Data Streams Classification: A Selective Ensemble with Adaptive Behavior

Abstract

Data streams classification represents an important and challenging task for a wide range of applications. The diffusion of new technologies, such as smartphones and sensor networks, related to communication services introduces new challenges in the analysis of streaming data. The latter requires the use of approaches that require little time and space to process a single item, providing an accurate representation of only relevant data characteristics for keeping track of concept drift. Based on these premises, this paper introduces a set of requirements related to the data streams classification proposing a new adaptive ensemble method. The outlined system employs two distinct structure, for managing both data aggregation and mining features. The latter are represented by a selective ensemble managed with an adaptive behavior. Our approach dynamically updates the threshold value for enabling the models directly involved in the classification step. The system is conceived to satisfy the proposed requirements even in the presence of concept drifting events. Finally, our method is compared with several existing systems employing both synthetic and real data. © Springer-Verlag Berlin Heidelberg 2013.