基于GEMS算法的潜变量高斯图模型结构学习

Latent-variable Gaussian graphical model structure learning by GEMS algorithm

考虑了潜变量高斯图模型下的结构学习(模型选择)问题,即存在潜变量时可观测变量间相互关系的估计问题.简要介绍了高斯图模型及潜变量高斯图模型下的LVglasso方法,给出了GEMS(广义期望模型选择)算法结合LVglasso下潜变量图模型选择的算法步骤.通过模拟,发现GEMS结合LVglasso方法在模型选择速度上比EM(期望最大化)算法有明显优势,并分析了拟南芥植物基因数据,估计了各基因间的条件相关性.

The latent-variable Gaussian graphical model structure learning problem is considered,that is,the estimation of the relationship among the observed variables when there are latent variables.The Gaussian graphical model and the method of LVglasso for latent-variable graphical model are briefly introduced,then the steps of GEMS(generalized Expectation Model Selection)algorithm combined with LVglasso are presented.Through simulations,it is found that GEMS performs much better than EM(Expectation Maximization)in the speed of model selection.And then Arabidopsis Thaliana gene data are analyzed and the conditional correlations between genes are estimated.