add LIBSVM

[ValdarT]

Added LIBSVM. A few notes:

• If I understand correctly then it needs to be added to the MLJRegistry first, before the loading and hence tests start working properly?
• I used the names as they are in LIBSVM but perhaps it would be better if they are the same as in the ScikitLearn port?
• I also included the one class SVM for anomaly/novelty detection but I'm not quite sure how that fits into MLJ currently.

[ablaom]

Thanks for that!!

*If I understand correctly then it needs to be added to the MLJRegistry first, before the loading and hence tests start working properly?

No. You don't need to register to test. You have simply forgotten to add import LIBSVM to your test file (before using MLJModels.LIBSVM_) to trigger lazy loading of your glue code. Travis still fails, but points to some error internal to LIBSVM.jl. Can you investigate?

I used the names as they are in LIBSVM but perhaps it would be better if they are the same as in the ScikitLearn port?

I suggest you use names that are as consistent as possible with the LIBSVM docs. MLJ has no problem with duplicate model names. You just can't simultaneously load two models with the same name from different packages - you can load one or the other.

I also included the one class SVM for anomaly/novelty detection but I'm not quite sure how that fits into MLJ currently.

I expect we should regard this as unsupervised learning (unless there appears some dissent at JuliaAI/MLJ.jl#51 (comment)), so use

mutable struct OneClassSVM <: MLJBase.Unsupervised

instead of

mutable struct OneClassSVM <: MLJBase.Deterministic

And you should probably rename your predict(::OneClassSVM, ...) to transform.
Also, make a final conversion the bitarray predictions (aka transformations) to a CategoricalVector (just apply categorical to the answer) to ensure they have the right scitype, which you will articulate with a declaration

output_scitype_union(::Type{OneClassSVM})) = MLJBase.Finite{2}

(unsupervised learners have "outputs" rather than "targets").

Incidentally, the target_scitype_union declarations for your classifiers should be MLJBase.Finite. Only the regressors should have Continuous as the scitype.

[ablaom]

Brilliant!! Thanks for this contribution.

It has been argued JuliaAI/MLJ.jl#51 (comment) that one-class classification is neither supervised or unsupervised and I am happy to concur with this opinion. So we will probably create a new category. For now, I think we can safely leave this as unsupervised.

I will proceed with registering LIBSVM.jl with MLJRegistry, to make the new models findable by all MLJ users. This will also be our first opportunity to test loading model code for models with names duplicated in separate packages. Will close #19 when this testing is complete.

Again many thanks.

[ablaom]

Now registered. Appears to work:

julia> using MLJ

julia> models()["LIBSVM"]
6-element Array{Any,1}:
 "EpsilonSVR" 
 "LinearSVC"  
 "NuSVR"      
 "NuSVC"      
 "SVC"        
 "OneClassSVM"

julia> @load SVC
import MLJModels ✔
import LIBSVM ✔
import MLJModels.LIBSVM_.SVC ✔

julia> mach = machine(SVC(), shuffle!(load_iris()))
Machine{SVC} @ 8…90


julia> evaluate!(mach)
┌ Info: Evaluating using cross-validation. 
│ nfolds=6. 
│ shuffle=false 
│ measure=MLJ.misclassification_rate 
│ operation=StatsBase.predict 
└ Resampling from all rows. 
Cross-validating: 100%[=========================] Time: 0:00:03
6-element Array{Float64,1}:
 0.0 
 0.0 
 0.04
 0.12
 0.0 
 0.08

[ablaom]

Will now make a new release of MLJModels