Classifier ensemble algorithm for learning from non-stationary data stream

Abstract

Nowadays, many sources generate unbounded data streams at high incoming rates. It is impossible to store these large volumes of data and it is necessary to process them in real time. Because these data are acquired over time and the dynamism of many real-world situations, the target function to be learned can change over time, a problem commonly called concept drift. This paper presents a new ensemble algorithm called Classifier Ensemble Algorithm (CEA), able for learning from data streams with concept drift. CEA manipulates these changes using a change detector in each base classifier. When the detector estimates a change, the classifier in which the change was estimated is replaced by a new one. CEA combines the simplicity of the bagging algorithm to train base classifiers and Exponentially Weighted Moving Average (EWMA) control charts to estimate the weights of each base classifier. The proposed algorithm is compared empirically with several bagging family ensemble algorithms able to learn from concept-drifting data. The experiments show promising results from the proposed algorithm (regarding accuracy), handling different types of concept drifts.

Keywords

• Data stream
• concept drift
• classifier ensemble