摘要
在基于知识增强的交通预测方面,现有知识增强方法难以表征交通知识时间动态性。因此,本文提出关系-属性时序知识表示(RAT-KR),据此构建时序交通知识图谱(RAT-KG)。在此基础上,设计时间感知图注意力知识嵌入模块(TGA-KE),在注意力计算中显式引入时间信息以学习关系权重的动态变化,并将该模块以特征融合方式接入GWNet与STAEFormer,形成TGA-GWNet与TGA-STAEFormer两种知识增强模型。实验结果显示,所提方法在提升预测精度的同时,加快了模型收敛,并为预测结果提供了更直观的解释。
In the field of knowledge-enhanced traffic prediction,existing knowledge enhancement methods struggle to represent the temporal dynamics of traffic knowledge.Therefore,this paper proposes a Relation-Attribute Temporal Knowledge Representation(RAT-KR),based on which a temporal traffic knowledge graph(RAT-KG)is constructed.Building upon this,a Time-aware Graph Attention Knowledge Embedding module(TGA-KE)is designed,explicitly incorporating temporal information into attention calculations to learn the dynamic changes of relationship weights.This module is then integrated into GWNet and STAEFormer using feature fusion,resulting in two knowledge-enhanced models:TGA-GWNet and TGA-STAEFormer.Experimental results show that the proposed method improves prediction accuracy while accelerating model convergence and providing a more intuitive explanation for the prediction results.
作者
任斌
王佳伟
吴亮弘
何春红
REN Bin;WANG Jiawei;WU Lianghong;HE Chunhong(International School of Microelectronics,Dongguan University of Technology,Dongguan 523808,China;School of Computer Science and Technology(School of Software,School of Cyberspace Security),Dongguan University of Technology,Dongguan 523808,China;College of Electronics and Information Engineering,Shenzhen University,Shenzhen 518060,China;School of Urban Construction and Environment,Dongguan City University,Dongguan 523419,China)
出处
《东莞理工学院学报》
2026年第1期73-80,共8页
Journal of Dongguan University of Technology
基金
国家自然科学基金面上项目(62273096)
广东省普通高校重点领域项目(2025ZDZX3039、2025ZDZX3061)
广东省普通高校创新团队项目(2024KCXTD046)。
关键词
交通预测
知识增强
知识图谱
深度学习
traffic prediction
knowledge enhancement
knowledge graph
deep learning
