Class-Incremental Learning (CIL) for time series data addresses the challenge of incrementally
learning new classes from sequential data streams without forgetting previously acquired
knowledge. In practical scenarios, labeling all incoming data is costly, necessitating
efficient strategies for selective labeling. To address this, we introduce a novel
framework integrating Active Learning (AL) into CIL specifically for time series classification
tasks, enhancing label efficiency under budget constraints. Our proposed method employs
a dual-memory architecture comprising an active short-term memory for temporary storage
of newly labeled samples and a consolidated long-term memory utilizing a replay mechanism
based on Adversarial Shapley value Experience Replay (ASER) guided by approximate
Shapley values. At each incremental learning step, our active learning module strategically
selects informative and diverse samples for labeling, thereby optimizing the annotation
budget. Experiments conducted on benchmark datasets demonstrate significant improvements
in accuracy and label efficiency compared to previous methods.
