摘要
我们提出了一种适用于图结构数据的半监督学习的可扩展方法，该方法基于一种直接在图上运行的高效卷积神经网络变体。我们通过局部一阶近似谱图卷积（spectral graph convolutions）来解释选择这种卷积架构的原因。我们的模型在图边的数量上呈线性扩展，并且学习到的隐藏层表示能够同时编码局部图结构和节点特征。在多个引文网络和知识图数据集上的实验表明，我们的方法显著优于相关方法。
代码仓库
kiharalab/gnn_pocket
pytorch
GitHub 中提及
udlf/wsef
pytorch
GitHub 中提及
lipingcoding/pygcn
pytorch
GitHub 中提及
mahsa91/RA-GCN
pytorch
GitHub 中提及
deepchem/moleculenet
pytorch
GitHub 中提及
yangjun1994/CAGCN
GitHub 中提及
tk-rusch/graphcon
pytorch
GitHub 中提及
alexOarga/haiku-geometric
jax
GitHub 中提及
fanzhenliu/dagad
pytorch
GitHub 中提及
tkipf/gcn
tf
GitHub 中提及
asif-salim/GCN-Graph_Convolutional_Network
tf
GitHub 中提及
emartinezs44/SparkGCN
GitHub 中提及
basiralab/Reproducible-Generative-Learning
pytorch
GitHub 中提及
KimMeen/GCN
pytorch
GitHub 中提及
thanhtrunghuynh93/pygcn
pytorch
GitHub 中提及
clin366/pygcn
pytorch
GitHub 中提及
tkipf/pygcn
官方
pytorch
GitHub 中提及
negarhdr/PGCN
pytorch
GitHub 中提及
toinesayan/node-classification-and-label-dependencies
pytorch
GitHub 中提及
jiangboahu/glcn-tf
tf
GitHub 中提及
bcsrn/gcn
pytorch
GitHub 中提及
HoganZhang/pygcn_python3
pytorch
GitHub 中提及
HamzaGbada/GCN-Numpy
pytorch
yCobanoglu/infinite-width-gnns
jax
GitHub 中提及
basiralab/reproduciblefedgnn
pytorch
GitHub 中提及
mahendrathapa/graph-convolution-network
pytorch
GitHub 中提及
LouisDumont/GCN---re-implementation
pytorch
GitHub 中提及
darnbi/pygcn
pytorch
GitHub 中提及
1075225782/GCN
pytorch
GitHub 中提及
hazdzz/GCN
pytorch
GitHub 中提及
tuananh0305/GCN_ANOMALY_DETECTION
GitHub 中提及
zhiqiang00/Hon-GCN
pytorch
GitHub 中提及
bepping/non-oversmoothing-gcns
GitHub 中提及
meliketoy/graph-cnn.pytorch
pytorch
GitHub 中提及
basiralab/RG-Select
pytorch
GitHub 中提及
ajbisberg/gcn
tf
GitHub 中提及
cybermonic/cage-4-submission
pytorch
GitHub 中提及
tkipf/keras-gcn
tf
GitHub 中提及
mahsa91/GKD
pytorch
GitHub 中提及
nieci2024/pyglcn
pytorch
GitHub 中提及
abi-kothapalli/minimumdominatingsets
tf
GitHub 中提及
nnaakkaaii/g2-MLP
pytorch
GitHub 中提及
ykrmm/sota_gnn
pytorch
GitHub 中提及
IllinoisGraphBenchmark/IGB-Datasets
pytorch
GitHub 中提及
selmiss/gp-tlstgcn
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-dd | GCN | Accuracy: 78.151±3.465 |
| graph-classification-on-enzymes | GCN | Accuracy: 73.466±4.372 |
| graph-classification-on-imdb-b | GCN | Accuracy: 79.500±3.109 |
| graph-classification-on-nci1 | GCN | Accuracy: 84.185±0.644 |
| graph-classification-on-nci109 | GCN | Accuracy: 83.140±1.248 |
| graph-classification-on-proteins | GCN | Accuracy: 75.536±1.622 |
| graph-property-prediction-on-ogbg-code2 | GCN | Ext. data: No Number of params: 11033210 Test F1 score: 0.1507 ± 0.0018 Validation F1 score: 0.1399 ± 0.0017 |
| graph-property-prediction-on-ogbg-code2 | GCN+virtual node | Ext. data: No Number of params: 12484310 Test F1 score: 0.1595 ± 0.0018 Validation F1 score: 0.1461 ± 0.0013 |
| graph-property-prediction-on-ogbg-molhiv | GCN (in Julia) | Ext. data: No Number of params: 527701 Test ROC-AUC: 0.7549 ± 0.0163 Validation ROC-AUC: 0.8042 ± 0.0107 |
| graph-property-prediction-on-ogbg-molhiv | GCN+virtual node | Ext. data: No Number of params: 1978801 Test ROC-AUC: 0.7599 ± 0.0119 Validation ROC-AUC: 0.8384 ± 0.0091 |
| graph-property-prediction-on-ogbg-molhiv | GCN | Ext. data: No Number of params: 527701 Test ROC-AUC: 0.7606 ± 0.0097 Validation ROC-AUC: 0.8204 ± 0.0141 |
| graph-property-prediction-on-ogbg-molpcba | GCN | Ext. data: No Number of params: 565928 Test AP: 0.2020 ± 0.0024 Validation AP: 0.2059 ± 0.0033 |
| graph-property-prediction-on-ogbg-molpcba | GCN+virtual node | Ext. data: No Number of params: 2017028 Test AP: 0.2424 ± 0.0034 Validation AP: 0.2495 ± 0.0042 |
| graph-property-prediction-on-ogbg-ppa | GCN | Ext. data: No Number of params: 479437 Test Accuracy: 0.6839 ± 0.0084 Validation Accuracy: 0.6497 ± 0.0034 |
| graph-property-prediction-on-ogbg-ppa | GCN+virtual node | Ext. data: No Number of params: 1930537 Test Accuracy: 0.6857 ± 0.0061 Validation Accuracy: 0.6511 ± 0.0048 |
| graph-regression-on-esr2 | GCN | R2: 0.642±0.000 RMSE: 0.528±0.642 |
| graph-regression-on-f2 | GCN | R2: 0.878±0.000 RMSE: 0.355±0.878 |
| graph-regression-on-kit | GCN | R2: 0.814±0.000 RMSE: 0.469±0.814 |
| graph-regression-on-lipophilicity | GCN | R2: 0.800±0.008 RMSE: 0.565±0.011 |
| graph-regression-on-parp1 | GCN | R2: 0.912±0.000 RMSE: 0.372±0.912 |
| graph-regression-on-pcqm4mv2-lsc | GCN | Test MAE: 0.1398 Validation MAE: 0.1379 |
| graph-regression-on-pgr | GCN | R2: 0.658±0.000 RMSE: 0.565±0.658 |
| graph-regression-on-zinc-full | GCN | Test MAE: 0.152±0.023 |
| heterogeneous-node-classification-on-acm | GCN | Macro-F1: 92.17 Micro-F1: 92.12 |
| heterogeneous-node-classification-on-dblp-2 | GCN | Macro-F1: 90.84 Micro-F1: 91.47 |
| heterogeneous-node-classification-on-freebase | GCN | Macro-F1: 27.84 Micro-F1: 60.23 |
| heterogeneous-node-classification-on-imdb | GCN | Macro-F1: 57.88 Micro-F1: 64.82 |
| link-property-prediction-on-ogbl-citation2 | Full-batch GCN | Ext. data: No Number of params: 296449 Test MRR: 0.8474 ± 0.0021 Validation MRR: 0.8479 ± 0.0023 |
| link-property-prediction-on-ogbl-collab | GCN (val as input) | Ext. data: No Number of params: 296449 Test Hits@50: 0.4714 ± 0.0145 Validation Hits@50: 0.5263 ± 0.0115 |
| link-property-prediction-on-ogbl-collab | GCN | Ext. data: No Number of params: 296449 Test Hits@50: 0.4475 ± 0.0107 Validation Hits@50: 0.5263 ± 0.0115 |
| link-property-prediction-on-ogbl-ddi | GCN+JKNet | Ext. data: No Number of params: 1421571 Test Hits@20: 0.6056 ± 0.0869 Validation Hits@20: 0.6776 ± 0.0095 |
| link-property-prediction-on-ogbl-ddi | GCN | Ext. data: No Number of params: 1289985 Test Hits@20: 0.3707 ± 0.0507 Validation Hits@20: 0.5550 ± 0.0208 |
| link-property-prediction-on-ogbl-ppa | GCN | Ext. data: No Number of params: 278529 Test Hits@100: 0.1867 ± 0.0132 Validation Hits@100: 0.1845 ± 0.0140 |
| molecular-property-prediction-on-esol | GCN | R2: 0.936±0.006 RMSE: 0.520±0.024 |
| molecular-property-prediction-on-freesolv | GCN | R2: 0.957±0.009 RMSE: 0.815±0.086 |
| node-classification-on-brazil-air-traffic | GCN_cheby (Kipf and Welling, 2017) | Accuracy: 0.516 |
| node-classification-on-chameleon-60-20-20 | GCN | 1:1 Accuracy: 64.18 ± 2.62 |
| node-classification-on-citeseer | GCN | Accuracy: 70.3 |
| node-classification-on-citeseer-60-20-20 | GCN | 1:1 Accuracy: 81.39 ± 1.23 |
| node-classification-on-cora | GCN | Accuracy: 81.5% |
| node-classification-on-cora-60-20-20-random | GCN | 1:1 Accuracy: 87.78 ± 0.96 |
| node-classification-on-cornell-60-20-20 | GCN | 1:1 Accuracy: 82.46 ± 3.11 |
| node-classification-on-europe-air-traffic | GCN_cheby (Kipf and Welling, 2017) | Accuracy: 46.0 |
| node-classification-on-europe-air-traffic | GCN (Kipf and Welling, 2017) | Accuracy: 37.1 |
| node-classification-on-facebook | GCN_cheby (Kipf and Welling, 2017) | Accuracy: 64.6 |
| node-classification-on-facebook | GCN (Kipf and Welling, 2017) | Accuracy: 57.5 |
| node-classification-on-film-60-20-20-random | GCN | 1:1 Accuracy: 35.51 ± 0.99 |
| node-classification-on-flickr | GCN_cheby (Kipf and Welling, 2017) | Accuracy: 0.479 |
| node-classification-on-flickr | GCN (Kipf and Welling, 2017) | Accuracy: 0.546 |
| node-classification-on-genius | GCN | Accuracy: 87.42 ± 0.37 |
| node-classification-on-nell | GCN | Accuracy: 66.0 |
| node-classification-on-non-homophilic | GCN | 1:1 Accuracy: 82.46 ± 3.11 |
| node-classification-on-non-homophilic-1 | GCN | 1:1 Accuracy: 75.5 ± 2.92 |
| node-classification-on-non-homophilic-13 | GCN | 1:1 Accuracy: 82.47 ± 0.27 |
| node-classification-on-non-homophilic-14 | GCN | 1:1 Accuracy: 87.42 ± 0.37 |
| node-classification-on-non-homophilic-15 | GCN | 1:1 Accuracy: 62.18 ± 0.26 |
| node-classification-on-non-homophilic-2 | GCN | 1:1 Accuracy: 83.11 ± 3.2 |
| node-classification-on-non-homophilic-4 | GCN | 1:1 Accuracy: 64.18 ± 2.62 |
| node-classification-on-non-homophilic-6 | GCN | 1:1 Accuracy: 62.23±0.53 |
| node-classification-on-penn94 | GCN | Accuracy: 82.47 ± 0.27 |
| node-classification-on-pubmed | GCN | Accuracy: 79.0 |
| node-classification-on-pubmed-60-20-20-random | GCN | 1:1 Accuracy: 88.9 ± 0.32 |
| node-classification-on-squirrel-60-20-20 | GCN | 1:1 Accuracy: 44.76 ± 1.39 |
| node-classification-on-texas-60-20-20-random | GCN | 1:1 Accuracy: 83.11 ± 3.2 |
| node-classification-on-wiki-vote | GCN_cheby (Kipf and Welling, 2017) | Accuracy: 49.5 |
| node-classification-on-wiki-vote | GCN (Kipf and Welling, 2017) | Accuracy: 32.9 |
| node-classification-on-wisconsin-60-20-20 | GCN | 1:1 Accuracy: 75.5 ± 2.92 |