v0.21.0

MLJ v0.21.0

Diff since v0.20.9

• (new models) Add the following models from MLJTransforms.jl and make them immediately available to the MLJ user (she does not need to use @load to load them): OrdinalEncoder, FrequencyEncoder, TargetEncoder, ContrastEncoder, CardinalityReducer, MissingnessEncoder.
• (mildly breaking) Have MLJTransforms.jl, instead of MLJModels.jl, provide the following built-in models, whose behaviour is unchanged: ContinuousEncoder, FillImputer, InteractionTransformer, OneHotEncoder, Standardizer, UnivariateBoxCoxTransformer, UnivariateDiscretizer, UnivariateFillImputer, UnivariateTimeTypeToContinuous, Standardizer.

Guide for possible source of breakage: While it was never necessary to use @load to load one of the models in the last list (assuming you have first run using MLJ) this is frequently not realised by users, and one sees things like @load OneHotEncoder pkg=MLJModels, which this release will break. If such a call is preceded by using MLJ or using MLJTransforms you can remove the loading command altogether (OneHotEncoder() already works), and in any case you can instead use @load OneHotEncoder pkg=MLJTransforms.

Merged pull requests:

• Make updates to reflect code reorganisation around addition of MLJTransforms.jl (#1177) (@ablaom)
• For a 0.21 release (#1180) (@ablaom)

Closed issues:

• Decision trees from ScikitLearn.jl not available (#545)
• Document RecursiveFeatureElimination (#1162)