Traffic flow prediction based on a multi-task pivotal attention network

Short-term traffic flow prediction remains challenging because traffic dynamics are nonlinear, non-stationary, and governed by both local topology and long-range directional dependencies. Existing approaches often rely on static or uniformly learned graphs, do not explicitly separate slow trends from rapid fluctuations, and insufficiently exploit correlated modalities such as flow, speed, and traffic type. To address these issues, we propose the Pivotal Multi-Modal Spatio-Temporal Network (PMST-Net), which combines a Pivotal Node Identification Module (PNIM), a Temporal Feature Disentanglement (TFD) layer, a Pivotal-Informed Spatial Attention (PISA) encoder, and a Cross-Modal Multi-Task Learning (CM-MTL) prediction head. Experiments on four public PeMS benchmarks demonstrate consistent improvements over strong baselines. On PEMS08, PMST-Net reduces mean absolute error (MAE) and root mean squared error (RMSE) by 42.3% and 37.2%, respectively, relative to the strongest baseline. Ablation, robustness, and efficiency analyses further confirm the contribution of each module and the practical value of the overall design.