A unified traffic flow prediction model considering node differences, spatio-temporal features, and local-global dynamics

Abstract

Traffic flow prediction is one of the core technologies in Intelligent Transportation Systems (ITS) and has extensive application value. The primary challenge lies in efficiently modeling the complex spatio-temporal dependencies within traffic data. Although spatio-temporal graph neural network models are regarded as effective solutions, their performance is limited by incomplete graph connectivity and the use of identical modeling approaches for all nodes, which not only hinders the learning of dynamic traffic patterns but also overlooks the heterogeneity between nodes. To address these limitations, a novel traffic flow prediction model based on dynamic spatio-temporal modeling with node differences is proposed. Specifically, an exogenous node selection module is designed to identify nodes highly correlated with the endogenous node (i.e., the node to be predicted) to assist in prediction. Subsequently, differentiated modeling approaches are employed: the endogenous node is represented using local–global embedding to capture its local–global features. In contrast, exogenous nodes are modeled using global embedding to obtain their global representations, thereby achieving comprehensive feature characterization. Finally, a spatio-temporal attention network is utilized to capture the spatio-temporal interactions among nodes. Extensive experiments on three real-world traffic datasets demonstrate that the proposed model achieves significant performance improvements over state-of-the-art baseline methods. The experimental results reveal that the proposed framework not only achieves superior predictive accuracy but also maintains highly competitive computational efficiency.