Time series classification using a hybrid feature space generation method

Abstract Time series classification (TSC) faces two fundamental challenges: (1) the growing scale of data leads to exponential increases in computational cost for whole-series methods, and (2) dimensionality reduction or transformation techniques often sacrifice critical information, compromising classification accuracy. To address these issues, this paper proposes MDKF, a hybrid feature space generation method that enhances existing machine learning classifiers without introducing a new classifier architecture. MDKF constructs a comprehensive feature representation by combining local features-motifs (repeating patterns) and discords (anomalous subsequences)–extracted from each time series, with global distance-based features derived from K-nearest neighbor statistics. This hybrid feature space preserves both local structural information and global distribution information while achieving effective dimensionality reduction. Experimental results on benchmark datasets demonstrate that MDKF significantly improves classification performance across multiple metrics (accuracy, precision, recall, and F1-score) for various machine learning classifiers, achieving competitive results compared to state-of-the-art TSC methods with substantially lower computational overhead.