KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Information Leakage
Risk in stacking where the meta-model is trained on the same data as the base models, leading to overfitting which blending seeks to avoid via hold-out validation.
Prediction Weighting
Technique in blending where the meta-model learns optimal weights to combine the predictions of the base models, often via simple linear regression.
Hold-out Stratification
Stratified split of the hold-out validation set in blending to ensure the class distribution is preserved, essential for imbalanced classification problems.
Multi-level Blending
Extension of blending where the predictions of the first meta-model become inputs for a second meta-model, creating a hierarchy of prediction combinations.
Cross-Blending
Variant of blending using multiple hold-out splits and averaging the predictions of the corresponding meta-models to reduce variance related to the specific hold-out choice.
Prediction Calibration
Step in blending where the output probabilities of the base models are recalibrated before being fed to the meta-model to ensure consistency in the prediction scale.
Stochastic Blending
Approach where the hold-out validation set is randomly selected over multiple iterations, training several meta-models whose predictions are then averaged.
Temporal Blending
Application of blending to time series data where the hold-out respects the chronological order, using recent periods to train the meta-model on past predictions.
Optimisation du Ratio Hold-out
Processus de détermination de la proportion optimale de données à réserver pour la validation hold-out en blending, équilibrant la qualité d'entraînement des modèles de base et du méta-modèle.
Blending Adaptatif
Méthode où le méta-modèle ajuste dynamiquement sa combinaison de prédictions en fonction de la performance observée de chaque modèle de base sur différents segments de données.