Discovering aroma patterns in food products using Latent Dirichlet Allocation and Jensen Shannon divergence

Aroma Extract Dilution Analysis (AEDA) evaluates volatile compounds most likely contributing to the overall aroma of a food sample by means of flavour dilution (FD) factors. In the food industry, this can be useful to compare aroma-active profiles of raw materials or finished products and to select those that are statistically similar. When multiple samples are analysed, the high number of variables makes it difficult to take conclusions. Principal Component Analysis (PCA) should not be applied to FD values as they are discrete numbers. To our knowledge, there are no appropriate methods available to interpret AEDA results from multiple samples. In this study, a new rapid methodology to interpret AEDA results was developed. Latent Dirichlet Allocation (LDA) was developed in the context of text analysis as a mean of dimensionality reduction and has been successfully applied for the analysis of AEDA outcomes. Furthermore, Jensen Shannon divergence measure was a useful tool to compare the distribution of volatile compounds with similar descriptions ("berries", "cheese" or "fruits") among different samples.