Efficient Time Series Classification Based on Learning Similar Trend Features in the Same Class Sequences
Guohui Ding, Zhaoyi Yuan, Wenjing Tang, Chao Jiang · 2023
Time series shapelets have been widely studied for their optimal interpretability and accuracy in solving time series classification (TSC) tasks. This study determine that similar trends of the same type of time series can be used to reduce the size of the candidate solution set and improve the efficiency of shapelets discovery. However, only a few studies have been conducted on this. Therefore, this study proposes a shapelets discovery algorithm that learns the common trend of the same type of time series to reconstruct the sequence. First, by constructing an autoencoder to fit time series of the same type with similar trends, we reconstruct a sequence that integrates the features of this type, thereby reducing the size of the candidate set. Thereafter, to compress the regenerated sequence and reduce its dimensionality, it is represented by a Symbolic Aggregate approXimation. Furthermore, filtering out impossible and low-quality candidate solutions further reduces the size of the candidate solution set. Finally, this study proposes a candidate solution evaluation method based on the Matrix Profile structure to find the discriminative shapelets that can best represent each type of time series and uses them to construct a classifier to complete the classification task. Extensive experiments demonstrate that the proposed algorithm effectively reduces the candidate solution set size and has advantages in efficiency, accuracy, and interpretability compared to traditional methods.