STOG: A Traffic Prediction Scheme Based on Spatio-Temporal Optimized Graph Neural Networks

How to alleviate the traffic congestion and improve the traffic capacity of road networks through smart prediction has become a top priority for the realization of Intelligent Transportation Systems (ITS). It is necessary to capture the complex spatio-temporal correlation through the traffic data of road networks to achieve an accurate traffic prediction. In this paper, we propose a prediction method of spatio-temporal optimal graph neural network (STOG). It can obtain the spatio-temporal features of road networks through diffusion graph convolution (DGC) and recurrent neural network (RNN). We further leverage a new spatial attention mechanism to gain the aggregated features of the sampled nodes through pooling operations. It not only avoids excessive parameters, but also makes the model pays more attention to the sampled nodes, thereby reducing the prediction error. Through comparison with various baseline methods on METR-LA dataset, the results show that the proposed model can achieve higher prediction accuracy.