Spatiotemporal feature augmentation serial-fast parallel learning for traffic prediction

Accurate traffic flow prediction is a key technology in Intelligent Transportation Systems (ITS) and is essential for planning future travel, improving urban traffic efficiency and supporting decision-making by traffic management authorities. However, the current mainstream spatio-temporal approaches usually ignore the interactive learning between spatio-temporal dependencies of road networks and fail to fully consider the different spatio-temporal patterns of each node due to the limitation of spatial heterogeneity. To address this problem, this paper proposes a Spatiotemporal Feature Augmentation Serial-Fast Parallel Learning (SFA-SFPL) traffic flow prediction model, which combines the intersection of Embedding Module, Feature Augmentation Layer, Fast Parallel Learning (FPL) and Serial Learning (SL) to achieve interactive learning of the model and to balance the sensory field and fitting ability of the model. Feature Augmentation Layer is used to enhance the information interaction between the channel and time dimensions; FPL achieves simultaneous and efficient mining of spatio-temporal features by significantly reducing the number of parameters. SL extracts spatial and temporal features in a serial manner to extend the spatial and temporal sensory field and avoid information coverage. Experiments on four real urban traffic datasets show that SFA-SFPL outperforms existing state-of-the-art methods in terms of prediction accuracy.