Traffic Flow Forecasting with Spatial-Temporal Graph Diffusion Network

Accurate forecasting of citywide traffic flow has been playing critical role in a variety spatial-temporal mining applications, such as intelligent control and public risk assessment. While previous work made significant efforts to learn temporal dynamics spatial dependencies, two key limitations exist current models. First, only the neighboring correlations among adjacent regions are considered most existing methods, global interregion dependency is ignored. Additionally, these methods fail encode complex transition regularities exhibited with time-dependent multi-resolution nature. To tackle challenges, we develop new prediction framework–Spatial-Temporal Graph Diffusion Network (ST-GDN). In particular, ST-GDN hierarchically structured graph neural architecture which learns not local region-wise geographical but also semantics from perspective. Furthermore, multi-scale attention network developed empower capability capturing multi-level dynamics. Experiments on four real-life datasets demonstrate that outperforms different types state-of-the-art baselines.