Using diversity to handle concept drift in on-line learning

A recent study of diversity using online ensembles of learning machines on the presence of concept drift shows that different diversity levels are required before and after a drift. Besides, studies from the dynamic optimisation problems area suggest that, if the best solution for a particular time step is adopted, it may lead to a future scenario in which low accuracy is obtained. Based on that, we propose in this paper a new online ensemble learning approach to handle concept drift, which uses ensembles containing different diversity levels. Even though a high diversity ensemble may have low accuracy while the concept is stable, it may present better accuracy after a drift. The proposed approach successfully chooses the ensemble to be used when a concept drift occurs and shows to obtain better accuracy than a system which adopts the strategy of learning a new classifier from scratch when a drift is detected (strategy adopted by many of the current approaches that explicitly use a drift detection method).