Add sample-weight interface point?

[ablaom]

Some supervised and unsupervised algorithms allow one to weight instances and currently there is no obvious interface for passing this information to theses algorithms

[fkiraly]

Dare I say ... tasks?

Of course one can also just add it to signatures wherever an X appears, and check whether row lengths agree.

[ablaom]

I'm thinking:

• add model trait supports_sample_weights to MLJBase, with default value false
• when true, MLJBase.fit has extended signature: fit(model, verbosity, X, y, w)
• when constructing a machine for such models the w is optional: machine(model, X, y, w) or machine(model, X, y) as before (for uniform weights).

As far as I can tell this breaks nothing. Note the length of mach.args for a machine is not constrained - it can be 1, 2 or 3 or whatever, so no problems there.

If people want to discuss contriving sample-weight support for non-supporting models (oversampling, and so forth), please open a new thread.

A test case already exists: the SVM models at https://github.com/alan-turing-institute/MLJModels.jl/blob/master/src/LIBSVM.jl supports weights (currently weights passed as a hyperparameter).

[fkiraly]

Small comment: I think the default/baseline case for the weighted version should be "ignore the weights". That way, every learner would be able to take weights as additional input, which should make building of learners easier and avoid an interface case distinction.

[fkiraly]

If you want oversampling (and so forth), I'd consider these reduction strategies, hence wrappers (or more generally, first-order compositors).

[ablaom]

@fkiraly I like your suggestion to make the fit signature the same for alll cases (w gets ignored when not used) but this will break all existing model implementations.

[ablaom]

Evaluation (as opposed to training) now supports per-observation weights #206

[ablaom]

Resolved a while ago.