用于交通流预测的多图融合与延迟对齐动态图卷积网络
Multi-Graph Fusion with Delay-Aligned Dynamic Graph Convolutional Network for Traffic Flow Prediction

交通流预测的根本挑战是如何高效捕获复杂而又动态的时空相关性。大多数研究所采用的时空图神经网络以静态方式捕获道路节点间的时空相关性，忽视了交通状况在不同节点间传播具有时间延迟的特性，难以建模隐藏动态关联。为此，提出多图融合与延迟对齐动态图卷积网络(MGF-DA-DGCN)，该网络设计了一个延迟对齐模块来捕获由于传播延迟而损失的空间信息。其次引入一种新的多图融合方法，在构造的多图基础上生成新的动态图结构，增强模型对空间异质性的捕获能力，模拟节点间动态关联。最后将多图融合动态图卷积嵌入到交互式学习结构中，在多个时间分辨率下同步捕获并共享时空相关性。三个公共交通流数据集与11种基线模型的实验结果表明，MGF-DA-DGCN具有最优的预测性能。
The principal challenge in predicting traffic flow is efficiently capturing the complex and dynamic spatial-temporal dependence. The majority of studies that have utilized spatial-temporal graph neural networks have focused on capturing the spatial-temporal correlations between road nodes in a static manner. However, this approach overlooks the fact that there is a time delay in the propagation of traffic conditions between different nodes, and as such can make modeling the hidden dynamic correlations between them difficult. To address these challenges, a multi-graph fusion and delay-aligned dynamic graph convolutional network (MGF-DA-DGCN) was proposed. The model designed a delay-aligned module to capture the spatial information lost due to propagation delay. Furthermore, a new multi-graph fusion method was introduced to generate a dynamic graph structure based on the constructed multi-graph. This approach enhanced the model’s ability to capture spatial heterogeneity and simulate the dynamic association between nodes. Finally, the multi-graph fusion dynamic graph convolution was embedded in the interactive learning structure to synchronously capture and share spatiotemporal correlation at multiple time resolutions. Experimental results from three public transportation flow datasets and 11 baseline models indicate that MGF-DA-DGCN yields the most accurate predictions.