Page 173 - Contributed Paper Session (CPS) - Volume 4

CPS2166 Divo Dharma Silalahi et al.
in the data pre-processing step. To examine the performance, the proposed
method was compared with the classical VIP (Oussama et al., 2012) and
MCUVE (Cai et al., 2008) using different datasets. These statistical measures
use Desirability Indices (Trautmann, 2004) such as Root Mean Squared Error
2
(RMSE), Coefficient of Determination (R ), Bias, and standard Error (SE) by
contrasting the actual (measured y) with the results given in the model
prediction. This study provides a development for process control in the
vibrational spectroscopy technique through wavelengths selection method.
Particularly in the chemical analysis, this will assist to a better understanding
on the main chemical compositions of the target sample.

2. Input Scaling of Filter-Wrapper Method
Following the OPLS-VIP (Galindo‐Prieto et al., 2014), the VIP score
measures not only the contribution of each j th wavelength in multivariate
models based on the projections to PLS components but also include the
orthogonal components. Statistically says the VIP score in the OPLS model
considers two amounts of variations that are in response variable y (SSY)

and in predictor variable X ( SSX ). There are four versions of OPLS-VIP

developed by Galindo‐Prieto et al. (Galindo‐Prieto et al., 2014), the fourth
variants is the recommended one due to its interpretative information ability
at the wavelengths that are more relevant both in predictive and orthogonal
components. In this study, the fourth variants OPLS-VIP score which considers
the combinationsSSX, SSY  in the weighting parameters and normalized

loadings v is preferred to be used for the scaling. In line with the earlier PLSR
g
theory which sensible only on the predictive components, with related to the
OPLS model then the variations in predictor variable X is also integrated in

the formulation. Here, there are two VIP scores proceed separately for the final
OPLS-VIP score which are VIP pred (predictive components) and VIP ortho

(orthogonal components). Let redefine g as the predictive component and
g o as the orthogonal component, then l stands for total number of
predictive components and l stands for total number of orthogonal
o
components with m and m are the total numbers of variables used in the
o
predictive and orthogonal components, respectively. The calculation for OPLS-
VIP score both in predictive and orthogonal can be written as below
   (v 2 x SSX )  (v 2 x SSY ) 
l
l
m g comp; g g comp; g  (1)
=
VIP = x  g 1 + g = 1
pred
2  SSX SSY 
  cum cum  
 
162 | I S I W S C 2 0 1 9