Advances in network-based ensemble classifiers for evolving data streams
Heitor Murilo Gomes · 2016
Data stream classification has grown in importance in recent years due to the large amount of data rapidly generated in a multitude of domains. Many algorithms were proposed to deal with the problems associated with data stream classification (e.g. concept drifts) with special attention to ensemble methods. Ensembles are often preferred due to their exibility and ability to generate accurate results without much configuration in comparison to other approaches that rely on single strong learners (e.g. SVMs or ANNs). We propose an ensemble-learning framework for evolving data stream classification in which ensemble classifiers are arranged in a network structure, namely the network-based ensembles. In this approach relations between component classifiers are explicitly defined and used in a variety of ways to improve the ensemble overall performance. Recent works, such as the Social Adaptive Ensemble (SAE) algorithm, already apply some of the concepts we use in our framework; therefore besides formalizing the characteristics of a network-based ensemble, we also discuss existing work and propose novel ensemble techniques.