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.

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.

Edit: Reverted.

The response_method_name commit has been reverted, and CI has passed.