A follow-up to my comment on the Array API: I don't think we can support the Array API here, as scipy.optimize.minimize does not appear to support it.

If I missed anything, please let me know—I'd be happy to investigate further.

Thanks for the PR. Here is a first pass of feedback:

I'm still working on addressing the feedback, but I also wanted to share some findings related to it and provide an update.

I few computational things seem off.

There are some CI failures—I'll fix those shortly.

Also considering adding a verbose parameter to CalibratedClassifierCV to optionally display convergence info when optimising the inverse temperature beta.

CI passed!

Close to the finish line.

I updated the user guide and the docstrings in calibration.py. I also modified the test to check that the temperature parameter is close to 1 when the temperature scaler is fitted on the training set of the LogisticRegression classifier.

There are still some comments that need to be addressed, and I'll work on them later.

I've refactored the part for checking reponse_method_name:

if len(classes) == 2 and predictions.shape[-1] == 1:
    response_method_name = _check_response_method(
        clf,
        ["decision_function", "predict_proba"],
    ).__name__
    if response_method_name == "predict_proba":
        predictions = np.hstack([1 - predictions, predictions])

I think this only needs to be applied in two places: : _fit_calibrator and _CalibratedClassifier.predict_proba. But please let me know if there's a better way to handle this.

I've also moved _temperature_scaling inside _TemperatureScaling.fit.

CI has passed, so it's ready for review!

I updated _convert_to_logits to handle the conversion to a 2D array in _TemperatureScaling.fit and _TemperatureScaling.predict when the input is 1D.

_TemperatureScaling now has a new parameter, response_method_name ("decision_function" or "predict_proba"), which indicates whether it should fit or predict based on the output of decision_function or predict_proba, respectively. The default value is "decision_function".

In the CalibratedClassifierCV workflow, this value is computed in _fit_calibrator, then passed to _TemperatureScaling when the calibrator is initialised. If the input is 1D, _convert_to_logits will interpret it as either probabilities or decision values, accordingly.

This is to address an edge case where the output of predict_proba is 1D and was incorrectly converted to logits using [-x, x], which was first caught by the test_calibration_prefit function.

Previously, I attempted something similar in 2769eab, but I thought it was awkward for _TemperatureScaling to store response_method_name, so I didn't finalise it at the time. We'll see how this version goes.