Scholarlink

<p>The link prediction in knowledge graphs plays a significant role in citation network applications, recommendation systems, and social network analysis. Although graph-based machine learning has improved, conventional models are unable to provide dynamic time variations and capture the underlying semantics in real world graphs. In this paper, authors provide a framework that incorporates domain-oriented regularizations into graph neural networks (GNNs) to increase link prediction performance. It proposes centrality-driven, research-domain similarity, and temporal loss regularization to enhance node embedding and allows the model to learn structural and semantic patterns natural to knowledge graphs. The framework is tested on the Cora, Citeseer, and PubMed datasets, where it is observed to greatly enhance prediction levels, achieving ROC-AUC scores of 90%, 91%, and 97%, along with better accuracy, F1 score, and recall. Scalability is also evident in the model, as demonstrated on PubMed, which contains more than 19,000 nodes and 44,000 edges. This publication underscores the importance of using structural and semantic regularization to solve real-world problems in such areas as social networks, ecommerce, and scholarly studies. The paper also argues for the introduction of regularization techniques in GNNs during the process of link prediction and compares the results of different models. It also covers the existing constraints and potential research directions for the future, including refining model generalization and scalability for large, dynamic graphs of knowledge.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing knowledge graph embeddings for improved link prediction using graph neural networks

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Osamah Khaled,
  • Shaista Habib,
  • Daniyal Adeeb,
  • Abdulmuneem Alselwi

摘要

The link prediction in knowledge graphs plays a significant role in citation network applications, recommendation systems, and social network analysis. Although graph-based machine learning has improved, conventional models are unable to provide dynamic time variations and capture the underlying semantics in real world graphs. In this paper, authors provide a framework that incorporates domain-oriented regularizations into graph neural networks (GNNs) to increase link prediction performance. It proposes centrality-driven, research-domain similarity, and temporal loss regularization to enhance node embedding and allows the model to learn structural and semantic patterns natural to knowledge graphs. The framework is tested on the Cora, Citeseer, and PubMed datasets, where it is observed to greatly enhance prediction levels, achieving ROC-AUC scores of 90%, 91%, and 97%, along with better accuracy, F1 score, and recall. Scalability is also evident in the model, as demonstrated on PubMed, which contains more than 19,000 nodes and 44,000 edges. This publication underscores the importance of using structural and semantic regularization to solve real-world problems in such areas as social networks, ecommerce, and scholarly studies. The paper also argues for the introduction of regularization techniques in GNNs during the process of link prediction and compares the results of different models. It also covers the existing constraints and potential research directions for the future, including refining model generalization and scalability for large, dynamic graphs of knowledge.