Fast and economic integration of new classes on the fly in evolving fuzzy classifiers using class decomposition

In this paper, we propose a fast and economic strategy for the integration of new classes on the fly into evolving fuzzy classifiers (EFC) during data stream mining processes. Fastness addresses the assurance that a newly arising class in the stream can be integrated in a way such that the classifier is able to correctly return the new class after receiving only a few training samples of it. Economic means that the classifier update cycles are decreased to a minimum amount of time, as these require operator's feedback for obtaining the ground truth labels, which are usually costly to obtain. The former is achieved by a class-decomposition approach, which splits up multi-class classification problems into several less imbalanced and less complex binary sub-problems. The latter is achieved by a single-pass active learning selection scheme which selects the most informative samples based on sample-wise criteria. The approach is compared with conventional single model architecture for EFC (EFC-SM) based on two data streams from a real-world application in the field of surface inspection. The comparison shows that the class decomposition approach can significantly reduce the delay of class integration, and this with a lower # of samples used for model updates than EFC-SM.