Scholarlink

<p>Graph neural networks (GNNs) have exhibited remarkable performance in addressing diverse graph learning tasks. However, inevitable missing information in graph networks hinders GNNs from aggregating more abundant feature information, limiting GNNs’ performance. Moreover, missing information further exacerbates the risk of overfitting in GNNs. In this manuscript, we devote to presenting a novel framework, i.e., <b>G</b>raph <b>N</b>eural <b>N</b>etworks via Adaptive Feature <b>P</b>erturbation and High-way <b>L</b>inks (PLGNN), to tackle these challenges. We introduce an efficient high-way links strategy to augment the graph, which enhances the features aggregation of GNNs, thereby improving the performance of PLGNN. Subsequently, an adaptive feature perturbation strategy is proposed to reduce model’s overfitting and also improve robustness of PLGNN. Then, we perform experiments on ten real-world datasets to reveal the superiority of PLGNN, with the corresponding performance being compared with that of state-of-the-art ones. Specifically, the <i>Accuracy</i> improved by an average of 2.6% on five node classification datasets, and an average of 2.1% on five graph classification datasets.</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

PLGNN: graph neural networks via adaptive feature perturbation and high-way links

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月12日

查看全文
  • Meixia He,
  • Peican Zhu,
  • Yang Liu,
  • Keke Tang

摘要

Graph neural networks (GNNs) have exhibited remarkable performance in addressing diverse graph learning tasks. However, inevitable missing information in graph networks hinders GNNs from aggregating more abundant feature information, limiting GNNs’ performance. Moreover, missing information further exacerbates the risk of overfitting in GNNs. In this manuscript, we devote to presenting a novel framework, i.e., Graph Neural Networks via Adaptive Feature Perturbation and High-way Links (PLGNN), to tackle these challenges. We introduce an efficient high-way links strategy to augment the graph, which enhances the features aggregation of GNNs, thereby improving the performance of PLGNN. Subsequently, an adaptive feature perturbation strategy is proposed to reduce model’s overfitting and also improve robustness of PLGNN. Then, we perform experiments on ten real-world datasets to reveal the superiority of PLGNN, with the corresponding performance being compared with that of state-of-the-art ones. Specifically, the Accuracy improved by an average of 2.6% on five node classification datasets, and an average of 2.1% on five graph classification datasets.