A Novel Method for Whole Time Series Clustering via Sub-Pattern Recognition

Whole time-series clustering (WTSC) methods often fail to treat sub-patterns as independent entities, limiting their ability to capture diverse behavioral characteristics in time series data. To address this limitation, this study proposes a sliding window-based whole time series clustering method (SW-WTSC) that enhances clustering performance through sub-pattern recognition. The method extracts sub-samples using a sliding window over training data, applies normalization and feature selection, and clusters the resulting sub-samples via k-means. Cluster labels are then aggregated at the sample level, enabling secondary clustering based on the highest similarity among sub-patterns. The proposed approach was evaluated on 85 datasets from the UCR repository. Statistical analysis using the non-parametric Friedman test demonstrated that SW-WTSC significantly outperforms state-of-the-art WTSC methods. These results indicate that incorporating subsequence-based analysis can substantially improve whole time-series clustering accuracy.