Introduction to BOSO

We present BOSO, an R package to perform feature selection in a linear regression problem. It implements a Bilevel Optimization Selector Operator.

Installation

BOSO can be installed from CRAN repository:

install.packages("BOSO")

Alternatively, it can be downloaded and installed from github:

library(devtools)
install_github(repo="lvalcarcel/BOSO", subdir="BOSO")

Introduction

The package package has been prepared to work like 'glmnet' and 'lasso', presented in the BestSubset package.

library(BOSO)

## Load the data prepared for this test
data("sim.xy", package = "BOSO")

Xtr <- sim.xy[['high-5']]$x
Ytr <- sim.xy[['high-5']]$y
Xval <- sim.xy[['high-5']]$xval
Yval <- sim.xy[['high-5']]$yval


## Perform BOSO
time <- Sys.time()
obj <- BOSO(x = Xtr, y = Ytr,
            xval = Xval, yval = Yval,
            IC = 'eBIC',
            nlambda=100,
            intercept= 0,
            standardize = 0,
            Threads=4, timeLimit = 60, verbose = 3, 
            seed = 2021)
time <- as.numeric(Sys.time() - time)

obj is a BOSO object, which have the following associated functions:

• coef(obj) returns the coefficients (betas) of the linear regression.
• predict(obj, xnew) returns the predicted outcome with a new X matrix.

betas <- coef(obj)
print(betas[betas!=0])

Ytr_predicted <- predict(obj, Xtr)
print(paste0("MSE for training set is ",  round(mean((Ytr_predicted-Ytr)^2),5)))

Yval_predicted <- predict(obj, Xval)
print(paste0("MSE for validation set is ", round(mean((Yval_predicted-Yval)^2),5)))