Scholarlink

<p>Knowledge-graph-enhanced recommendation can improve personalized retrieval, but existing methods often suffer from noise accumulation in multi-hop propagation and increased computational cost. To address these issues, we propose Quatnet, an interest propagation recommendation model based on quaternion knowledge graph embedding. Quatnet combines quaternion representation and convolutional feature extraction to model multi-relational triplets, and it performs lightweight recommendation using 1-hop interest propagation. In this way, the model strengthens semantic representation while reducing the influence of distant noisy entities. Experiments on MovieLens-1M and Last.FM show that Quatnet achieves competitive recommendation performance. On MovieLens-1M, its AUC and ACC are improved by 0.7% and 1.1% compared with RippleNet, with a 32 times reduction in computational load. On Last.FM, it maintains competitive performance with a 64 times computational load reduction. These results indicate that quaternion-based representation and localized interest propagation provide an effective trade-off between accuracy and computational efficiency.</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

Quatnet: an interest propagation recommendation model using quaternion-based knowledge graph embedding representation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Wei Xiong,
  • Fei Yang,
  • Xue Ouyang,
  • Mengyu Ma

摘要

Knowledge-graph-enhanced recommendation can improve personalized retrieval, but existing methods often suffer from noise accumulation in multi-hop propagation and increased computational cost. To address these issues, we propose Quatnet, an interest propagation recommendation model based on quaternion knowledge graph embedding. Quatnet combines quaternion representation and convolutional feature extraction to model multi-relational triplets, and it performs lightweight recommendation using 1-hop interest propagation. In this way, the model strengthens semantic representation while reducing the influence of distant noisy entities. Experiments on MovieLens-1M and Last.FM show that Quatnet achieves competitive recommendation performance. On MovieLens-1M, its AUC and ACC are improved by 0.7% and 1.1% compared with RippleNet, with a 32 times reduction in computational load. On Last.FM, it maintains competitive performance with a 64 times computational load reduction. These results indicate that quaternion-based representation and localized interest propagation provide an effective trade-off between accuracy and computational efficiency.