GeezerVison

Reference Issues/PRs

Fixes #34673.

What does this implement/fix? Explain your changes.

Since 1.8, a single zero in sample_weight makes IsolationForest (on its
defaults) and both bagging estimators raise
ValueError: Fewer non-zero entries in p than size when bootstrap=False.

The cause is in _generate_bagging_indices: without replacement,
random_state.choice(..., p=normalized_sample_weight) cannot draw more rows
than the number of non-zero weights, while max_samples defaults to the full
number of rows. With replacement there is no such bound, which is why the bug
only exists on the bootstrap=False path.

My fix only touches that path: when bootstrap_samples is False, clamp
max_samples to np.count_nonzero(sample_weight) before drawing. A
zero-weighted row has probability zero and is never drawn, so it is excluded
exactly as if it had been removed from the data, which is what a zero weight
means. If the user asked for fewer samples than the number of non-zero
weights, min(...) leaves their value unchanged.

The regression test checks more than the absence of the exception: it also
asserts that the zero-weighted row never appears in estimators_samples_,
so the statistical meaning of the fix is verified, not just the crash.

Introduce yourself

I am a high-school student from Tunisia studying computer science. I recently
completed the Machine Learning Specialization (Stanford Online /
DeepLearning.AI), where I used scikit-learn throughout, and I am contributing
this fix as a way to learn the library from the inside.

AI usage disclosure

I used AI assistance for:

• Code generation (e.g., when writing an implementation or fixing a bug)
• Test/benchmark generation
• Documentation (including examples)
• Research and understanding

This pull request includes code written with the assistance of AI. I have
personally reviewed, understood and tested the code.

Any other comments?

Before submitting, I checked the two call sites of
_generate_bagging_indices: max_samples is passed by value, so the clamp
only affects the local draw, and both callers only consume the returned
indices. estimators_samples_ regenerates the indices from the stored
_max_samples and _sample_weight, and since the clamp depends only on the
number of non-zero weights, the regenerated indices are identical to the
ones used during fitting.