AdaTM: Fine-grained Urban Flow Inference with Adaptive Knowledge Transfer across Multiple Cities

Inferring the fine-grained urban traffic flows based on the coarse-grained traffic flow observations is practically important to many real applications for smart city. Existing approaches mostly rely on a large number of high quality urban flow data, but neglect the data sparsity issue which is common in real-world scenarios. Therefore, the performance of existing methods may not be promising towards cities that lack sufficient traffic flow data. How to design a more generalizable urban flow inference model that is able to effectively transfer knowledge across multiple cities is challenging and remains as an open research problem. In this paper, we propose a novel fine-grained urban flow inference model named AdaTM, which leverages the city-specific and city-invariant knowledge extracted from multiple cities. Specifically, we first propose a transformer-based urban feature extraction network named UBFormer to comprehensively extract the spatial-temporal features of multiple source cities. Then, we incorporate a learnable integrator to fuse the city-invariant and city-specific feature representations for the target city with sparse traffic flow data. Finally, we construct the feature representation of the target city through adaptive feature fusion and infer its fine-grained urban flows through the designed urban flow upsampler. Extensive experiments conducted on four large real-world datasets demonstrate the effectiveness of our approach.