Use binary relevance method to create a multilabel learner.
Every learner which is implemented in mlr and which supports binary classification can be converted to a wrapped binary relevance multilabel learner. The multilabel classification problem is converted into simple binary classifications for each label/target on which the binary learner is applied.
Models can easily be accessed via getLearnerModel.
Note that it does not make sense to set a threshold in the used base learner
when you predict probabilities.
On the other hand, it can make a lot of sense, to call setThreshold
on the MultilabelBinaryRelevanceWrapper for each label indvidually;
Or to tune these thresholds with tuneThreshold; especially when you face very
unabalanced class distributions for each binary label.
makeMultilabelBinaryRelevanceWrapper(learner)
learner |
(Learner | |
Tsoumakas, G., & Katakis, I. (2006) Multi-label classification: An overview. Dept. of Informatics, Aristotle University of Thessaloniki, Greece.
Other wrapper:
makeBaggingWrapper(),
makeClassificationViaRegressionWrapper(),
makeConstantClassWrapper(),
makeCostSensClassifWrapper(),
makeCostSensRegrWrapper(),
makeDownsampleWrapper(),
makeDummyFeaturesWrapper(),
makeExtractFDAFeatsWrapper(),
makeFeatSelWrapper(),
makeFilterWrapper(),
makeImputeWrapper(),
makeMulticlassWrapper(),
makeMultilabelClassifierChainsWrapper(),
makeMultilabelDBRWrapper(),
makeMultilabelNestedStackingWrapper(),
makeMultilabelStackingWrapper(),
makeOverBaggingWrapper(),
makePreprocWrapperCaret(),
makePreprocWrapper(),
makeRemoveConstantFeaturesWrapper(),
makeSMOTEWrapper(),
makeTuneWrapper(),
makeUndersampleWrapper(),
makeWeightedClassesWrapper()
d = getTaskData(yeast.task)
# drop some labels so example runs faster
d = d[seq(1, nrow(d), by = 20), c(1:2, 15:17)]
task = makeMultilabelTask(data = d, target = c("label1", "label2"))
lrn = makeLearner("classif.rpart")
lrn = makeMultilabelBinaryRelevanceWrapper(lrn)
lrn = setPredictType(lrn, "prob")
# train, predict and evaluate
mod = train(lrn, task)
pred = predict(mod, task)
performance(pred, measure = list(multilabel.hamloss, multilabel.subset01, multilabel.f1))
# the next call basically has the same structure for any multilabel meta wrapper
getMultilabelBinaryPerformances(pred, measures = list(mmce, auc))
# above works also with predictions from resample!Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.