Multinomial sparse group lasso

Multiclass classification with feature and parameter selection using sparse group lasso for the multinomial model. Suitable for high dimensional problems.

This is the R package msgl version 2.3.8.

R-package Overview

This package implements procedures for working with multinomial logistic regression models using sparse group lasso. This includes procedures for fitting and cross validating sparse models in a high dimensional setup. See the Getting started with msgl (predict primary cancer site based on microRNA measurements) for an example of a workflow consisting of 1) model selection and assessment using cross validation, 2) estimation of a final model and 3) using the selected model for carrying out predictions on new data.

Classification of cancer site. Error estimated by 10-fold cross validation on a data set consisting of microRNA expression measurements of laser dissected primary cancers.

Package highlights:

• Feature and parameter selection
• Fast coordinate gradient descent algorithm
• Suitable for high dimensional multiclass classification
• Support for lasso, group lasso and sparse group lasso
• Supports custom grouping of features
• Supports sample weighting
• Supports individual weighting of the group and parameter penalties

The penalized maximum likelihood estimator for multinomial logistic regression is computed using a coordinate gradient descent algorithm via the sglOptim optimizer. Use of parallel computing for cross validation and subsampling is supported through the foreach and doParallel packages.

Installation

Installation from souce requires installation of the necessary compiler tools for Mac or Windows.

Install the package from GitHub:

# install.packages("remotes") # uncomment if remotes is not already installed
remotes::install_github("nielsrhansen/sglOptim", build_vignettes = TRUE)
remotes::install_github("nielsrhansen/msgl", build_vignettes = TRUE)

If you don't want to build the vignettes when installing, just remove the build_vignettes = TRUE argument.

Minimal Example

library(msgl)

# Load some data
data(PrimaryCancers)

# Setup 2 parallel units
cl <- makeCluster(2)
registerDoParallel(cl)

# Do 10-fold cross validation on 100 models with increasing complexity, using the 2 parallel units
fit.cv <- msgl::cv(
  x = x,
  classes = classes,
  alpha = 0.5,
  lambda = 0.5,
  use_parallel = TRUE
)

## Running msgl 10 fold cross validation (dense design matrix)
## 
##  Samples:  Features:  Classes:  Groups:  Parameters: 
##        165        372         9      372       3.348k

stopCluster(cl)

# Print information about models
# and cross validation errors
fit.cv

## 
## Call:
## msgl::cv(x = x, classes = classes, alpha = 0.5, lambda = 0.5, 
##     use_parallel = TRUE)
## 
## Models:
## 
##  Index:  Lambda:  Features:  Parameters:  Error: 
##        1     1.00        1.6         12.1    0.96
##       20     0.88          5         31.9    0.72
##       40     0.76        8.7         51.1    0.62
##       60     0.66       11.1           64    0.56
##       80     0.58       15.7         88.9    0.46
##      100     0.50       21.7        118.5    0.35
## 
## Best model:
## 
##  Index:  Lambda:  Features:  Parameters:  Error: 
##       96     0.51       20.8        113.7    0.35

Documentation

• R package documentation
• Getting started with msgl (Predict primary cancer site based on microRNA measurements)
• Sparse group lasso and high dimensional multinomial classification paper in Computational Statistics & Data Analysis

Author

Martin Vincent wrote the package. Niels Richard Hansen is the current maintainer.

License

GPL (>=2)