Traffic prediction is a crucial task in intelligent transportation systems, requiring the effective modeling of both spatial and temporal dependencies in traffic data. Most existing methods rely on predefined and fixed graph structures to capture spatial dependencies, while only limited short-term temporal patterns are modeled using RNNs or their variants. However, due to the dynamic changes of the traffic network structure, the static predefined graph structure often cannot accurately reflect the real spatial dependencies. Additionally, these methods struggle to effectively capture long-range temporal patterns. To address these limitations, we propose the Adaptive Semantic-Enhanced Spatial-Temporal Graph Network (ASSGN), a novel spatial-temporal modeling framework for traffic prediction. ASSGN dynamically learns hidden spatial dependencies through an adaptive dependency matrix. Additionally, it captures both long-term periodic patterns and short-term local information in the temporal dimension through a dual-branch design to enhance the semantic information. These components are seamlessly integrated and trained in an end-to-end manner. Experimental results on two real-world datasets demonstrate that ASSGN achieves superior performance in both short-term prediction and long-term prediction scenarios, highlighting its effectiveness and versatility.
Yin et al. (Fri,) studied this question.