LISTS: Local Instance Selection for Time Series Classification

Deep time series classification models commonly aggregate temporal representations into a single global representation, which may obscure class-discriminative information contained in local regions. To address this limitation, we propose LISTS, a Local Instance Selection framework for Time Series classification. LISTS sequentially selects a small number of local temporal representations that can complement the global context and explicitly retains them alongside the global representation for classification. We formulate this selection process as a sequential decision-making problem and optimize the selector using reinforcement learning with the classification objective as its learning signal. We further provide a theoretical analysis that motivates the proposed global-local formulation. Experiments on all 30 UEA benchmark datasets demonstrate strong classification performance, with LISTS achieving 74.4% average accuracy and the best average rank among ten methods, while a controlled synthetic study verifies that LISTS identifies class-informative local regions.