SeleÃÃo de protÃtipos: combinando auto-geraÃÃo de protÃtipos e mistura de gaussianas

AUTOR(ES)

Cristiano de Santana Pereira

DATA DE PUBLICAÇÃO

2008

RESUMO

Prototype selection is a machine learning scheme in which the main purpose is to choose pattern vectors from the training dataset that achieve a better fit to data point distributions and maintain class separation. The aim of this kind of technique is to find the smallest possible prototype set that minimizes the classification error rate. Prototype-based strategies have been used at real-world applications in several domains and promising results have been achieved by these techniques. The purpose of this work was to investigate prototype selection schemes based on self-generating of prototypes and Gaussian mixtures comparing the performance with some classical prototype selection algorithms. The respective advantages and problems are discussed. Resulting of this study, a novel hybrid model combining self-generating and Gaussian mixtures was proposed. This hybrid model overcomes difficulties of the methods studied and improves accuracy. This model have few parameters and uses a Gaussian Mixture ansatz to describe the data point distributions and to better fit the prototypes to the classification boundaries. To evaluate the hybrid method, experiments were performed over real-world databases. These chosen databases have been a widely used benchmark in machine learning. The hybrid model achieved superior performance in terms of classification accuracy when compared to the other evaluated methods. The second part of this work presents an architecture for a system to detect handwritten connected digits using the hybrid method proposed here. Besides, a set of 12 numeric features was introduced. The experiments performed over a real-world handwritten digits dataset achieved very good accuracy rates

ASSUNTO(S)

quantizaÃÃo vetorial ciencia da computacao vector quantization machine learning self-generating prototypes mistura de gaussianas seleÃÃo de protÃtipos gaussian mixtures connected handwritten digits aprendizagem de mÃquina auto-geraÃÃo de protÃtipos prototype selection

Documentos Relacionados

• ComunicaÃÃo intermediada por protÃtipos
• AvaliaÃÃo heurÃstica para protÃtipos de jogos digitais: adaptaÃÃo do mÃtodo de heurÃsticas para a aplicaÃÃo no primeiro protÃtipo funcional de jogos digitais
• Algoritmo adaptativo de seleÃÃo do esquema de codificaÃÃo no GPRS
• Estudo de processos para modelagem, avaliaÃÃo, alteraÃÃo e seleÃÃo de estruturas de proteÃnas
• Um processo para seleÃÃo de metodologias de desenvolvimento de software