DQeND Improves Early Time Series Classification in Dynamic Environments

Early Classification of Time Series (ECTS) demands accurate and timely decisions in environments that are inherently online and constantly changing. However, most existing ECTS methods assume data stationarity and rely on separable designs, where classification and triggering mechanisms are optimized independently. This fundamental limitation hinders their adaptability when data drift occurs. This research challenges the conventional separable paradigm by studying ECTS under non-stationary conditions. It provides the first systematic comparison between separable and end-to-end approaches across various controlled drifting scenarios. Introducing DQeND, a unified architecture built on Reinforcement Learning, the framework jointly learns data representation, classification, and the decision to trigger an early classification. DQeND consistently outperforms state-of-the-art separable baselines across a wide range of drift scenarios, demonstrating superior robustness. An ablation study confirms that jointly updating all modules is critical to these performance gains, suggesting that end-to-end learning offers improved adaptation for ECTS in dynamic environments.