Scholarlink

<p>Knowledge Graph Embedding (KGE) technology transforms entities and relations into low-dimensional space. However, in practical applications, the large and sparse atlas data make the prediction of low-frequency entities difficult. To this end, we design the Knowledge Graph Alleviating Sparsity Learning (KGASL) framework, which takes Knowledge Relation Attention Network (KRAN) and Knowledge Contrastive Learning (KCL) as two pillars. By deeply mining the graph context information of entities, KRAN cleverly projects adjacent triples to multiple semantic levels, and uses the attention mechanism to accurately integrate this information. On the other hand, KCL combines the contrast loss and cross-entropy loss, and uses the negative sampling strategy to introduce more negative samples, which makes the model more discriminative and robust in the face of noisy triples. Experimental results show that KGASL achieves state-of-the-art performance with a 4.3% MRR improvement over CompGCN on WN18RR, and maintains robustness under 40% noise injection.</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 density through graph relation attention and contrastive learning

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Chunyu Lu,
  • Tianran Chen,
  • Duo Shang,
  • Jun Luo,
  • Xin Hui,
  • Ruhui Shi

摘要

Knowledge Graph Embedding (KGE) technology transforms entities and relations into low-dimensional space. However, in practical applications, the large and sparse atlas data make the prediction of low-frequency entities difficult. To this end, we design the Knowledge Graph Alleviating Sparsity Learning (KGASL) framework, which takes Knowledge Relation Attention Network (KRAN) and Knowledge Contrastive Learning (KCL) as two pillars. By deeply mining the graph context information of entities, KRAN cleverly projects adjacent triples to multiple semantic levels, and uses the attention mechanism to accurately integrate this information. On the other hand, KCL combines the contrast loss and cross-entropy loss, and uses the negative sampling strategy to introduce more negative samples, which makes the model more discriminative and robust in the face of noisy triples. Experimental results show that KGASL achieves state-of-the-art performance with a 4.3% MRR improvement over CompGCN on WN18RR, and maintains robustness under 40% noise injection.