When operating a classifier in a production setting (i.e. predictive phase), practitioners are interested in potentially two different outputs: a "hard" decision used to leverage a business decision or/and a "soft" decision to get a confidence score linked to each potential decision (e.g. usually related to class probabilities).

Scikit-learn does not provide any flexibility to go from "soft" to "hard" predictions: it uses a cut-off point at a confidence score of 0.5 (or 0 when using decision_function) to get class labels. However, optimizing a classifier to get a confidence score close to the true probabilities (i.e. a calibrated classifier) does not guarantee to obtain accurate "hard" predictions using this heuristic. Reversely, training a classifier for an optimum "hard" prediction accuracy (with the cut-off constraint at 0.5) does not guarantee obtaining a calibrated classifier.

In this talk, we will present a new scikit-learn meta-estimator allowing us to get the best of the two worlds: a calibrated classifier providing optimum "hard" predictions. This meta-estimator will land in a future version of scikit-learn: https://github.com/scikit-learn/scikit-learn/pull/26120.

We will provide some insights regarding the way to obtain accurate probabilities and predictions and also illustrate how to use in practice this model on different use cases: cost-sensitive problems and imbalanced classification problems.