change stats.mode to collections.Counter, for better performance.

[DavidKatz-il]

Reference Issues/PRs

Fixes #18978

What does this implement/fix? Explain your changes.

As suggested in the issue we replaced scipy.stats.mode with collections.Counter since it has better performance.

[thomasjpfan]

Thank you for the PR @DavidKatz-il !

May you provide a benchmark as a comment to compare this implementation to main repo?

[DavidKatz-il]

# string/object
import string
import pandas as pd
import numpy as np
from sklearn.impute import SimpleImputer

imputer = SimpleImputer(strategy='most_frequent')
np.random.seed(42)
length = 10000
arr = np.random.choice(list(string.ascii_lowercase),  size=(length, length))
flat = arr.flatten()
np.put(flat, np.random.choice(flat.shape[0], size=int(flat.shape[0] * 0.1), replace=False), np.nan)
arr = np.reshape(flat, (length, length))
print('simple imputer string/object:')
%time imputer.fit(pd.DataFrame(arr))

# numeric
import pandas as pd
import numpy as np
from sklearn.impute import SimpleImputer

imputer = SimpleImputer(strategy='most_frequent')
np.random.seed(42)
length = 10000
arr = np.random.uniform(size=(length, length))
flat = arr.flatten()
np.put(flat, np.random.choice(flat.shape[0], size=int(flat.shape[0] * 0.1), replace=False), np.nan)
arr = np.reshape(flat, (length, length))
print('simple imputer numeric:')
%time imputer.fit(pd.DataFrame(arr))

ArraySize	stats numeric	counter numeric	stats string	counter string
100 X 100	16 ms	7 ms	20 ms	7 ms
1000 X 1000	250 ms	373 ms	689 ms	132 ms
10000 X 10000	11.5 s	38.4 s	1min 23s	16.4 s

[glemaitre]

Indeed, as unique in the encoder, we should use Counter only on object dtype and otherwise the mode strategy:

scikit-learn/sklearn/utils/_encode.py

Lines 30 to 33 in c9677d6

	if values.dtype == object:
	return _unique_python(values, return_inverse=return_inverse)
	# numerical
	out = np.unique(values, return_inverse=return_inverse)

[amueller]

LGTM

[glemaitre]

This looks good.

Please add an entry to the change log at doc/whats_new/v*.rst. Like the other entries there, please reference this pull request with :pr: and credit yourself (and other contributors if applicable) with :user:.

[glemaitre]

Last nitpicks

[glemaitre]

Thanks @DavidKatz-il

So this a fix that should be back-ported in 0.24.X for the release. I will label it.