✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: a6b5aa8. Link to the linter CI: here}

@ThorbenMaa Can you make sure to respect the linter and provide a proper title to the PR.
It will help at getting reviewer attention.

sorry, now it should be good to go.

cc: @lucyleeow

A couple of thoughts here:

• PrecisionRecallCurveDisplay has the advantage to call _get_response_values and make the transform as expected and pass the pos_label. So we should encourage using this class.
• when it comes to precision_recall_curve:
  • we need to amend the y_score docstring: we need to emphasize the "positive class" part for the probability and mention that when passing the decision_function, we expect that the sign of the decision function is aligned with the positive class as well (so multiplying by -1 in case of flipping the positive class).
  • I don't think that we should make the docstring of the function more complex. However, we should improve in the user guide. I'm almost thinking that we should have a small public utility function that could post-process the output of decision function of predict-proba. But right now we could explain this in the documentation.

Somehow, I would expect the following function:

def post_process_classification_scores(y_scores, response_method, classes, pos_label=None):
    """Post-process predictions from classifiers.

    This function is used to post-process predictions from classifiers.
    It is used to convert the predictions to take into account the positive class and
    compatible with the metrics.

    .. versionadded:: 1.7

    Parameters
    ----------
    y_scores : ndarray of shape (n_samples, n_classes) or (n_samples,)
        The predictions to post-process. It could be the output of
        :term:`predict_proba`, :term:`predict_log_proba`, or
        :term:`decision_function`.

    response_method : {"predict_proba", "predict_log_proba", "decision_function"}
        The response method to use to post-process the predictions.

    classes : ndarray of shape (n_classes,)
        The classes of the classifier.

    pos_label : int, float, bool or str, default=None
        The positive label.

    Returns
    -------
    y_scores : ndarray of shape (n_samples,), (n_samples, n_classes) or \
            (n_samples, n_output)
        Compressed predictions format as requested by the metrics.
    """
    target_type = type_of_target(classes)

    if target_type in ("binary", "multiclass"):
        if pos_label is not None and pos_label not in classes.tolist():
            raise ValueError(
                f"pos_label={pos_label} is not a valid label: It should be "
                f"one of {classes}"
            )
        elif pos_label is None and target_type == "binary":
            pos_label = classes[-1]

    if response_method in ("predict_proba", "predict_log_proba"):
        y_scores = _process_predict_proba(
            y_pred=y_scores,
            target_type=target_type,
            classes=classes,
            pos_label=pos_label,
        )
    elif response_method == "decision_function":
        y_scores = _process_decision_function(
            y_pred=y_scores,
            target_type=target_type,
            classes=classes,
            pos_label=pos_label,
        )

    return y_scores

But I don't know if such utility make sense to live in the scikit-learn codebase. Somehow, it could leave in the metrics module but it really feel that we did something wrong.

Thanks for your comment and your thoughts! I reset everything I did before to re-start out clean.

PrecisionRecallCurveDisplay has the advantage to call _get_response_values and make the transform as expected and pass the pos_label. So we should encourage using this class

Where would you adress this?

we need to amend the y_score docstring: we need to emphasize the "positive class" part for the probability and mention that when passing the decision_function, we expect that the sign of the decision function is aligned with the positive class as well (so multiplying by -1 in case of flipping the positive class).

I updated the y_score field docu accordingly. I hope I understood this correctly? I'm very open for suggestions :-)