stacks (version 1.0.0)

blend_predictions: Determine stacking coefficients from a data stack

Description

Evaluates a data stack by fitting a regularized model on the assessment predictions from each candidate member to predict the true outcome.

This process determines the "stacking coefficients" of the model stack. The stacking coefficients are used to weight the predictions from each candidate (represented by a unique column in the data stack), and are given by the betas of a LASSO model fitting the true outcome with the predictions given in the remaining columns of the data stack.

Candidates with non-zero stacking coefficients are model stack members, and need to be trained on the full training set (rather than just the assessment set) with fit_members(). This function is typically used after a number of calls to add_candidates().

Usage

blend_predictions(
  data_stack,
  penalty = 10^(-6:-1),
  mixture = 1,
  non_negative = TRUE,
  metric = NULL,
  control = tune::control_grid(),
  times = 25,
  ...
)

Value

A model_stack object—while model_stacks largely contain the same elements as data_stacks, the primary data objects shift from the assessment set predictions to the member models.

Arguments

data_stack

A data_stack object

penalty

A numeric vector of proposed values for total amount of regularization used in member weighting. Higher penalties will generally result in fewer members being included in the resulting model stack, and vice versa. The package will tune over a grid formed from the cross product of the penalty and mixture arguments.

mixture

A number between zero and one (inclusive) giving the proportion of L1 regularization (i.e. lasso) in the model. mixture = 1 indicates a pure lasso model, mixture = 0 indicates ridge regression, and values in (0, 1) indicate an elastic net. The package will tune over a grid formed from the cross product of the penalty and mixture arguments.

non_negative

A logical giving whether to restrict stacking coefficients to non-negative values. If TRUE (default), 0 is passed as the lower.limits argument to glmnet::glmnet() in fitting the model on the data stack. Otherwise, -Inf.

metric

A call to yardstick::metric_set(). The metric(s) to use in tuning the lasso penalty on the stacking coefficients. Default values are determined by tune::tune_grid() from the outcome class.

control

An object inheriting from control_grid to be passed to the model determining stacking coefficients. See tune::control_grid() documentation for details on possible values. Note that any extract entry will be overwritten internally.

times

Number of bootstrap samples tuned over by the model that determines stacking coefficients. See rsample::bootstraps() to learn more.

...

Additional arguments. Currently ignored.

Example Data

This package provides some resampling objects and datasets for use in examples and vignettes derived from a study on 1212 red-eyed tree frog embryos!

Red-eyed tree frog (RETF) embryos can hatch earlier than their normal 7ish days if they detect potential predator threat. Researchers wanted to determine how, and when, these tree frog embryos were able to detect stimulus from their environment. To do so, they subjected the embryos at varying developmental stages to "predator stimulus" by jiggling the embryos with a blunt probe. Beforehand, though some of the embryos were treated with gentamicin, a compound that knocks out their lateral line (a sensory organ.) Researcher Julie Jung and her crew found that these factors inform whether an embryo hatches prematurely or not!

Note that the data included with the stacks package is not necessarily a representative or unbiased subset of the complete dataset, and is only for demonstrative purposes.

reg_folds and class_folds are rset cross-fold validation objects from rsample, splitting the training data into for the regression and classification model objects, respectively. tree_frogs_reg_test and tree_frogs_class_test are the analogous testing sets.

reg_res_lr, reg_res_svm, and reg_res_sp contain regression tuning results for a linear regression, support vector machine, and spline model, respectively, fitting latency (i.e. how long the embryos took to hatch in response to the jiggle) in the tree_frogs data, using most all of the other variables as predictors. Note that the data underlying these models is filtered to include data only from embryos that hatched in response to the stimulus.

class_res_rf and class_res_nn contain multiclass classification tuning results for a random forest and neural network classification model, respectively, fitting reflex (a measure of ear function) in the data using most all of the other variables as predictors.

log_res_rf and log_res_nn, contain binary classification tuning results for a random forest and neural network classification model, respectively, fitting hatched (whether or not the embryos hatched in response to the stimulus) using most all of the other variables as predictors.

See ?example_data to learn more about these objects, as well as browse the source code that generated them.

Details

Note that a regularized linear model is one of many possible learning algorithms that could be used to fit a stacked ensemble model. For implementations of additional ensemble learning algorithms, see h2o::h2o.stackedEnsemble() and SuperLearner::SuperLearner().

See Also

Other core verbs: add_candidates(), fit_members(), stacks()

Examples

if (FALSE) { # rlang::is_installed("kernlab")
# \donttest{
# see the "Example Data" section above for
# clarification on the objects used in these examples!

# put together a data stack
reg_st <- 
  stacks() %>%
  add_candidates(reg_res_lr) %>%
  add_candidates(reg_res_svm) %>%
  add_candidates(reg_res_sp)
  
reg_st

# evaluate the data stack
reg_st %>%
  blend_predictions()

# include fewer models by proposing higher penalties
reg_st %>% 
  blend_predictions(penalty = c(.5, 1))

# allow for negative stacking coefficients 
# with the non_negative argument
reg_st %>% 
  blend_predictions(non_negative = FALSE)
  
# use a custom metric in tuning the lasso penalty
library(yardstick)
reg_st %>% 
  blend_predictions(metric = metric_set(rmse))
  
# pass control options for stack blending
reg_st %>% 
  blend_predictions(
    control = tune::control_grid(allow_par = TRUE)
  )
 
# to speed up the stacking process for preliminary
# results, bump down the `times` argument:
reg_st %>% 
  blend_predictions(times = 5)
  
# the process looks the same with 
# multinomial classification models
class_st <-
  stacks() %>%
  add_candidates(class_res_nn) %>%
  add_candidates(class_res_rf) %>%
  blend_predictions()
  
class_st

# ...or binomial classification models
log_st <-
  stacks() %>%
  add_candidates(log_res_nn) %>%
  add_candidates(log_res_rf) %>%
  blend_predictions()
  
log_st
# }
}