4 个月前

大规模图上的归纳表示学习

大规模图上的归纳表示学习

摘要

低维节点嵌入在大型图中的多种预测任务中已被证明极其有用,从内容推荐到识别蛋白质功能。然而,大多数现有的方法要求在训练嵌入时所有节点都必须存在于图中;这些先前的方法本质上是半监督的,无法自然地推广到未见过的节点。本文介绍了一种名为GraphSAGE的一般归纳框架,该框架利用节点特征信息(例如,文本属性)高效生成之前未见过的数据的节点嵌入。与为每个节点单独训练嵌入不同,我们学习了一个函数,通过采样和聚合节点局部邻域的特征来生成嵌入。我们的算法在三个归纳节点分类基准测试中优于强大的基线方法:我们在基于引用和Reddit帖子数据的演化信息图中对未见过的节点进行分类,并展示了我们的算法可以使用多图数据集中的蛋白质-蛋白质相互作用数据推广到完全未见过的图。

代码仓库

weiyinwei/mmgcn
pytorch
GitHub 中提及
silent-code/deep-trace
tf
GitHub 中提及
pmlg/pytorch_GCN
pytorch
GitHub 中提及
weiyinwei/huign
pytorch
GitHub 中提及
chatterjeeayan/upna
pytorch
GitHub 中提及
erfmah/answering_graph_queries
pytorch
GitHub 中提及
zxhhh97/ABot
pytorch
GitHub 中提及
arangoml/fastgraphml
pytorch
GitHub 中提及
stellargraph/stellargraph
tf
GitHub 中提及
qema/orca-py
pytorch
GitHub 中提及
hmcreamer/hackRice19
pytorch
GitHub 中提及
williamleif/GraphSAGE
官方
tf
GitHub 中提及
isotlaboratory/ml4vrp
pytorch
GitHub 中提及
massquantity/LibRecommender
tf
GitHub 中提及

基准测试

基准方法指标
graph-classification-on-cifar10-100kGraphSage
Accuracy (%): 66.08
graph-regression-on-zinc-500kGraphSage
MAE: 0.398
graph-regression-on-zinc-fullGraphSAGE
Test MAE: 0.126±0.003
link-property-prediction-on-ogbl-citation2NeighborSampling (SAGE aggr)
Ext. data: No
Number of params: 460289
Test MRR: 0.8044 ± 0.0010
Validation MRR: 0.8054 ± 0.0009
link-property-prediction-on-ogbl-citation2Full-batch GraphSAGE
Ext. data: No
Number of params: 460289
Test MRR: 0.8260 ± 0.0036
Validation MRR: 0.8263 ± 0.0033
link-property-prediction-on-ogbl-collabGraphSAGE (val as input)
Ext. data: No
Number of params: 460289
Test Hits@50: 0.5463 ± 0.0112
Validation Hits@50: 0.5688 ± 0.0077
link-property-prediction-on-ogbl-collabGraphSAGE
Ext. data: No
Number of params: 460289
Test Hits@50: 0.4810 ± 0.0081
Validation Hits@50: 0.5688 ± 0.0077
link-property-prediction-on-ogbl-collabGraphSAGE (val as input)
Number of params: 460289
link-property-prediction-on-ogbl-ddiGraphSAGE
Ext. data: No
Number of params: 1421057
Test Hits@20: 0.5390 ± 0.0474
Validation Hits@20: 0.6262 ± 0.0037
link-property-prediction-on-ogbl-ppaGraphSAGE
Ext. data: No
Number of params: 424449
Test Hits@100: 0.1655 ± 0.0240
Validation Hits@100: 0.1724 ± 0.0264
node-classification-on-brazil-air-trafficGraphSAGE (Hamilton et al., [2017a])
Accuracy: 0.404
node-classification-on-chameleon-60-20-20GraphSAGE
1:1 Accuracy: 62.15 ± 0.42
node-classification-on-citeseer-05GraphSAGE
Accuracy: 33.8%
node-classification-on-citeseer-1GraphSAGE
Accuracy: 51.0%
node-classification-on-citeseer-60-20-20GraphSAGE
1:1 Accuracy: 78.24 ± 0.30
node-classification-on-citeseer-fullGraphSAGE
Accuracy: 71.40%
node-classification-on-citeseer-with-publicGraphSAGE
Accuracy: 67.2
node-classification-on-cora-05GraphSAGE
Accuracy: 37.5%
node-classification-on-cora-1GraphSAGE
Accuracy: 49.0%
node-classification-on-cora-3GraphSAGE
Accuracy: 64.2%
node-classification-on-cora-60-20-20-randomGraphSAGE
1:1 Accuracy: 86.58 ± 0.26
node-classification-on-cora-full-supervisedGraphSAGE
Accuracy: 82.2%
node-classification-on-cora-with-public-splitGraphSAGE
Accuracy: 74.5%
node-classification-on-cornell-60-20-20GraphSAGE
1:1 Accuracy: 71.41 ± 1.24
node-classification-on-europe-air-trafficGraphSAGE (Hamilton et al., [2017a])
Accuracy: 27.2
node-classification-on-facebookGraphSAGE (Hamilton et al., [2017a])
Accuracy: 38.9
node-classification-on-film-60-20-20-randomGraphSAGE
1:1 Accuracy: 36.37 ± 0.21
node-classification-on-flickrGraphSAGE (Hamilton et al., [2017a])
Accuracy: 0.641
node-classification-on-non-homophilicGraphSAGE
1:1 Accuracy: 71.41 ± 1.24
node-classification-on-non-homophilic-1GraphSAGE
1:1 Accuracy: 64.85 ± 5.14
node-classification-on-non-homophilic-2GraphSAGE
1:1 Accuracy: 79.03 ± 1.20
node-classification-on-non-homophilic-4GraphSAGE
1:1 Accuracy: 62.15 ± 0.42
node-classification-on-pattern-100kGraphSage
Accuracy (%): 50.516
node-classification-on-ppiGraphSAGE
F1: 61.2
node-classification-on-pubmed-003GraphSAGE
Accuracy: 45.4%
node-classification-on-pubmed-005GraphSAGE
Accuracy: 53.0%
node-classification-on-pubmed-01GraphSAGE
Accuracy: 65.4%
node-classification-on-pubmed-60-20-20-randomGraphSAGE
1:1 Accuracy: 86.85 ± 0.11
node-classification-on-pubmed-full-supervisedGraphSAGE
Accuracy: 87.1%
node-classification-on-pubmed-with-publicGraphSAGE
Accuracy: 76.8%
node-classification-on-redditGraphSAGE
Accuracy: 94.32%
node-classification-on-squirrel-60-20-20GraphSAGE
1:1 Accuracy: 41.26 ± 0.26
node-classification-on-texas-60-20-20-randomGraphSAGE
1:1 Accuracy: 79.03 ± 1.20
node-classification-on-usa-air-trafficGraphSAGE (Hamilton et al., [2017a])
Accuracy: 31.6
node-classification-on-wiki-voteGraphSAGE (Hamilton et al., [2017a])
Accuracy: 24.5
node-classification-on-wisconsin-60-20-20GraphSAGE
1:1 Accuracy: 64.85 ± 5.14

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
大规模图上的归纳表示学习 | 论文 | HyperAI超神经