Scholarlink

In the last few years, Graph Neural Networks (GNNs) have accomplished greater performance in graph-related tasks like link prediction, node classification and graph generation. Although they achieve significant results, they still suffer from overfitting, lack of robustness and over-smoothing. To address these issues, we propose an innovative dropout scheme named INDRS. At its core, INDRS first performs two iterations of aggregation and then propagates the information through several branches using different configurations. The effectiveness of our proposed scheme has been proven by many experiments using two well-known datasets, including PubMed and Amazon Photo. Indeed, our suggested scheme allows avoiding the over-smoothing issue as well as enhancing GNNs robustness comparable to two recent competitive dropout techniques.

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

New Dropout Approach for Robust and Smooth Graph Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ali Boufssasse,
  • El houssaine Hssayni,
  • Nour-Eddine Joudar,
  • Mohamed Ettaouil

摘要

In the last few years, Graph Neural Networks (GNNs) have accomplished greater performance in graph-related tasks like link prediction, node classification and graph generation. Although they achieve significant results, they still suffer from overfitting, lack of robustness and over-smoothing. To address these issues, we propose an innovative dropout scheme named INDRS. At its core, INDRS first performs two iterations of aggregation and then propagates the information through several branches using different configurations. The effectiveness of our proposed scheme has been proven by many experiments using two well-known datasets, including PubMed and Amazon Photo. Indeed, our suggested scheme allows avoiding the over-smoothing issue as well as enhancing GNNs robustness comparable to two recent competitive dropout techniques.