How to adapt to breaking change in fit and predict for DecisionTree.jl models

[ablaom]

Version 0.3.1 (to be yanked) and version 0.4 introduced a breaking change to MLJModelInterface.fit and MLJModelInterface.predict for all 5 models. This change only effects developers who directly call those functions. Regular MLJ users who interact through the usual "machine" interface are not affected.

What changed is that the models mentioned now implement the MLJModelInterface data-front end. The most likely reason for breakage is that fit and predict are not being called with the model-specific form of data generated by the reformat method. This post describes a backwards-compatible fix.

Adding the model-specific pre-processor reformat to your fit/predict calls

If you are not already using reformat in your fit and predict calls then, where you previously made a call

MMI.fit(model, verbosity, data...)    # MMI = MLJModelInterface

you instead want

MMI.fit(model, verbosity, MMI.reformat(model, data...)...)

And instead of

MMI.predict(model, fitresult, Xnew)

you want

MMI.predict(model, fitresult, MMI.reformat(model, Xnew)...)

You have backwards compatibility because the fallback for reformat just slurps the data. This also means you can change these calls for all models (not just the DecisionTree ones).

Subsampling

If you subsample reformatted data before passing to fit or predict, you should always use subsampled_data = selectrows(model, I, reformatted_data...) where I is the indices for subsampling.