`This study presents a novel calibration assessment framework for ML models, designed to address the limitations of existing popular metrics, particularly the ECE. Our framework enables a more fine- grained evaluation of calibration by assessing model performance locally, for different confidence regions or classes, providing a com- prehensive understanding of the model’s behavior.`
https://boa.unimib.it/retrieve/3998da14-0e54-49d1-8a14-4af87f9226c7/Famiglini-2023-ECAI-VoR.pdf

|
0W/0Wh


