Abstract: In many machine learning tasks，it is necessary to investigate the statistical relationship among a set of high dimensional random variables.As an effective approach to solve this problem，sparse Gaussian Graphical Models(GGMs)has been widely applied to data mining，bioinformatics and financial analysis applications.Due to the high scale of model parameters as a square of data dimensionality，sparse GGMs learning remains a challenge problem especially in high dimensional settings.To address this problem，we propose in this paper a novel coordinate descent procedure to recover the sparse graph structure in a decentralized way.The core idea is based on the fact that sparse GGMs can be recovered by neighbor selection Lasso programs.We propose to apply coordinate descent optimization to each individual Lasso subproblem.When the samples are distributed on different machines，the coordinate descent procedure can be efficiently implemented in a parallel way using the MPI(Message­Passing Interface)framework.Experimental results show that our algorithm is able to significantly improve the running time efficiency，at almost no cost of structure estimation accuracy.