Multipatch Augmentation Learning Based on Dual-Policy Model for Time Series Classification

Multiple instance learning (MIL)-based time series classification (TSC) predicts instance-level labels and aggregates them to derive bag-level labels; however, this approach encounters instability as anomalous instances significantly impact the bag-level label predictions. Although well-chosen series transformations and patch lengths improve instance discriminability and stability, determining appropriate configurations remains challenging. To address this, a dual-policy model is adopted to adaptively select a suitable transformation combination and patch length from a predefined transformation space according to the preference distribution of the training set. Building on this, a novel TSC algorithm, multipatch augmentation learning based on dual-policy model (MPAL-DPM), is proposed. MPAL-DPM is trained in two stages and comprises three functional modules. First, the policy model determines a transformation combination, encoded with Gray codes that assign adjacent values to combinations differing by a single transformation and thereby promote smooth policy optimization. This combination drives policy series augmentation, emphasizing subtle yet significant variations. Then, a scalable patch fine-tunes the policy-preset patch length to match the augmented data, alleviating the unstable feature extraction caused by uneven information density. Finally, multipatch classification (MPC) extracts features from each patch, weighs the importance of instances, and yields more consistent predictions. Experiments on the UCR archive demonstrate that MPAL-DPM achieves performance comparable to state-of-the-art classifiers, confirming its effectiveness, robustness, and scalability.