Reproduce “FilterNet: Harnessing Frequency Filters for Time Series Forecasting” 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

Kun Yi, Jingru Fei, Qi Zhang, Hui He, Shufeng Hao, Defu Lian, Wei Fan

FilterNet introduces a novel architecture for time series forecasting that leverages learnable all-pass frequency filters as its fundamental building blocks. This approach achieves superior forecasting accuracy and computational efficiency over existing deep learning models across various real-world benchmarks, effectively utilizing the full frequency spectrum and robustly modeling time series components.

Key insights

  • Learnable all-pass frequency filters can directly replace traditional deep learning blocks (MLPs, Transformers) as fundamental components for robust and efficient time series forecasting.
  • Designing networks explicitly around frequency filtering allows for adaptive processing of the entire frequency spectrum, mitigating the vulnerability to high-frequency signals observed in some Transformer models.
  • The use of channel-shared filters for multivariate time series yields better performance, indicating that generalized temporal patterns can be learned and applied across different variables.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.