[ENH] Enable Grid-Search for TableVectorizer

[Vincent-Maladiere]

What does this PR fix/address?

Apply Gaël's suggestions and the outputs of discussion #796 to make grid-search possible.

from skrub import TableVectorizer
from skrub.datasets import fetch_employee_salaries

from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import make_pipeline
from sklearn.ensemble import HistGradientBoostingRegressor

dataset = fetch_employee_salaries()
X, y = dataset.X.head(500), dataset.y.head(500)

pipe = make_pipeline(
    TableVectorizer(),
    HistGradientBoostingRegressor()
)

param_grid = {"tablevectorizer__high_cardinality_transformer__n_components": [10, 20]}

cv = GridSearchCV(pipe, param_grid)
cv.fit(X, y)

What does it change?

• Replace the default None of the transformers with their global default
• When the user actively sets transformers to None, they are turned into "passthrough" during fit (e.g. high_cardinality_encoder = None will result in high_cardinality_encoder_ = passthrough
• Some revamp to make cloning during init and during fit more readable

[Vincent-Maladiere]

I need to address a small docstring error

[jeromedockes]

LGTM, thanks!

[jeromedockes]

we expose the ColumnTransformer's sparse_threshold parameter but with our default transformers the default will always be dense (even if toe onehot encoder yields many zeros)

we could consider

• pointing out in the doc that users need to change the transformers if they want sparse output
• not exposing the sparse_threshold and always returning dense data
• making the onehot encoder sparse by default

(not in this PR)

[jeromedockes]

not super important but you could propagate the change 'card_cat' -> 'cardinality' to local variables

along that line it would be nice if we picked one of "numeric" or "numerical" and used it all the time :)

[Vincent-Maladiere]

numeric, since it's shorter?

[jeromedockes]

sounds good! it's also the choice made by polars.selectors.numeric. pandas select_dtypes uses "number", "category"

[jeromedockes]

I guess this one is ready to merge?

[Vincent-Maladiere]

I think so!

[GaelVaroquaux]

Very nice. Congratulations!