Scholarlink

<p>Accurate traffic flow forecasting has long been a challenge in intelligent transportation systems (ITS). Along the temporal dimension, it is difficult to simultaneously capture both local fine-grained patterns and global long-term dependencies of traffic flow; along the spatial dimension, unified modeling remains insufficient for both pairwise relationships among traffic network nodes and coordinated behaviors across multiple nodes. To address this challenge, a novel Dual-Attention Dynamic Graph–Hypergraph Convolutional Network (DADGHCN) is proposed. Specifically, along the temporal dimension, temporal local convolutional attention and temporal global aware attention jointly model multi-scale temporal dependencies in traffic flow; the temporal global aware attention employs a dynamic time-aware masking matrix to enhance long-term dependency modeling. Along the spatial dimension, a dual-channel graph–hypergraph architecture is designed to characterize complex spatial correlations in traffic flow: the graph channel generates data-driven periodic dynamic graphs and fuses them with static graph masks to dynamically adjust the weights of pairwise relationships between nodes, while the hypergraph channel dynamically constructs hyperedge weights to explicitly model collaborative patterns among multiple nodes, forming complementarity with the graph branch at the edge–node level. Large-scale traffic networks produce high-dimensional and dense spatiotemporal sequences, and operations such as graph–hypergraph convolution, diffusion propagation, and attention often exhibit quadratic or higher complexity with respect to the number of nodes and the temporal length, imposing stringent requirements on parallel compute capability and memory bandwidth; therefore, efficient training and inference are required with support from high-performance computing (HPC) platforms. Experiments on four real-world datasets demonstrate that DADGHCN outperforms various representative methods in forecasting accuracy, validating the effectiveness and robustness of the dual-attention and dynamic graph–hypergraph collaborative modeling framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-attention dynamic graph–hypergraph convolutional network for traffic flow forecasting

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Hong Zhang,
  • Ting Chen,
  • Linbiao Chen,
  • Liang Hou

摘要

Accurate traffic flow forecasting has long been a challenge in intelligent transportation systems (ITS). Along the temporal dimension, it is difficult to simultaneously capture both local fine-grained patterns and global long-term dependencies of traffic flow; along the spatial dimension, unified modeling remains insufficient for both pairwise relationships among traffic network nodes and coordinated behaviors across multiple nodes. To address this challenge, a novel Dual-Attention Dynamic Graph–Hypergraph Convolutional Network (DADGHCN) is proposed. Specifically, along the temporal dimension, temporal local convolutional attention and temporal global aware attention jointly model multi-scale temporal dependencies in traffic flow; the temporal global aware attention employs a dynamic time-aware masking matrix to enhance long-term dependency modeling. Along the spatial dimension, a dual-channel graph–hypergraph architecture is designed to characterize complex spatial correlations in traffic flow: the graph channel generates data-driven periodic dynamic graphs and fuses them with static graph masks to dynamically adjust the weights of pairwise relationships between nodes, while the hypergraph channel dynamically constructs hyperedge weights to explicitly model collaborative patterns among multiple nodes, forming complementarity with the graph branch at the edge–node level. Large-scale traffic networks produce high-dimensional and dense spatiotemporal sequences, and operations such as graph–hypergraph convolution, diffusion propagation, and attention often exhibit quadratic or higher complexity with respect to the number of nodes and the temporal length, imposing stringent requirements on parallel compute capability and memory bandwidth; therefore, efficient training and inference are required with support from high-performance computing (HPC) platforms. Experiments on four real-world datasets demonstrate that DADGHCN outperforms various representative methods in forecasting accuracy, validating the effectiveness and robustness of the dual-attention and dynamic graph–hypergraph collaborative modeling framework.