Métodos ensemble online para classificação binária

Abstract

Ensemble methods stand out as highly effective approaches for classifying data streams, often surpassing individual models in terms of prediction errors. By combining individual models, the ensemble can enhance predictions, capturing different perspectives and nuances of the data. However, in dynamic and challenging scenarios, where data streams are voluminous and constantly changing, and often there is no access to training data in advance, ensemble methods based on traditional, especially batch, learning are not suitable. In these contexts, online machine learning emerges as the ideal solution. In this paradigm, models are continuously trained with each new observation, allowing them to dynamically adapt to changes in the data, where conditions and patterns can evolve rapidly. This monograph explores the gap between theory and practice in the realm of online machine learning, focusing on the ensemble method approach for binary classification. The monograph examines the inherent challenges in the selection and implementation of individual models, as well as explores some techniques for combining the predictions of these base models. Additionally, this work presents the results of experiments conducted on eight datasets, which simulate real-world scenarios. The results highlight the proposed ensemble method’s superiority over individual approaches considering the most common prediction metrics.