name for tabular learner factory

[jeromedockes]

#926 introduces a function to easily get a learner that works on tabular data (by chaining a TableVectorizer, missing value imputation if necessary, and an estimator).

It is used like this:

>>> from skrub import table_learner
>>> table_learner('regressor')
Pipeline(steps=[('tablevectorizer',
                 TableVectorizer(high_cardinality_transformer=MinHashEncoder(),
                                 low_cardinality_transformer=ToCategorical())),
                ('histgradientboostingregressor',
                 HistGradientBoostingRegressor(categorical_features='from_dtype'))])
>>> from sklearn.linear_model import Ridge
>>> table_learner(Ridge(solver='lsqr'))
Pipeline(steps=[('tablevectorizer', TableVectorizer()),
                ('simpleimputer', SimpleImputer()),
                ('ridge', Ridge(solver='lsqr'))])

What should be the name of this function (make_tabular_pipeline in the example)?

I'm not using a poll because in polls adding options removes existing votes. Please upvote the comment containing the option you like and don't hesitate to add options!

[jeromedockes]

make_learner

[TheooJ]

Although I like this name, I wonder whether most users will be familiar with the term "learner" (or "règle d'apprentissage", something that takes data as an input and returns a predictor, section 1.3 of https://www.imo.universite-paris-saclay.fr/~sylvain.arlot/enseign/2020-21_M2/1-fondamentaux.pdf would be my definition)

[jeromedockes]

make_skrub_learner

[jeromedockes]

make_tabular_learner

[jeromedockes]

make_skrub_pipeline

[jeromedockes]

make_tabular_pipeline

[Vincent-Maladiere]

I agree that the function should have "pipe" or "pipeline" since this is what is returned in the end. Since we're dealing with defaults for tabular data, I'm +1 for this option.

[GaelVaroquaux]

I'm actually not that convinced. The fact that it is a pipeline is a technical detail. A pipeline still has a fit and a predict. A non advanced user doesn't know what a pipeline is, and that function is specifically geared to a non advanced user.

[GaelVaroquaux]

I agree with Theo that users can probably relate "learner" more to their needs. make_tabular_learner is too long, in my opinion (long names impose a real toll on users). Thinking in different directions: - "tabularize(model)" - "prep_learner(model)" The last one is fun, because "to prep oneself" can involve putting on clothes, which is what we are doing to the learner.

…

On Jun 11, 2024, 14:27, at 14:27, "Théo Jolivet" ***@***.***> wrote: Although I like this name, I wonder whether most users will be familiar with the term "learner" (or "règle d'apprentissage", something that takes data as an input and returns a predictor section 1.3 of https://www.imo.universite-paris-saclay.fr/~sylvain.arlot/enseign/2020-21_M2/1-fondamentaux.pdf would be my definition) -- Reply to this email directly or view it on GitHub: #938 (reply in thread) You are receiving this because you are subscribed to this thread. Message ID: ***@***.***>

[jeromedockes]

tabularize

[jeromedockes]

prep_learner

[GaelVaroquaux]

wrap_learner

[GaelVaroquaux]

pipe_learner

[GaelVaroquaux]

Added this in case we want to make it explicit that it creates a pipeline (which could be a good idea)

[jeromedockes]

to me

• this sounds like something that learns pipes 😅
• once I see it's a verb, I still find it a bit weird because I would say we pipe some data to the learner (as we pipe a stream to a process), not that we pipe the learner itself

[glemaitre]

What about vectorized_learner(learner)?

[glemaitre]

I think I'm also fine with the short vectorize

[jeromedockes]

to me vectorizing code is making it applicable to multiple data, as in numpy.vectorize or simd instructions. then there is vectorizing data, ie feature extraction / tablevectorizer. So I'm not strongly opposed but I think vectorize(learner) could be a bit confusing

[TheooJ]

I wasn't aware of this kind of vectorization
In that case doesn't this make the TableVectorizer confusing too ?

[jeromedockes]

not really -- vectorizing data usually means feature extraction, as done in the tablevectorizer so that's fine. but here learner is not data, so the meaning of vectorize(learner) is less clear IMO

[ogrisel]

I think "vectorize" has ambiguous meanings in ML / array computing / HPC communities. I would much prefer "tabularize".

[jeromedockes]

skrubbify(learner) 😂

[glemaitre]

If we call the class Skrubber, then why not :)

[LeoGrin]

make_default_pipeline, make_easy_pipepline,make_simple_pipeline, or make_adapted_pipeline?

[jeromedockes]

I like make_default_pipeline

[GaelVaroquaux]

Too long

[GaelVaroquaux]

The name was chosen with a Twitter poll to consolidate the above discussion: https://x.com/skrub_data/status/1802687002657124557?t=RKbDh_hjE9-lxQiaC0oFcg&s=19

Name: tabular_learner