Machine learning based classification of yogurt aroma types with flavoromics

Abstract

Traditional sensory evaluation, relying on human assessors, is vulnerable to subjective error and lacks automation. Nonetheless, the complexity of human sensation makes it challenging to develop a computational method in place of human sensory evaluation. To tackle this challenge, this study constructed logistic regression classification models that could predict yogurt aroma types based on aroma-active compound concentrations with high classification accuracy (AUC ROC > 0.8). Furthermore, indicator compounds discovered from feature importance analysis of classification models led to the derivation of classification criteria of yogurt aroma types. Through constructing and analyzing machine learning models on yogurt aroma types, this study provides an automated pipeline to monitor sensory properties of yogurts.