MC-HSTA: A multi-source cross-domain hybrid spatio-temporal attention network for traffic flow prediction

Conventional representation learning methods have achieved remarkable performance in traffic flow forecasting when data is sufficient, while they struggle in scenarios with limited data. Existing works typically utilize transfer learning as a promising solution in scenarios with insufficient traffic data, which involves two-staged training: pre-training and fine-tuning. Nonetheless, neglecting inter-domain dependencies in multi-source domain scenarios and the presence of domain shift during the inference stage are two distinct challenges. Inter-domain dependencies refer to the complex interactions among multiple source domains that are often overlooked by traditional models, and domain shift refers to distributional differences between the source and target domains that impede effective adaptation. In the face of these challenges, we propose a novel multi-source cross-domain model incorporating a hybrid spatio-temporal attention mechanism and a domain adaptation module. The attention mechanism is designed to explicitly capture temporal and spatial dependencies, enabling effective modeling of inter-domain interactions across multiple source domains. Simultaneously, the domain adaptation module employs a domain adaptive classifier to align the feature distributions between source and target domains, effectively mitigating domain shift and enhancing cross-domain generalization. Extensive experiments on three public datasets demonstrate that our method consistently outperforms popular baseline models in traffic flow prediction. Concretely, it achieves better MAE, RMSE, and MAPE across the test periods, with an average reduction of 33.55 %, 25.74 %, and 38.42 %, respectively. This highlights the importance of addressing both inter-domain dependencies and domain shift in transfer learning with limited traffic flow data.