Kennard-Stone method outperforms the Random Sampling in the selection of calibration samples in SNPs and NIR data

ABSTRACT: Splitting the whole dataset into training and testing subsets is a crucial part of optimizing models. This study evaluated the influence of the choice of the training subset in the construction of predictive models, as well as on their validation. For this purpose we assessed the Kennard-Stone (KS) and the Random Sampling (RS) methods in near-infrared spectroscopy data (NIR) and marker data SNPs (Single Nucleotide Polymorphisms). It is worth noting that in SNPs data, there is no knowledge of reports in the literature regarding the use of the KS method. For the construction and validation of the models, the partial least squares (PLS) estimation method and the Bayesian Lasso (BLASSO) proved to be more efficient for NIR data and for marker data SNPs, respectively. The evaluation of the predictive capacity of the models obtained after the data partition occurred through the correlation between the predicted and the observed values, and the corresponding square root of the mean squared error of prediction. For both datasets, results indicated that the results from KS and RS methods differ statistically from each other by the F test (P-value < 0.01). The KS method showed to be more efficient than RS in practically all repetitions. Also, KS method has the advantage of being easy and fast to be applied and also to select the same samples, which provides excellent benefits in the following analyses.