Domain adversarial graph neural network with cross-city graph structure learning for traffic prediction

Deep learning models have emerged as a promising way for traffic prediction. However, the requirement for large amounts of training data remains a significant issue for achieving well-performing models. Data scarcity in real-world scenarios, caused by costly collection or privacy policies, can severely impede the performance of existing deep learning models. Transfer learning aims to leverage knowledge learned from data-sufficient cities to improve prediction performance in data-scarce cities. Unfortunately, most existing methods solely focus on transferring knowledge at the city level, neglecting fine-grained node-level correlations and distribution discrepancies between cities. In this paper, we propose DAGN, a domain adversarial graph neural network that mines inter-city spatial–temporal correlations and alleviates domain distribution discrepancies to address the data scarcity problem in traffic prediction. Specifically, DAGN comprises three key modules: (1) A cross-city graph structure learning module is developed to capture node-pair adjacent relationships across cities, enabling the dynamic aggregation of inter-city spatial–temporal information. Additionally, a graph reconstruction loss is proposed to enforce structural consistency between the learned and priori graphs. (2) A domain adversarial strategy is integrated with a spatial–temporal module, which jointly extracts domain-invariant spatial and temporal features to reduce the distribution discrepancies between cities. (3) To adaptively extract transferable knowledge from a global perspective, a global spatial–temporal attention module is designed. Extensive experiments on six traffic flow and traffic speed prediction benchmarks demonstrate that DAGN consistently outperforms state-of-the-art methods.