Interpretable machine learning for yak milk feeding pattern discrimination: Integrating XGBoost with multidimensional explainability analysis

Accurately identifying grazing (GZ) and supplementary feeding (SF) patterns in yak milk is important for product authentication; however, current methodologies are often expensive and time consuming. In this study, we examined 523 milk samples of lactating yaks at four stages of SF and tested 21 machine learning algorithms to develop a rapid and cost-effective classification method using routine compositional parameters. Ensemble learning techniques performed better than others, with XGBoost having the best accuracy (92%) and AUC (0.94). Multidimensional interpretability analyses, including SHAP, PDP, and ICE, identified fat content (27.8%) and lactose (23.1%) as the most important discriminators, along with biologically meaningful interactions, such as between fat, lactose, and freezing point. This interpretable framework provides a practical, low-cost method for milk authentication of yaks using ordinary dairy analyzers, providing a methodological foundation for the establishment of standardized GZ certification systems in milk production from yaks. • XGBoost achieved 92% accuracy for high-precision yak milk feeding pattern detection. • SHAP, PDP, and ICE methods explained XGBoost model decision mechanisms. • Routine milk composition data enables cost-effective feeding pattern classification. • SHAP revealed fat-freezing point and lactose-freezing point interaction effects.