[QST] Evaluate TargetEncoder implementation vs. scikit-learn

[beckernick]

Support for TargetEncoder was just merged into scikit-learn inspired by the cuML implementation.

We should evaluate the API alignment and functionality across the libraries and how the results compare on various datasets.

[jcrist]

API differences:

• sklearn includes a categories parameter where the user can explicitly pass in categories to use. This defaults to "auto", where they're discovered from the input data. cuml only supports the auto behavior.
• sklearn includes a target_type ({"auto", "continuous", "binary", "multiclass"}) for specifying the type of target. IIUC cuml's implementation matches the continuous version only.
• Both implementations include a smooth parameter, but they differ in defaults (and maybe in implementation, hard to tell just from reading the code).
• cuml has n_folds=4, sklearn has cv=5; both control the fold behavior when fitting.
• cuml has split_method ({"random", "continuous", "interleaved", "customize"}) for configuring how the data is partitioned into folds. sklearn has a shuffle parameter (defaults to True) to control whether the data is shuffled first before splitting into folds. shuffle=False is equal to split_method="continuous", and shuffle=True is close to split_method="random". cuml also supports passing in the folds as part of the fit call, sklearn doesn't.
• cuml has a seed parameter (defaults to 42), sklearn has a random_state parameter (defaults to None).
• cuml has a stat parameter for configuring the statistic to use (defaults to mean). sklearn only supports mean.

From reading through the code, if we want our TargetEncoder to match sklearn's behavior, this looks like a fair amount of API changes here. There might be a way to handle some of these changes through deprecations, but the implementations differ enough that a clean break may be easier to manage. Not sure about our deprecation-vs-breakage policy as a project.