fix(ensemble): take exactly the requested fraction when subsampling

[aniruddhaadak80]

Issue

Fixes #1347

Motivation and Context

In preprocessing/ensemble.py the two branches that resolve subsample_samples disagree with each other and with the parameter's own docstring:

python if isinstance(subsample_samples, int): ... size = min(subsample_samples, n_samples) # exactly what was asked else: ... size = int(subsample_samples * n_samples) + 1 # one row more

The docstring says "If float, subsample that fraction of samples", so the +1 is simply wrong. Two observable defects follow:

• Wrong size. SUBSAMPLE_SAMPLES=0.5 on 1000 rows subsamples 501 rows. The two spellings of the same request disagree: the float .5 and the int 500 build different model contexts.
• Crash on a valid request. SUBSAMPLE_SAMPLES=0.01 on 100 rows with 3 classes resolves to size=2, which trips the guard subsample_size (2) must be >= number of classes (3). The +1 is what pushes a legal value into the illegal range - dropping it gives 1, and the only reason that would still be short is a dataset with more classes than rows sampled, which is the guard's real purpose.

No opt-in is needed to hit this: "stratified" is the resolved default for classifiers.

The fix takes the fraction exactly, truncating rather than rounding up, and keeps the floor at one row via max(1, ...) so a tiny fraction of a tiny dataset cannot resolve to an empty subsample.

Two existing tests encoded the old +1 arithmetic in their assertions and comments, and have been updated to the documented contract - one changes rac=0.2 to rac=0.25 to keep the same size=5 that its balanced-coverage assertion depends on.

Public API Changes

• No Public API changes
• Yes, Public API changes (Details below)

No signature changes. A float SUBSAMPLE_SAMPLES now yields exactly int(frac * n_samples) rows, matching its documentation and the int spelling.

How Has This Been Tested?

New tests in ests/test_preprocessing/test_ensemble.py:

• est__get_subsample_indices_for_estimators__float_fraction_exact_size - asserts .5 of 1000 rows subsamples exactly 500, that the float and int spellings draw the same rows, and that .333 truncates to 333.
• est__get_subsample_indices_for_estimators__float_fraction_below_class_count - the class-count case above.

The first fails on the unfixed source and passes with the fix:

`

unfixed source

FAILED ...::test__get_subsample_indices_for_estimators__float_fraction_exact_size
1 failed, 1 passed

with the fix

12 passed, 89 deselected
`

The full file gives 5 failed / 96 passed with the fix and 5 failed / 94 passed on a pristine ensemble.py - the same five failures either way, all differentiable_input and majority_downsample tests that fail without model weights, and the +2 difference is the new tests.

Worth noting: the second new test passes both before and after. It is kept as a guard on the crash rather than as a regression test for the fix.
uff check and
uff format --check are clean.

Checklist

• The changes have been tested locally.
• Documentation has been updated (if the public API or usage changes).
• A changelog entry has been added (see changelog/README.md), or "no changelog needed" label requested.
• The code follows the project's style guidelines.
• I have considered the impact of these changes on the public API.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit be7753d. Configure here.}

[cursor]

Changelog retells internal incident

Low Severity

The towncrier fragment narrates the old one-row overshoot and names the internal class-count guard. It also claims small valid fractions no longer fail that guard, but dropping the +1 makes more of those requests fail it.

 

^{Triggered by learned rule: Towncrier fragments: user-facing change only; keep concise}

^{Reviewed by Cursor Bugbot for commit be7753d. Configure here.}

[cursor]

Test docstring narrates prior behavior

Low Severity

The new test docstring recounts the previous int(frac * n_samples) + 1 arithmetic and why it was wrong, rather than stating only the current size contract.

 

^{Triggered by learned rule: Docstrings describe current behavior only — no call sites or code history}

^{Reviewed by Cursor Bugbot for commit be7753d. Configure here.}