Scholarlink

<p>Link prediction is a fundamental challenge in graph representation learning, with significant applications in recommendation systems and image semantic relationship reasoning. Traditional graph auto-encoder methods often face limitations in structural exploitation, high-order feature extraction, and information aggregation. To address these issues, we introduce VG-GIN, a novel link prediction framework that deeply integrates Variational Graph Auto-Encoders (VGAE) and Graph Isomorphism Networks (GIN) through deep fusion. The VG-GIN model employs a multi-layer GIN encoder with multi-head attention mechanisms and gated residual connections for adaptive neighbor weighting and stable cross-layer information flow. We enhance the model's robustness with a dual augmentation strategy involving Gaussian noise injection and edge perturbation. Furthermore, our multimodal hybrid decoder combines random forest feature enhancement with an inner product decoder for optimized link prediction. Extensive experiments on six benchmark datasets demonstrate that VG-GIN outperforms state-of-the-art baseline models in AUC and AP metrics, confirming its effectiveness and providing new insights into structural optimization of graph neural networks for link prediction tasks. Our codes are available at <a href="https://github.com/gorwt/VG-GIN">https://github.com/gorwt/VG-GIN</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

Enhancing link prediction accuracy with VG-GIN: a fusion of variational graph auto-encoders and graph isomorphism networks

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Wentao Gao,
  • Lijuan Guo,
  • Jian Zhang,
  • Yutong Zhang

摘要

Link prediction is a fundamental challenge in graph representation learning, with significant applications in recommendation systems and image semantic relationship reasoning. Traditional graph auto-encoder methods often face limitations in structural exploitation, high-order feature extraction, and information aggregation. To address these issues, we introduce VG-GIN, a novel link prediction framework that deeply integrates Variational Graph Auto-Encoders (VGAE) and Graph Isomorphism Networks (GIN) through deep fusion. The VG-GIN model employs a multi-layer GIN encoder with multi-head attention mechanisms and gated residual connections for adaptive neighbor weighting and stable cross-layer information flow. We enhance the model's robustness with a dual augmentation strategy involving Gaussian noise injection and edge perturbation. Furthermore, our multimodal hybrid decoder combines random forest feature enhancement with an inner product decoder for optimized link prediction. Extensive experiments on six benchmark datasets demonstrate that VG-GIN outperforms state-of-the-art baseline models in AUC and AP metrics, confirming its effectiveness and providing new insights into structural optimization of graph neural networks for link prediction tasks. Our codes are available at https://github.com/gorwt/VG-GIN.