Variable Selection for Multi-Purpose Multivariate Data Analysis

TL;DR: In this paper, the authors proposed a method for selecting a subset of variables from a given set of p input variables, by the criterion of minimum trace of partial variances of unselected variables unexplained by selected variables.

Abstract: Recently we frequently analyze multivariate data with quite large number of variables. In such data sets, virtually duplicated variables may exist simultaneously even though they are conceptually distinguishable. Duplicate variables may cause problems such as the distortion of principal axes in principal component analysis and factor analysis and the distortion of the distances between observations, i.e. the input for cluster analysis. Also in supervised learning or regression analysis, duplicated explanatory variables often cause the instability of fitted models. Since real data analyses are aimed often at multiple purposes, it is necessary to reduce the number of variables to a parsimonious level. The aim of this paper is to propose a practical algorithm for selection of a subset of variables from a given set of p input variables, by the criterion of minimum trace of partial variances of unselected variables unexplained by selected variables. The usefulness of proposed method is demonstrated in visualizing the relationship between selected and unselected variables, in building a predictive model with very large number of independent variables, and in reducing the number of variables and purging/merging categories in categorical data.