Multicriteria selection of uncorrelated variables for modeling

• A variable selection method to retain the least correlated from a set of variables
• The variable subsets are chosen by calculation of the intrinsic criteria and procrustes criterion (similarity between the initial dataset and the subset)

Selection methods are commonly used to retain only the least correlated variables when data are described by a large number of variables. However, most of the currently available variable selection methods do not take the intrinsic quality of the subset retained into account. In this paper, we propose a new approach based on a multicriteria selection of variables. The intrinsic quality of the selected subset was assessed based on both criteria calculated from the model matrix and the procrustes analysis. This verification guarantees a good estimation of the coefficients for the model and a good representativity. This approach was applied to two cases: a benchmark dataset known as Coffee data and a real dataset produced by a study of quantitative structure–activity relationship. In both cases, the solutions were representative of the initial set and displayed good intrinsic quality, these solutions will therefore be useable in the modeling step.