Please join us for a UW Data Science Seminar featuring UW iUTS lab PhD student Zepu Wang on Thursday, March 12th from 4:30 to 5:20 p.m. PT. The seminar will be held in IEB G109.

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

Abstract: Standard Direct Forecast models rely on point-wise objectives (e.g., Mean Squared Error) that overlook the complex spatio-temporal dependencies inherent in graph signals. While recent frequency-domain methods like FreDF mitigate temporal autocorrelation, they neglect spatial and cross-spatio-temporal structures. To address this, we propose FreST Loss (Frequency-enhanced Spatio-Temporal Loss), which extends supervision to the joint spatio-temporal spectrum. By leveraging the Joint Spatio-temporal Fourier Transform, FreST Loss aligns predictions with ground truth in a unified spectral domain, effectively decorrelating complex dependencies. Theoretical analysis confirms that this approach reduces the estimation bias in time-domain training objectives. Extensive experiments on six real-world datasets demonstrate that FreST Loss is model-agnostic, consistently improving state-of-the-art baselines by capturing holistic spatio-temporal dynamics.

Speaker Bio: Zepu Wang is a second-year PhD student at the iUTS lab, led by Prof. Xuegang (Jeff) Ban. He is an interdisciplinary AI researcher focusing on deep learning in time series and spatio-temporal data analysis, uncertainty quantification, and their applications in transportation and urban systems.