Glosarium AI
Kamus lengkap Kecerdasan Buatan
Information Leakage (Blending)
Specific risk in blending where the meta-model can overfit the base models' predictions if the hold-out set is not sufficiently representative or is too small.
Blending Weights
Coefficients or parameters learned by the meta-model (often simple linear regression) to weight the predictions of each base model in the final combination.
Two-Level Training
Sequential process in blending where base models are trained first, followed by training the meta-model on their respective predictions.
Stacked Cross-Validation
Alternative to blending where predictions for the meta-model are generated via cross-validation on the training set, reducing overfitting risk but increasing complexity.
Model Diversity
Key principle in blending involving the use of base models with different algorithms (e.g., decision tree, SVM, neural network) to capture varied patterns and improve overall performance.
Out-of-Fold Predictions
Predictions generated by a model on the validation data of each fold in cross-validation, used in stacking but avoided in blending in favor of a hold-out set.
Meta-Model Overfitting
Phenomenon where the meta-model memorizes the base models' predictions on the hold-out set instead of generalizing their combination, often due to a too small hold-out set or an overly complex meta-model.
Linear Blending
Simplified form of blending where the meta-model is a linear regression, simply finding an optimal linear combination of the base models' predictions.
Stratified Split for Blending
Technique for splitting the dataset into training and hold-out sets for blending, preserving the distribution of target classes to avoid bias in the meta-model predictions.
Prediction Fusion
Action of combining the outputs of multiple estimators, which constitutes the core of blending and other ensemble methods to produce a more robust final prediction.
Weighted Blending
Variant of blending where the weights assigned to the base model predictions are defined manually or by a heuristic, rather than learned by a meta-model.
Generalization in Blending
Ability of the final blending model to perform correctly on new unseen data, depending on the robustness of the base models and the meta-model's ability to generalize their combination.