Reference Issues/PRs

Fixes #31290

What does this implement/fix? Explain your changes.

This PR fixes an issue where _safe_indexing triggers a SettingWithCopyWarning when used with a slice on pandas DataFrames.

The problem:

When _safe_indexing is used with a slice (e.g., slice(0, 2)) on a pandas DataFrame, it returns a view rather than a copy. If the user then modifies this view, pandas raises a SettingWithCopyWarning because it's unclear whether the modification should affect the original DataFrame.

The fix:

I modified the _pandas_indexing function in sklearn/utils/_indexing.py to explicitly create a copy when a slice is used:

# When using a slice, make a copy to avoid SettingWithCopyWarning
if isinstance(key, slice):
    result = result.copy()

This ensures that when _safe_indexing is used with a slice, it always returns a copy of the data, not a view. This prevents the SettingWithCopyWarning from being raised when the result is modified.

This change maintains the existing behavior of _safe_indexing (it still returns the same data), but ensures that the returned object is always a copy when using slices, which is consistent with the behavior when using other types of indexers.

Any other comments?

• I debated on whether to use .copy(deep=True) or .copy() but decided on .copy() because of the performance overhead of deep copying. The trade off is that deep copy would be safer but at the cost of performance.
• I saw a mention of this being a private function so the fix may not be needed but still decided to submit the PR since its a defensive programming move and shouldn't cause any issues
• I saw another guy comment on this issue (30+ days ago) and debated on whether to work on it (didn't want to step on toes) but didn't see any activity after that.