On ensemble components selection in data streams scenario with gradual concept-drift

Abstract

In the paper we study the issue of components selection of an ensemble for data stream classification. Decision about adding or removing single component has significant meaning not only for an accuracy in the current instant, but can be also significant for the further stream processing. The algorithm proposed in this paper is an enhanced version of the ASE (Automatically Sized Ensemble) algorithm which guarantees that a new component will be added to the ensemble only if it increases the accuracy not only for the current data chunk but also for the whole data stream. The algorithm is designed to improve data stream processing in the case when one concept is gradually replaced by the other. The Hellinger distance is applied to allow adding a new component, if its predictions differ significantly from the rest of the ensemble, even though that component does not increase accuracy of the whole ensemble.