GeezerVison

Reference Issues/PRs

Fixes #35347

What does this implement/fix? Explain your changes.

partial_fit is documented to return self for both of these estimators, and neither did. In both cases the override dropped the return that the implementation it mirrors already has, so the documented chaining idiom est.partial_fit(X, y).predict(X) fails with AttributeError: 'NoneType' object has no attribute 'predict'.

• MultiOutputRegressor.partial_fit (sklearn/multioutput.py:430) calls super().partial_fit(...) and discards the result. Its sibling MultiOutputClassifier inherits _MultiOutputEstimator.partial_fit (sklearn/multioutput.py:203), which does return self — the two siblings in the same hierarchy disagree, and only the classifier gets it right.
• BernoulliRBM.partial_fit (sklearn/neural_network/_rbm.py:310) ends with self._fit(...) and falls off the end of the function, while BernoulliRBM.fit (sklearn/neural_network/_rbm.py:437) ends with return self.

Both docstrings have a Returns section promising a fitted instance, so this makes the code match the documented contract. Production diff is 2 lines; the rest is one non-regression test per estimator.

Deliberately out of scope:

• partial_fit of other estimators — I scanned all 24 partial_fit definitions in the code base against their documented Returns section and these are the only two that disagree.
• The AttributeError vs NotFittedError question for chaining on an unfitted estimator — unrelated.

Testing

Red on clean main @ 6e37063 (source build, macOS, Python 3.13.12), the two new tests fail with assert None is <MultiOutputRegressor instance> and assert None is <BernoulliRBM instance>; the same script prints:

MOR.partial_fit -> None
MOC.partial_fit -> True
RBM.partial_fit -> None
RBM.fit -> True

Green on the branch. Both new tests keep a control assertion that passes before and after the change (MultiOutputClassifier.partial_fit(...) is clf and BernoulliRBM.fit(...) is rbm), so the red is attributable to the missing return rather than to the test scaffolding.

Changelog fragment added under many-modules, since the fix touches both sklearn.multioutput and sklearn.neural_network; happy to split it per module or move it if you prefer.

Introduce yourself

I use scikit-learn mainly for tabular modeling. I was reading the partial_fit docstrings while working with incremental training and noticed that the Returns section promised an instance that two implementations never returned; the asymmetry with MultiOutputClassifier / BernoulliRBM.fit looked like an oversight rather than a deliberate choice.

AI usage disclosure

I used AI assistance for:

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

Everything reported above (red/green, control assertions, test counts, lint) I ran and checked locally myself.

Any other comments?

The issue was triaged after the earlier attempt (#35349) had already been auto-closed by the bot for being opened while the issue still carried Needs Triage; this PR is opened after triage, as the contributing guide asks.