“In the realm of smart city development, a novel transfer learning approach has been introduced to address the scalability issue of city flow prediction in data-scarce areas. This method, leveraging spatiotemporal graph convolution, constructs a co-occurrence space to align data mappings between source and target domains, facilitating the transfer of a city flow prediction model. The technique, which captures concurrent spatiotemporal features and employs a Mahalanobis distance loss for feature alignment, has been proven superior to existing methods through evaluations on public bike flow datasets from Chicago, New York, and Washington in 2015.”