Spatiotemporal Graph Mixture of Experts for Highway Traffic Flow Prediction

With excellent learning ability, the pretrained large model is challenging the mainstream traffic prediction paradigm. However, the pretraining process of large spatiotemporal models still faces the problems of high training cost and fixed graph size limitation, which hinders the practical application of more flexible prediction models in intelligent transportation systems. To address these challenges, this article proposes the Spatiotemporal Graph Mixture of Experts (STGMoE), a novel framework that integrates dynamic graph message passing with an MoE mechanism. The proposed STGMoE framework takes multivariate time-series data as input and enables efficient conditional computation and flexible topological adaptation, ultimately facilitating accurate spatiotemporal feature extraction for downstream traffic prediction tasks. Experiments on the California PeMS and Beijing datasets demonstrate that the model outperforms mainstream methods in fully supervised prediction and zero-shot prediction, validating its generalization capability in complex and dynamic traffic environments.