family = "multinomial" is
specified, previously bolasso did not correctly support variable selection, plotting, etc.
This version addresses this. In general, when family = "multinomial", all outputs become a list
where each list element corresponds to a unique outcome in the multinomial response vector. However,
the content of each list element will be the same as it originally was (e.g. a tibble of
coefficients, or a vector of variable names, etc.). In the case of plotting, it will create a
facetted plot with one facet per unique outcome. The only limitations are that family = "mgaussian"
is not currently supported, and when predicting with family = "multinomial", only type = "class"
is currently supported. These may be added at a later date.Allows the user to extract the bootstrap indices with bootstrap_samples().
library(bolasso)
model <- bolasso(mpg ~ hp + wt, data = mtcars, n.boot = 10)
bootstrap_samples(model)
Fixes #12
Fixes #13
Fast Estimation Mode:
bolasso() gains a fast argument which optimizes computation by using a single cross-validated regression on the entire dataset to determine the optimal regularization parameter (lambda). This approach bypasses the need for cross-validation within each bootstrap replicate, drastically reducing computation time, especially beneficial for large datasets or when using a high number of bootstrap replicates.
# Fast mode reduces computation time by using a single cross-validated lambda
model_fast <- bolasso(
diabetes ~ .,
data = train,
n.boot = 1000,
progress = FALSE,
family = "binomial",
fast = TRUE
)
Enhanced Variable Selection Methods:
selected_vars() is now a shorthand for selected_variables().selected_variables()/selected_vars() supports two variable selection algorithms via the method argument: the Variable Inclusion Probability (VIP) method and the Quantile (QNT) method. The VIP method selects variables that appear in a high percentage of bootstrap models, while the QNT method selects variables based on bootstrap confidence intervals. Set method = "vip" or method = "qnt", respectively.
# Select variables using the VIP method with a 95% threshold
selected_vars_vip <- selected_variables(model, threshold = 0.95, method = "vip")
# Select variables using the QNT method
selected_vars_qnt <- selected_variables(model, threshold = 0.95, method = "qnt")
Tidy Method for Bolasso Objects:
tidy() extracts a tidy tibble summarizing bootstrap-level coefficients for each covariate. This method provides a clean and organized way to inspect model coefficients.
# Extract a tidy tibble of coefficients
tidy_coefs <- tidy(model, select = "lambda.min")
Variable Selection Visualization:
plot_selection_thresholds() provides a visual representation of the selection thresholds for each variable. This visualization helps users understand the stability and robustness of variable selection across different thresholds and methods.
# Visualize selection thresholds for variables
plot_selection_thresholds(model, select = "lambda.min")
plot_selected_variables() visualizes the coefficient distributions for only the selected variables. This function provides a focused view of the most relevant variables in the model.
# Plot coefficient distributions for selected variables
plot_selected_variables(
model,
threshold = 0.95,
method = "vip",
select = "lambda.min"
)
plot visualizes the coefficient distributions for all model covariates.
# Plot coefficient distributions for selected variables
plot(model, select = "lambda.min")
NEWS.md file to track changes to the package.bolasso() argument form has been renamed to formula to reflect common naming conventions in R statistical modeling packages.predict() and coef() methods are now implemented using future.apply::future_lapply allowing for computing predictions and extracting coefficients in parallel. This may result in slightly worse performance (due to memory overhead) when the model/prediction data is small but will be significantly faster when e.g. generating predictions on a very large data-set.bolasso() argument formula. The user-supplied value of formula is handled via deparse() which has a default width.cutoff value of 60. This was causing issues with formulas by splitting them into multi-element character vectors. It has now been set to the maximum value of 500L which will correctly parse all lengths of formulas.predict() now forces evaluation of the formula argument in the bolasso() call. This resolves an issue where, if a user passes a formula via a variable, predict() would pass the variable name to the underlying prediction function as opposed to the actual formula.