HyGL-STNet: hybrid graph learning with parallel decoupling for spatio-temporal forecasting

Abstract

With the proliferation of the Internet of Things (IoT) and wireless sensor networks (WSNs) in modern cities, massive amounts of real-time traffic monitoring data are continuously collected from distributed sensors. Accurate spatio-temporal forecasting on graph-structured data remains challenging due to the coexistence of localized interactions, long-range dependencies, and multi-scale temporal dynamics, requiring advanced machine learning and data mining techniques. In traffic flow prediction, most graph neural network (GNN) approaches rely heavily on predefined physical topologies, which limits their ability to capture latent semantic correlations among distant nodes and often leads to interference between short-term fluctuations and long-term trends. To address these issues, we propose HyGL-STNet, a hybrid graph learning framework with a parallel decoupling architecture. The proposed model consists of a Local-Physical Route that models neighborhood interactions based on physical connectivity and a Global-Semantic Route that discovers latent cross-regional dependencies through structure-enhanced attention. In addition, a self-supervised pre-trained history encoder is introduced to extract long-term temporal representations without information leakage. Extensive experiments on the benchmark PeMS04 and PeMS08 datasets demonstrate that HyGL-STNet consistently outperforms competitive baseline methods. In particular, the proposed model achieves an average MAE improvement of approximately 3.6% on PeMS04 and reduces prediction errors by up to 5.5% in the 60-minute long-term forecasting horizon on PeMS08. These results indicate that the proposed parallel decoupling paradigm effectively captures complex spatio-temporal dependencies and provides a robust framework for graph-based forecasting tasks.