Learning the effect of latent variables in Gaussian Graphical models with unobserved variables