Scholarlink

<p>In recent years, recommender systems have increasingly focused on providing personalized suggestions based on individual user preferences and interaction history. However, many existing approaches rely on fixed or mechanical sampling strategies, which fail to capture the rich diversity in sample distributions. As a result, these methods lack flexibility and adaptability, limiting their overall performance. In addition, traditional denoising techniques often involve the random removal of nodes or edges, which can lead to the loss of important structural information and user-item interaction patterns. This hinders the model’s ability to generalize and weakens its robustness. To address these issues, we propose <b>K</b>nowledge-<b>A</b>ware <b>G</b>raph <b>A</b>ugmentation for <b>L</b>earning to <b>R</b>ecommend (KAGLR). The KAGLR framework consists of two main components. First, it uses variational graph reconstruction to estimate a Gaussian distribution for each node in the interaction graph. By sampling from these distributions, the model generates multiple contrastive views that preserve the original graph structure and enable the discovery of hidden user-item relationships. Second, we introduce an automated graph augmentation module for denoising. This module adaptively removes noisy elements from the interaction graph while preserving essential properties, allowing the model to capture deeper interaction patterns. We evaluate KAGLR through extensive experiments on benchmark datasets. The results show that KAGLR consistently outperforms existing methods, achieving state-of-the-art performance in recommendation tasks.</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 graph augmentation for learning to recommend

  • 期刊论文

  • 发表日期：

    2026年2月6日

查看全文
  • Xinyue Zhang,
  • Shirong Ma,
  • Jiale Chen

摘要

In recent years, recommender systems have increasingly focused on providing personalized suggestions based on individual user preferences and interaction history. However, many existing approaches rely on fixed or mechanical sampling strategies, which fail to capture the rich diversity in sample distributions. As a result, these methods lack flexibility and adaptability, limiting their overall performance. In addition, traditional denoising techniques often involve the random removal of nodes or edges, which can lead to the loss of important structural information and user-item interaction patterns. This hinders the model’s ability to generalize and weakens its robustness. To address these issues, we propose Knowledge-Aware Graph Augmentation for Learning to Recommend (KAGLR). The KAGLR framework consists of two main components. First, it uses variational graph reconstruction to estimate a Gaussian distribution for each node in the interaction graph. By sampling from these distributions, the model generates multiple contrastive views that preserve the original graph structure and enable the discovery of hidden user-item relationships. Second, we introduce an automated graph augmentation module for denoising. This module adaptively removes noisy elements from the interaction graph while preserving essential properties, allowing the model to capture deeper interaction patterns. We evaluate KAGLR through extensive experiments on benchmark datasets. The results show that KAGLR consistently outperforms existing methods, achieving state-of-the-art performance in recommendation tasks.