SAF-PMNet: A Spatial Adaptive-Focused and Pattern Memory Network for Traffic Flow Prediction

In recent years, deep learning has made significant progress in traffic flow prediction. Nevertheless, existing models still exhibit notable limitations. On the one hand, most approaches rely on attention mechanisms with a computational complexity of O(N2), which makes it difficult to meet the lightweight requirements in large-scale road network scenarios. On the other hand, current methods make insufficient use of the inherent periodic and repetitive patterns in traffic flow, thereby constraining the capability to model long-term dependencies and limits prediction accuracy. To address these issues, this paper proposes the SAF-PMNet model, which integrates an adaptive spatial focusing mechanism with a learnable pattern memory bank. This design reduces the complexity to O(N) while enhancing the representation of spatiotemporal dependencies.The adaptive focusing mechanism dynamically captures critical spatial dependencies through learnable focusing factors and kernel function approximation, while the pattern memory bank stores and retrieves representative traffic patterns to strengthen the generalization of historical regularities.