Reference Issues/PRs

Fixes #29722 and #27747.

This makes predict(), predict_proba(), and score() of KNeighborsClassifier and RadiusNeighborsClassifier take X=None. In this case predictions for all training set points are returned, and points are NOT included into their own neighbors. This makes it consistent with the current behaviour of KNeighborsRegressor and RadiusNeighborsRegressor.

New functionality is described in docstrings (and added to the docstrings of KNeighborsRegressor and RadiusNeighborsRegressor).

Tests are added confirming that knn.fit(X, y).predict(None) is equivalent to cross_val_predict(knn, X, y, cv=LeaveOneOut()).

Sanity checking

@ogrisel asked me to add a notebook showing that this is a statistically sane thing to do (#29722 (comment)):

It would also be great to check in a side simulation study (in a notebook) to show that the variance of the LOO CV scheme is not significantly larger than the one obtained fors k-fold CV with k in {5, 10, 50} and for n_neighbors in {1, 10, 100} to confirm the outcome of the literature review in #27747 (comment). This side notebook should not part of the scikit-learn repo but linked in the PR to help assert that the new feature exposed in the PR is a methodologically sane thing to do.

I did this here: #29722 (comment). Is it sufficient to simply leave it there?