AI-ordlista
Den kompletta ordlistan över AI
Dynamic Ensemble Selection (DES)
Aggregation method where the subset of models to use for prediction is dynamically selected for each new instance, based on its specific characteristics and the models' competence in the local region of the input space.
Static Ensemble Selection (SES)
Approach where a fixed subset of classifiers is chosen once on the validation set, then used identically for all future instances, unlike dynamic selection which adapts instance by instance.
Local Region of Competence
Neighborhood around a test instance, defined by a distance metric, in which the competence of base models is evaluated to decide which models to use for predicting this instance.
K-Nearest Oracle (KNORA)
Family of DES algorithms that select models based on their performance on the k nearest neighbors of the test instance, with variants like KNORA-E (elimination) and KNORA-U (union).
Multiple Classifier Behaviour (MCB)
Technique that analyzes the behavior of classifiers on labeled instances to identify regions where certain models are more reliable, thus guiding dynamic selection.
Dynamic Selection vs. Dynamic Weighting
Distinction where dynamic selection chooses a subset of models, while dynamic weighting assigns weights to all models for the final prediction, both adapting instance by instance.
Homogeneous vs. Heterogeneous DES
Differentiation where homogeneous DES use base models of the same type (e.g., all decision trees), while heterogeneous ones combine different types of algorithms for increased diversity.
Region of Competence Estimation
Process of determining the relevant neighborhood for a test instance, crucial for evaluating the local competence of models, often based on metrics like Euclidean distance or cosine similarity.
DES with Pre-Processing
Approach where pre-processing techniques, such as dimensionality reduction or oversampling, are applied to improve the definition of competence regions before the dynamic selection of models.
Online DES
Variant of dynamic ensemble selection designed for data streams, where the competence of the models and the selection are continuously updated as new instances arrive.
DespeRt Algorithm
DES algorithm that evaluates the competence of classifiers based on the distribution of confidence levels of predictions on the k nearest neighbors, favoring models that are both competent and diverse.
A Priori vs. A Posteriori DES
Distinction where a priori methods select the models before seeing their predictions for the test instance, whereas a posteriori methods select after observing these predictions.
Dynamic Ensemble Selection for Imbalanced Data
Adaptation of DES methods to handle imbalanced datasets, where the model selection can be biased to improve the detection of the minority class.
Competence-Based Dynamic Selection
Central paradigm in DES where the decision to select a model relies solely on an estimate of its local competence for the considered instance, rather than on its global performance.