Estudios de medidas de diversidad en sistemas multiclasificadores
Fecha
2013-05-20
Autores
Cabrera Hernández, Leidys
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas. Facultad de Matemática, Física y Computación. Departamento de Computación
Resumen
En la actualidad las técnicas de clasificación existentes ayudan a resolver muchos problemas en el campo del aprendizaje automatizado y la búsqueda de patrones. Dentro de estas técnicas se encuentran los clasificadores individuales y los sistemas multiclasificadores, estos últimos combinan la salida de varios clasificadores con el objetivo de mejorar el resultado de la clasificación. Resulta intuitivo pensar que el resultado de combinar un grupo de clasificadores idénticos no va a ser mejor que el resultado de uno solo de sus miembros. Al contrario, resultaría más conveniente si se combinara un grupo de clasificadores diferentes entre sí, dado que al menos uno de ellos debe dar la respuesta correcta cuando el resto falle. Dicha diferencia es conocida principalmente como diversidad. Comprender y cuantificar la diversidad que existe en un ensamblado de clasificadores es un aspecto importante. En el presente trabajo se pretende analizar los resultados de medidas de diversidad usadas para tal propósito. Para esto se diseñan y ejecutan experimentos en los que se calculan los valores de las medidas para varias combinaciones de clasificadores, los cuales son entrenados con rasgos diferentes. El multiclasificador utilizado fue el “Vote_S_R” y para la combinación de las salidas de los clasificadores se utilizan dos funciones matemáticas: promedio de probabilidades y voto mayoritario.
Luego se aplican técnicas estadísticas como análisis descriptivos, análisis de correlación, entre otras, que demuestran la variabilidad de los resultados de las medidas de diversidad. Para el cálculo de estas medidas se utilizó el software “Diversidad”, al cual se le incorporó la implementación de la mayoría de las medidas. Además con su ayuda se obtienen resultados para aplicaciones reales.
Currently existing classification techniques help to solve many problems in the field of machine learning and pattern matching. Among these techniques are the individual classifiers multiclasifier systems. The latter combine the output of several classifiers to improve the classification result. It is intuitive to think that the result of combining a group of identical classifiers will not be better than the result of one of its members. On the contrary, it would be more convenient if we combine a group of different classifiers, given that at least one of them must give correct answer when the rest fail. This difference is primarily known as diversity. Understand and quantify the diversity that exists in an assembly of classifiers is an important aspect. In this paper we analyze the results of diversity measures used for this purpose. For doing this, experiments are designed and implemented in which the values are calculated for several combinations of steps classifiers, which are trained with different traits. The multiclasifier used was "Vote_S_R" and for the combination of the outputs of classifiers two mathematical functions: average odds and majority vote are used. Then statistical techniques as descriptive analysis, correlation analysis, among others were applied, which show the variability of the measurement results of the calculation of this diversity. The software "Diversity," was used with incorporated implementing measures. And with its help, real applications are obtained.
Currently existing classification techniques help to solve many problems in the field of machine learning and pattern matching. Among these techniques are the individual classifiers multiclasifier systems. The latter combine the output of several classifiers to improve the classification result. It is intuitive to think that the result of combining a group of identical classifiers will not be better than the result of one of its members. On the contrary, it would be more convenient if we combine a group of different classifiers, given that at least one of them must give correct answer when the rest fail. This difference is primarily known as diversity. Understand and quantify the diversity that exists in an assembly of classifiers is an important aspect. In this paper we analyze the results of diversity measures used for this purpose. For doing this, experiments are designed and implemented in which the values are calculated for several combinations of steps classifiers, which are trained with different traits. The multiclasifier used was "Vote_S_R" and for the combination of the outputs of classifiers two mathematical functions: average odds and majority vote are used. Then statistical techniques as descriptive analysis, correlation analysis, among others were applied, which show the variability of the measurement results of the calculation of this diversity. The software "Diversity," was used with incorporated implementing measures. And with its help, real applications are obtained.
Descripción
Palabras clave
Vote_S_R, Sistemas Multiclasificadores