Scholarlink

<p>Nowadays, graph neural networks (GNNs) have achieved remarkable success in various graph-related tasks. However, they still face several challenges, such as over-smoothing, lack of robustness, and overfitting. To address these issues, we propose a novel dropout technique called A-DropNode. This method conducts a message-passing layer to prevent the degradation of input data while maintaining node connectivity. Afterward, it randomly drops a portion of the node features. Building on A-DropNode, we introduce a new architecture named NGRA (Novel Graph Random Architecture), which consists of multiple A-DropNode branches with different random configurations applied after one iteration of aggregation. Additionally, consistency regularization is employed to facilitate self-supervised learning. Experiments on four commonly used data sets demonstrate the effectiveness of our method in preventing over-smoothing and improving robustness.</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

Addressing over-smoothing in graph neural networks: novel approach for enhanced stability

  • 期刊论文

  • 发表日期：

    2025年11月11日

查看全文
  • Ali Boufssasse,
  • El houssaine Hssayni,
  • Nour-Eddine Joudar,
  • Mohamed Ettaouil

摘要

Nowadays, graph neural networks (GNNs) have achieved remarkable success in various graph-related tasks. However, they still face several challenges, such as over-smoothing, lack of robustness, and overfitting. To address these issues, we propose a novel dropout technique called A-DropNode. This method conducts a message-passing layer to prevent the degradation of input data while maintaining node connectivity. Afterward, it randomly drops a portion of the node features. Building on A-DropNode, we introduce a new architecture named NGRA (Novel Graph Random Architecture), which consists of multiple A-DropNode branches with different random configurations applied after one iteration of aggregation. Additionally, consistency regularization is employed to facilitate self-supervised learning. Experiments on four commonly used data sets demonstrate the effectiveness of our method in preventing over-smoothing and improving robustness.