Empleo de relaciones de similaridad borrosa para el cálculo de pesos en algoritmos de aprendizaje
Fecha
2015-06-10
Autores
Coello Blanco, Lenniet
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
La Medida de Calidad de la Similaridad basada en la Teoría de los Conjuntos
Aproximados constituye una alternativa eficaz para el análisis de conjuntos de datos ya
que representa el grado de similaridad entre los objetos de un sistema de decisión
heterogéneo. No obstante esta medida presenta la limitante de usar umbrales en la
construcción de las relaciones de semejanza. Estos umbrales constituyen parámetros a
ajustar, haciendo los métodos derivados muy sensible a pequeñas variaciones. En esta
investigación se propone usar las relaciones borrosas de la Teoría de los Conjuntos
Borrosos para eliminar los umbrales y flexibilizar la construcción de las relaciones de
similaridad. Con esta modificación se redefine una nueva métrica denominada Medida
de la Calidad de la Similaridad Borrosa. La nueva medida sustituye la relación de
semejanza entre dos objetos por una relación borrosa binaria, que cuantifica la fortaleza
de la relación en un rango de [0,1]. La relación borrosa está caracterizada por una
función de pertenencia, que en este caso se define por una función de semejanza basada
en la Teoría de los Conjuntos Aproximados Extendida. La nueva medida incluye el
cálculo de los pesos de los rasgos; y se estudia su impacto al igual que su precursora
para mejorar el desempeño de los k-Vecinos más Cercanos y el Multilayer Perceptron.
Los resultados experimentales en bases de datos internacionales y aplicados en dos áreas
del conocimiento: la Ingeniería Civil y el Rendimiento académico; muestran un
comportamiento semejante en cuanto a precisión entre ambas medida.
The Similarity Quality Measure based on Rough Sets Theory is an effective alternative for intelligent analysis of heterogeneous data sets. This measures represents the degree of similarity between objects of a mixed decision system. However this measure has the limitation of use thresholds building similarity relations. These thresholds are parameters to adjust the method, making it very sensitive to small variations. In this research it use fuzzy relations to eliminate the thresholds in the Measure Quality of Similarity and make it more flexible building the relationships of similarity. With this modification it redefines a new metric called Fuzzy Similarity Quality measure. The new measure replaces the relation of similarity between two objects by a binary fuzzy relation, which quantify the strength of the relationship in a range of [0.1]. The fuzzy relation is characterized by a membership function, which in this case is defined by a similarity function based on extended Rough Set theory. The new measure includes calculate features weights; and its impact as its predecessor studying for improve the performance of the k-Nearest Neighbors and Multilayer Pereceptron. Experimental results in International database and it applied in two areas of knowledge: Engineering Civil and academic performance; the result show a similar behavior between them for accuracy measure.
The Similarity Quality Measure based on Rough Sets Theory is an effective alternative for intelligent analysis of heterogeneous data sets. This measures represents the degree of similarity between objects of a mixed decision system. However this measure has the limitation of use thresholds building similarity relations. These thresholds are parameters to adjust the method, making it very sensitive to small variations. In this research it use fuzzy relations to eliminate the thresholds in the Measure Quality of Similarity and make it more flexible building the relationships of similarity. With this modification it redefines a new metric called Fuzzy Similarity Quality measure. The new measure replaces the relation of similarity between two objects by a binary fuzzy relation, which quantify the strength of the relationship in a range of [0.1]. The fuzzy relation is characterized by a membership function, which in this case is defined by a similarity function based on extended Rough Set theory. The new measure includes calculate features weights; and its impact as its predecessor studying for improve the performance of the k-Nearest Neighbors and Multilayer Pereceptron. Experimental results in International database and it applied in two areas of knowledge: Engineering Civil and academic performance; the result show a similar behavior between them for accuracy measure.
Descripción
Palabras clave
Medida de la Calidad, Conjuntos Aproximados, K Vecinos más Cercanos, Multilayer Perceptron, Algoritmos de Aprendizaje, Similaridad Borrosa, Ingeniería Civil, Rendimiento Académico