EOCD: An ensemble optimization approach for concept drift applications q

This paper addresses pattern classification problems in the data stream context. In this context, the learning concepts may change or evolve and this is called concept drift. These drifts are divided into real or virtual, depending on the variables in which a drift occurs. The real concept drifts consist of a change in the instance labels. In other words, an instance may have its label changed during its processing. Similarly, a virtual concept drift consists of changes in the attributes of instances over time, without changing their labels [10]. In this paper, we will work with real drifts. Independently of being real or virtual, drifts can also be divided into abrupt and gradual, depending on the speed that Data streams applications generate a continuous stream of data in a high rate that it is not possible to store all data in available memory. Hence, it is important to apply techniques that are capable of learning concepts according to data presentation, taking into consider-ation available time, processing and memory resources. This paper presents a new ensemble-based approach to detect concept drift in the data stream context. This approach uses an explicit mechanism to adapt to concept drifts using a genetic algorithm in order to define the best ensemble configuration for the current scenario (detected drifts). The main aim of this approach is to provide an efficient structure to detect and to adapt to concept drifts and, as a consequence, to improve the performance of this system. Our findings show that the proposed method delivers outstanding results, outperforming the results delivered by the most efficient methods presented in the literature. (c) 2021 Elsevier Inc. All rights reserved.