Abstract
Honey variety authentication is critical for ensuring market integrity and protecting consumer rights, especially for high-value unifloral honeys, such as acacia honey, which are frequently adulterated with low-value alternatives such as rape honey due to their similar visual appearance. The aim of this study was to develop a method for precise discrimination between rape honey and acacia honey using their chemical profiles combined with machine learning. A total of 542 honey samples were collected from major beekeeping regions in China. Targeted quantification of 12 sugars and 20 amino acids was performed using UPLC-MS/MS. Multivariate analysis revealed significant differences in sugar and amino acid compositions between the two honey types, though partial samples overlapped due to chemical similarity. Six machine learning algorithms, including the Multilayer Perceptron, were employed for classification. Optimization was performed via 10-fold cross-validation and ADASYN oversampling, yielding optimal performance of 98% and 100% prediction accuracies for rape honey and acacia honey, respectively, on the independent test set. SHAP (Shapley Additive Explanations) analysis identified key differential markers, including fructose, turanose, glucose, and GABA, which contributed most to the classification. Furthermore, a user-friendly web application was developed to facilitate rapid on-site authentication. This study provides an innovative technical framework for honey variety discrimination, with potential applications in quality control and anti-fraud practices.