STC-CPR: Lightweight Spatio-Temporal Compression with Consistency-Regularized Prototype Routing for Traffic Data Imputation

Missing observations in large-scale traffic sensor networks can degrade downstream traffic monitoring and forecasting, especially when some sensors are entirely unobserved. Existing spatio-temporal imputation methods often rely on predefined road graphs or model dependencies in the full node–time space, resulting in sensitivity to noisy relations and high computational cost. We propose STC-CPR, a lightweight compress–route–reconstruct framework for unknown-node traffic data imputation. STC-CPR learns global dependencies in a compact spatio-temporal field and uses shared prototypes to exchange recurring traffic information among its latent components. A perturbation-based consistency objective improves the stability of prototype routing, while shallow bidirectional graph propagation provides complementary local information for node-level reconstruction. Experiments on four real-world traffic datasets show that STC-CPR consistently achieves strong imputation accuracy under different unknown-node ratios while maintaining favorable robustness and computational efficiency. Further analyses demonstrate the effectiveness of its compact representation and prototype routing.