Accuracy based weighted aging ensemble (AB-WAE) — Algorithm for data stream classification

Nowadays, most of the data comes continuously and its distribution may change over the time. Unfortunately, most of classifiers assume that statistical characteristics of used predicting model are not being changed. This work presents modification of the previously proposed Weighted Aging Classifier Ensemble (WAE), called Accuracy Based WAE (AB-WAE), which can easily adapt to the probability characteristic changes caused by so-called concept drift. The proposed model does not exploit a drift detection mechanism, but AB-WAE tries to change the line-up of the classifier ensemble and weights assigned to base classifiers. Thus, one individual classifier is trained on the basis of the each incoming data chunk, then AB-WAE chooses the most valuable ensemble taking into consideration: fixed ensemble size, the previously trained models and new trained classifier. The discussed WAE modification uses the ensemble of homogeneous classifiers only, what allows to employ more sophisticated combination rule based on support functions, what should boost the classification accuracy, what was confirmed by the computer experiments.