Original scientific paper

Procrustes Rotation and Pair-wise Correlation: a Parametric and a Non-parametric Method for Variable Selection

Károly Héberger ; Institute of Chemistry, Chemical Research Center, Hungarian Academy of Sciences, P.O. Box 17, H-1525, Budapest, Hungary
José M. Andrade ; Department of Analytical Chemistry, University of A Corunna, Campus da Zapateira s/n, E-15071, A Corunna, Spain

Abstract

In this work attention is focused on selecting a small set of original (independent) variables, which take into account the most important information present in the data matrix. The performance of two already implemented methods is compared from a practical point of view: the Procrustes Rotation algorithm, which is parametric, and the Pair-wise Correlation (PCM), which is nonparametric, because the tests used to discriminate between the variables are nonparametric. Using a well-documented data set (aphid data), both methods gave comparable results. Procrustes Rotation selected four variables, whereas four or five variables were retained using the pair-wise correlation method (depending on the test used). Three variables were common to both approaches and the main structure in the original data set was retained in both cases. Therefore, both methods are appropriate for variable selection. Selection criterion for ranking the variables using PCM was further developed to retain the highest »dissimilarity« among similar variables.

Keywords

Procrustes analysis; robust method; feature selection; classification; pair-wise correlation; similarity measures

Hrčak ID:

102654

URI

https://hrcak.srce.hr/102654

Publication date:

31.5.2004.

Visits: 1.575 *