README.md

In aplantin/CSGL: Compositional Sparse Group Lasso

CSGL

Fits the Compositional Sparse Group Lasso, an L1- and L2-penalized linear log-contrast model.

Installation instructions:

library(devtools)
install_github("aplantin/CSGL")
library(CSGL)

Sample usage with simulated data

# Function to generate compositional data 
library(mvtnorm) 
gen.comp.data <- function(p, n, tb, rho, sigma, seed){
    set.seed(seed)
    theta <- c(rep(log(0.5*p), 5), rep(0, (p-5)))
    H <- abs(outer(1:p, 1:p, "-"))
    Sigma <- rho^H
    W <- rmvnorm(n, theta, Sigma)
    X <- W
    for(i in 1:n){ X[i,] <- exp(W[i,])/sum(exp(W[i,])) }
    Z <- log(X)
    err <- rnorm(n, mean=0, sd=sigma)
    y <- Z%*%tb + err
    return(list(X=X, Z=Z, y=y))
}

# Generate a small compositional dataset 
dat <- gen.comp.data(p = 30, n = 50, tb = c(1, -1, rep(0, 28)), 
                     rho = 0.2, sigma = 0.5, seed = 1)

# Run CSGL with a (very short) sequence of lambda values 
res <- csgl(dat$Z, dat$y, groups = rep(1:6, each = 5), mu = 1, theta = 0.95, nlam = 3, min.frac = 0.1)
res$beta

# Run CSGL with a sequence of lambda values, choosing lambda by 5-fold cross-validation 
res.cv <- cv.csgl(dat$Z, dat$y, groups = rep(1:6, each = 5), mu = 1, theta = 0.95, nlam = 10, min.frac = 0.1, nfolds = 5)
res.cv$cv.beta

# Run CSGL with a sequence of lambda values, choosing lambda by GIC 
res.gic <- gic.csgl(dat$Z, dat$y, groups = rep(1:6, each = 5), mu = 1, theta = 0.95, nlam = 10, min.frac = 0.1) 
res.gic$gic.beta

aplantin/CSGL documentation built on May 21, 2019, 12:08 p.m.