Abstract

Food and beverage authentication is the process by which food or beverages are verified as complying with their label descriptions (Winterhalter, 2007). A common way to deal with an authentication process is to measure attributes, such as, groups of chemical compounds on samples of food, and then use these as input for a classification method. In many applications there may be several types of measurable attributes. An important problem thus consists of determining which of these would provide the best information, in the sense of achieving the highest possible classification accuracy at low cost. We approach the problem under a decision theoretic strategy, by framing it as the selection of an optimal test (Geisser and Johnson, 1992) or as the optimal dichotomization of screening tests variables (Wang and Geisser, 2005), where the 'test' is defined through a classification model applied to different groups of chemical compounds. The proposed methodology is motivated by data consisting of measurements of 19 chemical compounds (Anthocyanins, Organic Acids and Flavonols) on samples of Chilean red wines. The main goal is to determine the combination of chemical compounds that provides the best information for authentication of wine varieties, considering the losses associated to wrong decisions and the cost of the chemical analysis. The proposed methodology performs well on simulated data, where the best combination of responses is known beforehand.