DynaKey-GNN: An efficient dynamic key-node multi-graph neural network for spatio-temporal traffic flow forecasting

Accurate traffic flow prediction is crucial for effective management of urban transportation but remains challenging due to rapidly changing network conditions. We introduce a dynamic key-node multi-graph neural network (DynaKey-GNN) that identifies both long-term vital nodes and short-term critical nodes, enabling real-time adaptation to traffic shifts. An incremental update strategy selectively processes network changes, boosting computational efficiency without sacrificing accuracy. We also propose a dual-stream architecture that fuses global patterns with targeted key-node processing, capturing both stable and fast-evolving dependencies. Experiments on four real-world traffic datasets show that our approach achieves up to 12.37% higher accuracy than state-of-the-art baselines under dynamic scenarios. Case studies on volatile nodes further confirm the model’s ability to handle abrupt fluctuations in traffic flow, providing consistent and reliable forecasts.