4 个月前

图神经网络有多强大?

图神经网络有多强大?

摘要

图神经网络(GNNs)是一种有效的图表示学习框架。GNNs 采用邻域聚合方案,通过递归地聚合和转换节点的邻域节点表示向量来计算该节点的表示向量。许多 GNN 变体已被提出,并在节点分类和图分类任务中取得了最先进的结果。然而,尽管 GNNs 在图表示学习领域带来了革命性的变化,但对其表示性质和局限性的理解仍然有限。本文提出了一种理论框架,用于分析 GNNs 捕捉不同图结构的表达能力。我们的研究结果表征了流行的 GNN 变体(如图卷积网络和 GraphSAGE)的判别能力,并表明它们无法学会区分某些简单的图结构。随后,我们开发了一种简单架构,该架构在 GNN 类中具有最高的表达能力,并且其威力等同于 Weisfeiler-Lehman 图同构测试。我们在多个图分类基准数据集上对我们的理论发现进行了实证验证,并展示了我们的模型达到了最先进的性能。

基准测试

基准方法指标
graph-classification-on-bp-fmri-97GIN
Accuracy: 45.4%
F1: 42.3%
graph-classification-on-cifar10-100kGIN
Accuracy (%): 53.28
graph-classification-on-collabGIN-0
Accuracy: 80.2%
graph-classification-on-cox2GIN-0
Accuracy(10-fold): 81.13
graph-classification-on-ddGIN
Accuracy: 77.311±2.223
graph-classification-on-enzymesGIN
Accuracy: 68.303±4.170
graph-classification-on-hiv-dti-77GIN
Accuracy: 55.1%
F1: 53.6%
graph-classification-on-hiv-fmri-77GIN
Accuracy: 52.5%
F1: 35.6%
graph-classification-on-imdb-bGIN-0
Accuracy: 75.1%
graph-classification-on-imdb-bGIN
Accuracy: 81.250±3.775
graph-classification-on-imdb-mGIN-0
Accuracy: 52.3%
graph-classification-on-mutagGIN-0
Accuracy: 89.4%
graph-classification-on-nci1GIN
Accuracy: 84.818±0.936
graph-classification-on-nci1GIN-0
Accuracy: 82.7%
graph-classification-on-nci109GIN
Accuracy: 84.155±0.812
graph-classification-on-peptides-funcGIN
AP: 0.6043±0.0216
graph-classification-on-proteinsGIN
Accuracy: 75.536±1.851
graph-classification-on-proteinsGIN-0
Accuracy: 76,2%
graph-classification-on-ptcGIN-0
Accuracy: 64.40%
graph-classification-on-re-m5kGIN-0
Accuracy: 57.5%
graph-classification-on-reddit-bGIN-0
Accuracy: 92.4
graph-property-prediction-on-ogbg-code2GIN+virtual node
Ext. data: No
Number of params: 13841815
Test F1 score: 0.1581 ± 0.0026
Validation F1 score: 0.1439 ± 0.0020
graph-property-prediction-on-ogbg-code2GIN
Ext. data: No
Number of params: 12390715
Test F1 score: 0.1495 ± 0.0023
Validation F1 score: 0.1376 ± 0.0016
graph-property-prediction-on-ogbg-molhivGIN+virtual node
Ext. data: No
Number of params: 3336306
Test ROC-AUC: 0.7707 ± 0.0149
Validation ROC-AUC: 0.8479 ± 0.0068
graph-property-prediction-on-ogbg-molhivGIN
Ext. data: No
Number of params: 1885206
Test ROC-AUC: 0.7558 ± 0.0140
Validation ROC-AUC: 0.8232 ± 0.0090
graph-property-prediction-on-ogbg-molpcbaGIN+virtual node
Ext. data: No
Number of params: 3374533
Test AP: 0.2703 ± 0.0023
Validation AP: 0.2798 ± 0.0025
graph-property-prediction-on-ogbg-molpcbaGIN
Ext. data: No
Number of params: 1923433
Test AP: 0.2266 ± 0.0028
Validation AP: 0.2305 ± 0.0027
graph-property-prediction-on-ogbg-ppaGIN+virtual node
Ext. data: No
Number of params: 3288042
Test Accuracy: 0.7037 ± 0.0107
Validation Accuracy: 0.6678 ± 0.0105
graph-property-prediction-on-ogbg-ppaGIN
Ext. data: No
Number of params: 1836942
Test Accuracy: 0.6892 ± 0.0100
Validation Accuracy: 0.6562 ± 0.0107
graph-regression-on-esr2GIN
R2: 0.668±0.000
RMSE: 0.509±0.668
graph-regression-on-f2GIN
R2: 0.887±0.000
RMSE: 0.342±0.887
graph-regression-on-kitGIN
R2: 0.833±0.000
RMSE: 0.444±0.833
graph-regression-on-lipophilicityGIN
R2: 0.819±0.007
RMSE: 0.537±0.010
graph-regression-on-parp1GIN
R2: 0.922±0.000
RMSE: 0.349±0.922
graph-regression-on-pcqm4mv2-lscGIN
Test MAE: 0.1218
Validation MAE: 0.1195
graph-regression-on-pgrGIN
R2: 0.696±0.000
RMSE: 0.532±0.696
graph-regression-on-zinc-500kGIN
MAE: 0.526
graph-regression-on-zinc-fullGIN
Test MAE: 0.068±0.004
molecular-property-prediction-on-esolGIN
R2: 0.938±0.011
RMSE: 0.509±0.044
molecular-property-prediction-on-freesolvGIN
R2: 0.964±0.008
RMSE: 0.744±0.083
node-classification-on-pattern-100kGIN
Accuracy (%): 85.590

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
图神经网络有多强大? | 论文 | HyperAI超神经