R6 class for super learners

Super learner (stacked ensemble) implementation of the learner R6 class. learnerSL$new() takes the same arguments as learner$new(). The base learners are provided via estimate.args$learners, and by default the ensemble is estimated with superlearner(). Objects are usually created with the constructor function learner_sl(). The arguments formula and formula.keep.specials of learnerSL$new() are not used, and a warning is raised if they are provided.

The formula of the super learner is the formula of its first base learner. Hence, the methods design() and response() return the design and response of the first base learner. The formula is kept in sync with the first base learner by the update() method.

The update() method changes the response variable of the super learner and of all base learners. The covariates of the base learners are never modified by update().

The base learners are deep cloned when a learnerSL object is created and when it is cloned with clone(deep = TRUE). Updating a super learner therefore neither modifies the learner objects used to create it nor the base learners of any other super learner object.

See also

learner_sl, superlearner, cv.learner_sl

Super class

learner -> learner_sl

Active bindings

learners

Return the list of base learners (read-only). Use learnerSL$update() to update the response variable of the base learners.

Methods

Public methods

• learner_sl$new()

• learner_sl$estimate()

• learner_sl$update()

• learner_sl$summary()

• learner_sl$opt()

• learner_sl$clone()

Inherited methods

• learner$design()
• learner$predict()
• learner$print()
• learner$response()

---

learner_sl$new()

Create a new super learner object. The arguments are the same as for learner$new().

Usage

learner_sl$new(
  formula = NULL,
  estimate = superlearner,
  predict = predict.superlearner,
  predict.args = NULL,
  estimate.args = NULL,
  info = NULL,
  specials = c(),
  formula.keep.specials = FALSE,
  predict.filter = function(data) function(pred, newdata) pred,
  intercept = FALSE
)

Arguments

formula

(formula) Not used by learnerSL objects, because the formula of a super learner is the formula of its first base learner. A warning is raised if provided. Use learnerSL$update() to update the response variable.

estimate

(function) Estimation method of the ensemble. Defaults to superlearner. A user-defined function must have the arguments data and learners (base learners), e.g. function(data, learners, ...).

predict

(function) Prediction method. Defaults to predict.superlearner, which matches the default estimate method. A user-defined estimate method that returns a different model object requires a matching predict method.

predict.args

optional arguments to prediction function

estimate.args

(list) Arguments to the estimate method. Must contain the element learners, a list of learner objects that define the ensemble. All base learners must be defined with a formula and share the same response variable. The remaining elements (e.g., nfolds, meta.learner, model.score) are passed on to the estimate method. Hence, the defaults of superlearner apply to arguments that are not provided.

info

(character) Optional description of the model. Defaults to a listing of the names of the base learners.

specials

optional specials terms (weights, offset, id, subset, ...) passed on to design

formula.keep.specials

(logical) Not used by learnerSL objects.

predict.filter

function to post-process predictions. Useful to bound predictions or handle NAs. The argument is experimental and its behavior may change in the future.

intercept

(logical) include intercept in design matrix

---

learner_sl$estimate()

Estimation method. Estimates the super learner with the base learners of the object and the estimate function (by default superlearner).

Usage

learner_sl$estimate(data, ..., store = TRUE)

Arguments

data

(data.frame) Data used to estimate the super learner.

...

Additional arguments to the estimate function (e.g., nfolds), which take precedence over estimate.args. The base learners cannot be changed. Create a new super learner with learner_sl() to use other base learners.

store

(logical) If TRUE, the estimated model is stored inside the object.

---

learner_sl$update()

Update the response variable of the super learner and all its base learners. Each base learner keeps its covariates.

Usage

learner_sl$update(formula)

Arguments

formula

(character) Name or specification of the new response variable (e.g., "z" or "I(a == 1)").

---

learner_sl$summary()

Summary method. See learner$summary().

Usage

learner_sl$summary()

---

learner_sl$opt()

Get options

Usage

learner_sl$opt(arg)

Arguments

arg

(character) Name of option to get value of. Use "learners" to get the base learners.

---

learner_sl$clone()

The objects of this class are cloneable with this method.

Usage

learner_sl$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

n <- 200
d <- data.frame(x1 = rnorm(n), x2 = rnorm(n))
d$y <- d$x1 + rnorm(n)
d$z <- d$x1 - d$x2 + rnorm(n)

lrs <- list(
  "mean" = learner_glm(y ~ 1),
  "glm" = learner_glm(y ~ x1 + x2)
)
sl <- learnerSL$new(estimate.args = list(learners = lrs, nfolds = 2))
sl$formula # formula of the first base learner
#> y ~ 1
#> <environment: 0x636f67780ac0>

# update the response variable of the super learner and all base learners
sl$update("z")
sl$formula
#> z ~ 1
#> <environment: 0x636f67780ac0>
lapply(sl$learners, \(lr) lr$formula)
#> $mean
#> z ~ 1
#> <environment: 0x636f67780ac0>
#> 
#> $glm
#> z ~ x1 + x2
#> <environment: 0x636f67780ac0>
#> 

sl$estimate(d)
sl$predict(head(d))
#> [1]  0.05368882 -0.32306826 -1.02974860  0.57548211  0.37640144 -0.39584702