基于高斯分布和Householder flow的无监督图嵌入算法
Unsupervised graph embedding algorithm based on Gaussian distribution and Householder flow
刘国军 1范天祥 1王乃正 1张正达 1齐广智1
作者信息
- 1. 哈尔滨工业大学计算学部,黑龙江 哈尔滨 150001
- 折叠
摘要
为更好地表示节点,提出一种新的图嵌入方法,将节点表示为由均值和方差构成的高斯分布,通过应用一系列可逆Householder变换,将相对简单的分布转换为更灵活的分布,可以更好地捕获关于其表示的不确定性.为提高稳定性,采用Wasserstein距离进行分布之间的度量.试验结果表明,在多个基准数据集上,使用Householder变换的Graph2Gauss(G2G)算法比原始模型的链接预测表现更好.通过节点分类的效果可以看出,对于节点信息缺失的图,使用Wasserstein距离可以大幅增加节点分类的F1 分数.
Abstract
In order to express nodes better,a novel graph embedding method was proposed,each node was represented as a Gaussian distribution composed of mean and variance.By applying a series of reversible Householder transformations,the relatively simple distribution was transformed into a more flexible distribution,so that the uncertainty about its representation could be better captured.To improve stability,Wasserstein distance was used to measure between distributions.The experimental results showed that on multiple benchmark datasets,Graph2Gauss(G2G)algorithm using Householder transformation performed better than the original model on link prediction.It could be seen from the effectiveness of node classification that using Wasserstein distance could significantly increase the F1 score for graphs with missing node information.
关键词
无监督学习/图嵌入/高斯分布/Householder/flow/Wasserstein距离Key words
unsupervised learning/graph embedding/Gaussian distribution/Householder flow/Wasserstein distance引用本文复制引用
基金项目
国家自然科学基金资助项目(61976071)
黑龙江省联合基金资助项目(LH2020F012)
出版年
2024
NETL
NSTL
万方数据