AhaTrans: A Hierarchical Adaptive Transfer Learning Framework for Cross-City Traffic Flow Prediction

ICLR 2026 Conference Submission16381 Authors

Keywords: Urban Computing, Traffic Flow Prediction, Transfer Learning, Contrastive Learning

Abstract: Accurate prediction of urban traffic flow is essential for optimizing traffic management, enhancing urban planning, and promoting the development of smart cities. Due to the difficulty of data acquisition in many cities, data scarcity arises, significantly impeding the practical application of deep learning techniques. Consequently, researchers have turned to transfer learning for mitigating data scarcity through cross-city knowledge interaction. However, existing transfer learning methods lack precision and discrimination in spatio-temporal feature extraction, thereby restricting the predictive performance. Moreover, these approaches frequently fail to adequately account for the disparities between the source and target cities, resulting in the loss of essential knowledge and, at times, the introduction of detrimental knowledge into the target city. To overcome these challenges, we novelly introduce **A** **h**ierarchical **a**daptive **Trans**fer Learning Framework (**AhaTrans**), which ensures precise feature learning as well as effective, non-detrimental knowledge transfer in cross-city traffic flow prediction by focusing on three key levels: model architecture, feature representation, and data adaptation. Specifically, AhaTrans consists of the following three core modules: i) Guarded Transfer Experts Network (GTEN), which clearly distinguishes between shared and city-specific experts, enabling the target city to access beneficial knowledge from the source city while preventing harmful knowledge; ii) Spatial-Temporal Contrastive Embedding Module (STCE), which enhances the representation of spatio-temporal features through contrastive learning; iii) Transfer-Based Reweighting Module (TBR), which dynamically adjusts source city samples to extract knowledge most relevant for the target city's traffic patterns. Extensive experiments demonstrate that AhaTrans significantly outperforms existing methods, substantially improving the accuracy of traffic flow prediction while exhibiting excellent robustness and generalization capabilities.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 16381