Scholarlink

<p>Knowledge graph link prediction aims to predict missing triples in a graph, with graph neural network (GNN)-based methods gaining attention for their strong performance. However, existing approaches often face challenges with high computational complexity when handling large-scale heterogeneous graphs. Additionally, they suffer from the oversmoothing problem in deep networks, where node embeddings become overly uniform, failing to capture complex graph structures and diverse relational features. To address these issues, we propose the Residual Gated Dual-stage Fusion (RGDF) model. By introducing a lightweight gating mechanism to dynamically integrate residual information and topological features, RGDF effectively preserves historical information and significantly enhances node embedding quality. Furthermore, a dual-stage modeling approach (local interaction and global modulation) strengthens the capture of nonlinear dependencies between nodes and relations. Extensive experiments on multiple benchmark datasets demonstrate that RGDF achieves substantial improvements in inference accuracy for complex relations and training efficiency compared to existing state-of-the-art methods.</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

RGDF: a structure-aware lightweight fusion for efficient and accurate knowledge graph completion

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Qiang Cai,
  • Yuxiao Wu,
  • Yanzhao Ren,
  • Haisheng Li

摘要

Knowledge graph link prediction aims to predict missing triples in a graph, with graph neural network (GNN)-based methods gaining attention for their strong performance. However, existing approaches often face challenges with high computational complexity when handling large-scale heterogeneous graphs. Additionally, they suffer from the oversmoothing problem in deep networks, where node embeddings become overly uniform, failing to capture complex graph structures and diverse relational features. To address these issues, we propose the Residual Gated Dual-stage Fusion (RGDF) model. By introducing a lightweight gating mechanism to dynamically integrate residual information and topological features, RGDF effectively preserves historical information and significantly enhances node embedding quality. Furthermore, a dual-stage modeling approach (local interaction and global modulation) strengthens the capture of nonlinear dependencies between nodes and relations. Extensive experiments on multiple benchmark datasets demonstrate that RGDF achieves substantial improvements in inference accuracy for complex relations and training efficiency compared to existing state-of-the-art methods.