# Classifier Calibration: How to assess and improve predicted class probabilities: a survey

**Authors:** Telmo Silva Filho, Hao Song, Miquel Perello-Nieto, Raul Santos-Rodriguez, Meelis Kull, Peter Flach  
**arXiv:** [2112.10327](https://arxiv.org/abs/2112.10327)  
**Date:** 2021-12-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.10327v1/

## Key points

- Classifier calibration is essential for accurately assessing uncertainty and confidence levels associated with predicted class probabilities.
- The paper provides a comprehensive overview of the history and recent developments in classifier calibration, including evaluation metrics, visualization techniques, post-hoc calibration methods for binary and multiclass classification, and advanced concepts.
- Challenges exist in navigating the complex landscape of classifier calibration, but the authors offer insights into key methodologies to address them.
- Visual representations like histograms and reliability diagrams are used to illustrate how calibration techniques can impact prediction accuracy and reliability in various scenarios.
- The survey serves as a valuable resource for researchers and practitioners seeking to enhance their classifiers' performance through effective calibration strategies.

## Summary

This paper delves into the topic of classifier calibration and its importance in accurately assessing uncertainty and confidence levels associated with predicted class probabilities. The authors provide a comprehensive overview of the history and recent developments in this field, covering various topics such as evaluation metrics, visualization techniques, post-hoc calibration methods for both binary and multiclass classification, and advanced concepts. They also discuss the challenges involved in navigating this complex landscape and offer insights into key methodologies. Visual representations such as histograms and reliability diagrams are used to demonstrate how calibration techniques can impact prediction accuracy and reliability in different scenarios. Overall, this survey serves as a valuable resource for researchers and practitioners looking to improve their classifiers' performance through effective calibration strategies.