Dynamic weighted fusion of adaptive classifier ensembles based on changing data streams

Adapting classification systems according to new input data streams raises several challenges in changing environments. Although several adaptive ensemble-based strategies have been proposed to preserve previously-acquired knowledge and reduce knowledge corruption, the fusion of multiple classifiers trained to represent different concepts can increase the uncertainty in prediction level, since only a sub-set of all classifier may be relevant. In this paper, a new score-level fusion technique, called Swavgh, is proposed where each classifier is dynamically weighted according to the similarity between an input pattern and the histogram representation of each concept present in the ensemble. During operations, the Hellinger distance between an input and the histogram representation of every previously-learned concept is computed, and the score of every classifier is weighted dynamically according to the resemblance to the underlying concept distribution. Simulation produced with synthetic problems indicate that the proposed fusion technique is able to increase system performance when input data streams incorporate abrupt concept changes, yet maintains a level of performance that is comparable to the average fusion rule when the changes are more gradual