Scholarlink

<p>In recent years, knowledge graph (KG) has been widely used in knowledge-aware recommendation systems. It provides rich information to enhance item representation learning and improve recommendation performance. However, existing methods usually model user preference from a single perspective, neglecting the diversity of user interests. To address this limitation, we propose a recommendation model named user preference representation and dual contrastive learning knowledge-aware recommendation (UPDCL). Specifically, we construct a neighborhood representation learning module, which integrates user–item interaction information and combines popularity with preference matching to encode user and item representations. Meanwhile, we design a user diverse preference representation learning module, which captures multiple user preferences based on distance. Additionally, we design dual contrastive learning to optimize both user and item representations by extracting feature information from similar nodes and capturing structural information across different graphs. Through these steps, the recommendation system can fully utilize information from user–item interactions and knowledge graph, thereby helping to mitigate the problems of data sparsity and limited user representation. Finally, extensive experiments on the three real-world datasets demonstrate that our method outperforms state-of-the-art methods in terms of AUC, F1 and diversity metrics.</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

User preference representation and dual contrastive learning for knowledge-aware recommendation

  • 期刊论文

  • 发表日期：

    2025年7月31日

查看全文
  • Ting Wei,
  • Li-e Wang,
  • Xingcheng Fu,
  • Xianxian Li

摘要

In recent years, knowledge graph (KG) has been widely used in knowledge-aware recommendation systems. It provides rich information to enhance item representation learning and improve recommendation performance. However, existing methods usually model user preference from a single perspective, neglecting the diversity of user interests. To address this limitation, we propose a recommendation model named user preference representation and dual contrastive learning knowledge-aware recommendation (UPDCL). Specifically, we construct a neighborhood representation learning module, which integrates user–item interaction information and combines popularity with preference matching to encode user and item representations. Meanwhile, we design a user diverse preference representation learning module, which captures multiple user preferences based on distance. Additionally, we design dual contrastive learning to optimize both user and item representations by extracting feature information from similar nodes and capturing structural information across different graphs. Through these steps, the recommendation system can fully utilize information from user–item interactions and knowledge graph, thereby helping to mitigate the problems of data sparsity and limited user representation. Finally, extensive experiments on the three real-world datasets demonstrate that our method outperforms state-of-the-art methods in terms of AUC, F1 and diversity metrics.