Ensembles of Active Adaptive Incremental Classifiers

Abstract

An important and ubiquitous feature of the data stream generating process is its nonstationarity. Therefore, the models trained on such data streams have to be adaptive in order to react correctly on appearing concept drift. There is a number of concept drift detection methods that can be combined with a learning method to create active adaptive learner. However, none of the approaches was reported to be unambiguously the best. Within the presented work two ensemble approaches combining the adaptive base learners are applied in order to achieve higher classification quality. The analysis shows diversity of the utilised base adaptive learners justifying application of the proposed solution. The quality of results confirms that creating the ensemble of drift detectors can improve the classification quality.