Reproduce “FreDN: Spectral Disentanglement for Time Series Forecasting via Learnable Frequency Decomposition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhongde An, Jinhong You, Jiyanglin Li, Yiming Tang, Wen Li, Heming Du, Shouguo Du
FreDN introduces a frequency decomposition network for time series forecasting that effectively addresses spectral entanglement and the computational overhead of complex-valued learning. The model achieves superior forecasting accuracy, outperforming FreDF by 10% and FreTS by 14% on average MSE, while its ReIm Block reduces training time by up to 75%.
Key insights
- Adaptive spectral disentanglement directly in the frequency domain, through a learnable mechanism, is more effective than fixed filters or Top-K selection in separating complex time series components.
- Complex-valued neural network operations can be efficiently approximated by sharing real-valued MLPs between the real and imaginary parts of complex inputs, significantly reducing computational burden without sacrificing representational capacity.
- The gradient properties of frequency-domain MAE loss induce structured, long-range interactions in the time domain, which provides a theoretical explanation for its effectiveness in capturing global temporal patterns and aiding optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.