New multivariate strategy for panel evaluation using principal component similarity

Abstract

Advantages of principal component similarity (PCS) as an unsupervised classification techniques compared to supervised methods are the detectability of outliers or anomalies, a capacity to classify continuum, easy identification of causes for grouping, and a potential of discovering new groups. The ability of detecting outliers or anomalies was utilized to eliminate panel members who were not appropriate for the classification purposes by using PCS scattergrams. After eliminating the unqualified panelists, the evaluation score tables were realigned from the one based on panelists to that based on samples. Accumulated principal component score were computed for the samples in a form of SPC_i × S_i, where PC_i was the principal component score of sample i and S_i was its proportion within the total variation. The conventional averaging algorithms of evaluated scores are useful, as this techniques absorbs the effect of unjustifiable negative scores reported by outliers. However, more reasonable summary scores for the samples could be obtained by eliminating unreasonable evaluation scores made by the outliers as well as by using the above new summary values. Differences in deviations of evaluation among attributes, such as those in likenesses of color and taste of meat products, were normalized. Specific training or selecting qualified panelists prior to panel evaluation is unnecessary when the information obtained from panel evaluation is required to reflect the variable broad patterns of consumer preference. Information obtained from consumer preference test based on cluster analysis, which was applied to a beverage, might be more readily recovered by using this new strategy.