Employment of Uninformative Variable Elimination (UVE) as a robust variable selection method is reported in this study. Each regression coefficient represents the contribution of the corresponding variable in the established model, but in the presence of uninformative variables as well as collinearity reliability of the regression coefficient's magnitude is suspicious. Successive Projection Algorithm (SPA) and Gram-Schmidt Orthogonalization (GSO) were implemented as pre-selection technique for removing collinearity and redundancy among variables in the model. Uninformative variable elimination-partial least squares (UVE-PLS) was performed on the pre-selected data set and C-value's were calculated for each descriptor. In this case the C-value's of LIVE assisted by SPA or GSO could be used in order to rank the variables according to their importance. Leave-many-out cross-validation (LMO-CV) was applied to ordered descriptors for selecting optimal number of descriptors. Selwood data including 31 molecules and 53 descriptors, and anti-HIV data including 107 molecules and 160 descriptors were utilized in this study. When GSO pre-selection method is used for the Selwood data and SPA for the anti-HIV data set, obtained results were desired not only in the prediction ability of the constructed model but also in the number of selected informative descriptors. By applying GSO-UVE-PLS to the Selwood data, in an optimized condition, seven descriptors out of 53 were selected with q(2) = 0.769 and R-2 = 0.915. Also applying SPA-UVE-PLS on the anti-HIV data, nine descriptors were selected out of 160 with q(2) = 0.81, R-2 = 0.84 and Q(F3)(2) = 0.8. (C) 2013 Elsevier B.V. All rights reserved.
Chemometrics and Intelligent Laboratory Systems, 2013, Vol 128, p. 56-65