Semi-Supervised Classification with Graph Convolutional Networks

9 Sep 2016  ·  Thomas N. Kipf, Max Welling ·

We present a scalable approach for semi-supervised learning on graph-structured data that is based on an efficient variant of convolutional neural networks which operate directly on graphs. We motivate the choice of our convolutional architecture via a localized first-order approximation of spectral graph convolutions. Our model scales linearly in the number of graph edges and learns hidden layer representations that encode both local graph structure and features of nodes. In a number of experiments on citation networks and on a knowledge graph dataset we demonstrate that our approach outperforms related methods by a significant margin.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Graph Classification BP-fMRI-97 GCN Accuracy 60.7% # 4
F1 61.2% # 5
Node Classification Brazil Air-Traffic GCN (Kipf and Welling, 2017) Accuracy 0.432 # 5
Graph Classification CIFAR10 100k GCN Accuracy (%) 54.46 # 8
Node Classification Citeseer GCN Accuracy 70.3 # 54
Training Split fixed 20 per node # 3
Validation YES # 1
Node Classification CiteSeer (0.5%) GCN Accuracy 43.6% # 13
Node Classification CiteSeer (1%) GCN Accuracy 55.3% # 12
Node Classification Citeseer Full-supervised GCN Accuracy 79.34% # 3
Node Classification CiteSeer with Public Split: fixed 20 nodes per class GCN Accuracy 68.7 # 28
Node Classification Cora GCN Accuracy 83.0% # 40
Validation YES # 1
Document Classification Cora Graph-CNN Accuracy 81.5% # 5
Node Classification Cora (0.5%) GCN Accuracy 50.9% # 11
Node Classification Cora (1%) GCN Accuracy 62.3% # 11
Node Classification Cora (3%) GCN Accuracy 76.5% # 10
Node Classification Cora Full-supervised GCN Accuracy 86.64% # 5
Node Classification Cora with Public Split: fixed 20 nodes per class GCN Accuracy 80.5% # 25
Graph Classification HIV-DTI-77 GCN Accuracy 57.7% # 4
F1 54.4% # 5
Graph Classification HIV-fMRI-77 GCN Accuracy 58.3 # 2
F1 56.4& # 2
Graph Classification IPC-grounded GCN Accuracy 80.7% # 1
Graph Classification IPC-lifted GCN Accuracy 87.6% # 1
Graph Regression Lipophilicity GCN RMSE@80%Train 1.05 # 1
Node Classification NELL GCN Accuracy 66.0 # 2
Graph Property Prediction ogbg-code2 GCN+virtual node Test F1 score 0.1595 ± 0.0018 # 6
Validation F1 score 0.1461 ± 0.0013 # 7
Number of params 12484310 # 4
Ext. data No # 1
Graph Property Prediction ogbg-code2 GCN Test F1 score 0.1507 ± 0.0018 # 12
Validation F1 score 0.1399 ± 0.0017 # 12
Number of params 11033210 # 7
Ext. data No # 1
Graph Property Prediction ogbg-molhiv GCN+virtual node Test ROC-AUC 0.7599 ± 0.0119 # 34
Validation ROC-AUC 0.8384 ± 0.0091 # 12
Number of params 1978801 # 11
Ext. data No # 1
Graph Property Prediction ogbg-molhiv GCN Test ROC-AUC 0.7606 ± 0.0097 # 33
Validation ROC-AUC 0.8204 ± 0.0141 # 28
Number of params 527701 # 18
Ext. data No # 1
Graph Property Prediction ogbg-molhiv GCN (in Julia) Test ROC-AUC 0.7549 ± 0.0163 # 36
Validation ROC-AUC 0.8042 ± 0.0107 # 33
Number of params 527701 # 18
Ext. data No # 1
Graph Property Prediction ogbg-molpcba GCN+virtual node Test AP 0.2424 ± 0.0034 # 22
Validation AP 0.2495 ± 0.0042 # 22
Number of params 2017028 # 19
Ext. data No # 1
Graph Property Prediction ogbg-molpcba GCN Test AP 0.2020 ± 0.0024 # 28
Validation AP 0.2059 ± 0.0033 # 28
Number of params 565928 # 25
Ext. data No # 1
Graph Property Prediction ogbg-ppa GCN+virtual node Test Accuracy 0.6857 ± 0.0061 # 12
Validation Accuracy 0.6511 ± 0.0048 # 11
Number of params 1930537 # 8
Ext. data No # 1
Graph Property Prediction ogbg-ppa GCN Test Accuracy 0.6839 ± 0.0084 # 13
Validation Accuracy 0.6497 ± 0.0034 # 12
Number of params 479437 # 13
Ext. data No # 1
Link Property Prediction ogbl-citation2 Full-batch GCN Test MRR 0.8474 ± 0.0021 # 4
Validation MRR 0.8479 ± 0.0023 # 4
Number of params 296449 # 7
Ext. data No # 1
Link Property Prediction ogbl-collab GCN (val as input) Test Hits@50 0.4714 ± 0.0145 # 16
Validation Hits@50 0.5263 ± 0.0115 # 15
Number of params 296449 # 13
Ext. data No # 1
Link Property Prediction ogbl-collab GCN Test Hits@50 0.4475 ± 0.0107 # 18
Validation Hits@50 0.5263 ± 0.0115 # 15
Number of params 296449 # 13
Ext. data No # 1
Link Property Prediction ogbl-ddi GCN Test Hits@20 0.3707 ± 0.0507 # 12
Validation Hits@20 0.5550 ± 0.0208 # 12
Number of params 1289985 # 11
Ext. data No # 1
Link Property Prediction ogbl-ddi GCN+JKNet Test Hits@20 0.6056 ± 0.0869 # 10
Validation Hits@20 0.6776 ± 0.0095 # 7
Number of params 1421571 # 8
Ext. data No # 1
Link Property Prediction ogbl-ppa GCN Test Hits@100 0.1867 ± 0.0132 # 10
Validation Hits@100 0.1845 ± 0.0140 # 9
Number of params 278529 # 6
Ext. data No # 1
Node Property Prediction ogbn-arxiv GCN+residual+node2vec Test Accuracy 0.7278 ± 0.0013 # 33
Validation Accuracy 0.7414 ± 0.0008 # 30