Um estudo exploratório sobre o uso de diferentes algoritmos de classificação, de seleção de métricas, e de agrupamento na construção de modelos de predição cruzada de defeitos entre projetos

• Ano de defesa: 2015
• Autor(a) principal: Satin, Ricardo Francisco de Pierre
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Cornelio Procopio; Brasil; Programa de Pós-Graduação em Informática; UTFPR
• Palavras-chave em Português: Teoria da previsão; Software - Desenvolvimento; Falhas de sistemas de computação; Prediction theory; Computer software - Development; Computer system failures; CNPQ::CIENCIAS EXATAS E DA TERRA; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/2552

To predict defects in software projects is a complex task, especially for those projects that are in early stages of development by, often, providing few data for prediction models. The use of cross-project defect prediction is indicated in such a situation because it allows reuse data of similar projects. This work proposes an exploratory study on the use of different classification algorithms, of selection metrics, and grouping to build cross-project defect predictions models. This model was built using a performance measure, obtained by applying classification algorithms aim to find and group similar projects. Therefore, it was studied the application of 8 classification algorithms, 6 feature selection, and a cluster in a data set with 1283 projects, resulting in the construction of 61584 different prediction models. The classification algorithms and feature selection had their performance evaluated through different statistical tests showed that: the Naive Bayes was the best performance classifier, as compared with other 7 algorithms; the pair of feature selection algorithms that performed better was formed by CFS attribute evaluator and search method Genetic Search, compared with 6 other pairs. Considering the clustering algorithm, this proposal seems to be promising, since the results shows evidence that the predictions were best grouping using the predictions performed without any similarity clustering, and shows the decrease in training cost and testing during the prediction process.