摘要
在本研究中，我们提出了一种新颖且统一的图神经网络架构——图星网（GraphStar），该架构利用消息传递中继和注意力机制来完成多个预测任务，包括节点分类、图分类和链接预测。图星网解决了许多早期图神经网络面临的挑战，并在不增加模型深度或承担沉重计算成本的情况下实现了非局部表示。我们还提出了一种基于节点分类和文本分类作为图分类的新方法来解决特定主题的情感分析问题。我们的研究表明，“星节点”可以学习有效的图数据表示，并在这三项任务上改进了现有方法。具体而言，在图分类和链接预测方面，图星网在几个关键基准测试中的表现优于当前最先进的模型2-5%。
代码仓库
graph-star-team/graph_star
pytorch
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-dd | GraphStar | Accuracy: 79.60% |
| graph-classification-on-enzymes | GraphStar | Accuracy: 67.1% |
| graph-classification-on-mutag | GraphStar | Accuracy: 91.2% |
| graph-classification-on-proteins | GraphStar | Accuracy: 77.90% |
| link-prediction-on-citeseer-biased-evaluation | GraphStar (double weight on positive examples) | AP: 97.93 AUC: 97.47 Accuracy: 97.7 |
| link-prediction-on-cora-biased-evaluation | GraphStar (double weight on positive examples) | AP: 96.15 AUC: 95.65 Accuracy: 95.9 |
| link-prediction-on-pubmed-biased-evaluation | GraphStar (double weight on positive examples) | AP: 98.64 AUC: 97.67 Accuracy: 98.16 |
| node-classification-on-citeseer | GraphStar | Accuracy: 71.0 |
| node-classification-on-cora | GraphStar | Accuracy: 82.1% |
| node-classification-on-ppi | GraphStar | F1: 99.4 |
| node-classification-on-pubmed | GraphStar | Accuracy: 77.2% |
| sentiment-analysis-on-imdb | GraphStar | Accuracy: 96.0 |
| sentiment-analysis-on-mr | GraphStar | Accuracy: 76.6 |
| text-classification-on-20news | GraphStar | Accuracy: 86.9 |
| text-classification-on-ohsumed | GraphStar | Accuracy: 64.2 |
| text-classification-on-r52 | GraphStar | Accuracy: 95.00 |
| text-classification-on-r8 | GraphStar | Accuracy: 97.4 |