Decoupled Traffic Spatial-Temporal Graph Neural Network for Traffic Flow Prediction

Abstract

Predicting traffic flow is a crucial component of the Intelligent Transportation System. However, the complicated spatiotemporal correlation of traffic road network nodes proposes challenges for traffic flow prediction. To improve modeling performance, we propose an innovative deep learning-based model named DTSTGNN. Decoupled Traffic Spatial-Temporal Graph Neural Network (DTSTGNN). In DTSTGNN, the original traffic signal is decoupled into an instantaneous fusion signal and a long-term dependent signal, which are captured by two well-designed modules, the instantaneous fusion module and the long-term dependency module. To capture the varying dependencies between nodes, we design an adaptive dynamic adjacency matrix in the instantaneous fusion module. Long-term dependencies are caught by introducing multi-head self-attention layers. The effectiveness of our model is demonstrated by extensive experiments on two real traffic datasets.