Scholarlink

<p>Few-shot knowledge graph completion is a critical task when each relation is supported by only a limited number of samples. Existing approaches often struggle with incomplete information from local neighborhoods, which can exacerbate the effects of noise and undermine performance. To tackle the challenge, a novel SimSiam network-based model for few-shot knowledge graph completion, named HFIB, is proposed in this paper. HFIB utilizes the SimSiam network to learn robust triple representations, ensuring both alignment and uniformity in the learned embeddings. Moreover, HFIB incorporates an attention mechanism and model-agnostic meta-learning (MAML) to effectively unify semantic meanings in sparse relational neighborhoods, especially in noisy settings. In addition, MTranSparse enhances meta-relational representations using relation-specific sparse projection matrices, which dynamically manage projection complexity with adjustable sparsity. This approach aligns related entities in the projected space and separates irrelevant ones, mitigating entity sparsity from long-tailed distributions and improving generalization for few-shot relations. Extensive comparative experiments on the well-known Wiki-One and NELL-One datasets demonstrate that HFIB outperforms 11 competitive models across nearly all evaluation metrics. Furthermore, ablation study validates the effectiveness and efficiency of HFIB’s components. Additionally, the weighted SimSiam network significantly enhances the stability of HFIB. These findings confirm that HFIB is a robust and effective solution for few-shot knowledge graph completion.</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

HFIB: a novel SimSiam network-based model for few-shot knowledge graph completion

  • 期刊论文

  • 发表日期：

    2025年9月18日

查看全文
  • Haoran Li,
  • Chaoqun Zhang,
  • Weidong Tang,
  • Yuanbin Mo,
  • Wanqiu Li,
  • Zhilin Zeng

摘要

Few-shot knowledge graph completion is a critical task when each relation is supported by only a limited number of samples. Existing approaches often struggle with incomplete information from local neighborhoods, which can exacerbate the effects of noise and undermine performance. To tackle the challenge, a novel SimSiam network-based model for few-shot knowledge graph completion, named HFIB, is proposed in this paper. HFIB utilizes the SimSiam network to learn robust triple representations, ensuring both alignment and uniformity in the learned embeddings. Moreover, HFIB incorporates an attention mechanism and model-agnostic meta-learning (MAML) to effectively unify semantic meanings in sparse relational neighborhoods, especially in noisy settings. In addition, MTranSparse enhances meta-relational representations using relation-specific sparse projection matrices, which dynamically manage projection complexity with adjustable sparsity. This approach aligns related entities in the projected space and separates irrelevant ones, mitigating entity sparsity from long-tailed distributions and improving generalization for few-shot relations. Extensive comparative experiments on the well-known Wiki-One and NELL-One datasets demonstrate that HFIB outperforms 11 competitive models across nearly all evaluation metrics. Furthermore, ablation study validates the effectiveness and efficiency of HFIB’s components. Additionally, the weighted SimSiam network significantly enhances the stability of HFIB. These findings confirm that HFIB is a robust and effective solution for few-shot knowledge graph completion.