Application of machine learning for computational gastronomy

Abstract

Using the FlavorDB database, this study creates a machine learning model for categorizing ingredients based on their molecular composition. The study aims to classify 936 food components into 35 flavor categories by evaluating their chemical structures. The methodology used several critical techniques, including data preparation with one-hot encoding of molecular makeup, class imbalance resolution with SMOTE (Synthetic Minority Over-sampling Technique), and data standardization with StandardScaler. Several machine learning methods were tested, including Random Forest, Extra Trees, Neural Networks, LightGBM, and others. The Random Forest Classifier outperformed expectations, with an F1-score of 81.02% and an accuracy of 79.52%. A stacking ensemble strategy was also investigated, which combined numerous base models with a Random Forest meta-model and achieved an average accuracy of 71.52%. Advanced analytical techniques that complement the primary classification include, Principal Component Analysis (PCA) to reduce dimensionality, SHAP (SHapley Additive Explanations) analysis is used to interpret feature importance and, the ROC-AUC curve analysis revealed a nearly flawless integrated AUC score of 0.99. The study emphasizes the potential of machine learning in computational gastronomy by providing a novel way to ingredient classification using molecular data analysis.