Cross-city transfer learning for traffic forecasting via incremental distribution rectification

High-quality traffic forecasting is critical to facilitating urban intelligent transformation. In deep learning enabled urban intelligence era, accurate traffic forecasting requires large-scale data, obtaining which is not always feasible because of limitations posed by scarce equipment resources in cross-city scenarios. To address this problem, existing methods propose to learn transferable meta-knowledge from rich source city data so as to guide the forecasting on the target city. However, they mainly focus on obtaining a source distribution- centric or source and target shared knowledge, rather than target distribution-centric transferable knowledge, which results in unavoidable disturbances due to migration noise. In this case, we propose a cross-city transfer learning method based on Incremental Distribution Rectification from the perspectives of distribution discrepancy quantification and calibration, called Cross-IDR. Specifically, we leverage the Koopman enabled Optimal Transport method to measure the transfer process between distributions, thus modifying the source distribution-centric meta-knowledge to be target distribution-centric. In addition, a Spatio-Temporal Interaction alignment method is proposed to enhance the mining of cross-city interactions between spatial and temporal information. We verify the effectiveness of Cross-IDR on real-world traffic datasets, and the results demonstrate that it outperforms state-of-the-art models.