Disentangled diffusion models for probabilistic spatio-temporal traffic forecasting

With the widespread deployment of traffic sensors and the continuous advancement of sensor technology, accurate traffic prediction has become increasingly feasible. However, existing traffic prediction methods face significant challenges in addressing the inherent uncertainty present in traffic data. These methods typically rely on deterministic models, which struggle to capture and represent the random fluctuations and sudden changes within traffic data, limiting the reliability and accuracy of their predictions. To overcome this issue, we introduce a non-autoregressive diffusion model specifically designed for traffic prediction. Our model aims to capture the inherent uncertainty and complex spatio-temporal dynamics of traffic data by quantifying uncertainty through a diffusion model and learning spatio-temporal dependencies via a spatio-temporal graph neural network. Additionally, to better capture the relationship between frequency components and uncertainty, we employ discrete wavelet decomposition. This technique separates the high-frequency and low-frequency information, which are processed independently within a dual-channel spatio-temporal network. By decoupling these components, we can apply targeted denoising strategies to each frequency band during the diffusion process. This enables more accurate modeling of uncertainty and improved predictions. We evaluate our model on multiple real-world traffic datasets, and the experimental results demonstrate that it outperforms state-of-the-art probabilistic models.