提示
基于图神经网络与知识图谱数据增强的无监督知识图谱嵌入方法
首发时间：2025-03-14
摘要：随着数据规模的增加，构建大规模知识图谱的需求日益增长，但传统方法在进行大规模知识图谱嵌入时面临计算成本高、依赖大量标注数据等问题。为了解决上述问题，本文设计了一个基于图神经网络与图数据增强的无监督知识图嵌入方法。具体而言，将子图输入两个不同结构的gnn编码器进行图数据增强，得到正样本特征表示；同时破坏子图拓扑结构生成负样本并输入上述相同的编码器，获得负样本特征。将正样本表示相加送入读出函数，得到全局向量，通过最大化节点特征与全局向量的互信息进行无监督训练。该方法有效提升了模型性能，降低了训练成本，增强了泛化能力，并在大规模知识图谱节点分类任务上取得良好效果，尤其在flickr和yelp数据据集上表现优异。
关键词：
for information in english, please click here
unsupervised knowledge graph embedding model based on graph sampling and graph data augmentation
abstract：with the increasing scale of data, the demand for constructing large-scale knowledge graphs is growing. however, traditional methods face issues such as high computational costs and reliance on large amounts of labeled data when embedding large-scale knowledge graphs. to address these challenges, this thesis proposes an unsupervised knowledge graph embedding method based on graph neural networks and graph data augmentation. specifically, the subgraphs are fed into two gnn encoders with different structures for graph data augmentation to obtain positive sample feature representations. meanwhile, negative samples are generated by disrupting the topological structure of the subgraphs and fed into the same encoders to obtain negative sample features. the positive sample representations are then summed and fed into a readout function to obtain a global vector. the model is trained in an unsupervised manner by maximizing the mutual information between node features and the global vector. this method effectively improves model performance, reduces training costs, enhances generalization ability, and achieves good results in large-scale knowledge graph node classification tasks on the flickr and yelp datasets. ?????
keywords：
基金：
论文图表：
引用
导出参考文献
no.****
动态公开评议
共计0人参与
勘误表
基于图神经网络与知识图谱数据增强的无监督知识图谱嵌入方法
评论
全部评论