English

Decorrelating the Future: Joint Frequency Domain Learning for Spatio-temporal Forecasting

Machine Learning 2026-03-06 v1 Artificial Intelligence Machine Learning

Abstract

Standard direct forecasting models typically rely on point-wise objectives such as Mean Squared Error, which fail to capture the complex spatio-temporal dependencies inherent in graph-structured signals. While recent frequency-domain approaches such as FreDF mitigate temporal autocorrelation, they often overlook spatial and cross spatio-temporal interactions. To address this limitation, we propose FreST Loss, a frequency-enhanced spatio-temporal training objective that extends supervision to the joint spatio-temporal spectrum. By leveraging the Joint Fourier Transform (JFT), FreST Loss aligns model predictions with ground truth in a unified spectral domain, effectively decorrelating complex dependencies across both space and time. Theoretical analysis shows that this formulation reduces estimation bias associated with time-domain training objectives. Extensive experiments on six real-world datasets demonstrate that FreST Loss is model-agnostic and consistently improves state-of-the-art baselines by better capturing holistic spatio-temporal dynamics.

Keywords

Cite

@article{arxiv.2603.04418,
  title  = {Decorrelating the Future: Joint Frequency Domain Learning for Spatio-temporal Forecasting},
  author = {Zepu Wang and Bowen Liao and Jeff and Ban},
  journal= {arXiv preprint arXiv:2603.04418},
  year   = {2026}
}
R2 v1 2026-07-01T11:03:38.950Z