Navegando por Palavras-chave "Pruning"

Agora exibindo 1 - 1 de 1
  • Nenhuma Miniatura disponível
    Item
    Somente Metadadados
    Um arcabouço para seleção e combinação de classificadores baseado em algoritmos evolutivos
    (Universidade Federal de São Paulo (UNIFESP), 2019-11-29) Ferreira Junior, Alvaro Roberto [UNIFESP]; Faria, Fabio Augusto [UNIFESP]; Universidade Federal de São Paulo (UNIFESP)
    Classification is one of the most studied learning tasks in the area of machine learning and aims to find a hypothesis (model) that best fits and generalizes the behavior of data during the training stage. Once the classification model has been learned, it will assign classes to new examples in the test phase. In the literature, many studies have been carried out to solve classification problems in several knowledge domains (e.g., medicine, biology, safety and remote sensing). As there is not a single classifier that achieves satisfactory results in any application, a good alternative is to adopt information fusion strategies. Among the information fusion strategies are those related to the combination of classifiers or the so-called multiple classifier systems (MCSs). An MCS aims to combine classifiers that have complementary information to each other to improve the results of effectiveness in the target application. In these systems, a very important concept for measuring the degree of agreement/disagreement among classifiers is so-called diversity. Although many authors have adopted diversity in their work, it was noted that its use alone is not sufficient to obtain improvements in the effectiveness of MCSs. An essential factor for the success of these approaches is the combination of diversity and accuracy of the classifiers belonging to the MCS. As there are infinite classifiers in the literature, a challenge lies in the choice of classifiers that will compose the final classification system, thus arising the need to develop new strategies for classifier selection. In this sense, this work proposes the development of a framework for selection and combination of classifiers that uses different optimization techniques based on evolutionary algorithms combining measures of diversity and accuracy of classifiers, selecting a subset of these classifiers based on many other classifiers available and finally, create an MCS to improve the results of effectiveness in the classification task. In the experiments carried out it was possible to analyze the impact of each of the four steps of the CIF-E protocol (Classifiers, Initialization, Fitness of individuals and Evolutionary technique) that make up the proposed framework, totaling 24 different methods implemented and tested. In addition, a comparative analysis between the best methods proposed in this work and the methods in the literature was performed. Finally, the experiments show that the method based on Univariate Marginal Distribution Algorithm (UMDA) can be better than the seven state-of-the-art literature methods in eleven of the nineteen UCI datasets tested in this work.