Decoupled Diffusion Models for Efficient Spatio-Temporal Graph Forecasting

Juyeob Lee, Seongyeon Yang, Eunil Park

Keywords: spatio-temporal graph forecasting, probabilistic forecasting, diffusion model, decoupled graph neural network

Abstract: Graph-based diffusion models suffer from a critical computational bottleneck, limiting their use in practical applications such as spatio-temporal graph forecasting. We argue that this inefficiency stems from the fusion of information propagation and feature transformation within standard GNNs. In this paper, we introduce a design principle that decouples these two operations, enabling a highly efficient and linear architecture. Instantiating this principle, Decoupled Spatio-Temporal Diffusion Model (DSTD) leverages the principle alongside a dynamic multi-scale aggregation mechanism to achieve remarkable performance. On widely-used spatio-temporal graph forecasting benchmarks, DSTD not only outperforms existing probabilistic methods but also surpasses top-performing deterministic models, while demonstrating a significant reduction in inference time. Our results validate that decoupling is a powerful and effective strategy for building scalable and high-performing generative models for graph-structured data.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 3291