Scholarlink

<p>Graph analysis has become much more efficient and expressive due to the Graph Neural Networks (GNNs). However, traditional GNNs often face challenges such as oversmoothing feature sparsity and global structural sensitivity. The Distance-Driven Graph Neural Network (DDGNN) is a new method introduced in this paper that makes the representation of node features better by using marker nodes that are picked at random and in strategic way to find the shortest paths. With these marker nodes, DDGNN effectively stops features from becoming too smooth while keeping their uniqueness across multiple layers. It also encodes the global structural information explicitly. The experimental results indicate that DDGNN keeps its high level of classification accuracy at deeper levels with little performance loss. Compared to standard GNN architectures, DDGNN achieves a performance improvement of 1.5% on CORA and CiteSeer and 3% on PubMed. The model’s robustness and ability to generalize have been tested across a number of hyperparameter settings and two dataset splits (public and complete). These findings show that DDGNN has the potential to handle complex graph structures while greatly improving the accuracy of node classification and the expressiveness of models.</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

Distance driven graph neural network for advanced node classification through feature augmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月7日

查看全文
  • Imran Khan,
  • Mohammad Ubaidullah Bokhari,
  • Shahnwaz Afzal,
  • Shadab Alam

摘要

Graph analysis has become much more efficient and expressive due to the Graph Neural Networks (GNNs). However, traditional GNNs often face challenges such as oversmoothing feature sparsity and global structural sensitivity. The Distance-Driven Graph Neural Network (DDGNN) is a new method introduced in this paper that makes the representation of node features better by using marker nodes that are picked at random and in strategic way to find the shortest paths. With these marker nodes, DDGNN effectively stops features from becoming too smooth while keeping their uniqueness across multiple layers. It also encodes the global structural information explicitly. The experimental results indicate that DDGNN keeps its high level of classification accuracy at deeper levels with little performance loss. Compared to standard GNN architectures, DDGNN achieves a performance improvement of 1.5% on CORA and CiteSeer and 3% on PubMed. The model’s robustness and ability to generalize have been tested across a number of hyperparameter settings and two dataset splits (public and complete). These findings show that DDGNN has the potential to handle complex graph structures while greatly improving the accuracy of node classification and the expressiveness of models.