Integrating multivariate statistics and interpretable machine learning for the quantitative profiling of tea...

The accurate classification and quality control of the six tea categories remain critical challenges in tea science. The complex transformation of tea polyphenols and catechins during processing makes traditional chemical methods inefficient for classification. 1381 tea samples were analyzed to quantify 13 key variables including tea polyphenols, ester-type catechins, non-ester-type catechins, and critical ratios. We observed a gradient decrease in ester-type catechins correlated with fermentation intensity. Multivariate statistics revealed fermentation driven gradients in the composition of tea polyphenols and catechins. Yellow tea and white tea were relatively poorly discriminated in the PLS-DA model, with area under the curve value (AUC) values of 0.802 and 0.879, respectively. To enhance classification robustness, SMOTE oversampling (SMOTE), L1-based feature selection (L1), and eXtreme Gradient Boosting (XGBoost) modeling was integrated. The optimized SMOTE-L1-XGB model achieved in Macro Average metrics of Accuracy of 0.975, Precision of 0.878, Recall of 0.878, F1-score of 0.876, and AUC of 0.983 on the testing set. SHapley Additive exPlanations (SHAP) identified EGCG emerges as the most phenotypically significant contributor, with its effects showing sharp threshold dependence in green tea and black tea. The discriminative logic for white tea relies more on the ratio of the ratio of total catechins to tea polyphenols. ECG serves as a central regulator in yellow, oolong, and dark teas. This approach providing a robust, data driven strategy for precise tea authentication and quality control.