BUG sample-weight ignored by min_sample_leaf in HGBDT #34745

Description

In the following course, we are using the weight of HGBDT to correct probability estimate. We know that it should be equivalent with a post-hoc correction. However, they are not equivalent for HGBDT while it works for logistic regression. In the past, we thought that it was due to the binning but it has been fixed (cf. #29641).

Potential cause (to be check further)

Splitter in sklearn/ensemble/_hist_gradient_boosting/splitting.pyx applies min_samples_leaf to histogram counts (n_samples_left / n_samples_right), not to the sum of sample weights (or hessians). Loss gradients/hessians (and bin edges after #29641) are weighted correctly.

Minimal reproducer

import numpy as np
from sklearn.base import clone
from sklearn.datasets import make_classification
from sklearn.ensemble import HistGradientBoostingClassifier

X, y = make_classification(
    n_samples=1000,
    n_features=10,
    n_informative=5,
    n_redundant=0,
    weights=[0.75, 0.25],
    random_state=0,
)
rng = np.random.RandomState(0)
sw = rng.randint(1, 4, size=len(y))
X_rep, y_rep = np.repeat(X, sw, axis=0), np.repeat(y, sw)


def max_raw_diff(min_samples_leaf):
    est = HistGradientBoostingClassifier(
        min_samples_leaf=min_samples_leaf,
        early_stopping=False,
        max_iter=50,
        random_state=0,
    )
    weighted = clone(est).fit(X, y, sample_weight=sw)
    repeated = clone(est).fit(X_rep, y_rep)
    return np.max(np.abs(weighted._raw_predict(X) - repeated._raw_predict(X)))


print(max_raw_diff(1))   # ~1e-8
print(max_raw_diff(20))  # ~2.4, unexpected

Observed on main (1.10.dev0):

• min_samples_leaf=1: max |raw_pred| diff ≈ 4.6e-08
• min_samples_leaf=20: max |raw_pred| diff ≈ 2.42

Expected behavior

Expect equivalence.

Notes

• Documented limitation in test_sample_weight_effect (forces min_samples_leaf=1).
• Related: HistGradientBoosting counts and sample weights #26128 (histogram counts vs weights). Not a regression of Added sample weight handling to BinMapper under HGBT #29641; remaining gap after that fix.