Scholarlink

<p>The recommendation system is designed to offer personalized suggestions based on user interests, with the key challenge being the effective learning of user and item representations from historical interactions information. Graph Neural Networks (GNNs) have emerged as a powerful tool for addressing this challenge, owing to their ability to process graph-structured data. Despite their advantages, most GNN-based recommendation models fail to distinguish between the features of users and items during processing, which prevents them from capturing the unique local patterns of each node. Recently, the ApeGNN has introduced a graph diffusion mechanism to enhance neighborhood propagation, addressing issues such as poor adaptability of neighborhood types and unclear node importance. However, the adjacency matrix of this model suffers from noise and matrix sparsity, and ignores the correlation between users and items, the model failed to fully utilize these relationships, resulting in further loss of recommendation effectiveness. In addition, the traditional negative sampling method used in this model usually only selects negative samples with fixed hardness, which can easily lead to false positive and false negative problems. To overcome these shortcomings, we propose an <b>A</b>daptive Hardness <b>N</b>egative Sampling and Enhanced <b>G</b>raph <b>A</b>djacency matrix based recommendation model (ANGDA). This method enhances the adjacency matrix by designing a user-item interaction matrix and combining the correlation between users and items; Meanwhile, adaptively select negative samples with appropriate hardness during the training process. We conducted experiments on three recommendation datasets, and the results showed that this method outperforms existing GNN-based recommendation methods in terms of recommendation performance.</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

Enhancing recommendation with adaptive negative sampling and graph adjacency matrix optimization

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Wenjun Quan,
  • Wenjie Gao

摘要

The recommendation system is designed to offer personalized suggestions based on user interests, with the key challenge being the effective learning of user and item representations from historical interactions information. Graph Neural Networks (GNNs) have emerged as a powerful tool for addressing this challenge, owing to their ability to process graph-structured data. Despite their advantages, most GNN-based recommendation models fail to distinguish between the features of users and items during processing, which prevents them from capturing the unique local patterns of each node. Recently, the ApeGNN has introduced a graph diffusion mechanism to enhance neighborhood propagation, addressing issues such as poor adaptability of neighborhood types and unclear node importance. However, the adjacency matrix of this model suffers from noise and matrix sparsity, and ignores the correlation between users and items, the model failed to fully utilize these relationships, resulting in further loss of recommendation effectiveness. In addition, the traditional negative sampling method used in this model usually only selects negative samples with fixed hardness, which can easily lead to false positive and false negative problems. To overcome these shortcomings, we propose an Adaptive Hardness Negative Sampling and Enhanced Graph Adjacency matrix based recommendation model (ANGDA). This method enhances the adjacency matrix by designing a user-item interaction matrix and combining the correlation between users and items; Meanwhile, adaptively select negative samples with appropriate hardness during the training process. We conducted experiments on three recommendation datasets, and the results showed that this method outperforms existing GNN-based recommendation methods in terms of recommendation performance.