Learning when drop_zeros=False

[ColdTeapot273K]

Versions

river version: recent main (d606d7b)
Python version: Python 3.11.8
Operating system: Fedora Linux

Describe the bug

These 4 interesting lines effectively stop OneHotEncoder from learning when drop_zeros=False:

• https://github.com/online-ml/river/blob/main/river/preprocessing/one_hot.py#L203-L204
• https://github.com/online-ml/river/blob/main/river/preprocessing/one_hot.py#L234-L235

Steps/code to reproduce

Setup:

# Sample code to reproduce the problem
>>> from pprint import pprint
>>> import random
>>> import string

>>> random.seed(42)
>>> alphabet = list(string.ascii_lowercase)
>>> X = [
...     {
...         'c1': random.choice(alphabet),
...         'c2': random.choice(alphabet),
...     }
...     for _ in range(4)
... ]

>>> from river import preprocessing
>>> oh = preprocessing.OneHotEncoder(drop_zeros=True)
>>> for x in X:
...:     oh.learn_one(x)
...:     pprint(oh.transform_one(x))
...: 
{'c1_u': 1, 'c2_d': 1}
{'c1_a': 1, 'c2_x': 1}
{'c1_i': 1, 'c2_h': 1}
{'c1_h': 1, 'c2_e': 1}

Actual result:

>>> oh.values
defaultdict(set, {})

Expected result:

>>> oh.values
defaultdict(set, {'c1': {'a', 'h', 'i', 'u'}, 'c2': {'d', 'e', 'h', 'x'}})

[MaxHalford]

Hey there @ColdTeapot273K!

Would you say this is a bug? The idea here is that there is no need to keep track of values when drop_zeros=True. Indeed, if drop_zeros=False, then we need to keep track of values in order to add zeros. But if not, there is no need to keep track, which is why we return directly in learn_one and learn_many.

What am I missing? Maybe it would be preferable if oh.values was a private property?