AdaKAN: Kolmogorov-Arnold Networks with Adaptive Spectral Decomposition for Time Series Forecasting

Yuhan Wu, Xiyu Meng, Zehao Liu, Yuhua Zhou, Junru Zhang, Pengfei Jiao, Canran Xiao, Yabo Dong, Dongming Lu

Keywords: Time Series Forecasting, Kolmogorov-Arnold Network, Adaptive Spectral Filter, Frequency Decomposition

Abstract: Real-world time series typically contain intertwined frequency components: low-frequency (trends), mid-frequency (periodicities), and high-frequency (short-term fluctuations or unexpected events), posing a significant challenge for accurate time series forecasting. To address this, we propose AdaKAN, a novel time–frequency Kolmogorov–Arnold Network equipped with an Adaptive Spectral Filter Module. Specifically: (i) AdaKAN adaptively decomposes time series into low-, mid-, and high-frequency components via learnable spectral thresholds. (ii) A dual-KAN structure is then employed: Fourier KAN captures global dependencies and periodic patterns, while Temporal KAN focuses on local structures and temporal dependencies. (iii) Each frequency band is processed with KANs of different orders to better capture frequency-specific dynamics. Finally, time and frequency features are fused to form a comprehensive representation. Extensive experiments on multiple benchmarks demonstrate that AdaKAN consistently outperforms existing SOTA methods, offering a superior balance of accuracy and efficiency as an extremely lightweight architecture.

Primary Area: learning on time series and dynamical systems

Submission Number: 4433