SELEÇÃO DE CLASSIFICADORES EM FLUXOS DE DADOS BASEADA EM PERFORMANCE DE CURTO PRAZO II

INTRODUCTION: Dynamic classifier selection (DCS) are well-known machine learning techniques in the batch environment that improve the performance of ensembles. Most methods use kNN-type learners, generating high computational costs and making the application of these approaches to the stream environment unfeasible. AIMS: In this work the analysis of the MSTS algorithm, which selects classifiers based on their performance in the most recent instances in an efficient way for the context of data streams, is extended. MATERIALS AND METHODS: The impact of this selection strategy with base classifiers other than Hoeffding Trees is evaluated with ensembles from the classification literature on data streams. RESULTS: Results show that using learners based on different Hoeffding Trees (EFDT) can bring performance gains in certain situations. FINAL CONSIDERATIONS: The work showed that the method proposed in the 2021-2022 pibic project can improve the predictive quality of ensembles with base classifiers other than Hoeffding Trees and that it also presents similar behaviors reported to batch machine learning classifier selection techniques.

KEYWORDS: Data stream; Selection of Classifiers; Classification.