Słownik AI
Kompletny słownik sztucznej inteligencji
Class-balanced Calibration
Calibration strategy that weights the contributions of each class according to their inverse frequency, preventing the domination of majority classes in the calibration estimation.
Cost-sensitive Calibration
Approach that integrates misclassification costs directly into the calibration process, optimizing calibrated probabilities according to business impacts specific to the imbalance.
Focal Loss Calibration
Calibration method combining focal loss for training and specific post-hoc techniques, designed to maintain calibration on difficult examples of minority classes.
Label Smoothing Calibration
Technique that regularizes hard labels into smoother distributions during training, naturally improving calibration on imbalanced classes by reducing overconfidence.
Distribution Calibration
Extension of classical calibration aiming to align not only marginal predictions but also the joint prediction-label distribution, more robust to multi-class imbalance.
Threshold Moving Calibration
Technique that dynamically adjusts class decision thresholds after calibration, optimizing imbalance-specific metrics like F1-score or AUC-PR.