基于图表示学习的知识图谱时序推理模型  

A Knowledge Graph Temporal Reasoning Model Based on Graph Representation Learning

作　　者：张宇姣 徐健 吴迪 ZHANG Yujiao;XU Jian;WU Di(College of Computer Science and Technology,Taiyuan Normal University,Jinzhong 030619,Shanxi,China;School of Mechanical and Electrical Engineering,Guilin University of Electronic Technology,Guilin 541004,Guangxi,China;College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,Heilongjiang,China;College of Computer and Control Engineering,Qiqihar University,Qiqihar 161006,Heilongjiang,China)

机构地区：[1]太原师范学院计算机科学与技术学院,山西晋中030619 [2]桂林电子科技大学机电工程学院,广西桂林541004 [3]哈尔滨工程大学计算机科学与技术学院,哈尔滨150001 [4]齐齐哈尔大学计算机与控制工程学院,黑龙江齐齐哈尔161006

出　　处：《济南大学学报(自然科学版)》2025年第2期272-277,共6页Journal of University of Jinan(Science and Technology)

基　　金：国家自然科学基金项目(32060157)。

摘　　要：针对传统知识图谱推理模型在时间关联推理方面的局限性,以及现有模型仅通过在静态知识图谱中加入时间戳组合,而未充分考虑时间序列依赖关系的问题,提出基于图表示学习的知识图谱时序推理(KGTR_GRL)模型;针对图表示学习中的邻居结构信息,设计多关系图结构编码器,以解决当前大部分研究忽略的节点重要性问题;为了更深入地捕获时间信息,将注意力机制引入到时序编码器中,设计模型时序推理算法,通过解码器计算评分并转换为候选实体的概率;采用2个公开数据集测试模型的性能,并与多个现有模型的性能进行比较。结果表明,KGTR_GRL模型表现出更好的性能,实验中平均倒数排名,预测排名小于或等于1、10的三元组的平均占比指标均优于其他现有模型,证明了考虑多阶邻居特征信息的多关系编码器性能的优越性。To address the limitations of traditional knowledge graph reasoning models in temporal correlation reasoning,and the issue that existing models only incorporate timestamps into static knowledge graphs without fully considering the dependency of temporal sequences,a knowledge graph temporal reasoning model based on graph representation learning(KGTR_GRL)was proposed.In response to the neighbor structure information in graph representation learning,a multi-relational graph structure encoder was designed to address the issue of node importance,which was overlooked by most current studies.To capture temporal information more comprehensively,the attention mechanism was introduced into the temporal encoder,and a model temporal reasoning algorithm was designed,where scores were calculated through the decoder and converted into probabilities for candidate entities.The performance of the model was tested on two public datasets,and comparisons were made with several existing models.The results show that the proposed KGTR_GRL model exhibits superior performance,and the mean reciprocal ranking(MRR),the mean proportion of triples with predicted rankings less than or equal to 1(Hits@1)and 10(Hits@10)are all better than other existing models in the experiment,which proves the superiority of multi-relation encoder considering multi-order neighbor feature information.

关 键 词：时序推理 时序知识图谱 图表示学习 图卷积神经网络 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]