Chen Yuxin, Huo Jingyi, Lin Fangru, Yan Hui
School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing, 210094, Jiangsu, China.
School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing, 210094, Jiangsu, China.
Neural Netw. 2025 Mar;183:106950. doi: 10.1016/j.neunet.2024.106950. Epub 2024 Nov 28.
Traffic flow forecasting is a crucial yet complex task due to the intricate spatial-temporal correlations arising from road interactions. Recent methods model these interactions using message-passing Graph Convolution Networks (GCNs), which work for homophily graphs where connected nodes primarily exhibit close observations. However, relying solely on homophily graphs presents inherent limitations in traffic modeling, as road interactions can yield not only close but also distant observations over time, revealing diverse and dynamic node-wise correlations. We designate this phenomenon as homophily-heterophily dynamics, which has been largely overlooked in previous works. To address this gap, we propose a homophily-heterophily Spatial-Temporal Graph Convolution Network (HSTGCN) that exploits both homophily and heterophily components in the spatial-temporal domain. Specifically, we first adopt time-related node attributes to disentangle the diverse and dynamic node-wise relations across time, thereby obtaining homophily and heterophily Spatial-Temporal Graphs (STGs), which provide comprehensive insights into road interactions. Subsequently, we construct dual information propagation branches, each outfitted with a specific type of STG, to exploit multiple ranges of spatial-temporal correlations from distinct perspectives through dilated causal spatial-temporal graph convolution operations on STGs. Additionally, we introduce a Graph Collaborative Learning Module (GCLM) to capture the complementary information of these two branches via mutual information transfer. Experimental evaluation on four real-world traffic datasets reveals that our model outperforms state-of-the-art methods.
由于道路相互作用产生的复杂时空相关性，交通流预测是一项至关重要但又复杂的任务。最近的方法使用消息传递图卷积网络（GCN）对这些相互作用进行建模，该网络适用于同配性图，其中相连节点主要呈现出相近的观测值。然而，仅依赖同配性图在交通建模中存在固有限制，因为道路相互作用不仅会随着时间产生相近的观测值，还会产生距离较远的观测值，从而揭示出多样且动态的节点级相关性。我们将这种现象称为同配 - 异配动态，而这在以往的工作中很大程度上被忽视了。为了弥补这一差距，我们提出了一种同配 - 异配时空图卷积网络（HSTGCN），它在时空域中利用了同配性和异配性成分。具体来说，我们首先采用与时间相关的节点属性来解开不同时间的多样且动态的节点级关系，从而获得同配性和异配性时空图（STG），这些图能提供对道路相互作用的全面洞察。随后，我们构建双信息传播分支，每个分支配备一种特定类型的STG，通过对STG进行扩张因果时空图卷积操作，从不同角度利用多个范围的时空相关性。此外，我们引入了一个图协同学习模块（GCLM），通过互信息传递来捕捉这两个分支的互补信息。对四个真实世界交通数据集的实验评估表明，我们的模型优于现有方法。
