducted on 128 datasets from the UCR archive showed that RSAST achieves accuracy comparable to the baselines, along with competitive performance against other shapelet-based techniques.