DSTAN: attention-enhanced dynamic spatial-temporal network for traffic forecasting

Traffic forecasting is an enduring research topic in the design of intelligent transportation systems and spatial-temporal data mining. Accurate prediction can help facilitate urban resource optimization and improve road efficiency. However, the complex spatial-temporal dependencies and dynamic urban conditions make it extremely challenging. Although many spatial-temporal modeling approaches have been proposed recently, they still suffer from the following three problems: (1) Inadequate modeling of temporal correlations; (2) Ignoring the fundamental fact that the location dependence of road networks changes dynamically over time; (3) Difficulty in extracting deeper spatial-temporal features layer by layer. In this paper, we propose a novel Dynamic Spatial-Temporal Attention-enhanced Network called DSTAN for traffic prediction. In DSTAN, we combine gated temporal units with trend-aware multi-head temporal attention to jointly capture local and long-range temporal dependencies. We also employ learnable node embeddings to extract heterogeneous information and integrate this with the spatial attention module to learn dynamic spatial correlations without any expert knowledge. Structurally, we stack multiple spatial-temporal blocks to improve the model's capability to identify complex patterns. Extensive experiments have been conducted on four widely used datasets, demonstrating that our method surpasses all baseline methods while exhibiting strong interpretability.