AI Glossary
The complete dictionary of Artificial Intelligence
Adaptive Random Forest
Extension of the classical Random Forest for data streams, using Hoeffding trees as base learners and integrating mechanisms for concept drift detection and adaptation with dynamic classifier reweighting.
Online Bagging
Incremental variant of Bootstrap Aggregating where each instance in the stream is used to update base classifiers according to a Poisson(1) distribution, enabling continuous learning without requiring the entire dataset.
STAGGER Algorithm
One of the first concept drift demonstration algorithms using weighted hypotheses that dynamically adapt to concept changes by adjusting weights based on correct or incorrect predictions.
SAM-KNN (Self-Adjusting Memory KNN)
Variant of k-Nearest Neighbors for data streams that dynamically adjusts its memory size based on concept stability, automatically eliminating obsolete instances while preserving representative ones.
Leveraging Bagging
Improvement of Online Bagging introducing random reweighting strategies to increase base classifier diversity and accelerate adaptation to concept changes with theoretical performance guarantees.
OzaBagging
Specific implementation of the online bagging algorithm proposed by Oza and Russell, where each stream instance is weighted according to a Poisson distribution to simulate the bootstrap process in streaming environments.
HAT (Hoeffding Adaptive Tree)
Extension of the Hoeffding Tree integrating concept change detection mechanisms using ADWIN to evaluate node performance and dynamically replace inefficient subtrees with new adapted structures.
StreamDM
Open-source framework for data mining on data streams implemented in Scala for Spark, providing a collection of incremental algorithms optimized for distributed processing of large-scale streams.
MOA (Massive Online Analysis)
Environnement logiciel de référence pour l'évaluation et le développement d'algorithmes d'apprentissage incrémental, incluant des outils de génération de streams artificiels et des métriques d'évaluation adaptées au concept drift.
Accuracy Weighted Ensemble
Algorithme d'ensemble pour streams qui maintient une collection de classifieurs entraînés sur différentes périodes, avec pondération dynamique basée sur leur performance récente pour optimiser les prédictions face au drift conceptuel.
Concept Drift Velocity
Métrique quantifiant la vitesse à laquelle le concept sous-jacent évolue dans un flux de données, utilisée pour adapter dynamiquement la fréquence de mise à jour du modèle et l'allocation des ressources computationnelles.