Time-domain Decomposition Adaptive Spatio-temporal Graph Neural Networks for Traffic Forecasting

Traffic prediction is an essential part of intelligent transportation systems, and accurate prediction is a daunting task because traffic systems are highly nonlinear and complex. In recent years, spatio-temporal graph neural networks have been widely employed to capture the intricate spatio-temporal patterns of traffic flows and have made outstanding breakthroughs. Most current spatio-temporal graph neural networks, however, are based on pre-defined graph adjacency matrices and are not able to effectively describe the intricate information of real-world road networks. In this paper, we introduce a Time-domain Decomposition Adaptive Spatio-temporal Graph Neural Network (TDAGNN) model that adaptively learns graph structures from both the macroscopic and the microscopic levels. Besides, we specialize in designing a powerful spatio-temporal graph convolutional model with the capability to capture traffic data's temporal and spatial relationships. Extensive experiments are performed to evaluate two real-world traffic datasets to show that our model surpasses the current algorithms with respect to prediction accuracy and generalization capability and achieves new breakthroughs and enhancements in traffic prediction missions.