README.md

In matt-sutton/sgspls: Sparse group-subgroup partial least squares

Sparse group-subgroup PLS

sgspls is an R package that provides an implementation of an alternating convex search algorithm for computing partial least squares solutions with multiple grouping levels. Most of the linear algebra is written in C++ using Rcpp and the Armadillo C++ library.

A preliminary paper has been submitted describing the statistical properties of the sparse group-subgroup PLS with grouping structure and will be made avaliable when the manuscript is accepted.

Installation

Use devtools to install:

library(devtools)
install_github("matt-sutton/sgspls")

Example Usage (regression)

set.seed(1)
n = 50; p = 510; 
size.groups = 30; size.subgroups = 5
groupX <- ceiling(1:p / size.groups)
subgroupX <- ceiling(1:p / size.subgroups)

X = matrix(rnorm(n * p), ncol = p, nrow = n)

beta <- rep(0,p)
bSG <- -2:2; b0 <- rep(0,length(bSG))
betaG <- c(bSG, b0, bSG, b0, bSG, b0)
beta[1:size.groups] <- betaG

y = X %*% beta + 0.5*rnorm(n)

#--------------------------------------#
#-- Set up a basic model to tune --#
library(sgspls)

cv_pls <- list(X=X, Y=y, groupX=groupX, subgroupX=subgroupX)

#---------------------------------------------#
#-- Tune over 1 to 2 groups and multiple    --#
#-- sparsity levels for the first component --#

cv_pls_comp1 <- tune_sgspls(pls_obj = cv_pls, group_seq = 1:2, scale_resp = FALSE)

#-- MSEP is on the original scale for the response --#
cv_pls_comp1$results_tuning
cv_pls_comp1$best

# Use the optimal fit for the first component and tune the second component
cv_pls_comp2 <- tune_sgspls(pls_obj = cv_pls_comp1, group_seq = 1:2, scale_resp = FALSE)
cv_pls_comp2$results_tuning
cv_pls_comp2$best

# Use the optimal fit for the second component and tune the third component
cv_pls_comp3 <- tune_sgspls(pls_obj =  cv_pls_comp2, group_seq = 1:2, scale_resp = FALSE)
cv_pls_comp3$best

model <- do.call(sgspls, args = cv_pls_comp3$parameters)

model

model_coef <- coef(model, type = "coefficients")

cbind(beta, model_coef$B[,,2])

Example Usage (canonical)

set.seed(1)
n = 50; p = 500; q = 100
ncomp <- 3

Scores <- MASS::mvrnorm(n, rep(0, ncomp), Sigma = diag(1, ncomp, ncomp))
Tm <- Scores
Sm <- Scores + MASS::mvrnorm(n, rep(0, ncomp), Sigma = diag(0.1, ncomp, ncomp))


subgroupC <- -2:2; p_sg <- length(subgroupC) 
groupC <- c(subgroupC, rep(0,p_sg), subgroupC, rep(0,p_sg), subgroupC, rep(0,p_sg))
p_g <- length(groupC) 

groupX <- ceiling(1:p / p_g)
subgroupX <- ceiling(1:p / p_sg)


Cm <- matrix(0,nrow = p, ncol = ncomp)
Cm[groupX == 1, 1] <- groupC
Cm[groupX == 2, 2] <- groupC
Cm[groupX == 3, 3] <- groupC
Dm <- matrix(runif(q*ncomp), ncol = ncomp)

X <- Tm%*%t(Cm) + MASS::mvrnorm(n, rep(0, p), Sigma = diag(0.1,p,p))
Y <- Sm%*%t(Dm) + MASS::mvrnorm(n, rep(0, q), Sigma = diag(0.1,q,q))

model <- sgspls(X, Y, ncomp = 5, mode = "regression", keepX = 17,
                groupX = groupX, subgroupX = subgroupX,
                indiv_sparsity_x = 0.8, subgroup_sparsity_x = 0.15)

#-- Scree type plot for the sgspls canonical method --#
plot(model)

#-- Sparse Version --#
model <- sgspls(X, Y, ncomp = 3, mode = "regression", keepX = 1,
                groupX = groupX, subgroupX = subgroupX,
                indiv_sparsity_x = 0.8, subgroup_sparsity_x = 0.15)

# check the recovery
model$weights$X

matt-sutton/sgspls documentation built on June 22, 2019, 10:21 a.m.