Attention-Based Local-Global Hypergraph Convolutional Network for Traffic Flow Forecasting

Traffic flow forecasting is of significance for traffic control and management. In recent years, researchers have widely utilized data-driven spatio-temporal graph neural networks (GNNs) to capture the complex spatio-temporal characteristics of traffic flow and achieve accurate prediction outcomes. However, GNNs fail to capture non-paired relationships, and previous studies struggle to encompass both local and global dependencies. To address these limitations, we introduce a novel attention-based local-global hypergraph convolutional network (ALGHGCN) for traffic flow forecasting. Specifically, to handle spatial dependencies, the proposed model incorporates well-designed spatial convolution blocks that integrate the attention mechanism. These blocks capture non-paired dependencies using local hypergraphs constructed from road networks as well as global hypergraphs created from historical time series. Additionally, a temporal attention block is employed to exploit relationships for temporal dependencies. Comparative analysis against multiple baselines using real-world data sets demonstrates the superior performance of the proposed ALGHGCN model.