TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting

• DOI: 10.48550/arxiv.2502.06910
• Publication Date: 2025-02-09
• Authors: Songtao Huang; Zhen Zhao; Can Li; Lei Bai
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Machine Learning (cs.LG)

Real-world time series often have multiple frequency components that are intertwined with each other, making accurate forecasting challenging. Decomposing the mixed into single is a natural choice. However, information density of patterns varies across different frequencies, and employing uniform modeling approach for can lead to inaccurate characterization. To address this challenges, inspired by flexibility recent Kolmogorov-Arnold Network (KAN), we propose KAN-based Frequency Decomposition Learning architecture (TimeKAN) complex challenges caused mixtures. Specifically, TimeKAN mainly consists three components: Cascaded (CFD) blocks, Multi-order KAN Representation (M-KAN) blocks Mixing blocks. CFD adopt bottom-up cascading obtain representations band. Benefiting from high KAN, design novel M-KAN block learn represent specific temporal within Finally, used recombine bands original format. Extensive experimental results real-world datasets demonstrate achieves state-of-the-art performance as an extremely lightweight architecture. Code available at https://github.com/huangst21/TimeKAN.