Scholarlink

This paper introduces Virtual Adjacency Graph Neural Network (VAGNN). Compared to existing Graph Neural Networks (GNNs) that employ fixed schemes to construct neighborhoods for information aggregation, VAGNN leverages a virtual adjacency matrix to optimize neighborhoods by selectively excluding 1-hop neighbors while incorporating high-hop local neighbors and global nodes. Also, VAGNN allows for the selection of different attention mechanisms for aggregation and the incorporation of supplementary information into attention weights. The linear computational complexity of VAGNN makes it scalable for handling large graphs. Experimental evaluations on diverse real-world datasets validate the generalization and scalability capabilities of VAGNN. Parameter sensitivity analysis also reveals the importance of carefully balancing the inclusion of local and global information in VAGNN. This work lays the foundation for further exploration in developing more efficient techniques for virtual adjacency matrix construction and weighted aggregation functions, opening up possibilities for the design of more robust GNNs in the future.

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

VAGNN: Advancing the Generalization of Graph Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shuming Liang,
  • Yu Ding,
  • Bin Liang,
  • Zhidong Li,
  • Siqi Zhang,
  • Yang Wang,
  • Fang Chen

摘要

This paper introduces Virtual Adjacency Graph Neural Network (VAGNN). Compared to existing Graph Neural Networks (GNNs) that employ fixed schemes to construct neighborhoods for information aggregation, VAGNN leverages a virtual adjacency matrix to optimize neighborhoods by selectively excluding 1-hop neighbors while incorporating high-hop local neighbors and global nodes. Also, VAGNN allows for the selection of different attention mechanisms for aggregation and the incorporation of supplementary information into attention weights. The linear computational complexity of VAGNN makes it scalable for handling large graphs. Experimental evaluations on diverse real-world datasets validate the generalization and scalability capabilities of VAGNN. Parameter sensitivity analysis also reveals the importance of carefully balancing the inclusion of local and global information in VAGNN. This work lays the foundation for further exploration in developing more efficient techniques for virtual adjacency matrix construction and weighted aggregation functions, opening up possibilities for the design of more robust GNNs in the future.