CONCEPT DRIFT IN DATA STREAM CLASSIFICATION USING ENSEMBLE METHODS: TYPES, METHODS AND CHALLENGES

Abstract

Ensemble Methods grows along with Machine Learning and Computational Intelligence domain proves to be effective and versatile. It helps in reducing variance and improves accuracy. It effectively addresses many Machine Learning challenges such as data stream classification, class Imbalance in datasets and concept drift occurrence in non-stationary environments. Data stream refers to rapidly generated heterogeneous data in a continuous way. One of the key challenges considered in learning from data streams is the detection of concept drift, i.e., changes in data distribution underlying data streams, observed over time. Such changes in incoming data deteriorate the accuracy of the classifier since classifier has been learned over past data instances that are stable. Thus detection of concept drift is important task. The real life examples of drift are spam detection, credit card fraud detection, and weather predictions. This paper provides an overview of ensemble methods, data stream classification, concept drift, its types and drift detection methods and concludes with the related works in detecting concept drift that occurs during data stream classification.

Keywords: Ensemble, Classifier, Data Streams, Class Imbalance, Concept drift