FEATURE-SELECTION AND VALIDATION OF SIMCA MODELS - A CASE-STUDY WITH A TYPICAL ITALIAN CHEESE

A strategy for building class models by means of SIMCA (soft independent modelling of class analogy) is suggested. to be applied in the case of a small number of objects and a large number of variables. The strategy uses both the customary procedure, based on the selection of variables with both high modelling and discrimination powers, and a novel procedure. Here Monte Carlo simulations are used to obtain the significance level of the experimental Fisher weights. so that only the relevant variables are selected. avoiding the use of noisy information. The validation of SIMCA models is performed by means of a leave-one-out procedure: many validation parameters are suggested to evaluate the accuracy of the models obtained. Data on a typical Italian cheese have been used to show the feature selection and the validation procedures. The significance of the validation parameters has been tested by comparing the results of the 'cheese categories' with those obtained from artificial and real data sets (the variety Versicolor of the iris flower, and categories of typical wines and olive oils). The models computed for the typical cheese are shown to be reliable.