AnÃlise discriminante clÃssica e de nÃcleo: avaliaÃÃes e algumas contribuiÃÃes relativas aos mÃtodos Boosting e Bootstrap

AUTOR(ES)

Marcelo Rodrigo Portela Ferreira

DATA DE PUBLICAÇÃO

2007

RESUMO

Since information technology is essential for many activities of modern life and massive data sets come with it, data mining has become one of the most important research topics in statistical science. Even thought there are many fields related to data mining, the task of classification still remain as one of the most common in statistical literature. This dissertation reviews two classical classifications methods, linear and quadratic discriminant analysis, and a nonparametric method, the kernel discriminant analysis. Simulations experiments and real data sets are used to compare the three classification methods. It also presents some contributions which are related to boosting and bootstrap methods in the classification context. The first contribution is a new formulation to the boosting method in linear discriminant analysis. The numerical results show that this new formulation has similar performance to the previous one. However, the new boosting formulation is more appropriate from the conceptual point of view. Two bootstrap methods for classification problems are introduced and evaluated. The first bootstrap method is used to obtain a classification frontier. The concept of classification frontier can be understood as a region where it is difficult to assign one observation to one of some relevant populations. The second bootstrap method is a confidence interval for the classification error rate. Confidence intervals can be used to compare two or more classification methods in the inference framework

ASSUNTO(S)

fronteira de classificaÃÃo classification frontier boosting quadratic discriminant analysis kernel discriminant analysis classificaÃÃo anÃlise discriminante linear anÃlise discriminante de nÃcleo estatistica boosting anÃlise discriminante quadrÃtica classification linear discriminant analysis bootstrap intervalos de confianÃa bootstrap bootstrap bootstrap confidence intervals

Documentos Relacionados

• ContribuiÃÃo ao processo de regionalizaÃÃo de vazÃes : AnÃlise de mÃtodos e variÃveis explicativas
• Desenvolvimento local sustentÃvel no setor sucroalcooleiro de Pernambuco: anÃlise e contribuiÃÃes
• ContribuiÃÃes à anÃlise de desempenho e projeto de controladores lÃgicos nebulosos.
• ComparaÃÃo dos mÃtodos anÃlise por envoltÃria de dados e anÃlise por fronteira estocÃstica em avaliaÃÃo de eficiÃncia.
• Limite mÃnimo de detecÃÃo de mÃtodos de anÃlise de Salmonelas spp. em alimentos: uma contribuiÃÃo metodolÃgica