3 个月前

半跳:一种用于减缓消息传递的图上采样方法

半跳:一种用于减缓消息传递的图上采样方法

摘要

消息传递神经网络在图结构数据上已取得显著成功。然而,在许多实际场景中,消息传递机制可能导致过平滑问题,或在相邻节点属于不同类别时失效。本文提出一种简单但通用的框架,用于提升消息传递神经网络的学习性能。我们的方法通过在原始图的每条边上引入“慢节点”(slow nodes),对边进行上采样,从而实现源节点与目标节点之间的通信中介。该方法仅修改输入图结构,无需改动现有模型,具有即插即用的特性,易于集成。为深入理解减缓消息传递带来的优势,我们从理论和实证两个层面进行了分析。在多个监督学习与自监督学习基准任务上的实验结果表明,该方法在各类场景下均取得显著提升,尤其在异质性(heterophilic)条件下表现突出——即相邻节点更可能具有不同标签的情形。此外,我们进一步展示了该方法在自监督学习中的应用潜力:通过随机在图的不同边上引入慢节点,可生成具有不同路径长度的多尺度视图,从而为自监督学习提供有效的数据增强策略。

代码仓库

nerdslab/halfhop
官方
pytorch

基准测试

基准方法指标
node-classification-on-amz-compGraphSAGE
Accuracy: 84.79%
node-classification-on-amz-compGCN
Accuracy: 90.22%
node-classification-on-amz-compHH-GCN
Accuracy: 90.92%
node-classification-on-amz-compHH-GraphSAGE
Accuracy: 86.6%
node-classification-on-amz-photoGraphSAGE
Accuracy: 95.03%
node-classification-on-amz-photoHH-GraphSAGE
Accuracy: 94.55%
node-classification-on-amz-photoHH-GCN
Accuracy: 94.52%
node-classification-on-amz-photoGCN
Accuracy: 93.59%
node-classification-on-chameleon-60-20-20HH-GraphSAGE
1:1 Accuracy: 62.98 ± 3.35
node-classification-on-chameleon-60-20-20HH-GCN
1:1 Accuracy: 60.24 ± 1.93
node-classification-on-chameleon-60-20-20HH-GAT
1:1 Accuracy: 61.12 ± 1.83
node-classification-on-coauthor-csHH-GCN
Accuracy: 94.71%
node-classification-on-coauthor-csGCN
Accuracy: 94.06%
node-classification-on-coauthor-csGraphSAGE
Accuracy: 95.11%
node-classification-on-coauthor-csHH-GraphSAGE
Accuracy: 95.13%
node-classification-on-cornell-60-20-20HH-GAT
1:1 Accuracy: 72.7 ± 4.26
node-classification-on-cornell-60-20-20HH-GraphSAGE
1:1 Accuracy: 74.6 ± 6.06
node-classification-on-cornell-60-20-20HH-GCN
1:1 Accuracy: 63.24 ± 5.43
node-classification-on-squirrel-60-20-20HH-GCN
1:1 Accuracy: 47.19 ± 1.21
node-classification-on-squirrel-60-20-20HH-GraphSAGE
1:1 Accuracy: 45.25 ± 1.52
node-classification-on-squirrel-60-20-20HH-GAT
1:1 Accuracy: 46.35 ± 1.86
node-classification-on-texas-60-20-20-randomHH-GAT
1:1 Accuracy: 80.54 ± 4.80
node-classification-on-texas-60-20-20-randomHH-GCN
1:1 Accuracy: 71.89 ± 3.46
node-classification-on-texas-60-20-20-randomHH-GraphSAGE
1:1 Accuracy: 85.95 ± 6.42
node-classification-on-wiki-csHH-GraphSAGE
Accuracy: 82.81
node-classification-on-wiki-csGCN
Accuracy: 81.93
node-classification-on-wiki-csHH-GCN
Accuracy: 82.57
node-classification-on-wiki-csGraphSAGE
Accuracy: 83.67
node-classification-on-wisconsin-60-20-20HH-GCN
1:1 Accuracy: 79.8 ± 4.30
node-classification-on-wisconsin-60-20-20HH-GraphSAGE
1:1 Accuracy: 85.88 ± 3.99
node-classification-on-wisconsin-60-20-20HH-GAT
1:1 Accuracy: 83.53 ± 3.84

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
半跳:一种用于减缓消息传递的图上采样方法 | 论文 | HyperAI超神经