We use a multiclass-to-binary reduction principle, where multiple binary problems are created from the multiclass task. How these binary problems are generated is defined by an error-correcting-output-code (ECOC) code book. This also allows the simple and well-known one-vs-one and one-vs-rest approaches. Decoding is currently done via Hamming decoding, see e.g. here http://jmlr.org/papers/volume11/escalera10a/escalera10a.pdf.
Currently, the approach always operates on the discrete predicted labels of the binary base models (instead of their probabilities) and the created wrapper cannot predict posterior probabilities.
makeMulticlassWrapper(learner, mcw.method = "onevsrest")Learner | character(1)]
The learner.
If you pass a string the learner will be created via makeLearner.function(task) and which
returns a ECOC codematrix with entries +1,-1,0.
Columns define new binary problems, rows correspond to classes (rows must be named).
0 means class is not included in binary problem.
Default is “onevsrest”.Learner].
makeBaggingWrapper,
makeConstantClassWrapper,
makeCostSensClassifWrapper,
makeCostSensRegrWrapper,
makeDownsampleWrapper,
makeFeatSelWrapper,
makeFilterWrapper,
makeImputeWrapper,
makeMultilabelBinaryRelevanceWrapper,
makeMultilabelClassifierChainsWrapper,
makeMultilabelDBRWrapper,
makeMultilabelNestedStackingWrapper,
makeMultilabelStackingWrapper,
makeOverBaggingWrapper,
makePreprocWrapperCaret,
makePreprocWrapper,
makeRemoveConstantFeaturesWrapper,
makeSMOTEWrapper,
makeTuneWrapper,
makeUndersampleWrapper,
makeWeightedClassesWrapper