Traffic forecasting with meta attentive graph convolutional recurrent network

Traffic forecasting is essential for the development of intelligent transportation systems. However, existing forecasting models often struggle to effectively capture the complex spatial-temporal dependencies inherent in traffic data. Many current approaches are limited in their ability to model node-specific patterns and to simultaneously capture both short- and long-range dependencies. In this paper, we propose a novel traffic forecasting model, the Meta Attentive Graph Convolutional Recurrent Network (MAGCRN), which addresses these limitations through two key modules: (1) Node-Specific Meta Pattern Learning (NMPL) and (2) Node Attention Weight Generation (NAWG). The NMPL module captures the unique characteristics of each node in the traffic network by dynamically generating node-specific convolutional filters. The NAWG module enhances the model’s ability to capture both short- and long-range temporal dependencies by generating attention weights that connect node-specific features with those across the entire temporal dimension. Comprehensive experiments on six real-world traffic datasets demonstrate that MAGCRN consistently outperforms state-of-the-art baselines in both traffic flow and speed prediction tasks. The code is available at https://github.com/Aazeb/MAGCRN.