DTW-GBC Improves Noisy Time-Series Classification Efficiency.

Dynamic Time Warping (DTW)-based Nearest-Neighbor (NN) classifiers are effective for classifying time-series data due to their ability to handle varying speeds and alignments. However, they are vulnerable to mislabeled training data, which can significantly degrade performance. Additionally, their inference process is computationally intensive, requiring numerous DTW calculations. To address these limitations, the researchers introduce DTW-based Granular Ball Computing (DTW-GBC). This novel approach organizes temporally similar training samples into "granular balls," effectively grouping them. Classification is then performed at the granule level rather than comparing against every individual sample. The paper details two strategies for constructing these granular balls. Experiments on four benchmark datasets with symmetric label noise demonstrated that both DTW-GBC variants effectively mitigated performance degradation caused by noise. Crucially, they also required substantially fewer comparisons during inference compared to traditional DTW-based 1-NN, achieving a favorable balance between classification robustness and computational efficiency.