A compact Kolmogorov-Arnold network mixer for long-term time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Long-term time series forecasting (LTSF) is crucial for applications like energy management and weather prediction.
• Current LTSF models, often MLP- and Transformer-based, face challenges in multi-step-ahead accuracy.
• The need for more expressive nonlinear modeling cores in LTSF is recognized.

Purpose of the Study:

• To investigate Kolmogorov-Arnold Networks (KANs) as potential components for LTSF.
• To explore design choices that optimize KANs for LTSF tasks.
• To propose and evaluate KANMixer, a KAN-centered architecture for LTSF.

Main Methods:

• Developed KANMixer, a compact architecture featuring a multi-scale pooling frontend, KAN-based temporal mixing blocks, and KAN-based prediction heads.
• Evaluated KANMixer against LTSF baselines on seven standard benchmarks using a unified reproduction protocol.
• Conducted statistical tests, ablations, efficiency profiling, Gaussian-noise evaluation, and hyperparameter sensitivity analysis.

Main Results:

• KANMixer achieved competitive performance against representative LTSF baselines, particularly on ETT-style datasets.
• Dataset-dependent limitations were observed, indicating KANs are not universally superior.
• The practical utility of KANs was found to be contingent on basis-function selection, architectural integration, and computational resources.

Conclusions:

• KANs present a promising nonlinear modeling approach for LTSF.
• KANs are not plug-and-play solutions and require careful consideration of design and efficiency trade-offs.
• Further research should focus on optimizing KAN integration for robust and efficient LTSF.