Wrap Scikitlearn.jl Elastic Net algorithms

[ablaom]

• sklearn.linear_model.ElasticNet : this will be ElasticNet <: Deterministic{Any} , target_type(::ElasticNet) = MLJBase.Continuous, input_types(::ElasticNet) = MLJBase.Continous

• sklearn.linear_model.ElasticNetCV: this will be ElasticNetCV <: Deterministic{Any}, with same target_type and input_types as above. Note the sklearn model has verbose as a hyperparameter, which we will drop from the MLJ model. The MLJ fit will pass either 0 or 1 to the sklearn fit, according to whether the MLJ fit verbosity is <= 0 or > 0.

[ablaom]

Some extra details:

ElasticNetCV

• Return for the report a named tuple with keys (:l1_ratio, :alpha, :n_iter, :dual_gap) whose values are fitresult.l1_ratio_, fitresult.alpha_, etc

• Define fitted_params(::ElasticNetCV, fitresult) to be named tuple with keys (:intercept, :coef) and values given by fitresult.intercept_ and fitresult.coef_.

• If alphas are specified, clean! should ensure they are strictly positive (set to one otherwise). The value(s) of l1_ratio must be in the interval (0, 1] (and set to one otherwise).

[ablaom]

Plain ElasticNet

• Return for the report a named tuple with keys (:n_iter, :dual_gap) whose values are fitresult.n_iter_, fitresult.dual_gap_.

• Define fitted_params(::ElasticNet, fitresult) to be named tuple with keys (:intercept, :coef) and values given by fitresult.intercept_ and fitresult.coef_.

• The clean! method to set alpha=1 if alpha is not strictly positive. It should set l1_ratio=1 if l1_ratio is not in the interval (0, 1].

[ValdarT]

Just out of curiosity.. Why wrap scikit-learn when Julia already has a wrapper around the original Fortran code (https://github.com/JuliaStats/GLMNet.jl) as well as a native Julia implementation (https://github.com/JuliaStats/Lasso.jl)?

[ablaom]

Yes, I am aware of Lasso.jl, having implemented Koala's interface for it. We doing the sklearn because it looked quick and easy for @ysimillides, who has already done some sk wraps, and (I am busy with other things). It also struck me that Lasso.jl needed a little TLC, although maybe that's changed?

If you were interested in doing implementations for GLMNet or Lasso, I'm very happy to provide guidance.

[ablaom]

done