
摘要
数据增强通过扩充训练集,有助于神经网络提升泛化能力,但如何有效增强图数据以提升图神经网络(GNN)的性能,仍是当前尚未解决的关键问题。现有大多数图正则化方法主要通过增删边来调整图的拓扑结构,而本文提出一种新方法:通过增强节点特征来提升模型性能。我们提出了FLAG(Free Large-scale Adversarial Augmentation on Graphs),该方法在训练过程中基于梯度生成对抗性扰动,迭代地对节点特征进行增强。通过使模型对输入数据的微小波动保持不变性,该方法显著提升了模型在分布外样本上的泛化能力,并在测试阶段有效增强了模型性能。FLAG是一种通用的图数据增强方法,适用于节点分类、链接预测和图分类等多种任务。该方法具有高度的灵活性与可扩展性,可与任意GNN主干网络结合,并适用于大规模数据集。通过大量实验与消融研究,我们验证了该方法的有效性与稳定性。此外,我们还提供了直观的观察结果,以帮助深入理解该方法的工作机制。
代码仓库
ytchx1999/GCN_res-FLAG
pytorch
GitHub 中提及
devnkong/FLAG
官方
pytorch
GitHub 中提及
sangyx/gtrick/tree/main/benchmark/pyg
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-property-prediction-on-ogbg-molhiv | GCN+FLAG | Ext. data: No Number of params: 527701 Test ROC-AUC: 0.7683 ± 0.0102 Validation ROC-AUC: 0.8176 ± 0.0087 |
| graph-property-prediction-on-ogbg-molhiv | DeeperGCN+FLAG | Ext. data: No Number of params: 531976 Test ROC-AUC: 0.7942 ± 0.0120 Validation ROC-AUC: 0.8425 ± 0.0061 |
| graph-property-prediction-on-ogbg-molhiv | GIN+virtual node+FLAG | Ext. data: No Number of params: 3336306 Test ROC-AUC: 0.7748 ± 0.0096 Validation ROC-AUC: 0.8438 ± 0.0128 |
| graph-property-prediction-on-ogbg-molhiv | GIN+FLAG | Ext. data: No Number of params: 1885206 Test ROC-AUC: 0.7654 ± 0.0114 Validation ROC-AUC: 0.8225 ± 0.0155 |
| graph-property-prediction-on-ogbg-molpcba | GCN+virtual node+FLAG | Ext. data: No Number of params: 2017028 Test AP: 0.2483 ± 0.0037 Validation AP: 0.2556 ± 0.0040 |
| graph-property-prediction-on-ogbg-molpcba | GIN+FLAG | Ext. data: No Number of params: 1923433 Test AP: 0.2395 ± 0.0040 Validation AP: 0.2451 ± 0.0042 |
| graph-property-prediction-on-ogbg-molpcba | GCN+FLAG | Ext. data: No Number of params: 565928 Test AP: 0.2116 ± 0.0017 Validation AP: 0.2150 ± 0.0022 |
| graph-property-prediction-on-ogbg-molpcba | DeeperGCN+virtual node+FLAG | Ext. data: No Number of params: 5550208 Test AP: 0.2842 ± 0.0043 Validation AP: 0.2952 ± 0.0029 |
| graph-property-prediction-on-ogbg-molpcba | GIN+virtual node+FLAG | Ext. data: No Number of params: 3374533 Test AP: 0.2834 ± 0.0038 Validation AP: 0.2912 ± 0.0026 |
| graph-property-prediction-on-ogbg-ppa | DeeperGCN+FLAG | Ext. data: No Number of params: 2336421 Test Accuracy: 0.7752 ± 0.0069 Validation Accuracy: 0.7484 ± 0.0052 |
| graph-property-prediction-on-ogbg-ppa | GIN+FLAG | Ext. data: No Number of params: 1836942 Test Accuracy: 0.6905 ± 0.0092 Validation Accuracy: 0.6465 ± 0.0070 |
| graph-property-prediction-on-ogbg-ppa | GCN+virtual node+FLAG | Ext. data: No Number of params: 1930537 Test Accuracy: 0.6944 ± 0.0052 Validation Accuracy: 0.6638 ± 0.0055 |
| graph-property-prediction-on-ogbg-ppa | GIN+virtual node+FLAG | Ext. data: No Number of params: 3288042 Test Accuracy: 0.7245 ± 0.0114 Validation Accuracy: 0.6789 ± 0.0079 |