Design of a reference value-based sample-selection method and evaluation of its prediction capability

• A new sample selection method, reference value based selection method, is present
• Different selection performances are compared based on prediction results.
• Independent validation set is used to test model performance of different selection method
• Three constituents of soy sauce representative for real complicate sample used to make through comparison of different selection method

A calibration set comprises the multidimensional space that represents the samples for prediction. The representative ability of a calibration set is a major factor that affects the predictive performance of a multivariate regression. A new reference value (YR)-based sample-selection algorithm that assembles a dependent value (y-value) uniform distribution is presented to assure the representation. The existing typical sample-selection algorithm is used for comparison. A set of soy sauce data is used as a set of typical samples that have a complex solution. Comparing the prediction results, it is shown that YR sample-selection has similar prediction performance to that of sample set partitioning based on joint x–y distances (SPXY), but with a simpler algorithm. The calibration models of the y-reference-included sample sets (SPXY and YR) are more accurate than those of y-reference-excluded sample sets (RS and KS). After modeling with the selected representative samples, the performances of YR and SPXY are comparable to that of full sample modeling with fewer samples.