J. Graffelman

There exist many ways to create a graphical representation of a correlation matrix. Often correlations between variables are graphically approximated by the cosines of the angles between vectors. The relationship between cosines and correlations follows directly from multivariate sample geometry, and is widely used in practice in the interpretation of biplots obtained by principal component analysis. However, in the latter biplots the approximation by cosines is suboptimal in the least squares sense. Moreover, it is difficult to reliably infer the sample correlations by eye from these biplots. By developing plots that have correlations that are linear in the angle, both these aspects can be improved. In the examination of a set of data matrices the cosine based plots typically gave the poorest approximation, plots with a linear interpretation rule for the angle improved the fit, and the best fit was generally obtained by principal factor analysis and using scalar products.

Palabras clave: biplot, scalar product, interpretation function

Programado

JC3 Clasificación y análisis multivariante 2
19 de abril de 2012  12:00
Sala Londres

Logistic biplots for categorical data

J. L. Vicente Villardón

URL de la comunicación: http://biplot.usal.es/ClassicalBiplot/index.html

Fichero de la comunicación:

Aplicación a catálogos estelares de técnicas estadísticas de reducción de la dimensión y clasificación

A. Berihuete, A. Jimenez Jimenez, F. Álvarez González

Detección de interacciones relevantes para el problema de clasificación binaria en conjuntos de datos de alta dimensión

J. Martín Arevalillo, H. Navarro Veguillas