摘要
关系图卷积网络（Relational Graph Convolutional Network, R-GCN）的提出在语义网领域具有里程碑意义，作为一种被广泛引用的方法，它将端到端的层次化表示学习范式推广至知识图谱（Knowledge Graphs, KGs）领域。R-GCN通过反复聚合邻居节点经关系特定参数化变换后的表示，生成目标节点的嵌入表示。然而，在本文中，我们主张R-GCN的核心贡献并非其学习得到的权重，而在于其“消息传递”（message passing）的计算范式。为此，我们提出“随机关系图卷积网络”（Random Relational Graph Convolutional Network, RR-GCN），该方法不进行任何参数训练，而是通过聚合来自邻居节点的随机变换后的随机表示来构建节点嵌入，即完全不依赖学习参数。实验结果表明，RR-GCN在节点分类与链接预测任务中均能与完全训练的R-GCN相媲美。
代码仓库
predict-idlab/RR-GCN
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| link-prediction-on-fb15k-237 | RR-GCN-PPV | Hits@1: 0.157 Hits@10: 0.412 Hits@3: 0.256 MRR: 0.238 |
| node-classification-on-aifb | RR-GCN-PPV | Accuracy: 86.11 |
| node-classification-on-aifb | RR-GCN-PPV-CUT | Accuracy: 95.83 |
| node-classification-on-am | RR-GCN-PPV | Accuracy: 84.65 |
| node-classification-on-am | RR-GCN-PPV-CUT | Accuracy: 84.8 |
| node-classification-on-am | RR-GCN-PPV-CUT (Unimportant relations removed) | Accuracy: 91.31 |
| node-classification-on-amplus | RR-GCN-PPV | Accuracy: 84.54 |
| node-classification-on-amplus | R-GCN | Accuracy: 83.81 |
| node-classification-on-bgs | RR-GCN-PPV-CUT | Accuracy: 84.14 |
| node-classification-on-bgs | RR-GCN-PPV | Accuracy: 78.97 |
| node-classification-on-dblp | R-GCN | Accuracy: 68.51 |
| node-classification-on-dblp | RR-GCN-PPV | Accuracy: 70.61 |
| node-classification-on-dmg777k | R-GCN | Accuracy: 62.51 |
| node-classification-on-dmg777k | RR-GCN-PPV | Accuracy: 63.97 |
| node-classification-on-dmgfull | RR-GCN-PPV | Accuracy: 63.38 |
| node-classification-on-dmgfull | R-GCN | Accuracy: 57.52 |
| node-classification-on-mdgenre | R-GCN | Accuracy: 67.33 |
| node-classification-on-mdgenre | RR-GCN-PPV | Accuracy: 67.15 |
| node-classification-on-mutag | RR-GCN-PPV | Accuracy: 79.41 |