Data with shifting concept classification using simulated recurrence

Abstract

One of the serious problems of modern pattern recognition is concept drift i.e., model changing during exploitation of a given classifier. The paper proposes how to adapt a single classifier system to the new model without the knowledge of correct classes. The proposed simulated concept recurrence is implemented in the non-recurring concept shift scenario without the drift detection mechanism. We assume that the model could change slightly, what allows us to predict a set of possible models. Quality of the proposed algorithm was estimated on the basis of computer experiment which was carried out on the benchmark dataset. © 2012 Springer-Verlag.