Can a supervised model also implement the transform method.

[OkonSamuel]

Can a supervised model also implement the transform method or is it restricted to unsupervised models only.

[ablaom]

Yes, it can!

Curious about the use case.

In MLJFlux we are working on neural nets that learn feature embeddings along with the main layers. The transform method maps the categorical features into their learned embeddings for use in other models that only handle continuous data.

[OkonSamuel]

Oh I just thought may want to use lda to reduce the dimension of a dataset before applying another classification model

[tlienart]

right so you want to have a transform that uses the projection matrix learned in the LDA? I think you can "just" do that by defining a MLJBase.transform(::BayesianLDA ... )

[ablaom]

You should also declare an output_scitype. Currently this is not reported in the registry unless the model is unsupervised, and I don't think this is used. However, for future proofing, it would be good to have. So, you probably want something like:

using ScientificTypes
MLJBase.output_scitype(::Type{<:BayesianLDA}) = Table(Continuous)

This is assuming you convert your transformed inputs into tabular form (the recommendation for consistency with PCA, and so forth), which you do with MLJBase.table(Xreduced). This automatically adds labels to the columns (x1, x2, etc I think).