A Classifier Ensemble for Concept Drift Using a Constrained Penalized Regression Combiner

Abstract

Concept drift represents that the underlying data generating distribution changes over time and it is a common phenomenon in a stream of data sets. In particular, concept drift entails the change of the input-output dependency so that it makes predictive learning harder compared to ordinary static learning circumstances. Various learning algorithms have been proposed to tackle the concept drift inherent in data stream and ensemble methods have been verified as a best approach for learning a drifting concept in many cases. Here, we propose an ensemble method which utilizes constrained penalized regression as a combiner to track a drifting concept in a classification setting. We develop an efficient optimization algorithm to implement the proposed method and present numerical results verifying the promising aspects of the suggested method for a concept drift learning in changing environments.