Scholarlink

In recent years, Graph Neural Networks (GNNs) have gained prominence across a wide range of domains where data are graph-structured. However, despite the impressive performance of GNNs, the oversmoothing phenomenon has become a common issue faced by GNNs. In this paper, we investigate oversmoothing through the lens of neural collapse and shed light on the connection between oversmoothing and neural collapse. We demonstrate that in the context of GNNs, neural collapse precludes the occurrence of oversmoothing. Drawing upon this insight, we propose the Neural Collapse inspired Regularization (NCR), a pioneering methodology that alleviates oversmoothing by optimizing the feature to satisfy the simplex Equiangular Tight Frame (ETF). Our comprehensive experiments show that NCR outperforms existing methods of tackling oversmoothing across 9 real-world graph datasets.

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

Neural Collapse Inspired Regularization for Deep Graph Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhijian Zhuo,
  • Jinwen Ma

摘要

In recent years, Graph Neural Networks (GNNs) have gained prominence across a wide range of domains where data are graph-structured. However, despite the impressive performance of GNNs, the oversmoothing phenomenon has become a common issue faced by GNNs. In this paper, we investigate oversmoothing through the lens of neural collapse and shed light on the connection between oversmoothing and neural collapse. We demonstrate that in the context of GNNs, neural collapse precludes the occurrence of oversmoothing. Drawing upon this insight, we propose the Neural Collapse inspired Regularization (NCR), a pioneering methodology that alleviates oversmoothing by optimizing the feature to satisfy the simplex Equiangular Tight Frame (ETF). Our comprehensive experiments show that NCR outperforms existing methods of tackling oversmoothing across 9 real-world graph datasets.