Allow kernel to be callable, fix some sloppy implementation of the MLJ interface; add docstrings

[ablaom]

In this PR we:

• update the compatibility requirement of LIBSVM.jl to the latest version

• (enhancement) Allow the kernel hyper-parameter to be a function or other callable

• (breaking) Overload fitted_params of all models so as to be more user-friendly

• (breaking) Remove a number of superfluous hyper-parameters (hyper-parameters whose values are ignored internally):

  • p from LinearSVC (alias for ϵ in the cited paper) which is only defined for regression models
  • cost from NuSVC (alias for C in cited paper) which is redundant as nu is a re-parameterization of it
  • cost from OneClassSVM (again, nu is used instead)

• (breaking) Re-implement OneClassSVM as a UnsupervisedDetector satisfying MLJ's API for such models. In particular, transform(mach, Xnew) now returns raw outlier scores. To extract normalised probabilities or for threshold outlier classification, refer to the new document string.

• (breaking) Re-implement SVC and LinearSVC so that optional class weight dictionaries are specified along with the training data, and fix bugs in how these were interpreted internally. Remove weight support for NuSVC as LIBSVM.jl does not appear to support it.

• (enhancement) Add MLJ-compliant document strings to all models, with detailed examples

My apologies for combining so much in a single PR. The original intention was just to update document strings. The issues were revealed only progressively as I worked through these.

[codecov-commenter]

Codecov Report

Merging #13 (7b17864) into master (8a3693d) will increase coverage by 2.67%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #13      +/-   ##
==========================================
+ Coverage   88.23%   90.90%   +2.67%     
==========================================
  Files           1        1              
  Lines         153      176      +23     
==========================================
+ Hits          135      160      +25     
+ Misses         18       16       -2     

Impacted Files	Coverage Δ
src/MLJLIBSVMInterface.jl	90.90% <100.00%> (+2.67%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8a3693d...7b17864. Read the comment docs.

[ablaom]

Erstwhile maintainers of LIBSVM.jl: @iblis17 @barucden @till-m @mpastell.

This PR adds detailed document strings to the models constituting an MLJ interface to LIBSVM.jl . I wonder if any of you have time in the next week or so to look over these strings, which appear at the end of the file MLJLIBSVMInterface.jl . Familiarity with MLJ would be helpful, but hardly essential. I'm mainly interested in low-level editorial comments and that the descriptions of model hyper-parameters (taken mostly from scikit-learn) are correct. In return, I would be happy to update the LIBSVM.jl readme with guidance on using the MLJ interface. The MLJ doc-string standard is here.

[ablaom]

@OkonSamuel Perhaps, some time over the next week you can look over the code (don't worry about the doc-strings).

[barucden]

Besides that one missing parenthesis, I couldn't spot any apparent problem.

Consider adding a note regarding the serialization of the trained model when a user-defined kernel was used (see, e.g., my comment that shows the limitation). I believe this has not been targeted in LIBSVM.jl yet, so if it's not already covered in MLJ on the general level somehow, I think there should be a note for it.

[till-m]

I didn't really find anything noteworthy either. Comments are just small things that caught my eye.

I also noticed that attempting to use Kernel.Precomputed now throws an error. Is there a particular reason for not supporting pre-computed kernels?

[ablaom]

@barucden says:

Consider adding a note regarding the serialization of the trained model when a user-defined kernel was used

Done, thanks!

[ablaom]

@till-m says:

I also noticed that attempting to use Kernel.Precomputed now throws an error. Is there a particular reason for not supporting pre-computed kernels?

No, it is unlikely MLJ will support this soon as it requires adjustments to the API that would be special to SVM (at present).

[ablaom]

@till-m @barucden Thanks for your prompt review of the doc-strings!