Package 'bolasso'

Bootsrap-enhanced Lasso

Description

This function implements model-consistent Lasso estimation through the bootstrap. It supports parallel processing by way of the future package, allowing the user to flexibly specify many parallelization methods. This method was developed as a variable-selection algorithm, but this package also supports making ensemble predictions on new data using the bagged Lasso models.

Usage

bolasso(
  formula,
  data,
  n.boot = 100,
  progress = TRUE,
  implement = "glmnet",
  x = NULL,
  y = NULL,
  ...
)

Arguments

formula	An optional object of class formula (or one that can be coerced to that class): a symbolic description of the model to be fitted. Can be omitted when x and y are non-missing.
data	An optional object of class data.frame that contains the modeling variables referenced in form. Can be omitted when x and y are non-missing.
n.boot	An integer specifying the number of bootstrap replicates.
progress	A boolean indicating whether to display progress across bootstrap folds.
implement	A character; either 'glmnet' or 'gamlr', specifying which Lasso implementation to utilize. For specific modeling details, see glmnet::cv.glmnet or gamlr::cv.gamlr.
x	An optional predictor matrix in lieu of form and data.
y	An optional response vector in lieu of form and data.
...	Additional parameters to pass to either glmnet::cv.glmnet or gamlr::cv.gamlr.

Value

An object of class bolasso. This object is a list of length n.boot of cv.glmnet or cv.gamlr objects.

References

Bach FR (2008). “Bolasso: model consistent Lasso estimation through the bootstrap.” CoRR, abs/0804.1302. 0804.1302, https://arxiv.org/abs/0804.1302.

See Also

glmnet::cv.glmnet and gamlr::cv.gamlr for full details on the respective implementations and arguments that can be passed to ....

Examples

mtcars[, c(2, 10:11)] <- lapply(mtcars[, c(2, 10:11)], as.factor)
idx <- sample(nrow(mtcars), 22)
mtcars_train <- mtcars[idx, ]
mtcars_test <- mtcars[-idx, ]

## Formula Interface

# Train model
set.seed(123)
bolasso_form <- bolasso(
  form = mpg ~ .,
  data = mtcars_train,
  n.boot = 20,
  nfolds = 5,
  implement = "glmnet"
)

# Extract selected variables
selected_vars(bolasso_form, threshold = 0.9, select = "lambda.min")

# Bagged ensemble prediction on test data
predict(bolasso_form,
        new.data = mtcars_test,
        select = "lambda.min")

## Alternal Matrix Interface

# Train model
set.seed(123)
bolasso_mat <- bolasso(
  x = model.matrix(mpg ~ . - 1, mtcars_train),
  y = mtcars_train[, 1],
  data = mtcars_train,
  n.boot = 20,
  nfolds = 5,
  implement = "glmnet"
)

# Extract selected variables
selected_vars(bolasso_mat, threshold = 0.9, select = "lambda.min")

# Bagged ensemble prediction on test data
predict(bolasso_mat,
        new.data = model.matrix(mpg ~ . - 1, mtcars_test),
        select = "lambda.min")

---

Bolasso-selected Variables

Description

Identifies independent variables that are selected by the Bolasso algorithm at least the fraction of the time specified by the user-defined threshold. The typical value for this threshold is 0.9 and typically shouldn't be lower than that.

Usage

selected_vars(object, threshold = 0.9, summarise = TRUE, ...)

Arguments

object	An object of class bolasso.
threshold	A numeric between 0 and 1, specifying the fraction of bootstrap replicates for which Lasso must select a variable for it to be considered a selected variable.
summarise	A Boolean indicator where FALSE indicates returning the full set of coefficients at the selected variable/bootstrap replicate level and TRUE indicates taking the average of each variable's coefficient across bootstrap replicates. The default value is TRUE as it's more efficient and interpretable.
...	Additional arguments to pass to predict on objects with class cv.glmnet or cv.gamlr.

Value

A tibble with each selected variable and its respective coefficient for each bootstrap replicate.

See Also

glmnet::predict.glmnet() and gamlr:::predict.gamlr for details on additional arguments to pass to ....

---