Combining class-modelling and discriminant methods for improvement of products authentication

Abstract

Class- and discriminant modelling are two types of classification tools applied to construct models used for prediction of belongingness of samples to the classes studied. Although class- and discriminant methods have similar goals, the areas of their applicability are different. A class-model is constructed individually for each of the classes studied, based on the similarities among samples from the same class. A classical discriminant model is constructed based on differences among classes studied, and a new sample is always assigned to one of these classes. This characteristic differentiates the discriminant approach from class-modelling, where a new sample can be assigned to none of the classes studied. Due to this property, the class-modelling approach is widely applied for authentication purposes. Yet, if more classes similar to one another are authenticated, individual class-models constructed for these classes can lead to poor classification results. In such cases, the discriminant model can provide a better classification outcome since it takes advantage of differences between the classes. However, the discriminant approach alone is inappropriate for authentication, thus a possible solution is to use a method that benefits from both class- and discriminant modelling. In this study, several methods that combine class-modelling with a discriminant approach were tested, i.e., two-step approach, two soft discriminant methods based on PLS-DA (Partial Least Squares Discriminant Analysis), and ROC (Receiver Operating Characteristics) curve-based SIMCA (Soft Independent Modelling of Class Analogy). The methods were compared with respect to their pros, cons and the scope of their applicability for the example of authentication of three Cyclopia species. They are used for the production of honeybush tea (protected in the European Union as a Geographical Indication (GI)). Moreover, several different authentication scenarios were considered to test the methods thoroughly. It was revealed that the two-step approach, soft discriminant method by Calvini et al. and ROC curve-based SIMCA led to the most efficient models.