摘要
我们提出了一种基于Transformer的图神经网络模型，命名为UGformer，用于学习图结构表示。具体而言，本文介绍了UGformer的两种变体：第一种变体（于2019年9月发布）在每个输入节点的采样邻居集合上应用Transformer；第二种变体（于2021年5月发布）则在所有输入节点上应用Transformer。实验结果表明，第一种UGformer变体在图分类任务的归纳设置（inductive setting）以及无监督直推设置（unsupervised transductive setting）下的基准数据集上均达到了当前最优的准确率；第二种UGformer变体在归纳文本分类任务中也取得了当前最优的性能表现。代码已公开，获取地址为：\url{https://github.com/daiquocnguyen/Graph-Transformer}。
代码仓库
daiquocnguyen/Graph-Transformer
官方
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-collab | U2GNN (Unsupervised) | Accuracy: 95.62% |
| graph-classification-on-collab | U2GNN | Accuracy: 77.84% |
| graph-classification-on-dd | U2GNN (Unsupervised) | Accuracy: 95.67% |
| graph-classification-on-dd | U2GNN | Accuracy: 80.23% |
| graph-classification-on-imdb-b | U2GNN | Accuracy: 77.04% |
| graph-classification-on-imdb-b | U2GNN (Unsupervised) | Accuracy: 96.41% |
| graph-classification-on-imdb-m | U2GNN (Unsupervised) | Accuracy: 89.2% |
| graph-classification-on-imdb-m | U2GNN | Accuracy: 53.60% |
| graph-classification-on-mutag | U2GNN (Unsupervised) | Accuracy: 88.47% |
| graph-classification-on-mutag | U2GNN | Accuracy: 89.97% |
| graph-classification-on-proteins | U2GNN | Accuracy: 78.53% |
| graph-classification-on-proteins | U2GNN (Unsupervised) | Accuracy: 80.01% |
| graph-classification-on-ptc | U2GNN (Unsupervised) | Accuracy: 91.81% |
| graph-classification-on-ptc | U2GNN | Accuracy: 69.63% |