seagull

This package provides regularization paths for the lasso, (fitted) group lasso, and (fitted) sparse-group lasso. The underlying mathematical model is a mixed model, i.e., a model with fixed and random effects. (Whereas it is actually optional to include any fixed effect.)

The (fitted) sparse-group lasso contains two penalty terms, which are combined via a mixing parameter 0 <= alpha <= 1. Thus, if the parameter is set to either 1 or 0, the resulting regularization operator is the lasso or the (fitted) group lasso, respectively.

Key features:

• The lasso, (fitted) group lasso, and sparse-group lasso are implemented via proximal gradient descent.

• The fitted sparse-group lasso (fitSGL) is implemented via proximal-averaged gradient descent.

• By default, a grid search for the penalty parameter lambda is performed. Warm starts are implemented to effectively accelerate this procedure.

• The step size between consecutive iterations is automatically determined via backtracking line search - except for the fitSGL, which needs a fixed step size to be provided upon its call.

Installation

To get the current release version from CRAN, please type:

install.packages("seagull")

To get the current development version from github, please type:

# install.packages("devtools")
devtools::install_github("jklosa/seagull")

Components

A data set is included and can be loaded:

data("seagull_data")

Furthermore, the following functions are available to the user:

• seagull

• lambda_max__lasso

• lambda_max_group_lasso

• lambda_max_fitted_group_lasso

• lambda_max_sparse_group_lasso

Example

Please load the data as shown in the section above and get started:

## Call the lasso:
fit_l <- seagull(y = phenotypes[, 1], Z = genotypes, alpha = 1)

## Call the group lasso:
fit_gl <- seagull(y = phenotypes[, 1], Z = genotypes, groups = groups, alpha = 0)

## Call the sparse-group lasso:
fit_sgl <- seagull(y = phenotypes[, 1], Z = genotypes, groups = groups)