mlr3 Basics on “Iris” - Hello World!

Goals and Prerequisites

This use case shows how to use the basic mlr3 package on the iris Task, so it’s our “Hello World” example. It assumes no prior knowledge in ML or mlr3. You can find most of the content here also in the mlr3book in a more detailed way. Hence we will not make a lot of general comments, but keep it hands-on and short.

The following operations are shown:

• Creating Tasks and Learners
• Training and predicting
• Resampling / cross-validation
• Installing more Learners
• Benchmarking to compare multiple Learners

Loading basic packages

We load the mlr3verse package which pulls in the most important packages for this example. The mlr3learners package loads additional learners.

library(mlr3verse)
library(mlr3learners)

lgr::get_logger("mlr3")$set_threshold("warn")

Creating tasks and learners

Let’s work on the canonical, simple iris data set, and try out some ML algorithms. We will start by using a decision tree with default settings.

# creates mlr3 task from scratch, from a data.frame
# 'target' names the column in the dataset we want to learn to predict
task = as_task_classif(iris, target = "Species")
# in this case we could also take the iris example from mlr3's dictionary of shipped example tasks
# 2 equivalent calls to create a task. The second is just sugar for the user.
task = mlr_tasks$get("iris")
task = tsk("iris")
print(task)

<TaskClassif:iris> (150 x 5): Iris Flowers
* Target: Species
* Properties: multiclass
* Features (4):
  - dbl (4): Petal.Length, Petal.Width, Sepal.Length, Sepal.Width

# create learner from dictionary of mlr3learners
# 2 equivalent calls:
learner_1 = mlr_learners$get("classif.rpart")
learner_1 = lrn("classif.rpart")
print(learner_1)

<LearnerClassifRpart:classif.rpart>: Classification Tree
* Model: -
* Parameters: xval=0
* Packages: mlr3, rpart
* Predict Types:  [response], prob
* Feature Types: logical, integer, numeric, factor, ordered
* Properties: importance, missings, multiclass, selected_features, twoclass, weights

Train and predict

Now the usual ML operations: Train on some observations, predict on others.

# train learner on subset of task
learner_1$train(task, row_ids = 1:120)
# this is what the decision tree looks like
print(learner_1$model)

n= 120 

node), split, n, loss, yval, (yprob)
      * denotes terminal node

1) root 120 70 setosa (0.41666667 0.41666667 0.16666667)  
  2) Petal.Length< 2.45 50  0 setosa (1.00000000 0.00000000 0.00000000) *
  3) Petal.Length>=2.45 70 20 versicolor (0.00000000 0.71428571 0.28571429)  
    6) Petal.Length< 4.95 49  1 versicolor (0.00000000 0.97959184 0.02040816) *
    7) Petal.Length>=4.95 21  2 virginica (0.00000000 0.09523810 0.90476190) *

# predict using observations from task
prediction = learner_1$predict(task, row_ids = 121:150)
# predict using "new" observations from an external data.frame
prediction = learner_1$predict_newdata(newdata = iris[121:150, ])
print(prediction)

<PredictionClassif> for 30 observations:
    row_ids     truth   response
          1 virginica  virginica
          2 virginica versicolor
          3 virginica  virginica
---                             
         28 virginica  virginica
         29 virginica  virginica
         30 virginica  virginica

Evaluation

Let’s score our Prediction object with some metrics. And take a deeper look by inspecting the confusion matrix.

head(as.data.table(mlr_measures))

              key                          label task_type          packages predict_type task_properties
1:            aic   Akaika Information Criterion      <NA>              mlr3         <NA>                
2:            bic Bayesian Information Criterion      <NA>              mlr3         <NA>                
3:    classif.acc        Classification Accuracy   classif mlr3,mlr3measures     response                
4:    classif.auc       Area Under the ROC Curve   classif mlr3,mlr3measures         prob        twoclass
5:   classif.bacc              Balanced Accuracy   classif mlr3,mlr3measures     response                
6: classif.bbrier             Binary Brier Score   classif mlr3,mlr3measures         prob        twoclass

scores = prediction$score(msr("classif.acc"))
print(scores)

classif.acc 
  0.8333333 

scores = prediction$score(msrs(c("classif.acc", "classif.ce")))
print(scores)

classif.acc  classif.ce 
  0.8333333   0.1666667 

cm = prediction$confusion
print(cm)

            truth
response     setosa versicolor virginica
  setosa          0          0         0
  versicolor      0          0         5
  virginica       0          0        25

Changing hyperpars

The Learner contains information about all parameters that can be configured, including data type, constraints, defaults, etc. We can change the hyperparameters either during construction of later through an active binding.

as.data.table(learner_1$param_set)[, .(id, class, lower, upper, nlevels)]

                id    class lower upper nlevels
 1:             cp ParamDbl     0     1     Inf
 2:     keep_model ParamLgl    NA    NA       2
 3:     maxcompete ParamInt     0   Inf     Inf
 4:       maxdepth ParamInt     1    30      30
 5:   maxsurrogate ParamInt     0   Inf     Inf
 6:      minbucket ParamInt     1   Inf     Inf
 7:       minsplit ParamInt     1   Inf     Inf
 8: surrogatestyle ParamInt     0     1       2
 9:   usesurrogate ParamInt     0     2       3
10:           xval ParamInt     0   Inf     Inf

learner_2 = lrn("classif.rpart", predict_type = "prob", minsplit = 50)
learner_2$param_set$values$minsplit = 50

Resampling

Resampling simply repeats the train-predict-score loop and collects all results in a nice data.table::data.table().

cv10 = rsmp("cv", folds = 10)
rr = resample(task, learner_1, cv10)
print(rr)

<ResampleResult> of 10 iterations
* Task: iris
* Learner: classif.rpart
* Warnings: 0 in 0 iterations
* Errors: 0 in 0 iterations

rr$score(msrs(c("classif.acc", "classif.ce")))[, .(iteration, task_id, learner_id, resampling_id, classif.ce)]

    iteration task_id    learner_id resampling_id classif.ce
 1:         1    iris classif.rpart            cv 0.06666667
 2:         2    iris classif.rpart            cv 0.00000000
 3:         3    iris classif.rpart            cv 0.06666667
 4:         4    iris classif.rpart            cv 0.06666667
 5:         5    iris classif.rpart            cv 0.20000000
 6:         6    iris classif.rpart            cv 0.06666667
 7:         7    iris classif.rpart            cv 0.06666667
 8:         8    iris classif.rpart            cv 0.00000000
 9:         9    iris classif.rpart            cv 0.00000000
10:        10    iris classif.rpart            cv 0.20000000

# get all predictions nicely concatenated in a table
prediction = rr$prediction()
as.data.table(prediction)

     row_ids      truth   response
  1:       5     setosa     setosa
  2:       7     setosa     setosa
  3:      26     setosa     setosa
  4:      34     setosa     setosa
  5:      55 versicolor versicolor
 ---                              
146:     100 versicolor versicolor
147:     113  virginica  virginica
148:     124  virginica  virginica
149:     136  virginica  virginica
150:     139  virginica versicolor

cm = prediction$confusion
print(cm)

            truth
response     setosa versicolor virginica
  setosa         50          0         0
  versicolor      0         45         6
  virginica       0          5        44

Populating the learner dictionary

mlr3learners ships out with a dozen different popular Learners. We can list them from the dictionary. If we want more, we can install an extension package, mlr3extralearners, from GitHub. Note how after loading mlr3extralearners the dictionary increases in size.

head(as.data.table(mlr_learners)[, c("key", "packages")])

                   key                 packages
1:   classif.cv_glmnet mlr3,mlr3learners,glmnet
2:       classif.debug                     mlr3
3: classif.featureless                     mlr3
4:      classif.glmnet mlr3,mlr3learners,glmnet
5:        classif.kknn   mlr3,mlr3learners,kknn
6:         classif.lda   mlr3,mlr3learners,MASS

library(mlr3extralearners)
print(as.data.table(mlr_learners)[, c("key", "packages")])

                    key                                                packages
  1: classif.AdaBoostM1                            mlr3,mlr3extralearners,RWeka
  2:        classif.C50                              mlr3,mlr3extralearners,C50
  3:        classif.IBk                            mlr3,mlr3extralearners,RWeka
  4:        classif.J48                            mlr3,mlr3extralearners,RWeka
  5:       classif.JRip                            mlr3,mlr3extralearners,RWeka
 ---                                                                           
130:     surv.penalized       mlr3,mlr3proba,mlr3extralearners,penalized,pracma
131:        surv.ranger                 mlr3,mlr3proba,mlr3extralearners,ranger
132:         surv.rfsrc mlr3,mlr3proba,mlr3extralearners,randomForestSRC,pracma
133:           surv.svm            mlr3,mlr3proba,mlr3extralearners,survivalsvm
134:       surv.xgboost                mlr3,mlr3proba,mlr3extralearners,xgboost

Benchmarking multiple learners

The benchmark function can conveniently compare `r ref(“Learner”, “Learners”) on the same dataset(s).

learners = list(learner_1, learner_2, lrn("classif.randomForest"))
grid = benchmark_grid(task, learners, cv10)
bmr = benchmark(grid)
print(bmr)

<BenchmarkResult> of 30 rows with 3 resampling runs
 nr task_id           learner_id resampling_id iters warnings errors
  1    iris        classif.rpart            cv    10        0      0
  2    iris        classif.rpart            cv    10        0      0
  3    iris classif.randomForest            cv    10        0      0

print(bmr$aggregate(measures = msrs(c("classif.acc", "classif.ce"))))

   nr      resample_result task_id           learner_id resampling_id iters classif.acc classif.ce
1:  1 <ResampleResult[21]>    iris        classif.rpart            cv    10   0.9266667 0.07333333
2:  2 <ResampleResult[21]>    iris        classif.rpart            cv    10   0.9266667 0.07333333
3:  3 <ResampleResult[21]>    iris classif.randomForest            cv    10   0.9533333 0.04666667

Conclusion

We left out a lot of details and other features. If you want to know more, read the mlr3book and the documentation of the mentioned packages.