4 个月前

自适应通用广义PageRank图神经网络

自适应通用广义PageRank图神经网络

摘要

在许多重要的图数据处理应用中,获取的信息既包括节点特征,也包括图拓扑结构的观察。图神经网络(GNNs)旨在利用这两种证据来源,但它们并未最优地权衡其效用,且未能以一种普遍的方式将它们集成。这里的“普遍性”指的是独立于同质性或异质性图假设。为了解决这些问题,我们引入了一种新的广义PageRank(GPR)GNN架构,该架构自适应地学习GPR权重,从而联合优化节点特征和拓扑信息的提取,无论节点标签的同质性或异质性程度如何。所学的GPR权重会自动调整以适应节点标签模式,而不受初始化类型的影响,从而保证对通常难以处理的标签模式具有出色的性能。此外,这些权重允许避免特征过度平滑的问题,这一过程会使特征信息失去区分能力,而无需网络保持浅层结构。我们对GPR-GNN方法进行了理论分析,并借助所谓的上下文随机块模型生成的新颖合成基准数据集来支持这一分析。我们还使用已知的同质性和异质性基准数据集,在节点分类问题上将我们的GNN架构与几种最先进的GNNs进行了性能比较。结果表明,与现有技术相比,GPR-GNN在合成数据和基准数据上均提供了显著的性能提升。

代码仓库

jianhao2016/GPRGNN
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
node-classification-on-actorGPRGCN
Accuracy: 35.16 ± 0.9
node-classification-on-chameleonGPRGCN
Accuracy: 62.59 ± 2.04
node-classification-on-chameleon-60-20-20GPRGNN
1:1 Accuracy: 67.48 ± 0.40
node-classification-on-citeseer-48-32-20GPRGCN
1:1 Accuracy: 77.13 ± 1.67
node-classification-on-citeseer-60-20-20GPRGNN
1:1 Accuracy: 67.63 ± 0.38
node-classification-on-cora-48-32-20-fixedGPRGCN
1:1 Accuracy: 87.95 ± 1.18
node-classification-on-cora-60-20-20-randomGPRGNN
1:1 Accuracy: 79.51 ± 0.36
node-classification-on-cornellGPRGCN
Accuracy: 78.11 ± 6.55
node-classification-on-cornell-60-20-20GPRGNN
1:1 Accuracy: 91.36 ± 0.70
node-classification-on-film-60-20-20-randomGPRGNN
1:1 Accuracy: 39.30 ± 0.27
node-classification-on-geniusGPRGCN
Accuracy: 90.05 ± 0.31
node-classification-on-non-homophilicMLP-2
1:1 Accuracy: 91.30 ± 0.70
node-classification-on-non-homophilicGPRGNN
1:1 Accuracy: 91.36 ± 0.70
node-classification-on-non-homophilic-1GPRGNN
1:1 Accuracy: 93.75 ± 2.37
node-classification-on-non-homophilic-11GPRGCN
1:1 Accuracy: 62.59 ± 2.04
node-classification-on-non-homophilic-12GPRGCN
1:1 Accuracy: 46.31 ± 2.46
node-classification-on-non-homophilic-13GPRGCN
1:1 Accuracy: 81.38 ± 0.16
node-classification-on-non-homophilic-14GPRGCN
1:1 Accuracy: 90.05 ± 0.31
node-classification-on-non-homophilic-15GPRGCN
1:1 Accuracy: 61.89 ± 0.29
node-classification-on-non-homophilic-2MLP-2
1:1 Accuracy: 92.26 ± 0.71
node-classification-on-non-homophilic-2GPRGNN
1:1 Accuracy: 92.92 ± 0.61
node-classification-on-non-homophilic-4GPRGNN
1:1 Accuracy: 67.48 ± 0.40
node-classification-on-non-homophilic-6GPRGNN
1:1 Accuracy: 66.90±0.50
node-classification-on-non-homophilic-7GPRGCN
1:1 Accuracy: 78.11 ± 6.55
node-classification-on-non-homophilic-8GPRGCN
1:1 Accuracy: 82.55 ± 6.23
node-classification-on-non-homophilic-9GPRGCN
1:1 Accuracy: 81.35 ± 5.32
node-classification-on-penn94GPRGCN
Accuracy: 81.38 ± 0.16
node-classification-on-pubmed-48-32-20-fixedGPRGCN
1:1 Accuracy: 87.54 ± 0.38
node-classification-on-pubmed-60-20-20-randomGPRGNN
1:1 Accuracy: 85.07 ± 0.09
node-classification-on-squirrelGPRGCN
Accuracy: 46.31 ± 2.46
node-classification-on-squirrel-60-20-20GPRGNN
1:1 Accuracy: 49.93 ± 0.53
node-classification-on-texasGPRGCN
Accuracy: 81.35 ± 5.32
node-classification-on-texas-60-20-20-randomGPRGNN
1:1 Accuracy: 92.92 ± 0.61
node-classification-on-wisconsinGPRGCN
Accuracy: 82.55 ± 6.23
node-classification-on-wisconsin-60-20-20GPRGNN
1:1 Accuracy: 93.75 ± 2.37

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
自适应通用广义PageRank图神经网络 | 论文 | HyperAI超神经