Traffic flow prediction via dynamic hypergraph learning

Abstract

In the field of intelligent transportation systems, efficiently and accurately predicting traffic flow and its evolution trends has become an important and urgent research task. Graph neural networks have been widely used in traffic flow prediction problems, but many methods ignore the high-order relationship of patterns between traffic flow and traffic nodes. To address the above issues, we propose a Transformer-based Hypergraph Convolutional Network (TSHGCN) for traffic flow prediction. Firstly, we adopt a hypergraph structure to more effectively capture the high-order nonlinear spatial correlations between traffic nodes. Then, an improved Transformer network is proposed, which accurately captured the global temporal features among various traffic nodes by combining the time distillation mechanism and the self-attention network. In addition, we integrate the above spatiotemporal modeling through efficient channel attention mechanism and multi-scale temporal information fusion mechanism, accurately extracting spatiotemporal features and achieving the final refined representation of traffic flow. Experiments on the California datasets (PeMSD4 and PeMSD8) with 5 independent random seed runs and strict statistical tests show that the TSHGCN model achieves the best performance on core metrics (MAE, RMSE, MAPE) under a unified experimental setting, and the performance improvement over state-of-the-art baselines is statistically significant.