Diversidade e similaridade como critério de seleção de classificadores em comitês de seleção dinâmica

Abstract

Pattern classification techniques are considered to be key activities in the area of pattern recognition, where seeks to assign a test sample to a class. The use of individual classifiers usually exhibits deficiencies in recognition rates when compared to the use of multiple classifiers to perform the same classification task. According to the literature, ensemble of classifiers provide better recognition rates when candidate classifiers present uncorrelated errors in different sub-spaces of the problem. In this context, this doctoral thesis explores several methods of selection of classifiers, based on dynamic selection, adding a selection criterion that prioritizes diversity and/or similarity between the base classifiers. In this way the experiments evaluated aim to empirically elucidate the relevance of diversity and/or similarity among the base classifiers of ensembles based on dynamic selection. Many papers explore diversity in ensemble systems based on static selection and indicate that diversity among the base classifiers is a factor that positively influences accuracy rates, however in the context of ensemble based on dynamic selection there is no enough related literature and few research that explore the influence of diversity and similarity among the base classifiers.