摘要
本文提出了一种统一的框架——深度图核（Deep Graph Kernels, DGK），旨在借鉴自然语言建模与深度学习领域的最新进展，学习图结构中子结构的潜在表示。该框架通过学习子结构的潜在表示，有效捕捉子结构之间的依赖关系。我们在三种主流图核方法上验证了该框架的实例化，分别为图小波核（Graphlet kernels）、Weisfeiler-Lehman子树核（Weisfeiler-Lehman subtree kernels）以及最短路径图核（Shortest-Path graph kernels）。在多个基准数据集上的实验结果表明，深度图核在分类准确率方面显著优于当前最先进的图核方法。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-collab | DGK | Accuracy: 73.09% |
| graph-classification-on-dd | DGK | Accuracy: 73.50% |
| graph-classification-on-enzymes | DGK | Accuracy: 53.43% |
| graph-classification-on-imdb-b | DGK | Accuracy: 66.96% |
| graph-classification-on-imdb-m | DGK | Accuracy: 44.55% |
| graph-classification-on-mutag | DGK | Accuracy: 87.44% |
| graph-classification-on-proteins | DGK | Accuracy: 75.68% |
| graph-classification-on-re-m12k | DGK | Accuracy: 32.22% |
| graph-classification-on-re-m5k | DGK | Accuracy: 41.27% |
| malware-detection-on-android-malware-dataset | Deep WL kernel | Accuracy: 98.16 |