GTGCN: A gated temporal-graph coupling network model based on contrastive learning for traffic prediction

Accurate traffic prediction is crucial for intelligent transportation systems but remains challenging due to complex spatiotemporal dependencies and inherent limitations of Graph Neural Networks (GNNs), such as over-smoothing and poor adaptability to node heterogeneity. This paper proposes a Gated Temporal-Graph Coupling Network Model (GTGCN) to address these issues. GTGCN adopts a decoupled architecture: a node-wise Transformer encoder captures deep temporal dependencies, while a shallow graph convolutional network (GCN) with a gated fusion mechanism aggregates spatial information, fundamentally mitigating over-smoothing. To enhance robustness against node heterogeneity, we introduce a node-level contrastive learning auxiliary task via feature masking and edge dropping, encouraging discriminative node representations. Extensive experiments on three real-world datasets (PEMS04, PEMS08, METR-LA) demonstrate that GTGCN consistently outperforms ten baselines across short-term (15 min), medium-term (30 min), and long-term (60 min) predictions. Specifically, compared to AGCRN, GTGCN reduces MAE by 9.6%-17.9% on 60-min predictions. Ablation studies validate the contribution of each component, and further analysis shows that the gated fusion mechanism adaptively balances self- and neighbor information for heterogeneous nodes, while the decoupled design effectively suppresses over-smoothing. The proposed model offers a robust and efficient solution for traffic prediction.