Transfer learning with spatiotemporal graph convolution networks improves city flow prediction

A new transfer learning method, TL-STGCN, improves city flow prediction in data-scarce areas by leveraging data from other cities. The method uses a spatiotemporal graph convolution network to capture shared features across city road networks and human travel habits, then aligns these features in a co-occurrence space to transfer knowledge effectively. Tested on bike flow data from Chicago, New York, and Washington, TL-STGCN outperformed ten baseline methods across various transfer scenarios, demonstrating its effectiveness.