Scholarlink

<p>Recommendation algorithms based on knowledge graphs (KGs) have been a research focus and hotspot in the field of recommendation systems in recent years. This is mainly because the introduction of a KG can yield auxiliary information about the item of interest and achieve more accurate recommendation effects for users. However, such a model faces two major challenges in predictive tasks. First, these methods find it difficult to capture the interaction information between users and items from a global perspective. Second, in most cases, noisy data, which result from users mistakenly clicking on items they are not interested in, exist in KGs, and these noisy data have a negative impact on the resulting recommendation effect. To address these challenges, this paper proposes a knowledge-aware recommendation approach based on hypergraph representation learning and transformer model optimization (KHRT), which employs a hypergraph that can be used to directly model higher-order relations, thereby enriching the interaction information between users and items. Owing to the lack of global interaction information between users and items in local graphs, a global hypergraph is constructed within the given local graph. Conversely, nonglobal graphs contain redundant information; thus, a nonglobal hypergraph is constructed, enabling the capture of more comprehensive interaction information between users and items. Moreover, the multihead attention mechanism of the transformer model is used to enhance the cooperative relationships between user nodes and item nodes, and more valuable preference information is mined from noisy user interaction data, such as items that users are not interested in. The embeddings of user and item nodes are optimized to alleviate noise interference and achieve improved recommendation performance based on user preferences. Experiments conducted on three real recommendation datasets indicate that the proposed approach outperforms the state-of-the-art traditional recommendation methods and KG-based methods in almost all comparisons.</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

Knowledge-aware recommendation based on hypergraph representation learning and transformer model optimization

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Yuqi Zuo,
  • Yunfeng Zhang,
  • Qiuyue Zhang,
  • Wenbo Zhang

摘要

Recommendation algorithms based on knowledge graphs (KGs) have been a research focus and hotspot in the field of recommendation systems in recent years. This is mainly because the introduction of a KG can yield auxiliary information about the item of interest and achieve more accurate recommendation effects for users. However, such a model faces two major challenges in predictive tasks. First, these methods find it difficult to capture the interaction information between users and items from a global perspective. Second, in most cases, noisy data, which result from users mistakenly clicking on items they are not interested in, exist in KGs, and these noisy data have a negative impact on the resulting recommendation effect. To address these challenges, this paper proposes a knowledge-aware recommendation approach based on hypergraph representation learning and transformer model optimization (KHRT), which employs a hypergraph that can be used to directly model higher-order relations, thereby enriching the interaction information between users and items. Owing to the lack of global interaction information between users and items in local graphs, a global hypergraph is constructed within the given local graph. Conversely, nonglobal graphs contain redundant information; thus, a nonglobal hypergraph is constructed, enabling the capture of more comprehensive interaction information between users and items. Moreover, the multihead attention mechanism of the transformer model is used to enhance the cooperative relationships between user nodes and item nodes, and more valuable preference information is mined from noisy user interaction data, such as items that users are not interested in. The embeddings of user and item nodes are optimized to alleviate noise interference and achieve improved recommendation performance based on user preferences. Experiments conducted on three real recommendation datasets indicate that the proposed approach outperforms the state-of-the-art traditional recommendation methods and KG-based methods in almost all comparisons.