Scholarlink

<p>The social network is usually utilized as auxiliary data to alleviate data sparsity and cold start problems of recommender systems. However, the social network obeys the power-law distribution and the social relations are indiscriminately utilized, resulting in providing insufficient and inaccurate information. Early efforts ignore these drawbacks and fail to exploit abundant information fully. In this paper, we propose a novel multi-neighbor social recommendation framework MNGCN based on graph convolutional network. The representations of user and item could be learned by iteratively integrating their multiple types of neighbor information. In addition, we apply a node-level attention mechanism to aggregate the same type of neighbors and a category-level attention mechanism to incorporate different categories of neighbors. A sampler is utilized to accurately select the social neighbors of users with regard to different items. Besides, the interactions and ratings are captured simultaneously in the user-item interactive network. Extensive experiments on two classical datasets illustrate that MNGCN achieves the best performance, and the ablation study demonstrates the necessity and the effectiveness of each component.</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

Multi-neighbor social recommendation with attentional graph convolutional network

  • 期刊论文

  • 发表日期：

    2025年3月20日

查看全文
  • Min Zhang,
  • Xiao Liao,
  • Xinlei Wang,
  • Xiaojuan Wang,
  • Lei Jin

摘要

The social network is usually utilized as auxiliary data to alleviate data sparsity and cold start problems of recommender systems. However, the social network obeys the power-law distribution and the social relations are indiscriminately utilized, resulting in providing insufficient and inaccurate information. Early efforts ignore these drawbacks and fail to exploit abundant information fully. In this paper, we propose a novel multi-neighbor social recommendation framework MNGCN based on graph convolutional network. The representations of user and item could be learned by iteratively integrating their multiple types of neighbor information. In addition, we apply a node-level attention mechanism to aggregate the same type of neighbors and a category-level attention mechanism to incorporate different categories of neighbors. A sampler is utilized to accurately select the social neighbors of users with regard to different items. Besides, the interactions and ratings are captured simultaneously in the user-item interactive network. Extensive experiments on two classical datasets illustrate that MNGCN achieves the best performance, and the ablation study demonstrates the necessity and the effectiveness of each component.