Targeted prediction of sensory preference for fermented pomegranate juice based on machine learning

Abstract

The fermented pomegranate juice (FPJs) has become a promising alternative to satisfy the growing demands for pomegranate products, however the development of new products depending on sensory evaluation is time-consuming and labor-intensive. In this study, 90 FPJs were prepared to predict the key physiochemical features that affecting sensory preferences by machine learning (ML). The GradientBoosting algorithm performed the best among 9 ML prediction models with good robustness and predictive ability on the validation set, meanwhile weighted preferences score (WPS) model predicted sensory preference more accurately as compared to comprehensive preferences score (CPS) model. Based on SHapley Additive exPlanations (SHAP) analysis, TSS, CD and LAB were the top 3 key features affecting preferences scores in both CPS and WPS models, but ranked differently. ML and its interpretability provided a new method and theoretical basis for development of consumer-preferred products and improvement of food sensory quality in food industry.