Modeling Sensory Analysis Datasets: The Case of Italian Cheeses

Sensory characteristics as measured by Sensory Analysis (SA) play a key role in the definition of typicality markers for food products. These markers can be use, for example, to support a PDO (Protected Designation of Origin) or to implement quality control strategies. SA data, however, is characterised by high noise, poor data/feature ratio and high dimensionality, which poses a big challenge for classical statistical methods. We apply Random Forest (RF) and Support Vector Machines (SVM) as innovative modeling techniques in the field of sensory analysis, and compare them with the more classical Linear Discriminant Analysis (LDA). We apply these techniques to model a SA dataset from the analysis of 6 typical cheeses of the Trentino province (North Italy). Both new methods clearly improve the results obtained with standard LDA, and SVM also outperforms RF in this application. We also evaluate the relative importance of the diverse sensory attributes through the use of a new feature selection setup, adapted from a recent high-throughput molecular profiling methodology. Using the ‘out-of-bag’ error estimates provided by RF, we highlight the most informative variables that are effective for an optimum classification of the products while controlling the risk of overfitting. Our results indicate that the proposed selection method is efficient and robust in identifying informative parameters for product classification.