Ordered Predictors Selection: New Approaches for Variable Selection in Multivariate Calibration

The ordered predictors selection (OPS)1 is a variable selection method based on obtaining an informative vector that contains information about the location of the best independent variables for prediction. In this work, three new OPS approaches are proposed and applied: (1) AutoOPS, where is chosen automatically the vector that selecting the set of variables with the best prediction capability; (2) FeedOPS, wherein the pre-selected variables return to a new selection until obtain the best set of variables and (3) iOPS, where the OPS is applied at variable independent intervals. Multivariate calibration models using partial least squares (PLS) regression were built applying the old OPS version (OPS1.0), new approaches of OPS (OPS2.0) and other three algorithms of variable selection: genetic algorithm2 (GA), successive projections algorithm4 (SPA) and recursive weighted partial least squares3 (RPLS). A Raman spectroscopy dataset5 was used in this study. This dataset and additional information about it are available at http://www.models.ku.dk/datasets. The property analyzed was the iodine value, which ranged from 52 to 60 g I2/100g fat. The calibration and prediction sets were separated into 75 and 30 samples respectively. 5667 variables were used. The homemade OPS algorithms were written, tested and applied in Matlab R2017b. Statistical parameters of models are shown in Table 1. RMSEP values and relative error (%) are shown in Fig. 1. A Tukey test was applied to the RMSEP values with 95% confidence level. These results indicated that the OPS2.0 was able to improve statistical parameters regarding the other methods investigated in this work, providing substantially better predictions. For other different data sets, OPS2.0 has also shown better performance.