Confidence or Competence: A Calibration Study of Machine Learning Classifiers for Selective Prediction in High-Stakes Contexts

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The central claim of the paper is that, besides being capable of making correct predictions, a model must also have the ability to say "I do not know" in critical moments.To accomplish this, the indicator of model honesty should be given more significance than the performance measure modeled in terms of accuracy.Five classifiers have been trained on the COMPAS recidivism dataset, and their performance has been evaluated not only based on accuracy and F1 but also on their calibration quality and performance of selective predictions.Results indicate that the Multilayer Perceptron (MLP), despite its modest raw accuracy, achieves the best calibration.The Random Forest model, which is often the default recommendation, is the least well-calibrated regarding its uncertainty in the uncalibrated state but improves significantly under post-hoc calibration.Every single classifier improves accuracy when allowed to abstain on low-confidence cases.The conclusion reached in the paper is that selective prediction is a practical, model-independent approach that can lead to safer AI systems, along with accuracy evaluation not being enough for making deployment decisions.