Attention Based Multi-scale Spatial-temporal Fusion Propagation Graph Network for Traffic Flow Prediction

Timely and accurate traffic flow prediction holds significant value for public commuting and urban traffic management. However, the independent temporal and spatial components in recent methods struggle to fully capture the underlying spatial-temporal relationships and multi-scale temporal features. To overcome these limitations, we propose a novel attention based multi-scale spatial-temporal fusion propagation graph network to effectively model the spatial-temporal relationships in traffic data. Specifically, we design a multi-scale spatial-temporal mix-hop graph convolution module and a spatial-temporal attention-mechanism to explore the spatial-temporal relationships in traffic data through a cohesive modeling approach. Extensive experiments conducted on four real-world traffic flow datasets demonstrate that our model significantly outperforms baseline methods.