Selective Cross-City Transfer Model with Dual-Attention Graph Feature Extraction

Cross-city transfer learning relies on auxiliary city information (e.g., POI) to compute node features, which are then used for selective knowledge transfer. The effectiveness of this process depends on the quality of node features, but existing GNNbased methods fail to fully capture auxiliary information, leading to suboptimal feature representations. To address this, we propose DAGTrans, a Selective Cross-City Transfer Model with DualAttention Graph Feature Extraction. First, we decouple node embeddings using Node Attention (NA) and Edge Attention (EA) to separate feature representations from interference signals. Second, Dual-Channel Graph Convolution Network (DC-GCN) extracts fine-grained semantic information from the node feature graph, capturing both local and global perspectives. Third, inspired by counterfactual causal intervention, we introduce random perturbations to the interference graph and integrate it with the node feature graph, minimizing distributional shifts to improve node embedding learning. DAGTrans enhances feature expressiveness and better leverages auxiliary city information, achieving $0.89 \%$ improvement over state-of-the-art methods in cross-city traffic prediction.