From Retrieval to Recognition: Knowledge-Enhanced Foundation Models for Time Series Classification

Time series foundation models (TSFMs) have achieved notable progress in classification tasks through large-scale pretraining. However, existing methods rely on implicit parameter fine-tuning, where task-relevant information is entangled with abundant irrelevant knowledge in vast parameter spaces. This may hinder TSFMs from precisely activating useful knowledge for downstream classification. To this end, we propose REKIN, a plug-and-play knowledge-enhanced framework for time series classification built upon pretrained TSFMs within a retrieval-to-recognition paradigm. Unlike traditional parameter-only approaches, REKIN introduces a decoupled query encoder to extract discriminative patterns and retrieve instance-level numerical knowledge from a constructed knowledge base. The retrieved explicit knowledge is then fused with the TSFM’s implicit parametric embeddings, enabling joint knowledge transfer at both the parameter and feature levels. In this way, REKIN allows TSFMs to recall task-relevant patterns learned during pretraining, improving the efficiency of knowledge utilization in downstream transfer. Experiments on 158 time series datasets show that REKIN consistently improves the classification performance of state-of-the-art TSFMs. Ablation and knowledge-enhanced studies further validate the effectiveness of the proposed framework.