2.5 Resampling

Resampling strategies are usually used to assess the performance of a learning algorithm. mlr3 entails 6 predefined resampling strategies: Cross-validation, Leave-one-out cross validation, Repeated cross-validation, Out-of-bag bootstrap and other variants (e.g. b632), Monte-Carlo cross-validation and Holdout. The following sections provide guidance on how to set and select a resampling strategy and how to subsequently instantiate the resampling process.

Below you can find a graphical illustration of the resampling process:

2.5.1 Settings

In this example we use the iris task and a simple classification tree from the rpart package.

When performing resampling with a dataset, we first need to define which approach should be used. mlr3 resampling strategies and their parameters can be queried by looking at the data.table output of the mlr_resamplings dictionary:

##            key        params iters
## 1:   bootstrap repeats,ratio    30
## 2:      custom                   0
## 3:          cv         folds    10
## 4:     holdout         ratio     1
## 5: repeated_cv repeats,folds   100
## 6: subsampling repeats,ratio    30

Additional resampling methods for special use cases will be available via extension packages, such as mlr3spatiotemporal for spatial data (still in development).

The model fit conducted in the train/predict/score chapter is equivalent to a “holdout resampling”, so let’s consider this one first. Again, we can retrieve elements from the dictionary mlr_resamplings via $get() or with the convenience functionrsmp():

## <ResamplingHoldout> with 1 iterations
## * Instantiated: FALSE
## * Parameters: ratio=0.6667

Note that the $is_instantiated field is set to FALSE. This means we did not actually apply the strategy on a dataset yet. Applying the strategy on a dataset is done in the next section Instantiation.

By default we get a .66/.33 split of the data. There are two ways in which the ratio can be changed:

  1. Overwriting the slot in $param_set$values using a named list:
  1. Specifying the resampling parameters directly during construction:
## <ResamplingHoldout> with 1 iterations
## * Instantiated: FALSE
## * Parameters: ratio=0.8

2.5.2 Instantiation

So far we just set the stage and selected the resampling strategy.

To actually perform the splitting and obtain indices for the training and the test split the resampling needs a Task. By calling the method instantiate(), we split the indices of the data into indices for training and test sets. These resulting indices are stored in the Resampling object:

## [1] 3
##  int [1:100] 3 4 5 7 8 11 15 19 23 24 ...
##  int [1:50] 6 10 12 13 14 16 17 18 21 22 ...

2.5.3 Execution

With a Task, a Learner and a Resampling object we can call resample(), which fits the learner to the task at hand according to the given resampling strategy. This in turn creates a ResampleResult object.

Before we go into more detail, let’s change the resampling to a “3-fold cross-validation” to better illustrate what operations are possible with a ResampleResult. Additionally, when actually fitting the models, we tell resample() to keep the fitted models by setting the store_models option to TRUE:

## <ResampleResult> of 3 iterations
## * Task: pima
## * Learner: classif.rpart
## * Warnings: 0 in 0 iterations
## * Errors: 0 in 0 iterations

The following operations are supported with ResampleResult objects:

Calculate the average performance across all resampling iterations:

## classif.ce 
##     0.2773

Extract the performance for the individual resampling iterations:

##             task task_id               learner    learner_id     resampling
## 1: <TaskClassif>    pima <LearnerClassifRpart> classif.rpart <ResamplingCV>
## 2: <TaskClassif>    pima <LearnerClassifRpart> classif.rpart <ResamplingCV>
## 3: <TaskClassif>    pima <LearnerClassifRpart> classif.rpart <ResamplingCV>
##    resampling_id iteration prediction classif.ce
## 1:            cv         1     <list>     0.2891
## 2:            cv         2     <list>     0.2344
## 3:            cv         3     <list>     0.3086

Check for warnings or errors:

## Empty data.table (0 rows and 2 cols): iteration,msg
## Empty data.table (0 rows and 2 cols): iteration,msg

Extract and inspect the resampling splits:

## <ResamplingCV> with 3 iterations
## * Instantiated: TRUE
## * Parameters: folds=3
## [1] 3
##  int [1:256] 2 3 8 15 16 17 19 20 21 24 ...
##  int [1:512] 10 11 14 25 27 32 35 37 45 46 ...

Retrieve the learner of a specific iteration and inspect it:

## n= 512 
## 
## node), split, n, loss, yval, (yprob)
##       * denotes terminal node
## 
##  1) root 512 182 neg (0.3555 0.6445)  
##    2) glucose>=122.5 219  88 pos (0.5982 0.4018)  
##      4) mass>=29.85 161  48 pos (0.7019 0.2981) *
##      5) mass< 29.85 58  18 neg (0.3103 0.6897) *
##    3) glucose< 122.5 293  51 neg (0.1741 0.8259)  
##      6) pregnant>=6.5 53  23 neg (0.4340 0.5660)  
##       12) glucose>=96 34  13 pos (0.6176 0.3824) *
##       13) glucose< 96 19   2 neg (0.1053 0.8947) *
##      7) pregnant< 6.5 240  28 neg (0.1167 0.8833) *

Extract the predictions:

## <PredictionClassif> for 768 observations:
##     row_id truth response prob.pos prob.neg
##          2   neg      neg   0.1167  0.88333
##          3   pos      neg   0.3103  0.68966
##          8   neg      pos   0.6176  0.38235
## ---                                        
##        758   pos      neg   0.2101  0.78989
##        759   neg      neg   0.2101  0.78989
##        760   pos      pos   0.9231  0.07692
## <PredictionClassif> for 256 observations:
##     row_id truth response prob.pos prob.neg
##          2   neg      neg   0.1167   0.8833
##          3   pos      neg   0.3103   0.6897
##          8   neg      pos   0.6176   0.3824
## ---                                        
##        763   neg      neg   0.1053   0.8947
##        764   neg      pos   0.6176   0.3824
##        765   neg      neg   0.1167   0.8833

Note that if you want to compare multiple Learners in a fair manner, it is important to ensure that each learner operates on the same resampling instance. This can be achieved by manually instantiating the instance before fitting model(s) on it.

Hint: If your aim is to compare different Task, Learner or Resampling, you are better off using the benchmark() function which is covered in the next section on benchmarking. It is a wrapper around resample(), simplifying the handling of large comparison grids.

If you discover this only after you’ve run multiple resample() calls, don’t worry. You can combine multiple ResampleResult objects into a BenchmarkResult (also explained in the section benchmarking).

2.5.4 Custom resampling

Sometimes it is necessary to perform resampling with custom splits. If you want to do that because you are coming from a specific modeling field, first take a look at the mlr3 extension packages, to check wheter your resampling method has been implemented already. If this is not the case, feel welcome to extend an existing package or create your own extension package.

A manual resampling instance can be created using the "custom" template.

## [1] 1
##  [1]   1   2   3   4   5   6   7   8   9  10  51  52  53  54  55  56  57  58  59
## [20]  60 101 102 103 104 105 106 107 108 109 110
##  [1]  11  12  13  14  15  16  17  18  19  20  61  62  63  64  65  66  67  68  69
## [20]  70 111 112 113 114 115 116 117 118 119 120

2.5.5 Plotting Resample Results

Again, mlr3viz provides a autoplot() method.

All available plot types are listed on the manual page of autoplot.ResampleResult().