A dynamic teacher-student gating framework for traffic prediction

Accurate traffic speed prediction is essential for efficient urban mobility, yet increasingly complex spatial models often yield diminishing returns. We propose the Dynamic Teacher–Student Gating Framework (DTSGF), which combines diverse spatial modelling strategies through knowledge distillation without increasing student complexity. Its predict–evaluate–teach paradigm uses fine-grained teacher routing to select the most suitable teacher’s hidden state at each time step, generating soft labels for a lightweight student. The resulting three-layer Transformer, DTSGFormer, outperforms baseline models on two real-world datasets while using only 5% of their parameters.