PipeOpEnsemble {mlr3pipelines}

Ensembling Base Class

Description

Parent class for PipeOps that aggregate predictions. Implements the private$.train() and private$.predict() methods necessary for a PipeOp and requires deriving classes to create the private$weighted_avg_predictions() function.

Format

Abstract R6Class inheriting from PipeOp.

Construction

Note: This object is typically constructed via a derived class, e.g. PipeOpClassifAvg or PipeOpRegrAvg.

PipeOpEnsemble$new(innum = 0, collect_multiplicity = FALSE, id, param_set = ps(), param_vals = list(), packages = character(0), prediction_type = "Prediction")

• innum :: numeric(1)
  Determines the number of input channels. If innum is 0 (default), a vararg input channel is created that can take an arbitrary number of inputs.

• collect_multiplicity :: logical(1)
  If TRUE, the input is a Multiplicity collecting channel. This means, a Multiplicity input, instead of multiple normal inputs, is accepted and the members are aggregated. This requires innum to be 0. Default is FALSE.

• id :: character(1)
  Identifier of the resulting object.

• param_set :: ParamSet
  ("Hyper"-)Parameters in form of a ParamSet for the resulting PipeOp.

• param_vals :: named list
  List of hyperparameter settings, overwriting the hyperparameter settings that would otherwise be set during construction. Default list().

• packages :: character
  Set of packages required for this PipeOp. These packages are loaded during ⁠$train()⁠ and ⁠$predict()⁠, but not attached. Default character(0).

• prediction_type :: character(1)
  The predict entry of the ⁠$input⁠ and ⁠$output⁠ type specifications. Should be "Prediction" (default) or one of its subclasses, e.g. "PredictionClassif", and correspond to the type accepted by private$.train() and private$.predict().

Input and Output Channels

PipeOpEnsemble has multiple input channels depending on the innum construction argument, named "input1", "input2", ... if innum is nonzero; if innum is 0, there is only one vararg input channel named "...". All input channels take only NULL during training and take a Prediction during prediction.

PipeOpEnsemble has one output channel named "output", producing NULL during training and a Prediction during prediction.

The output during prediction is in some way a weighted averaged representation of the input.

State

The ⁠$state⁠ is left empty (list()).

Parameters

• weights :: numeric
  Relative weights of input predictions. If this has length 1, it is ignored and weighs all inputs equally. Otherwise it must have length equal to the number of connected inputs. Initialized to 1 (equal weights).

Internals

The commonality of ensemble methods using PipeOpEnsemble is that they take a NULL-input during training and save an empty ⁠$state⁠. They can be used following a set of PipeOpLearner PipeOps to perform (possibly weighted) prediction averaging. See e.g. PipeOpClassifAvg and PipeOpRegrAvg which both inherit from this class.

Should it be necessary to use the output of preceding Learners during the "training" phase, then PipeOpEnsemble should not be used. In fact, if training time behaviour of a Learner is important, then one should use a PipeOpLearnerCV instead of a PipeOpLearner, and the ensemble can be created with a Learner encapsulated by a PipeOpLearner. See LearnerClassifAvg and LearnerRegrAvg for examples.

Fields

Only fields inherited from PipeOp.

Methods

Methods inherited from PipeOp as well as:

• weighted_avg_prediction(inputs, weights, row_ids, truth)
  (list of Prediction, numeric, integer | character, list) -> NULL
  Create Predictions that correspond to the weighted average of incoming Predictions. This is called by private$.predict() with cleaned and sanity-checked values: inputs are guaranteed to fit together, row_ids and truth are guaranteed to be the same as each one in inputs, and weights is guaranteed to have the same length as inputs.
  This method is abstract, it must be implemented by deriving classes.

See Also

https://mlr-org.com/pipeops.html

Other PipeOps: PipeOp, PipeOpImpute, PipeOpTargetTrafo, PipeOpTaskPreproc, PipeOpTaskPreprocSimple, mlr_pipeops, mlr_pipeops_boxcox, mlr_pipeops_branch, mlr_pipeops_chunk, mlr_pipeops_classbalancing, mlr_pipeops_classifavg, mlr_pipeops_classweights, mlr_pipeops_colapply, mlr_pipeops_collapsefactors, mlr_pipeops_colroles, mlr_pipeops_copy, mlr_pipeops_datefeatures, mlr_pipeops_encode, mlr_pipeops_encodeimpact, mlr_pipeops_encodelmer, mlr_pipeops_featureunion, mlr_pipeops_filter, mlr_pipeops_fixfactors, mlr_pipeops_histbin, mlr_pipeops_ica, mlr_pipeops_imputeconstant, mlr_pipeops_imputehist, mlr_pipeops_imputelearner, mlr_pipeops_imputemean, mlr_pipeops_imputemedian, mlr_pipeops_imputemode, mlr_pipeops_imputeoor, mlr_pipeops_imputesample, mlr_pipeops_kernelpca, mlr_pipeops_learner, mlr_pipeops_missind, mlr_pipeops_modelmatrix, mlr_pipeops_multiplicityexply, mlr_pipeops_multiplicityimply, mlr_pipeops_mutate, mlr_pipeops_nmf, mlr_pipeops_nop, mlr_pipeops_ovrsplit, mlr_pipeops_ovrunite, mlr_pipeops_pca, mlr_pipeops_proxy, mlr_pipeops_quantilebin, mlr_pipeops_randomprojection, mlr_pipeops_randomresponse, mlr_pipeops_regravg, mlr_pipeops_removeconstants, mlr_pipeops_renamecolumns, mlr_pipeops_replicate, mlr_pipeops_scale, mlr_pipeops_scalemaxabs, mlr_pipeops_scalerange, mlr_pipeops_select, mlr_pipeops_smote, mlr_pipeops_spatialsign, mlr_pipeops_subsample, mlr_pipeops_targetinvert, mlr_pipeops_targetmutate, mlr_pipeops_targettrafoscalerange, mlr_pipeops_textvectorizer, mlr_pipeops_threshold, mlr_pipeops_tunethreshold, mlr_pipeops_unbranch, mlr_pipeops_updatetarget, mlr_pipeops_vtreat, mlr_pipeops_yeojohnson

Other Multiplicity PipeOps: Multiplicity(), mlr_pipeops_classifavg, mlr_pipeops_featureunion, mlr_pipeops_multiplicityexply, mlr_pipeops_multiplicityimply, mlr_pipeops_ovrsplit, mlr_pipeops_ovrunite, mlr_pipeops_regravg, mlr_pipeops_replicate

Other Ensembles: mlr_learners_avg, mlr_pipeops_classifavg, mlr_pipeops_ovrunite, mlr_pipeops_regravg