Traffic flow prediction is a challenging task due to the intricate spatiotemporal dependencies among different traffic patterns. Previous approaches based on graph neural networks often modeled these dependencies as undirected graphs, which is not in line with reality. In fact, the dependencies between traffic flow sequences are inherently directional and dynamic over time. To address these issues, we propose a novel sparse directed graph convolution model, referred to as SDSTGCN. By incorporating self-attention mechanisms and asymmetric spatiotemporal convolutions, we accurately capture the directional dependencies between sensor nodes and effectively model the hidden spatiotemporal relationships. Furthermore, we apply a sparsification technique to eliminate the redundant noise introduced by the self-attention mechanisms. Extensive numerical evaluations on three real-world datasets demonstrate that our proposed method achieves state-of-the-art performance, significantly outperforming the baseline methods.
|