Dynamic spatio-temporal graph interaction attention network for traffic flow prediction

Abstract

Against the backdrop of rapid urbanization, traffic flow prediction has become pivotal in urban transportation management and road planning. However, traffic data exhibits complex spatio-temporal dependencies, including long-term periodic trends and abrupt short-term fluctuations. Moreover, traffic patterns differ markedly across regions due to variations in geographic topology and the dynamic nature of inter-node interactions. To address these challenges, we propose a traffic flow prediction model based on a dynamic spatio-temporal graph interaction attention network (DynSTGIA). The model integrates a Time Fusion Attention (TFA) module to jointly capture localized short-term fluctuations and global long-term temporal dependencies, while a Memory-Guided Spatio-temporal Graph Module (MG-STM) incorporates learnable memory with multi-head attention to adaptively generate dynamic graphs and capture evolving spatial correlations. Moreover, to overcome the limitation of modality separation in traditional spatio-temporal models and enhance spatio-temporal fusion, we introduce an interaction learning mechanism that enables deep integration of temporal and spatial representations. Extensive experiments on five real-world traffic datasets demonstrate that DynSTGIA achieves up to 2.1 % MAE and 9.8 % RMSE improvements over strong baselines, confirming its superior performance across diverse traffic scenarios.