LYN: Leveraging Pyramid Structures and Aggregating Attention for General Time Series Classification Tasks

Time series classification has become increasingly prominent due to its wide-ranging applications in domains such as meteorology, finance, industry, and healthcare. Traditional methods have demonstrated effectiveness but often require extensive preprocessing and are limited in handling complex temporal patterns. The deep learning approach has improved end-to-end classification but still faces challenges in local feature integration and parameter tuning. Transformer have recently emerged as a powerful tool for time series analysis. However, existing Transformer-based models often struggle with integrating local and global features efficiently and require significant preprocessing. This paper introduces a novel architecture named LYN, designed to enhance time series classification by leveraging a pyramid structure and an innovative aggregating attention mechanism. The model addresses the limitations of existing approaches by integrating local and global sequence features, dynamically balancing fine details and global context through a biologically inspired attention mechanism. Transformer pyramid architecture generates multi-scale feature maps, enabling comprehensive extraction of temporal information across different resolutions. We evaluate the performance of the proposed model on multiple datasets. The results demonstrates superior performance of our model across various datasets, setting a new benchmark in time series classification.