4 个月前

简化图卷积网络

简化图卷积网络

摘要

图卷积网络(GCNs)及其变体近年来受到了广泛关注,并已成为学习图表示的主流方法。GCNs主要受到近期深度学习方法的启发,因此可能继承了不必要的复杂性和冗余计算。在本文中,我们通过依次移除非线性层和合并连续层之间的权重矩阵来减少这种过度复杂性。我们对所得的线性模型进行了理论分析,表明其相当于一个固定的低通滤波器后接一个线性分类器。值得注意的是,我们的实验评估证明,这些简化措施在许多下游应用中并未对准确性产生负面影响。此外,所得模型可以扩展到更大的数据集,具有天然的可解释性,并且比FastGCN的速度提高了多达两个数量级。

代码仓库

hazdzz/SGC
pytorch
GitHub 中提及
ydtydr/hyla
pytorch
GitHub 中提及
changminwu/expandergnn
pytorch
GitHub 中提及
Tiiiger/SGC
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
graph-regression-on-lipophilicitySGC
RMSE: 0.998
node-classification-on-chameleon-60-20-20SGC-2
1:1 Accuracy: 62.67 ± 2.41
node-classification-on-chameleon-60-20-20SGC-1
1:1 Accuracy: 64.86 ± 1.81
node-classification-on-citeseer-60-20-20SGC-2
1:1 Accuracy: 80.75 ± 1.15
node-classification-on-citeseer-60-20-20SGC-1
1:1 Accuracy: 79.66 ± 0.75
node-classification-on-cora-60-20-20-randomSGC-1
1:1 Accuracy: 85.12 ± 1.64
node-classification-on-cora-60-20-20-randomSGC-2
1:1 Accuracy: 85.48 ± 1.48
node-classification-on-cornell-60-20-20SGC-1
1:1 Accuracy: 70.98 ± 8.39
node-classification-on-cornell-60-20-20SGC-2
1:1 Accuracy: 72.62 ± 9.92
node-classification-on-film-60-20-20-randomSGC-2
1:1 Accuracy: 28.81 ± 1.11
node-classification-on-film-60-20-20-randomSGC-1
1:1 Accuracy: 25.26 ± 1.18
node-classification-on-geniusSGC 2-hop
Accuracy: 82.10 ± 0.14
node-classification-on-geniusSGC 1-hop
Accuracy: 82.36 ± 0.37
node-classification-on-non-homophilicSGC-1
1:1 Accuracy: 70.98 ± 8.39
node-classification-on-non-homophilicSGC-2
1:1 Accuracy: 72.62 ± 9.92
node-classification-on-non-homophilic-1SGC-1
1:1 Accuracy: 70.38 ± 2.85
node-classification-on-non-homophilic-1SGC-2
1:1 Accuracy: 74.75 ± 2.89
node-classification-on-non-homophilic-13SGC 2-hop
1:1 Accuracy: 76.09 ± 0.45
node-classification-on-non-homophilic-13SGC 1-hop
1:1 Accuracy: 66.79 ± 0.27
node-classification-on-non-homophilic-14SGC 2-hop
1:1 Accuracy: 82.10 ± 0.14
node-classification-on-non-homophilic-14SGC 1-hop
1:1 Accuracy: 82.36 ± 0.37
node-classification-on-non-homophilic-15SGC 2-hop
1:1 Accuracy: 59.94 ± 0.21
node-classification-on-non-homophilic-15SGC 1-hop
1:1 Accuracy: 58.97 ± 0.19
node-classification-on-non-homophilic-2SGC-2
1:1 Accuracy: 81.31 ± 3.3
node-classification-on-non-homophilic-2SGC-1
1:1 Accuracy: 83.28 ± 5.43
node-classification-on-non-homophilic-4SGC-2
1:1 Accuracy: 62.67 ± 2.41
node-classification-on-non-homophilic-4SGC-1
1:1 Accuracy: 64.86 ± 1.81
node-classification-on-non-homophilic-6SGC-1
1:1 Accuracy: 59.73±0.12
node-classification-on-penn94SGC 2-hop
Accuracy: 76.09 ± 0.45
node-classification-on-penn94SGC 1-hop
Accuracy: 66.79 ± 0.27
node-classification-on-pubmed-60-20-20-randomSGC-1
1:1 Accuracy: 85.5 ± 0.76
node-classification-on-pubmed-60-20-20-randomSGC-2
1:1 Accuracy: 85.36 ± 0.52
node-classification-on-squirrel-60-20-20SGC-2
1:1 Accuracy: 41.25 ± 1.4
node-classification-on-squirrel-60-20-20SGC-1
1:1 Accuracy: 47.62 ± 1.27
node-classification-on-texas-60-20-20-randomSGC-2
1:1 Accuracy: 81.31 ± 3.3
node-classification-on-texas-60-20-20-randomSGC-1
1:1 Accuracy: 83.28 ± 5.43
node-classification-on-wisconsin-60-20-20SGC-2
1:1 Accuracy: 74.75 ± 2.89
node-classification-on-wisconsin-60-20-20SGC-1
1:1 Accuracy: 70.38 ± 2.85
relation-extraction-on-tacredC-SGC
F1: 67.0
sentiment-analysis-on-mrSGC
Accuracy: 75.9
sentiment-analysis-on-mrSGCN
Accuracy: 75.9
skeleton-based-action-recognition-on-sbuSGCConv
Accuracy: 94.0%
text-classification-on-20newsSGC
Accuracy: 88.5
text-classification-on-20newsSGCN
Accuracy: 88.5
text-classification-on-ohsumedSGC
Accuracy: 68.5
text-classification-on-ohsumedSGCN
Accuracy: 68.5
text-classification-on-r52SGCN
Accuracy: 94.0
text-classification-on-r52SGC
Accuracy: 94.0
text-classification-on-r8SGC
Accuracy: 97.2
text-classification-on-r8SGCN
Accuracy: 97.2

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
简化图卷积网络 | 论文 | HyperAI超神经