DTW-GBC: Robust Time-Series Classification with Granular Balls

A novel approach known as DTW-based Granular Ball Computing (DTW-GBC) has been introduced to enhance the efficiency and robustness of time-series classification utilizing Dynamic Time Warping (DTW). While traditional DTW-based Nearest-Neighbor (NN) classifiers perform well, they are susceptible to errors from mislabeled training data and necessitate numerous DTW calculations during inference. DTW-GBC addresses this by grouping temporally similar training samples into granular balls and classifying at the granule level, which decreases the number of necessary comparisons. Two strategies for constructing granular balls were created for DTW-GBC. Tests on four benchmark datasets with symmetric label noise indicated that both variants of DTW-GBC effectively reduce performance decline from label noise and require significantly fewer comparisons than DTW-based 1-NN during inference. The results imply that DTW-GBC strikes a beneficial balance between classification robustness and inference efficiency. This research is accessible on arXiv with the identifier 2608.11704 in the Computer Science > Machine Learning category.