基于图神经网络的测井领域知识图谱实体对齐方法  

Entity Alignment Method for Logging Domain Data Asset Knowledge Graph Based on Graph Neural Networks

作　　者：曹茂俊[1] 王瑞芳 张国良 CAO Mao-jun;WANG Rui-fang;ZHANG Guo-liang(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《计算机技术与发展》2025年第4期121-126,共6页Computer Technology and Development

基　　金：中石油科技技术开发项目(2021DJ4001);黑龙江省高等教育教学改革项目(SJGY20220253);东北石油大学特色领域团队专项项目(2022TSTD-03)。

摘　　要：针对测井领域数据资产知识图谱在数据融合时出现的命名规则呈现高度多样性、行业特性显著且语义实体繁多的情况,导致测井知识图谱中实体存在大量歧义、冗余及关联错误等问题,该文提出了一种基于结构嵌入与属性嵌入的知识图谱实体对齐方法。通过在图卷积神经网络中引入高速公路网络机制(Highway Networks)来捕捉图结构的深层次特征,在图注意力神经网络中聚合高速公路网络机制来有效提取实体的属性特征,并使用最小化基于边际的损失函数来优化模型参数。在测井领域数据资产知识图谱数据集中的2个知识图谱上进行实体对齐实验,实验结果表明,该方法在实体对齐的性能上超越了所有对比模型,其Hits@10值达84.8%,比表现最好的对比模型高约0.5百分点。In view of the fact that the naming rules of the data asset knowledge graph in the logging field are highly diverse,the industry characteristics are significant,and the semantic entities are numerous,which leads to a large number of ambiguities,redundancy and association errors in the logging knowledge graph,we propose a knowledge graph entity alignment method based on structural embedding and attribute embedding.By introducing the highway network mechanism in the graph convolutional network to capture the deep features of the graph structure,the highway network mechanism is aggregated in the graph attention network to effectively extract the attribute features of the entity,and the model parameters are optimized by using the minimization margin-based loss function.The experimental results show that the proposed method is better than all the comparison models for entity alignment,with a Hits@10 value of 84.8%,which is about 0.5 percentage points better than that of the best performance comparison model.

关 键 词：知识图谱 实体对齐 图卷积神经网络 图注意力神经网络 高速公路网络机制 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]