Exploring Idealized Regional Match for Cross-City Cross-Mode Traffic Flow Prediction

Guoliang Shi, Qiang Zhou, Jingjing Gu

Abstract: Traffic flow prediction often requires much training data for effective supervised training, which is a shackle for applying smart transportation management in many developing cities with deficient urban infrastructures. Although urban transfer learning has received some attention to date, there still remain unaddressed challenges, i.e., data scarcity of the target city and cross-city cross-mode data heterogeneity, which will cause the insufficient transferability of regional features. To alleviate theses challenges, we propose Cross-city Region Adaptive Prediction by Exploring Pattern Idealized Correlation named CrapEpic. Firstly, regarding the data scarcity, we investigate a data enhancement method that generates conjugate traffic flow data in both the target and source cities through flow pattern folding. Secondly, to mitigate the data heterogeneity issue, we design a multi-layer regional match function based on the shared knowledge to filter the transferable regional traffic flow. Finally, we propose a conjugate traffic flow transfer learning model to extract the shared knowledge from idealized regional match for cross-city cross-mode traffic flow prediction. Extensive experiments on real-world datasets of three cities demonstrate that CrapEpic outperforms the state-of-the-art baselines.