Interactive demo · Statistical Learning: Theory and Methods · Chapter 17

The graphical lasso on the flow-cytometry data

In a Gaussian graphical model a missing edge is a zero in the precision matrix Θ = Σ−1. The graphical lasso maximises log det Θ − trace(SΘ) − λ‖Θ‖1 (17.21) by solving one lasso problem per protein in turn (Algorithm 17.2), so a larger penalty λ sets more partial correlations exactly to zero. The data are the 11 proteins of Sachs et al. (2003), measured on 7466 cells; with the book's covariance matrix, λ = 14 gives Figure 17.1 and λ = 36, 27, 7, 0 give Figure 17.5.

Matrix that is penalised

Method

Penalty λ

Try these

Algorithm 17.2 with the diagonal penalised (so ŵjj = sjj + λ), cold start W = S + λI, coordinate-descent lasso (17.26) inside, sweeps until W stops changing. Neighbourhood selection fits the lasso ½βTS11β − s12Tβ + λ‖β‖1 of each protein on the others and joins j and k if either coefficient (‘or’) or both (‘and’) are nonzero; the graphical-lasso edges at the same λ are shown faintly for comparison. The covariance scale is the book's file sachs.covmatrix (the sample covariance divided by 1000); the correlation scale standardises every protein first.

In the lecture notes