• Proposes a transfer learning framework based on spatiotemporal graph convolution to overcome data scarcity in city flow prediction.
• Introduces a co-occurrence space to align source and target domain features, enabling cross-city model transfer.
• Designs a dynamic spatiotemporal graph convolution module with a temporal encoder to capture concurrent spatiotemporal features.
• Achieves state-of-the-art performance on public bike flow datasets, demonstrating significant improvements over existing methods.