[Response] Thanks for your comments.
34. How does the author know that ANN was not overfitting? According to
Beer’s law, the relationship between absorbance and the analyte of
interest is linear.
How
does the author justify the use of ANN that models non-linear
relationships for an application were the linear relationship is proven?
Is
it possible that he/she were modeling interferences in the frying
process?
[Response] Thanks for your comments. Multivariate approaches
suffer from overfitting, thus validation is an obligatory component of
any analysis (Cavanna, Righetti, Elliott, & Suman, 2018)). Typically,
cross-validation approaches are used, in which a proportion of the data
(e.g., 10–40%, the “validation set”) are randomly removed, and the
model is built with the remaining “training set”. This procedure is
repeated many times until each sample has been in the test set exactly
once (leave-n-out procedure). The accuracy of the model on these
left-out samples gives an estimate of the predictive power for unseen
samples and also the robustness of the model to perturbations of the
data. Therefore, we applied the validation model so as to evaluate the
model is not overfitting. There is no one-to-one correspondent
relationship between thesechemical indexes (polar
compounds, free fatty acids, hydroperoxide) and the selected peak areas,
therefore they may be not absolutely the linear relationships between
the chemical indexes and the selected peaks. Many researches using the
ANN to calibrate the model with the spectra and other chemical index.
indicating it is available for choosing ANN for this spectroscopic
application (Özbalci, et al., 2013; Wenning et al., 2010; Barmpalexis,
2018; Li, et al., 1999; Agatonovic-Kustrin, 2013).