
摘要
近年来,提出了将深度学习应用于图等结构化数据的先进方法。特别是,研究重点放在了将卷积神经网络推广到图数据上,这包括重新定义适用于图的卷积和下采样(池化)操作。将卷积操作推广到图的方法已被证明可以提高性能,并且得到了广泛应用。然而,将下采样应用于图的方法仍然难以实现,并且有改进的空间。在本文中,我们提出了一种基于自注意力机制的图池化方法。利用图卷积进行自注意力计算,使得我们的池化方法能够同时考虑节点特征和图拓扑结构。为了确保公平比较,现有的池化方法和我们的方法采用了相同的训练流程和模型架构。实验结果表明,我们的方法在基准数据集上使用合理的参数数量实现了优越的图分类性能。
代码仓库
inyeoplee77/SAGPool
官方
pytorch
GitHub 中提及
walidgeuttala/Synthetic-Benchmark-for-Graph-Classification
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| graph-classification-on-dd | SAGPool_h | Accuracy: 76.45% |
| graph-classification-on-dd | SAGPool_g | Accuracy: 76.19% |
| graph-classification-on-frankenstein | SAGPool_g | Accuracy: 62.57 |
| graph-classification-on-frankenstein | SAGPool_h | Accuracy: 61.73 |
| graph-classification-on-nci1 | SAGPool_g | Accuracy: 74.06% |
| graph-classification-on-nci1 | SAGPool_h | Accuracy: 67.45% |
| graph-classification-on-nci109 | SAGPool_h | Accuracy: 67.86 |
| graph-classification-on-nci109 | SAGPool_g | Accuracy: 74.06 |
| graph-classification-on-proteins | SAGPool_g | Accuracy: 70.04% |
| graph-classification-on-proteins | SAGPool_h | Accuracy: 71.86% |