Batch active learning for time-series classification with multi-mode exploration

Collecting a sufficient amount of labeled data is challenging in practice. To deal with this challenge, active learning, which selects informative instances for annotation, has been studied. However, for time series, the dataset quality is often quite poor, and its multi-modality makes it unsuited to conventional active learning methods. Existing time series active learning methods have limitations, such as redundancy among selected instances, unrealistic assumptions on datasets, and inefficient calculations. We propose a batch active learning method for time series (BALT), which efficiently selects a batch of informative samples. BALT performs efficient clustering and picks one instance with the maximum informativeness score from each cluster. Using this score, we consider in-batch diversity explicitly so as to effectively handle multi-modality by exploring unknown regions, even under an extreme lack of labeled data. We also apply an adaptive weighting strategy to emphasize exploration in the early stage of the algorithm but shift to exploitation as the algorithm proceeds. Through experiments on several time-series datasets under various scenarios, we demonstrate the efficacy of BALT in achieving superior classification performance with less computation time under a predetermined budget, compared to existing time-series active learning methods.