3 个月前

DiffWire:基于洛瓦兹界的风险图重连

DiffWire:基于洛瓦兹界的风险图重连

摘要

图神经网络(Graph Neural Networks, GNNs)在多种领域中已展现出与现有方法相媲美的性能,能够有效应对各类图相关任务,包括节点分类、图分类、链接预测以及节点和图聚类等。大多数GNN基于消息传递(message passing)框架,因此也被称为消息传递神经网络(MPNNs)。尽管取得了令人瞩目的成果,MPNNs仍被广泛报道存在过平滑(over-smoothing)、过压缩(over-squashing)和欠传播(under-reaching)等问题。为缓解这些局限性,学术界提出了图重连(graph rewiring)与图池化(graph pooling)等方法。然而,当前多数先进的图重连方法在保留图的全局拓扑结构方面表现不佳,且通常不具备可微性或归纳性,同时需要手动调参。本文提出一种名为DiffWire的新颖图重连框架,专为MPNN设计,其核心思想基于Lovász界(Lovász bound),具备原理严谨、完全可微且无需参数的特性。该方法通过引入两种新型、互补的MPNN层,构建了统一的图重连理论体系:CT-Layer(Commute Time Layer)能够学习节点间的通行时间(commute time),并将其作为边权重调整的相关性函数;GAP-Layer(Spectral Gap Optimization Layer)则根据网络特性和具体任务需求,优化图的谱间隙(spectral gap)。我们通过基准图分类数据集,分别实证验证了这两种层的有效性。此外,还开展了初步研究,探讨CT-Layer在同质性(homophilic)与异质性(heterophilic)节点分类任务中的适用性。DiffWire将通行时间的学习与曲率相关定义相结合,为构建更具表达能力的MPNN开辟了新路径。

基准测试

基准方法指标
graph-classification-on-collabCT-Layer
Accuracy: 69.87%
graph-classification-on-collabGAP-Layer (Rcut)
Accuracy: 64.47%
graph-classification-on-collabDiffWire
Accuracy: 72.24%
graph-classification-on-collabGAP-Layer (Ncut)
Accuracy: 65.89%
graph-classification-on-imdb-binaryGAP-Layer (Rcut)
Accuracy: 69.93
graph-classification-on-imdb-binaryGAP-Layer (Ncut)
Accuracy: 68.8
graph-classification-on-imdb-binaryCT-Layer
Accuracy: 69.84
graph-classification-on-mutagGAP-Layer (Ncut)
Accuracy: 86.9%
graph-classification-on-mutagCT-Layer
Accuracy: 87.58%
graph-classification-on-mutagGAP-Layer (Rcut)
Accuracy: 86.9%
graph-classification-on-proteinsDiffWire
Accuracy: 74.91%
graph-classification-on-proteinsGAP-Layer (Ncut)
Accuracy: 75.34%
graph-classification-on-proteinsCT-Layer
Accuracy: 75.38%
graph-classification-on-proteinsGAP-Layer (Rcut)
Accuracy: 75.03%
graph-classification-on-reddit-binaryGAP-Layer (Rcut)
Accuracy: 77.63
graph-classification-on-reddit-binaryGAP-Layer (Ncut)
Accuracy: 76
graph-classification-on-reddit-binaryDiffWire
Accuracy: 77.17
graph-classification-on-reddit-binaryCT-Layer
Accuracy: 78.45
node-classification-on-actorCT-Layer (PE)
Accuracy: 29.35
node-classification-on-actorCT-Layer
Accuracy: 31.98
node-classification-on-citeseerCT-Layer (PE)
Accuracy: 72.26
node-classification-on-citeseerCT-Layer
Accuracy: 66.71
node-classification-on-coraCT-Layer (PE)
Accuracy: 83.66%
node-classification-on-coraCT-Layer
Accuracy: 67.96%
node-classification-on-cornellCT-Layer
Accuracy: 69.04
node-classification-on-cornellCT-Layer (PE)
Accuracy: 58.02
node-classification-on-pubmedCT-Layer
Accuracy: 68.19
node-classification-on-pubmedCT-Layer (PE)
Accuracy: 86.07
node-classification-on-wisconsinCT-Layer
Accuracy: 79.05
node-classification-on-wisconsinCT-Layer (PE)
Accuracy: 69.25

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
DiffWire:基于洛瓦兹界的风险图重连 | 论文 | HyperAI超神经