Scholarlink

<p>Deep recommender systems with knowledge graphs (KG) as side information have attracted considerable interest recently. Although they have shown good performance in capturing graph structure information and improving explainability, knowledge-based models are unaware of external user knowledge and ignore the difference in importance between different relations. In this paper, we propose a User-Item Knowledge Graph Convolutional Network (UI-KGCN) model that comprehensively considers the importance of user features and relations, and utilizes different aggregators to integrate two-ends knowledge graph representation. The key contribution lies in three facets: (1) introducing user-end feature information and constructing the knowledge graph; (2) proposing a relation selection strategy to constructs a weighted graph by ranking the top <i>k</i> relations with the greatest impact; (3) designing three node aggregation strategies. Comparison experiments show that the proposed UI-KGCN model significantly outperforms state-of-the-art baselines in different recommendation scenarios. The self-neighbor aggregator of UI-KGCN performs best among all methods, especially on the Recall metric, achieving gains of 1.26%, 1.11% and 2.55% respectively compared with the best one among baselines on the Library, Book-Crossing and MovieLens dataset. Further ablation experiments verify the influence of different sub-modules and feature combinations for learning user and item representations, justifying the rationality and effectiveness of UI-KGCN.</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

Does user-end work? User-item-aware knowledge graph convolutional networks for recommendation

  • 期刊论文

  • 发表日期：

    2025年8月1日

查看全文
  • Xiao Gu,
  • Ling Jian

摘要

Deep recommender systems with knowledge graphs (KG) as side information have attracted considerable interest recently. Although they have shown good performance in capturing graph structure information and improving explainability, knowledge-based models are unaware of external user knowledge and ignore the difference in importance between different relations. In this paper, we propose a User-Item Knowledge Graph Convolutional Network (UI-KGCN) model that comprehensively considers the importance of user features and relations, and utilizes different aggregators to integrate two-ends knowledge graph representation. The key contribution lies in three facets: (1) introducing user-end feature information and constructing the knowledge graph; (2) proposing a relation selection strategy to constructs a weighted graph by ranking the top k relations with the greatest impact; (3) designing three node aggregation strategies. Comparison experiments show that the proposed UI-KGCN model significantly outperforms state-of-the-art baselines in different recommendation scenarios. The self-neighbor aggregator of UI-KGCN performs best among all methods, especially on the Recall metric, achieving gains of 1.26%, 1.11% and 2.55% respectively compared with the best one among baselines on the Library, Book-Crossing and MovieLens dataset. Further ablation experiments verify the influence of different sub-modules and feature combinations for learning user and item representations, justifying the rationality and effectiveness of UI-KGCN.