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

A research team including Zepu Wang, Bowen Liao, and Jeff Ban has introduced FreST Loss, a breakthrough training objective that addresses fundamental limitations in spatio-temporal forecasting. Traditional models rely on point-wise objectives like Mean Squared Error (MSE) that fail to capture complex dependencies in graph-structured signals, while recent frequency-domain approaches like FreDF only mitigate temporal autocorrelation while overlooking spatial and cross spatio-temporal interactions. The new approach extends supervision to the joint spatio-temporal spectrum through Joint Fourier Transform (JFT), effectively decorrelating dependencies across both space and time dimensions.

The technical innovation lies in aligning model predictions with ground truth in a unified spectral domain, which theoretical analysis shows reduces estimation bias associated with time-domain training objectives. In extensive experiments across six real-world datasets, FreST Loss demonstrated consistent improvements over state-of-the-art baselines by better capturing holistic spatio-temporal dynamics. The model-agnostic nature means it can be applied to various forecasting architectures, potentially revolutionizing applications from traffic prediction and weather forecasting to financial market analysis and epidemiological modeling where spatial and temporal patterns interact complexly.