Autogluon: Monotonic constraints for autogluon-tabular

[testvinder]

I would find it very useful if autogluon-tabular could add partial support for monotonic constraints. I would be happy to write a PR if that has interest?

Of the underlying models used only Lightgbm and Catboost support monotonic constraints but we could add a warning that the constraints will only be respected by these two models. Monotonic constraints can help in ability to generalize and are often very helpful in establishing user trust.

[Innixma]

I wasn't previously aware of this functionality, but it does seem useful!

The primary work involved to get this in will be tracking a feature as it undergoes transformation in the feature generator. This is because the features (and their order) can change after undergoing automated feature generation.

This special functionality would likely have to exist as a new variable in feature_types_metadata, which the models can reference in their .fit logic. Likely something such as a dictionary of 1, -1, 0 values with keys being the feature names. This would correspond to the original data given by the user, and then keys would be pruned out if the feature was removed from the feature list post-feature generation.

Contributions are welcome!

[testvinder]

Great! I'll start working on a PR.