研究人员开发了一种使用持久高斯扰动来对抗循环图神经网络（GNN）中过平滑的新颖方法。该技术在每次传播步骤后注入独立的斯诺伊噪声，将GNN转变为随机动力系统。分析表明，该方法保证节点表示不会崩溃到低维子空间，从而防止渐近过平滑并保持表示多样性。 AI
影响 引入了一种新颖的理论框架和实用方法，以增强图神经网络的性能和深度。
排序理由 学术论文，详细介绍了改进GNN的新理论方法和实验验证。
在 Hugging Face Daily Papers 阅读 →
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
arXiv:2607.28185v1 Announce Type: new Abstract: Oversmoothing is a fundamental limitation of deep graph neural networks (GNNs), where repeated message passing causes node representations to become increasingly similar, eventually collapsing toward a low-dimensional subspace. This…
Oversmoothing is a fundamental limitation of deep graph neural networks (GNNs), where repeated message passing causes node representations to become increasingly similar, eventually collapsing toward a low-dimensional subspace. This phenomenon limits the effective depth of messag…