Comparison of multivariate analysis methods to evaluate

VERDINI, RA; ZORRILLA, SE; RUBIOLO, AC; NAKAI, S

Villa de Merlo, San Luis, Argentina

Congreso; Asociación Argentina de Químicos Analíticos; 2005

Asociación Argentina de Químicos Analíticos

Classification of cheese variety and maturity, based on the contents of casein, peptide and amino acid data measured by electrophoretic and/or chromatographic analysis, is being continuously discussed. In addition, if compositional and textural analyses are also incorporated, it may produce so large amount of data that the objective assessment is essential. Structure within these large data sets can be determined when they are objectively analyzed using multivariate statistical methods. However, for successful classification or differentiation, analytical and statistical methods should be carefully selected and the quality of the data set must be high. In the present work, some unsupervised multivariate methods applied to chemical and physical data were compared to evaluate Port Salut Argentino cheese ripening. Principal component analysis was used for data mining, reducing data dimensionality from 35 to 7 input variables. Subsequently, principal component analysis (PCA), principal component similarity (PCS), Kohonen self-organizing artificial neural network (K-SOFM) were applied for grouping the samples in two-dimensional maps. Principal component analysis yielded two PCs that explained 91.0% of the data set variation (PC1 71.8% and PC2 19.2%). Samples were grouped according to ripening time and sampling site in the PC score plot. Four PCs were used for PCS computation. In general, the PCA and PCS mapping of control cheeses was comparable and two PCs accounted for more than 90% of the data variation. As a result, PCS was not superior to PCA for classifying samples. However, plots of adjusted factor score deviations from reference line were very useful for the understanding of the influence of the variables on the mapping of the samples. Several architectures of K-SOFM were tested: 3 × 3, 4 × 4, 5 × 5, and 6 × 6.  The sample grouping in the top-map, except only in few cases, was similar for all tested architectures. Samples were grouped according to ripening time and sampling site in the top-map. The three multivariate methods were very useful for the modeling of Port Salut Argentino cheese samples showing an adequate mapping of both the samples and the variables and providing complementary information. In addition, in each case, the relationship between the groups and the original variables was established.