Spatiotemporal heterogeneity-aware meta-parameter interaction learning for traffic flow forecasting

Traffic flow forecasting remains an active and enduring research focus in the field of intelligent transportation systems. Most state-of-the-art forecasting models concentrate on learning general spatiotemporal patterns shared across all nodes, often neglecting spatiotemporal heterogeneity. This oversight limits their capacity to fully capture complex and dynamic dependencies in traffic data. To address this issue, we propose a novel method named Spatiotemporal Heterogeneity-Aware Meta-Parameter Interaction Learning (SHAMPIL). Specifically, SHAMPIL first implicitly captures spatiotemporal heterogeneity by learning spatial and temporal embeddings, which effectively act as a clustering mechanism. Then, we introduce a new meta-parameter learning paradigm that derives modality-specific parameters from a meta-parameter pool, guided by the learned heterogeneity. Finally, a spatiotemporal interaction learning module is developed, which adaptively queries a heterogeneity-aware traffic pattern library to reconstruct data-driven dynamic graph structures, enabling interactive modeling from global to local scales. Extensive experiments on four real-world benchmark datasets demonstrate that SHAMPIL consistently achieves superior forecasting performance compared to baseline models.