Add selectors

[jeromedockes]

This adds the (private) skrub._selectors module.
It is kind of a weird PR because we decided to keep it private for now, and it is not yet used anywhere in skrub 😅

However it is a step towards column-wise transformations (#877), and in the longer term towards the skrub Pipe/Recipe/..., so selectors will have a reason to exist at some point.

Moreover, I think in the future we will probably want to expose them to advanced users and accept them in skrub interfaces that currently accept a list of column names such as SelectCols (in addition to the list of columns, not instead).

The selectors achieve 2 things:

1: Delayed selection: passing a selection rule that describes what columns should be selected, when the dataframe on which the rule is evaluated and expanded into a concrete list of column names is not available yet. For example, instantiating a SelectCols that selects "all numeric columns except the column named 'ID'", even though the actual dataframe (and concrete list of column names) is only known at fit time, not instantiation time

A consequence is 1b: have a good value to represent "all columns" as a default argument for interfaces that expect a column list (None to mean "all" is not great, _selectors.all() is better).

2: Easily expressing complex selection rules because (i) the _selectors module offers a range of useful pre-defined rules such as numeric(), cardinality_below(30) or glob('User_*'), and (ii) because selectors can be combined with the operators ~, |, &, -, ^.

The docstring of skrub/_selectors/__init__.py provides more details but here is a summary:

• Selectors have a .expand(dataframe) method that returns the list of column names that should be selected.
• make_selector turns a column name, list of column names, or selector into a selector.
• select returns the part of a dataframe matched by a selector (ie indexes the dataframe with the result of expand).

>>> from skrub import _selectors as s

>>> import pandas as pd
>>> df = pd.DataFrame(
...     {
...         "height_mm": [297.0, 420.0],
...         "width_mm": [210.0, 297.0],
...         "kind": ["A4", "A3"],
...         "ID": [4, 3],
...     }
... )

>>> s.all().expand(df)
['height_mm', 'width_mm', 'kind', 'ID']
>>> (s.all() - ['ID']).expand(df)
['height_mm', 'width_mm', 'kind']
>>> s.numeric().expand(df)
['height_mm', 'width_mm', 'ID']
>>> (s.glob('*_mm') | s.string()).expand(df)
['height_mm', 'width_mm', 'kind']
>>> s.cols('kind', 'ID').expand(df)
['kind', 'ID']
>>> s.make_selector(['kind', 'ID'])
cols('kind', 'ID')
>>> s.select(df, s.glob('*_mm'))
   height_mm  width_mm
0      297.0     210.0
1      420.0     297.0

[jeromedockes]

I'm not sure why codecov reports some lines as missing. They are covered by the tests, for example the reported line here is covered by this test; when running the tests locally I see 100% coverage for the _selectors module

[TheooJ]

Thanks @jeromedockes !
I mostly have small nitpicks about markdown, but it looks very good to me !

[TheooJ]

Small nitpick here, your function has a parameter column whereas some other funcs use col, like _to_list_pandas for instance

I'd harmonize this for consistency

[jeromedockes]

I 100% agree, however looking at the module there seems to be an almost balanced mix of col and columns (my bad). so let's harmonize it in another PR as this one isn't really related to the dataframe module so doing it here would obfuscate the diff with a ton of unrelated changes

[TheooJ]

Same comment as for _is_integer

[jeromedockes]

thanks a lot for the thorough review @TheooJ !

[TheooJ]

LGTM :) thanks