PDFormer introduces a novel approach to traffic flow prediction by capturing dynamic spatial dependencies, incorporating long-range views, and explicitly modeling time delays. The model outperforms existing methods on various datasets and enhances interpretability through attention maps.
The fundamental challenge in traffic flow prediction is effectively modeling complex spatial-temporal dependencies. Spatial-temporal Graph Neural Network (GNN) models are promising but have limitations. PDFormer addresses these limitations by introducing a novel Propagation Delay-aware dynamic long-range transFormer.
The model includes a spatial self-attention module to capture dynamic spatial dependencies, graph masking matrices for short- and long-range views, and a feature transformation module to model time delays. Experimental results show state-of-the-art performance and computational efficiency across real-world traffic datasets.
PDFormer's design allows it to capture both short- and long-range spatial dependencies simultaneously. By considering the time delay in spatial information propagation, the model achieves superior performance compared to existing methods. Visualization of attention maps enhances interpretability and highlights the model's effectiveness.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문