Σύγκριση πέντε μεθόδων μείωσης διαστάσεων στο πολυδιάστατο σύνολο δεδομένων «Forest Cover Type»

Abstract

A multidimensional and multivariate structure, with mixed data type gives to researchers the opportunity to use many statistical methods of dimensionality reduction, which aimed at a reduced representation of the original data set that will be smaller in “volume” but will still contain critical and useful information. In this study, the statistical dimensionality reduction methods that were compared with each other, through appropriate data set, were Principal Components Analysis, Factor Analysis, Correspondence Analysis, Categorical Principal Components Analysis and Factor Analysis for Mixed Data. For the comparisons of these methods, various strategies were applied. The aims of this study were to compare the results of five dimensionality reduction methods, to check the application of these methods to multidimensional mixed data type and to compare the results’ extraction times from different statistical softwares, with purpose to highlighting significant computational and interpretive disadvantages. The statistical softwares used were Python and SPSS. Important disadvantages of these methods were the “curse of dimensionality”, in the sense of determining the number of important dimensions, the increased computing power that was required, the lack of softwares’ code for certain methods, the differentiation in terms of calculations between the softwares and by extension the extraction of different results, the disability of some softwares in terms of the management of many pseudo-variables and the difficulty of highlighting the most appropriate method for reducing the mathematical dimensions.