Impute factorial features by adding a new level ".MISSING".
Impute numerical features by constant values shifted below the minimum or above the maximum by using \(min(x) - offset - multiplier * diff(range(x))\) or \(max(x) + offset + multiplier * diff(range(x))\).
This type of imputation is especially sensible in the context of tree-based methods, see also Ding & Simonoff (2010).
Learners expect input Tasks to have the same factor (or ordered) levels during
training as well as prediction. This PipeOp modifies the levels of factor and ordered features,
and since it may occur that a factor or ordered feature contains missing values only during prediction, but not
during training, the output Task could also have different levels during the two stages.
To avoid problems with the Learners' expectation, controlling the PipeOps' handling of this edge-case is necessary.
For this, use the create_empty_level hyperparameter inherited from PipeOpImpute.
If create_empty_level is set to TRUE, then an unseen level ".MISSING" is added to the feature during
training and missing values are imputed as ".MISSING" during prediction.
However, empty factor levels during training can be a problem for many Learners.
If create_empty_level is set to FALSE, then no empty level is introduced during training, but columns that
have missing values only during prediction will not be imputed. This is why it may still be necessary to use
po("imputesample", affect_columns = selector_type(types = c("factor", "ordered")))
(or another imputation method) after this imputation method.
Note that setting create_empty_level to FALSE is the same as setting it to TRUE and using PipeOpFixFactors
after this PipeOp.
Format
R6Class object inheriting from PipeOpImpute/PipeOp.
Construction
id::character(1)
Identifier of resulting object, default"imputeoor".param_vals:: namedlist
List of hyperparameter settings, overwriting the hyperparameter settings that would otherwise be set during construction. Defaultlist().
Input and Output Channels
Input and output channels are inherited from PipeOpImpute.
The output is the input Task with all affected features having missing values imputed as described above.
State
The $state is a named list with the $state elements inherited from PipeOpImpute.
The $state$model contains either ".MISSING" used for character and factor (also
ordered) features or numeric(1) indicating the constant value used for imputation of
integer and numeric features.
Parameters
The parameters are the parameters inherited from PipeOpImpute, as well as:
min::logical(1)
Shouldintegerandnumericfeatures be shifted below the minimum? Initialized toTRUE. IfFALSEthey are shifted above the maximum. See also the description above.offset::numeric(1)
Numerical non-negative offset as used in the description above forintegerandnumericfeatures. Initialized to1.multiplier::numeric(1)
Numerical non-negative multiplier as used in the description above forintegerandnumericfeatures. Initialized to1.
Internals
Adds an explicit new level() to factor and ordered features, but not to character features.
For integer and numeric features uses the min, max, diff and range functions.
integer and numeric features that are entirely NA are imputed as 0. factor and ordered features that are
entirely NA are imputed as ".MISSING".
Fields
Only fields inherited from PipeOp.
Methods
Only methods inherited from PipeOpImpute/PipeOp.
References
Ding Y, Simonoff JS (2010). “An Investigation of Missing Data Methods for Classification Trees Applied to Binary Response Data.” Journal of Machine Learning Research, 11(6), 131-170. https://jmlr.org/papers/v11/ding10a.html.
See also
https://mlr-org.com/pipeops.html
Other PipeOps:
PipeOp,
PipeOpEncodePL,
PipeOpEnsemble,
PipeOpImpute,
PipeOpTargetTrafo,
PipeOpTaskPreproc,
PipeOpTaskPreprocSimple,
mlr_pipeops,
mlr_pipeops_adas,
mlr_pipeops_blsmote,
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_decode,
mlr_pipeops_encode,
mlr_pipeops_encodeimpact,
mlr_pipeops_encodelmer,
mlr_pipeops_encodeplquantiles,
mlr_pipeops_encodepltree,
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_imputesample,
mlr_pipeops_kernelpca,
mlr_pipeops_learner,
mlr_pipeops_learner_pi_cvplus,
mlr_pipeops_learner_quantiles,
mlr_pipeops_missind,
mlr_pipeops_modelmatrix,
mlr_pipeops_multiplicityexply,
mlr_pipeops_multiplicityimply,
mlr_pipeops_mutate,
mlr_pipeops_nearmiss,
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_rowapply,
mlr_pipeops_scale,
mlr_pipeops_scalemaxabs,
mlr_pipeops_scalerange,
mlr_pipeops_select,
mlr_pipeops_smote,
mlr_pipeops_smotenc,
mlr_pipeops_spatialsign,
mlr_pipeops_subsample,
mlr_pipeops_targetinvert,
mlr_pipeops_targetmutate,
mlr_pipeops_targettrafoscalerange,
mlr_pipeops_textvectorizer,
mlr_pipeops_threshold,
mlr_pipeops_tomek,
mlr_pipeops_tunethreshold,
mlr_pipeops_unbranch,
mlr_pipeops_updatetarget,
mlr_pipeops_vtreat,
mlr_pipeops_yeojohnson
Other Imputation PipeOps:
PipeOpImpute,
mlr_pipeops_imputeconstant,
mlr_pipeops_imputehist,
mlr_pipeops_imputelearner,
mlr_pipeops_imputemean,
mlr_pipeops_imputemedian,
mlr_pipeops_imputemode,
mlr_pipeops_imputesample
Examples
library("mlr3")
set.seed(2409)
data = tsk("pima")$data()
data$y = factor(c(NA, sample(letters, size = 766, replace = TRUE), NA))
data$z = ordered(c(NA, sample(1:10, size = 767, replace = TRUE)))
task = TaskClassif$new("task", backend = data, target = "diabetes")
task$missings()
#> diabetes age glucose insulin mass pedigree pregnant pressure
#> 0 0 5 374 11 0 0 35
#> triceps y z
#> 227 2 1
po = po("imputeoor")
new_task = po$train(list(task = task))[[1]]
new_task$missings()
#> diabetes age pedigree pregnant glucose insulin mass pressure
#> 0 0 0 0 0 0 0 0
#> triceps y z
#> 0 0 0
new_task$data()
#> diabetes age pedigree pregnant glucose insulin mass pressure triceps
#> <fctr> <num> <num> <num> <num> <num> <num> <num> <num>
#> 1: pos 50 0.627 6 148 -819 33.6 72 35
#> 2: neg 31 0.351 1 85 -819 26.6 66 29
#> 3: pos 32 0.672 8 183 -819 23.3 64 -86
#> 4: neg 21 0.167 1 89 94 28.1 66 23
#> 5: pos 33 2.288 0 137 168 43.1 40 35
#> ---
#> 764: neg 63 0.171 10 101 180 32.9 76 48
#> 765: neg 27 0.340 2 122 -819 36.8 70 27
#> 766: neg 30 0.245 5 121 112 26.2 72 23
#> 767: pos 47 0.349 1 126 -819 30.1 60 -86
#> 768: neg 23 0.315 1 93 -819 30.4 70 31
#> y z
#> <fctr> <ord>
#> 1: .MISSING .MISSING
#> 2: l 9
#> 3: q 6
#> 4: f 3
#> 5: l 3
#> ---
#> 764: o 7
#> 765: n 5
#> 766: e 6
#> 767: c 8
#> 768: .MISSING 9
# recommended use when missing values are expected during prediction on
# factor columns that had no missing values during training
gr = po("imputeoor", create_empty_level = FALSE) %>>%
po("imputesample", affect_columns = selector_type(types = c("factor", "ordered")))
t1 = as_task_classif(data.frame(l = as.ordered(letters[1:3]), t = letters[1:3]), target = "t")
t2 = as_task_classif(data.frame(l = as.ordered(c("a", NA, NA)), t = letters[1:3]), target = "t")
gr$train(t1)[[1]]$data()
#> t l
#> <fctr> <ord>
#> 1: a a
#> 2: b b
#> 3: c c
# missing values during prediction are sampled randomly
gr$predict(t2)[[1]]$data()
#> t l
#> <fctr> <ord>
#> 1: a a
#> 2: b c
#> 3: c c