City pulse aware graph neural network for joint volume and speed estimation in urban traffic

Accurate perception of traffic states is crucial for intelligent urban transportation systems. However, in reality, due to the high cost of detector deployment, our observation is often limited to only a portion of road segments. To acquire the dynamics of all road segments, we propose a novel model named City Pulse Aware Graph Neural Network (CPA-Net), which estimates traffic volume and speed in urban transport networks. Multi-head self-attention layers are used to enhance temporal feature extraction. We introduce dynamic correlation graph and functional similarity graph to adaptively capture spatial correlations, and then fuse different graphs through a gated diffusion convolutional layer. Additionally, we design a tailored external feature extraction module to efficiently integrate various external data into the model. We develop a multi-task learning architecture to achieve joint estimation of traffic volume and speed. Experimental results on real-world urban traffic datasets demonstrate that CPA-Net achieves better performance than baseline models. Furthermore, extensive tests under different missing rates and time windows are performed to reveal the model’s strong robustness. Finally, we identify the significant effects of each module through ablation analysis.