4 个月前

Geom-GCN:几何图卷积网络

Geom-GCN:几何图卷积网络

摘要

消息传递神经网络(MPNNs)已在多种实际应用中成功应用于图结构数据的表示学习。然而,MPNNs的聚合器存在两个基本弱点,限制了其表示图结构数据的能力:在邻域中丢失节点的结构信息以及在非同配图中缺乏捕捉长程依赖关系的能力。鲜有研究从不同角度注意到这些弱点。基于对经典神经网络和网络几何的观察,我们提出了一种新颖的几何聚合方案,以克服这两个弱点。该方案的基本思想是,在图上的聚合可以从图所隐含的连续空间中获益。所提出的聚合方案具有置换不变性,并由三个模块组成:节点嵌入、结构邻域和双层聚合。我们还在图卷积网络中实现了该方案,称为Geom-GCN(几何图卷积网络),用于在图上进行归纳学习。实验结果表明,所提出的Geom-GCN在广泛的公开图数据集上达到了最先进的性能。代码可在https://github.com/graphdml-uiuc-jlu/geom-gcn 获取。注释:- “非同配图”是指节点之间连接模式不遵循同质性原则的图,即高度数节点倾向于与低度数节点相连。- “置换不变性”是指算法对节点顺序的变化不敏感,这是图神经网络的一个重要特性。- “双层聚合”是指在两个层次上进行信息聚合的过程,通常包括局部和全局层次。

代码仓库

bingzhewei/geom-gcn
pytorch
GitHub 中提及
graphdml-uiuc-jlu/geom-gcn
官方
pytorch
GitHub 中提及
alexfanjn/geomgcn_pyg
pytorch
GitHub 中提及

基准测试

基准方法指标
node-classification-on-actorGeom-GCN-S
Accuracy: 30.3
node-classification-on-actorGeom-GCN-P
Accuracy: 31.63
node-classification-on-actorGeom-GCN-I
Accuracy: 29.09
node-classification-on-chameleonGeom-GCN-P
Accuracy: 60.9
node-classification-on-chameleonGeom-GCN-S
Accuracy: 59.96
node-classification-on-chameleonGeom-GCN-I
Accuracy: 60.31
node-classification-on-chameleon-60-20-20Geom-GCN*
1:1 Accuracy: 60.9
node-classification-on-citeseer-48-32-20Geom-GCN
1:1 Accuracy: 78.02 ± 1.15
node-classification-on-citeseer-60-20-20Geom-GCN*
1:1 Accuracy: 77.99
node-classification-on-cora-48-32-20-fixedGeom-GCN
1:1 Accuracy: 85.35 ± 1.57
node-classification-on-cora-60-20-20-randomGeom-GCN*
1:1 Accuracy: 85.27
node-classification-on-cornellGeom-GCN-I
Accuracy: 56.76
node-classification-on-cornellGeom-GCN-S
Accuracy: 55.68
node-classification-on-cornellGeom-GCN-P
Accuracy: 60.81
node-classification-on-cornell-60-20-20Geom-GCN*
1:1 Accuracy: 60.81
node-classification-on-film-60-20-20-randomGeom-GCN*
1:1 Accuracy: 31.63
node-classification-on-non-homophilicGeom-GCN*
1:1 Accuracy: 60.81
node-classification-on-non-homophilic-1Geom-GCN*
1:1 Accuracy: 64.12
node-classification-on-non-homophilic-10Geom-GCN
1:1 Accuracy: 31.59 ± 1.15
node-classification-on-non-homophilic-11Geom-GCN
1:1 Accuracy: 60.00 ± 2.81
node-classification-on-non-homophilic-12Geom-GCN
1:1 Accuracy: 38.15 ± 0.92
node-classification-on-non-homophilic-2Geom-GCN*
1:1 Accuracy: 67.57
node-classification-on-non-homophilic-4Geom-GCN*
1:1 Accuracy: 60.9
node-classification-on-non-homophilic-7Geom-GCN
1:1 Accuracy: 60.54 ± 3.67
node-classification-on-non-homophilic-8Geom-GCN
1:1 Accuracy: 64.51 ± 3.66
node-classification-on-non-homophilic-9Geom-GCN
1:1 Accuracy: 66.76 ± 2.72
node-classification-on-pubmed-48-32-20-fixedGeom-GCN
1:1 Accuracy: 89.95 ± 0.47
node-classification-on-pubmed-60-20-20-randomGeom-GCN*
1:1 Accuracy: 90.05
node-classification-on-squirrelGeom-GCN-P
Accuracy: 38.14
node-classification-on-squirrelGeom-GCN-I
Accuracy: 33.32
node-classification-on-squirrelGeom-GCN-S
Accuracy: 36.24
node-classification-on-squirrel-60-20-20Geom-GCN*
1:1 Accuracy: 38.14
node-classification-on-texasGeom-GCN-S
Accuracy: 59.73
node-classification-on-texasGeom-GCN-P
Accuracy: 67.57
node-classification-on-texasGeom-GCN-I
Accuracy: 57.58
node-classification-on-texas-60-20-20-randomGeom-GCN*
1:1 Accuracy: 67.57
node-classification-on-wisconsinGeom-GCN-I
Accuracy: 58.24
node-classification-on-wisconsinGeom-GCN-P
Accuracy: 64.12
node-classification-on-wisconsinGeom-GCN-S
Accuracy: 56.67
node-classification-on-wisconsin-60-20-20Geom-GCN*
1:1 Accuracy: 64.12

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
Geom-GCN:几何图卷积网络 | 论文 | HyperAI超神经