Continuous spatio-temporal hybrid graph convolutional network for air pollutant prediction

Accurate prediction of air pollution offers significant benefits for safeguarding public health and enhancing urban safety management. Changes in air pollutant concentrations within urban networks are influenced by spatio-temporal interactions among cities. However, existing modeling approaches fail to comprehensively capture these intricate synergistic relationships, particularly the higher-order synergies and continuous processes. This study proposes a Continuous Spatio-temporal Hybrid Graph Convolutional Network (CSHGCN) to address the existing gaps. First, we introduce a hypergraph generation technique to explicitly identify and model higher-order synergistic relationships of air pollutants among multiple cities. Then, a hybrid graph convolutional module is developed to model multi-city interaction relationships by integrating both distance graph and hypergraph. Finally, we embed the hybrid graph structure into Neural Ordinary Differential Equations, enabling continuous spatio-temporal modeling via solving equations. Extensive experiments on four real-world datasets demonstrate that the CSHGCN significantly outperforms baseline models, achieving improvements in Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE), improved by 2.26%, 3.66%, and 2.66% on average compared to the optimal baseline model.