Dynamic spatiotemporal graph network with a temporal masked transformer for traffic flow forecasting

Accurate traffic flow forecasting requires a model to capture time-varying spatial correlations, nonlocal temporal dependencies, and recurring traffic patterns simultaneously. Existing spatiotemporal models often rely on fixed relational graphs or local temporal operators, which limits their ability to represent evolving dependencies in traffic networks. To address these limitations, we propose a Dynamic Spatiotemporal Graph Network with a Temporal Masked Transformer (DSGN-TMT). DSGN-TMT combines layer-specific independent blocks with cross-layer information-sharing blocks to learn complementary representations along the spatial, temporal, and feature dimensions. A time-aware dynamic graph constructor generates periodicity-conditioned spatial and temporal adjacency structures, while a temporal graph convolution module treats historical time slots as graph nodes to capture flexible temporal interactions. In addition, a Temporal Masked Transformer (TMT), a full-history contextual Transformer encoder with an optional masked-subsequence design, is integrated into the forecasting network and trained end-to-end using only forecasting loss; in both training and inference, the TMT processes the complete historical sequence without masking and produces context-aware temporal representations. Experiments on four PeMS benchmark datasets show that DSGN-TMT achieves the lowest MAE, RMSE, and MAPE among 23 comparison methods. Ablation results further demonstrate the complementary contributions of layer-specific learning, cross-layer sharing, dynamic graph construction, temporal graph convolution, and contextual temporal encoding. These results indicate that explicitly coordinating dynamic graph learning with contextual temporal representation is effective for traffic flow forecasting.