Detecting and adapting to drifting concepts

The importance of incremental learning in changing environments has been acknowledged in recent years. In this paper we present an ensemble learning method for supervised learning with drifting concepts. The method employs hypothesis test as mechanism for detecting concept drift and learns a base classifier for each new training data chunk. Former classifiers deemed as usable by the hypothesis test mechanism and the new classifiers are integrated to form the final classifiers ensemble for prediction. The main focus of the work is to identify the usability of base classifiers that representing the same or similar concept with the current one, make full use of the older valid information together with the newer examples to improve classification accuracy, and avoid the interference of classifiers representing conflictive concepts with the current one. Experiments with simulated concept drift scenarios compared the proposed method with other approaches. The results showed that the method could consistently recognize different types of drift, adapt quickly to these changes to maintain its performance level, and utilize the former knowledge to improve its performance for recurring context.