Scholarlink

<p>Traditional Graph Attention Networks (GAT) exhibit inherent limitations when deployed for knowledge reasoning tasks, especially in processing vast graph data and the high-performance computing requirements (HPC). To address these challenges, we propose a lightweight GAT model for knowledge reasoning. First, to mitigate the constraints imposed by graph scale and enable efficient node representation learning, we introduce an inverse index query method to efficiently select triples from a large knowledge base to construct the target subgraph. Then, to further boost the computational efficiency of GAT, we design a lightweight method that integrates multi-level knowledge distillation and post-training parameter quantization, while ensuring model performance and reducing model parameters and memory usage. Experimental results show that our model has superior performance and demonstrates remarkable improvements in real-time computing efficiency compared to other benchmark models.</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

A lightweight knowledge reasoning method for large-scale knowledge graphs

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Han Zhang,
  • Whenjun Zhou,
  • Bingzhi Xu,
  • Yuanbo Guo,
  • Chengfang Zhang

摘要

Traditional Graph Attention Networks (GAT) exhibit inherent limitations when deployed for knowledge reasoning tasks, especially in processing vast graph data and the high-performance computing requirements (HPC). To address these challenges, we propose a lightweight GAT model for knowledge reasoning. First, to mitigate the constraints imposed by graph scale and enable efficient node representation learning, we introduce an inverse index query method to efficiently select triples from a large knowledge base to construct the target subgraph. Then, to further boost the computational efficiency of GAT, we design a lightweight method that integrates multi-level knowledge distillation and post-training parameter quantization, while ensuring model performance and reducing model parameters and memory usage. Experimental results show that our model has superior performance and demonstrates remarkable improvements in real-time computing efficiency compared to other benchmark models.