Cross-City Traffic Prediction via Self-Supervised Graph Learning

Traffic prediction plays a critical role in intelligent transportation systems. Recently, data-driven methods for traffic prediction have become increasingly popular and have outperformed traditional approaches in terms of effectiveness. However, the data scarcity problem that arises from inadequate sensing and data storage infrastructure in underdeveloped urban areas and dilapidated districts impedes the deployment of existing deep learning-based models, which typically require substantial training data. To address these issues, we propose a novel self-supervised graph learning-based cross-city learning framework (SSGL-CTP) for traffic prediction across cities. The framework first pretrains a transferable spatial feature extractor on a data-abundant source city and then adapts it to the target city through parameter transfer, without requiring direct cross-city data sharing. Specifically, the framework combines a dynamic trajectory-transfer graph and an origin-destination graph to model short-range and long-range spatial-temporal dependencies, and learns transferable representations through masked node recovery before target-city fine-tuning. The effectiveness of SSGL-CTP is validated through extensive experiments utilizing real-world taxi data from Chengdu and Wenzhou, China. These experiments demonstrate its superiority over contemporary baseline methods, with improvements of up to 9.2%.