GODM: Graph Orthogonal Diffusion Models for Spatio-Temporal Forecasting

Abstract

Probabilistic spatio-temporal forecasting is a pivotal challenge in the data mining community. Recently, diffusion models have emerged as a powerful paradigm in this field, acting as conditional generative models. Beyond standard diffusion models, a significant research trend involves tailoring the diffusion model to incorporate domain-specific inductive biases. However, for complex spatio-temporal data, the high-dimensional coupling between spatial and temporal domains makes the design of such models exceptionally difficult. In the present work, we propose a novel framework (GODM) that adopts orthogonalization to decouple intricate temporal dynamics while employing spatial convolutions to simulate information flow across space. By embedding the spatio-temporal structure directly into the forward process, the GODM introduces crucial inductive biases with negligible computational overhead. Extensive experiments on multiple benchmark datasets demonstrate that the GODM framework consistently outperforms state-of-the-art spatio-temporal diffusion models.