A Dynamic Regional-Aggregation-Based Heterogeneous Graph Neural Network for Traffic Prediction

Traffic flow prediction, crucial for intelligent transportation systems, has seen advancements with graph neural networks (GNNs), yet existing methods often fail to distinguish between the importance of different intersections. These methods usually model all intersections uniformly, overlooking significant differences in traffic flow characteristics and influence ranges between ordinary and important nodes. To tackle this, this study introduces a dynamic regional-aggregation-based heterogeneous graph neural network (DR-HGNN). This model categorizes intersections into two types—ordinary and important—to apply tailored feature aggregation strategies. Ordinary intersections aggregate features based on local neighborhood information, whereas important intersections utilize deeper neighborhood diffusion and multi-hop dependencies to capture broader traffic influences. The DR-HGNN model also employs a dynamic graph structure to reflect temporal changes in traffic flows, alongside an attention mechanism for adaptive regional feature aggregation, enhancing the identification of critical traffic nodes. Demonstrating its efficacy, the DR-HGNN achieved 19.2% and 15.4% improvements in the RMSE over 50 min predictions in the METR-LA and PEMS-BAY datasets, respectively, offering a more precise prediction method for traffic management.