mSSL: The Multivariate Spike-and-Slab LASSO

Sameer K. Deshpande

An R package for simultaneous variable and covariance selection in multivariate linear regression with spike-and-slab LASSO (SSL) regularization. For further details about the method, please see Deshpande, Rockova, and George (2018).

Details

Installation

The package source files are contained in the sub-directory mSSL/. To install, you can either download that directory and then build and install the package from the command line (e.g. R CMD BUILD mSSL followed by R CMD INSTALL mSSL_1.0.tar.gz). You can also install the package using devtools::install_github as follows.

library(devtools)
devtools::install_github(repo = "skdeshpande91/multivariate_SSL/mSSL")

Examples

The following code chunk shows the basic usage of the basic functions.

library(SSLASSO)
library(MASS)
library(mSSL)

n <- 100
p <- 50
q <- 25
rho <- 0.9

set.seed(129)
SigmaX <- 0.7^(abs(outer(1:p, 1:p, "-")))
X <- mvrnorm(n, rep(0, times = p), SigmaX)
B_orig <- matrix(sample(c(runif(floor(p*q*0.2), -2,2),
                          rep(0, times = p*q - floor(p*q*0.2)))), nrow = p, ncol = q)
XB <- X %*% B_orig
Sigma_orig <- rho^(abs(outer(1:q, 1:q, "-")))
Omega_orig <- solve(Sigma_orig)
Omega_orig[abs(Omega_orig) < 1e-3] <- 0 # need to truncate the essentially negligible terms introduced by solve
set.seed(130)
E <- mvrnorm(n, rep(0, times = q), Sigma_orig)
Y <- XB + E

fit_mSSL_dpe <- mSSL_dpe(X,Y)
fit_mSSL_dcpe <- mSSL_dcpe(X,Y)
fit_sslg <- sep_sslg(X,Y)


# Compare variable selection performance
round(error_B(fit_mSSL_dpe$B, B_orig), digits = 4)
round(error_B(fit_mSSL_dcpe$B, B_orig), digits = 4)
round(error_B(fit_sslg$B, B_orig), digits = 4)

# Compare covariance selection performance
round(error_Omega(fit_mSSL_dpe$Omega, Omega_orig), digits = 4)
round(error_Omega(fit_mSSL_dcpe$Omega, Omega_orig), digits = 4)
round(error_Omega(fit_sslg$Omega, Omega_orig), digits = 4)

The sub-directory examples/ contains R scripts to replicate some of the examples in our paper.