More trainer related naming alignment

[sfilipi]

Take another pass over the trainers and the model parameter types, and align them, because now we have:

LogisticRegressionMulticlassClassificationTrainer but MulticlassLogisticRegressionModelParameters.

I think following the same principles on the ModelParams might make them more relatable; so change MulticlassLogisticRegressionModelParameters to LogisticRegressionMulticlassModelParameters

[sfilipi]

@shauheen would this be a candidate for Project 13?

[abgoswam]

In general, it seems there isn't a 1:1 mapping between trainer and model parameter types.

• SdcaNonCalibratedBinaryClassificationTrainer uses LinearBinaryModelParameters
• SdcaCalibratedBinaryClassificationTrainer uses CalibratedModelParametersBase as type of the model parameter.

However, there are some discrepancies that exists. Whenever possible we should align the trainer and model parameter types.

Listing some of the trainers where we can fix this :

• MulticlassLogisticRegressionModelParameters
• BinaryClassificationGamModelParameters
• RegressionGamModelParameters
• MultiClassNaiveBayesModelParameters
• PrincipleComponentModelParameters
• OrdinaryLeastSquaresRegressionModelParameters
• FastForestClassificationModelParameters

[abgoswam]

Update:

• To keep consistency between Class names of Trainers and ModelParameters, we will not use the word "Classification" in either the Trainer class or the ModelParameter class

• We feel its OK to drop the word "Classification" for 2 main reasons :

  • we have sufficient context to just use suffix : BinaryTrainer , BinaryModelParameters etc. without using the word "Classification"
  • adding the word "Classification" leads to long names

Here is a summary of the Trainer and ModelParameter class names

Updated Trainer Class Name
FastTreeBinaryTrainer
FastTreeRegressionTrainer
FastTreeRankingTrainer
FastTreeTweedieTrainer
FastForestRegressionTrainer
FastForestBinaryTrainer
MatrixFactorizationTrainer
GamBinaryTrainer
GamRegressionTrainer
LogisticRegressionBinaryTrainer
LogisticRegressionMulticlassTrainer (see comment below by @wschin )
AveragedPerceptronTrainer
OnlineGradientDescentTrainer
PoissonRegressionTrainer
KMeansTrainer
OlsTrainer
PriorTrainer
PairwiseCouplingTrainer
OneVersusAllTrainer
NaiveBayesMulticlassTrainer
SgdCalibratedTrainer
SgdNonCalibratedTrainer
FieldAwareFactorizationMachineTrainer
SymbolicSgdTrainer
LightGbmRegressionTrainer
LightGbmBinaryTrainer
LightGbmRankingTrainer
LightGbmMulticlassTrainer
LinearSvmTrainer
RandomizedPcaTrainer
SdcaCalibratedBinaryTrainer
SdcaNonCalibratedBinaryTrainer
SdcaMulticlassTrainer
SdcaRegressionTrainer
SdcaTrainerBase
LbfgsTrainerBase
GamTrainerBase

Updated ModelParameter Class Name
GamBinaryModelParameters
GamRegressionModelParameters
LightGbmBinaryModelParameters
LightGbmRankingModelParameters
LightGbmRegressionModelParameters
FastTreeBinaryModelParameters
FastTreeRegressionModelParameters
FastTreeRankingModelParameters
FastTreeTweedieModelParameters
FastForestBinaryModelParameters
FastForestRegressionModelParameters
NaiveBayesMulticlassModelParameters
OlsModelParameters
PcaModelParameters
FieldAwareFactorizationMachineModelParameters
KMeansModelParameters
PoissonRegressionModelParameters
MatrixFactorizationModelParameters
OneVersusAllModelParameters
PairwiseCouplingModelParameters
PriorModelParameters
LinearRegressionModelParameters
RegressionModelParameters
LinearBinaryModelParameters
CalibratedModelParametersBase
GamModelParametersBase
TreeEnsembleModelParameters
TreeEnsembleModelParametersBasedOnQuantileRegressionTree
TreeEnsembleModelParametersBasedOnRegressionTree

[wschin]

For multi-class LR trainer and its model, we will different names in #2976. Looks like they don't need multiclass.

• (rename) LogisticRegressionMulticlassClassificationTrainer ---> LbfgsMaximumEntropyTrainer
• (rename) MulticlassLogisticRegressionModelParameters ---> MaximumEntropyModelParameters

We can NOT have LogisticRegressionMulticlass because LogisticRegression is binary classification only.