The following classes provide a unified interface to all popular machine learning methods in R: (cost-sensitive) classification, regression, survival analysis, and clustering. Many are already integrated in mlr, others are not, but the package is specifically designed to make extensions simple.
Section integrated learners shows the already implemented machine learning methods and their properties. If your favorite method is missing, either open an issue or take a look at how to integrate a learning method yourself. This basic introduction demonstrates how to use already implemented learners.
Constructing a learner
A learner in mlr is generated by calling makeLearner(). In the constructor you need to specify which learning method you want to use. Moreover, you can:
- Set hyperparameters.
- Control the output for later prediction, e.g., for classification whether you want a factor of predicted class labels or probabilities.
- Set an ID to name the object (some methods will later use this ID to name results or annotate plots).
# Classification tree, set it up for predicting probabilities
classif.lrn = makeLearner("classif.randomForest", predict.type = "prob", fix.factors.prediction = TRUE)
# Regression gradient boosting machine, specify hyperparameters via a list
regr.lrn = makeLearner("regr.gbm", par.vals = list(n.trees = 500, interaction.depth = 3))
# Cox proportional hazards model with custom name
surv.lrn = makeLearner("surv.coxph", id = "cph")
# K-means with 5 clusters
cluster.lrn = makeLearner("cluster.kmeans", centers = 5)
# Multilabel Random Ferns classification algorithm
multilabel.lrn = makeLearner("multilabel.rFerns")The first argument specifies which algorithm to use. The naming convention is classif.<R_method_name> for classification methods, regr.<R_method_name> for regression methods, surv.<R_method_name> for survival analysis, cluster.<R_method_name> for clustering methods, and multilabel.<R_method_name> for multilabel classification.
Hyperparameter values can be specified either via the ... argument or as a list via par.vals. The first option is preferred as par.vals is mainly used to declare hyperparameters that are set differently in mlr compared to the defaults of the underlying model. If you want to change a hyperparameter in mlr by default that differs from the actual default, make sure to also add an entry in the "note" slot of the learner. This entry should describe the reason for the change. Common ones are turning off automatic parallelization or changing logical arguments of the learner to enable a more conservative memory management.
Occasionally, factor features may cause problems when fewer levels are present in the test data set than in the training data. By setting fix.factors.prediction = TRUE these are avoided by adding a factor level for missing data in the test data set.
Let’s have a look at two of the learners created above.
classif.lrn
## Learner classif.randomForest from package randomForest
## Type: classif
## Name: Random Forest; Short name: rf
## Class: classif.randomForest
## Properties: twoclass,multiclass,numerics,factors,ordered,prob,class.weights,oobpreds,featimp
## Predict-Type: prob
## Hyperparameters:
surv.lrn
## Learner cph from package survival
## Type: surv
## Name: Cox Proportional Hazard Model; Short name: coxph
## Class: surv.coxph
## Properties: numerics,factors,weights
## Predict-Type: response
## Hyperparameters:All generated learners are objects of class Learner (makeLearner()). This class contains the properties of the method, e.g., which types of features it can handle, what kind of output is possible during prediction, and whether multi-class problems, observations weights or missing values are supported.
As you might have noticed, there is currently no special learner class for cost-sensitive classification. For ordinary misclassification costs you can use standard classification methods. For example-dependent costs there are several ways to generate cost-sensitive learners from ordinary regression and classification learners. This is explained in greater detail in the section about cost-sensitive classification.
Accessing a learner
The Learner (makeLearner()) object is a list and the following elements contain information regarding the hyperparameters and the type of prediction.
# Get the configured hyperparameter settings that deviate from the defaults
cluster.lrn$par.vals
## $centers
## [1] 5
# Get the set of hyperparameters
classif.lrn$par.set
## Type len Def Constr Req Tunable Trafo
## ntree integer - 500 1 to Inf - TRUE -
## mtry integer - - 1 to Inf - TRUE -
## replace logical - TRUE - - TRUE -
## classwt numericvector <NA> - 0 to Inf - TRUE -
## cutoff numericvector <NA> - 0 to 1 - TRUE -
## strata untyped - - - - FALSE -
## sampsize integervector <NA> - 1 to Inf - TRUE -
## nodesize integer - 1 1 to Inf - TRUE -
## maxnodes integer - - 1 to Inf - TRUE -
## importance logical - FALSE - - TRUE -
## localImp logical - FALSE - - TRUE -
## proximity logical - FALSE - - FALSE -
## oob.prox logical - - - Y FALSE -
## norm.votes logical - TRUE - - FALSE -
## do.trace logical - FALSE - - FALSE -
## keep.forest logical - TRUE - - FALSE -
## keep.inbag logical - FALSE - - FALSE -
# Get the type of prediction
regr.lrn$predict.type
## [1] "response"Slot $par.set is an object of class ParamSet (ParamHelpers::makeParamSet()). It contains, among others, the type of hyperparameters (e.g., numeric, logical), potential default values and the range of allowed values.
Moreover, mlr provides function getHyperPars() or its alternative getLearnerParVals() to access the current hyperparameter setting of a Learner, (makeLearner()) and getParamSet() to get a description of all possible settings. These are particularly useful in case of wrapped Learner (makeLearner())s, for example if a learner is fused with a feature selection strategy, and both, the learner as well the feature selection method, have hyperparameters. For details see the section on wrapped learners.
# Get current hyperparameter settings
getHyperPars(cluster.lrn)
## $centers
## [1] 5
# Get a description of all possible hyperparameter settings
getParamSet(classif.lrn)
## Type len Def Constr Req Tunable Trafo
## ntree integer - 500 1 to Inf - TRUE -
## mtry integer - - 1 to Inf - TRUE -
## replace logical - TRUE - - TRUE -
## classwt numericvector <NA> - 0 to Inf - TRUE -
## cutoff numericvector <NA> - 0 to 1 - TRUE -
## strata untyped - - - - FALSE -
## sampsize integervector <NA> - 1 to Inf - TRUE -
## nodesize integer - 1 1 to Inf - TRUE -
## maxnodes integer - - 1 to Inf - TRUE -
## importance logical - FALSE - - TRUE -
## localImp logical - FALSE - - TRUE -
## proximity logical - FALSE - - FALSE -
## oob.prox logical - - - Y FALSE -
## norm.votes logical - TRUE - - FALSE -
## do.trace logical - FALSE - - FALSE -
## keep.forest logical - TRUE - - FALSE -
## keep.inbag logical - FALSE - - FALSE -We can also use getParamSet() or its alias getLearnerParamSet() to get a quick overview about the available hyperparameters and defaults of a learning method without explicitly constructing it (by calling makeLearner()).
getParamSet("classif.randomForest")
## Type len Def Constr Req Tunable Trafo
## ntree integer - 500 1 to Inf - TRUE -
## mtry integer - - 1 to Inf - TRUE -
## replace logical - TRUE - - TRUE -
## classwt numericvector <NA> - 0 to Inf - TRUE -
## cutoff numericvector <NA> - 0 to 1 - TRUE -
## strata untyped - - - - FALSE -
## sampsize integervector <NA> - 1 to Inf - TRUE -
## nodesize integer - 1 1 to Inf - TRUE -
## maxnodes integer - - 1 to Inf - TRUE -
## importance logical - FALSE - - TRUE -
## localImp logical - FALSE - - TRUE -
## proximity logical - FALSE - - FALSE -
## oob.prox logical - - - Y FALSE -
## norm.votes logical - TRUE - - FALSE -
## do.trace logical - FALSE - - FALSE -
## keep.forest logical - TRUE - - FALSE -
## keep.inbag logical - FALSE - - FALSE -Functions for accessing a Learner’s meta information are available in mlr. We can use getLearnerId(), getLearnerShortName() and getLearnerType() to get Learner’s ID, short name and type, respectively. Moreover, in order to show the required packages for the Learner, one can call getLearnerPackages().
# Get object's id
getLearnerId(surv.lrn)
## [1] "cph"
# Get the short name
getLearnerShortName(classif.lrn)
## [1] "rf"
# Get the type of the learner
getLearnerType(multilabel.lrn)
## [1] "multilabel"
# Get required packages
getLearnerPackages(cluster.lrn)
## [1] "stats" "clue"Modifying a learner
There are also some functions that enable you to change certain aspects of a Learner (makeLearner()) without needing to create a new Learner (makeLearner()) from scratch. Here are some examples.
# Change the ID
surv.lrn = setLearnerId(surv.lrn, "CoxModel")
surv.lrn
## Learner CoxModel from package survival
## Type: surv
## Name: Cox Proportional Hazard Model; Short name: coxph
## Class: surv.coxph
## Properties: numerics,factors,weights
## Predict-Type: response
## Hyperparameters:
# Change the prediction type, predict a factor with class labels instead of probabilities
classif.lrn = setPredictType(classif.lrn, "response")
# Change hyperparameter values
cluster.lrn = setHyperPars(cluster.lrn, centers = 4)
# Go back to default hyperparameter values
regr.lrn = removeHyperPars(regr.lrn, c("n.trees", "interaction.depth"))Listing learners
A list of all learners integrated in mlr and their respective properties is shown in the Appendix.
If you would like a list of available learners, maybe only with certain properties or suitable for a certain learning Task() use function listLearners().
# List everything in mlr
lrns = listLearners()
head(lrns[c("class", "package")])
## class package
## 1 classif.ada ada,rpart
## 2 classif.adaboostm1 RWeka
## 3 classif.binomial stats
## 4 classif.boosting adabag,rpart
## 5 classif.bst bst,rpart
## 6 classif.C50 C50
# List classifiers that can output probabilities
lrns = listLearners("classif", properties = "prob")
head(lrns[c("class", "package")])
## class package
## 1 classif.ada ada,rpart
## 2 classif.adaboostm1 RWeka
## 3 classif.binomial stats
## 4 classif.boosting adabag,rpart
## 5 classif.C50 C50
## 6 classif.cforest party
# List classifiers that can be applied to iris (i.e., multiclass) and output probabilities
lrns = listLearners(iris.task, properties = "prob")
head(lrns[c("class", "package")])
## class package
## 1 classif.adaboostm1 RWeka
## 2 classif.boosting adabag,rpart
## 3 classif.C50 C50
## 4 classif.cforest party
## 5 classif.ctree party
## 6 classif.cvglmnet glmnet
# The calls above return character vectors, but you can also create learner objects
head(listLearners("cluster", create = TRUE), 2)
## [[1]]
## Learner cluster.cmeans from package e1071,clue
## Type: cluster
## Name: Fuzzy C-Means Clustering; Short name: cmeans
## Class: cluster.cmeans
## Properties: numerics,prob
## Predict-Type: response
## Hyperparameters: centers=2
##
##
## [[2]]
## Learner cluster.Cobweb from package RWeka
## Type: cluster
## Name: Cobweb Clustering Algorithm; Short name: cobweb
## Class: cluster.Cobweb
## Properties: numerics
## Predict-Type: response
## Hyperparameters: