ABSTRACT
As unsupervised classifications, principal component similarity (PCS) and cluster analysis (CA) were compared for outlier detectability in panel evaluation. By rotating the reference, PCS can define outlying panelists based on the similarity of their evaluation patterns with that of the reference panelist. As a result, the outliers detected on PCS scattergrams are dependent on the reference selected, whereas, outliers detected by CA are based on dissimilarity, thus being rather unilateral. The definition of outliers in PCS is new as it is different from the currently most popular definitions based on dissimilarity. For verifying the outliers thus obtained, random-centroid optimization (RCO) was applied for selecting the best samples by each cluster of panelists. This combination of PCS/RCO may be useful in finding the likeness distribution among consumers and then in creating food products to correctly respond to the demands of different consumer groups.
ACKNOWLEDGMENTS
A technical assistance of Ms. Tomoe Fukushima is greatly appreciated. This work was financially supported by a grant from the Natural Sciences and Engineering Council of Canada.