mlr (version 2.10)

makeStackedLearner: Create a stacked learner object.

Description

A stacked learner uses predictions of several base learners and fits a super learner using these predictions as features in order to predict the outcome. The following stacking methods are available:

average

Averaging of base learner predictions without weights.

stack.nocv

Fits the super learner, where in-sample predictions of the base learners are used.

stack.cv

Fits the super learner, where the base learner predictions are computed by crossvalidated predictions (the resampling strategy can be set via the resampling argument).

hill.climb

Select a subset of base learner predictions by hill climbing algorithm.

compress

Train a neural network to compress the model from a collection of base learners.

Usage

makeStackedLearner(base.learners, super.learner = NULL, predict.type = NULL,
  method = "stack.nocv", use.feat = FALSE, resampling = NULL,
  parset = list())

Arguments

base.learners

[(list of) Learner] A list of learners created with makeLearner.

super.learner

[Learner | character(1)] The super learner that makes the final prediction based on the base learners. If you pass a string, the super learner will be created via makeLearner. Not used for method = 'average'. Default is NULL.

predict.type

[character(1)] Sets the type of the final prediction for method = 'average'. For other methods, the predict type should be set within super.learner. If the type of the base learner prediction, which is set up within base.learners, is

"prob"

then predict.type = 'prob' will use the average of all bease learner predictions and predict.type = 'response' will use the class with highest probability as final prediction.

"response"

then, for classification tasks with predict.type = 'prob', the final prediction will be the relative frequency based on the predicted base learner classes and classification tasks with predict.type = 'response' will use majority vote of the base learner predictions to determine the final prediction. For regression tasks, the final prediction will be the average of the base learner predictions.

method

[character(1)] “average” for averaging the predictions of the base learners, “stack.nocv” for building a super learner using the predictions of the base learners, “stack.cv” for building a super learner using crossvalidated predictions of the base learners. “hill.climb” for averaging the predictions of the base learners, with the weights learned from hill climbing algorithm and “compress” for compressing the model to mimic the predictions of a collection of base learners while speeding up the predictions and reducing the size of the model. Default is “stack.nocv”,

use.feat

[logical(1)] Whether the original features should also be passed to the super learner. Not used for method = 'average'. Default is FALSE.

resampling

[ResampleDesc] Resampling strategy for method = 'stack.cv'. Currently only CV is allowed for resampling. The default NULL uses 5-fold CV.

parset

the parameters for hill.climb method, including

replace

Whether a base learner can be selected more than once.

init

Number of best models being included before the selection algorithm.

bagprob

The proportion of models being considered in one round of selection.

bagtime

The number of rounds of the bagging selection.

metric

The result evaluation metric function taking two parameters pred and true, the smaller the score the better.

the parameters for compress method, including

k

the size multiplier of the generated data

prob

the probability to exchange values

s

the standard deviation of each numerical feature

Examples

  # Classification
  data(iris)
  tsk = makeClassifTask(data = iris, target = "Species")
  base = c("classif.rpart", "classif.lda", "classif.svm")
  lrns = lapply(base, makeLearner)
  lrns = lapply(lrns, setPredictType, "prob")
  m = makeStackedLearner(base.learners = lrns,
    predict.type = "prob", method = "hill.climb")
  tmp = train(m, tsk)
  res = predict(tmp, tsk)

  # Regression
  data(BostonHousing, package = "mlbench")
  tsk = makeRegrTask(data = BostonHousing, target = "medv")
  base = c("regr.rpart", "regr.svm")
  lrns = lapply(base, makeLearner)
  m = makeStackedLearner(base.learners = lrns,
    predict.type = "response", method = "compress")
  tmp = train(m, tsk)
  res = predict(tmp, tsk)