Scholarlink

<p>Graph Neural Networks (GNNs) exhibit excellent performance in extracting node features from graph-structured data. To enhance the representation of central node features and mitigate the over-smoothing issue, several models have refined their methods for acquiring information from distant neighbor nodes. However, most of these methods overlook the impact of distant same-type nodes on the central node and are unable to adequately mine the information contained in these distant neighbor nodes, which limits their performance. To address this, we propose a GNN model with a pre-clustering strategy, called PCE-GNN. Specifically, PCE-GNN enhances node representations through two collaborative modules: the local aggregation module effectively aggregates 1-hop neighbor information via a multi-head graph attention mechanism, while the long-distance similar neighbor aggregation module combines a pre-clustering strategy with GNN layers to utilize reconstructed star-shaped subgraphs for capturing information of distant neighbor nodes with similar features. Subsequently, these two parts of information are integrated via a max-pooling layer to form the final representation of the central node. Experimental results show that the dual-module collaborative approach of PCE-GNN possesses strong feature enhancement capabilities, outperforming baselines in node classification tasks on both public datasets and equipment maintenance datasets. The source code is available at <a href="http://github.cn/SanJinCabbage/PCE-GNN.">http://github.cn/SanJinCabbage/PCE-GNN.</a></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

PCE-GNN: a node feature-enhanced graph neural network with pre-clustering strategy

  • 期刊论文

  • 发表日期：

    2025年7月10日

查看全文
  • Yongbo Li,
  • Fangfang Xie,
  • Xi Li,
  • Kaiyan Chen,
  • Jiangyi Yao,
  • Xiongwei Li

摘要

Graph Neural Networks (GNNs) exhibit excellent performance in extracting node features from graph-structured data. To enhance the representation of central node features and mitigate the over-smoothing issue, several models have refined their methods for acquiring information from distant neighbor nodes. However, most of these methods overlook the impact of distant same-type nodes on the central node and are unable to adequately mine the information contained in these distant neighbor nodes, which limits their performance. To address this, we propose a GNN model with a pre-clustering strategy, called PCE-GNN. Specifically, PCE-GNN enhances node representations through two collaborative modules: the local aggregation module effectively aggregates 1-hop neighbor information via a multi-head graph attention mechanism, while the long-distance similar neighbor aggregation module combines a pre-clustering strategy with GNN layers to utilize reconstructed star-shaped subgraphs for capturing information of distant neighbor nodes with similar features. Subsequently, these two parts of information are integrated via a max-pooling layer to form the final representation of the central node. Experimental results show that the dual-module collaborative approach of PCE-GNN possesses strong feature enhancement capabilities, outperforming baselines in node classification tasks on both public datasets and equipment maintenance datasets. The source code is available at http://github.cn/SanJinCabbage/PCE-GNN.