Scholarlink

<p>In the evolving landscape of graph neural networks (GNNs), this work is focused on dealing with the inherent challenges posed by noise and adversarial interferences in network-structured data. We propose an innovative GNN model with feature fusion (FFGNN) designed to enhance the resilience and reliability of GNNs in the face of practical scenarios. FFGNN introduces a denoising module to enhance robustness and suppress excessive feature smoothing, while incorporating an attention mechanism to improve model performance. Experimental validation on benchmark datasets, including Cora, CiteSeer and PubMed, demonstrates the superiority of our algorithm framework in various scenarios. We evaluated the performance of FFGNN under different conditions, such as feature noise, adversarial attacks, and clean data, showing that the complementary denoising and attention modules significantly enhance the model’s robustness and accuracy compared to other baseline models. This work represents a paradigm shift in GNN design, offering a novel approach to graph signal denoising and ensuring stable performance across diverse applications.</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

Robust graph neural networks based on feature fusion

  • 期刊论文

  • 发表日期：

    2025年1月20日

查看全文
  • Yan Jin,
  • Haoyu Shi,
  • Huaiye Meng

摘要

In the evolving landscape of graph neural networks (GNNs), this work is focused on dealing with the inherent challenges posed by noise and adversarial interferences in network-structured data. We propose an innovative GNN model with feature fusion (FFGNN) designed to enhance the resilience and reliability of GNNs in the face of practical scenarios. FFGNN introduces a denoising module to enhance robustness and suppress excessive feature smoothing, while incorporating an attention mechanism to improve model performance. Experimental validation on benchmark datasets, including Cora, CiteSeer and PubMed, demonstrates the superiority of our algorithm framework in various scenarios. We evaluated the performance of FFGNN under different conditions, such as feature noise, adversarial attacks, and clean data, showing that the complementary denoising and attention modules significantly enhance the model’s robustness and accuracy compared to other baseline models. This work represents a paradigm shift in GNN design, offering a novel approach to graph signal denoising and ensuring stable performance across diverse applications.