作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• •    

基于异质图游走与Bi-Mamba的路网节点重要性排序学习

  • 发布日期:2026-11-20

Learning to Rank Road Network Node Importance via Heterogeneous Graph Walks and Bi-Mamba

  1. 针对现有道路关键路段识别方法对起讫点(origin-destination, OD)需求、路径选择行为和属性关联利用不足,难以准确刻画路段间功能依赖关系的问题,从而影响了排序结果的准确性与鲁棒性。为此,本文提出一种基于异质图游走与Bi-Mamba的路网节点重要性排序学习模型(HetGBMR)。该模型构建由OD、路径和路段组成的出行图,以表征交通需求驱动下的跨类型关联;同时引入属性引导图,补充同类型节点之间由属性特征诱导的潜在相似关系。然后,设计异质图联合随机游走算法,在出行图与属性引导图之间协同采样,生成包含丰富上下文信息的节点序列。随后,采用Bi-Mamba编码器建模随机游走序列中的长距离上下文依赖,并通过注意力加权聚合获得目标路段表示。最后,利用加权成对排序损失优化路段间的相对重要性关系,提高关键路段排序的判别能力。在SUMO生成的仿真路网RoadNet_SY和RoadNet_RD上的实验结果表明,模型在肯德尔(Kendall)系数上分别达到0.7123和0.6328,整体优于多种传统排序方法和图表示学习方法;与整体表现次优的MGL2Rank相比,Kendall系数分别提升了6.98%和2.78%。实验结果表明,本文模型能够充分挖掘道路网络中的出行语义与结构依赖关系,提升路段表示的判别能力,并在关键路段排序任务中取得较好的识别效果。
  • Published:2026-11-20

摘要: Existing methods for critical road segment identification often insufficiently exploit origin-destination (OD) demand, route choice behavior, and attribute-induced associations, making it difficult to accurately characterize functional dependencies among road segments, thereby affecting the accuracy and robustness of the ranking results. To address this issue, this paper proposes a learning-to-rank model for road network node importance based on heterogeneous graph walks and Bi-Mamba(HetGBMR). The model first constructs a trip graph composed of OD pairs, paths, and road segments to represent traffic-demand-driven cross-type associations. Meanwhile, an attribute-guided graph is introduced to supplement latent similarity relationships among nodes of the same type induced by attribute features. Then, a joint heterogeneous graph random walk algorithm is designed to perform collaborative sampling between the trip graph and the attribute-guided graph, generating node sequences with rich contextual information. Subsequently, a Bi-Mamba encoder is employed to model long-range contextual dependencies in the random walk sequences, and an attention-based weighted aggregation mechanism is used to obtain target road segment representations. Finally, a weighted pairwise ranking loss is adopted to optimize the relative importance relationships among road segments, thereby improving the discriminative capability of critical road segment ranking. Experimental results on the SUMO-generated simulated road networks RoadNet_SY and RoadNet_RD show that the proposed model achieves Kendall coefficients of 0.7123 and 0.6328, respectively, outperforming various traditional ranking methods and graph representation learning methods overall. Compared with MGL2Rank, the second-best method in terms of overall performance, the Kendall coefficients are improved by 6.98% and 2.78%. The results demonstrate that the proposed model can effectively exploit travel semantics and structural dependencies in road networks, enhance the discriminative capacity of road segment representations, and achieve promising performance in critical road segment ranking.