Spatial-Temporal Dynamic Graph Convolutional Neural Network for Traffic Prediction

Abstract

Due to the complexity and dynamics of transportation systems, traffic prediction has become a challenging task. The accuracy of prediction is influenced by the spatial-temporal correlation within the traffic system. Previous approaches mainly relied on a pre-defined static adjacency matrix combined with graph convolutional neural networks to capture spatial correlation, neglecting the dynamic relationships between nodes over time. In this study, we propose a novel prediction model called the spatial-temporal dynamic graph convolutional neural network (STDGCN). By fusing node embeddings and input features, we obtain a new node representation that incorporates both static and dynamic features. To capture the dynamic relationships, we introduce a similarity calculation to construct a dynamic adjacency matrix. This matrix contains rich spatial relationships that serve as a reference for subsequent prediction tasks. We further employ Graph Convolutional Networks (GCN) and Gated Recurrent Units (GRU) to capture the spatial-temporal correlation. By combining these components, we establish a comprehensive traffic volume prediction model. To evaluate the performance of our proposed method, we conduct experiments on two real datasets. The experimental results demonstrate that our model achieves state-of-the-art performance in accurately predicting traffic volumes.