Dynamic Spatial-Temporal Dual Graph Neural Networks for Urban Traffic Prediction

Accurate traffic prediction is a crucial aspect of intelligent transportation systems. However, existing methods typically rely on static graphs to learn correlations between different sensor in space, which ignores dynamic impact of latent factors on topology of the road network. To address this issue, we propose a traffic flow prediction method based on a dynamic spatial-temporal dual graph neural network that extracts deeper and finer-grained features from traffic data. Firstly, we propose a new data-driven strategy based on a dynamic spatial-temporal-aware graph to replace the commonly used predefined static graph in traditional graph convolutional networks. This strategy enables us to collect edge attributes (geographical proximity) and node attributes (spatial heterogeneity) between nodes. Secondly, we introduce the duality principle to construct the dual hypergraph of the traffic graph, which captures the correlations between edges of the traffic graph. In the process of dynamic graph convolutional iteration, we capture the dependencies between dynamic edge attributes and static node attributes on the basis of merging spatial relationships. Finally, an improved multi-head attention mechanism designed to represent dynamic spatial correlations. We conducted experiments on two real-world traffic prediction tasks, results demonstrate our method outperforms others.