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

Area of Science:

• Dairy Science
• Machine Learning Applications
• Animal Nutrition

Background:

• Accurate identification of yak milk feeding patterns (grazing vs. supplementary feeding) is crucial for product authentication.
• Existing methods for yak milk analysis are often costly and time-consuming.
• Developing rapid, cost-effective authentication techniques is essential for the dairy industry.

Purpose of the Study:

• To develop and validate a rapid, cost-effective machine learning-based method for classifying yak milk based on feeding patterns.
• To utilize routine compositional parameters for distinguishing between grazing (GZ) and supplementary feeding (SF) in yaks.
• To establish a foundation for standardized GZ certification systems in yak milk production.

Main Methods:

• Examined 523 lactating yak milk samples across four supplementary feeding stages.
• Tested 21 machine learning algorithms, focusing on ensemble techniques.
• Employed multidimensional interpretability analyses (SHAP, PDP, ICE) to identify key discriminators.

Main Results:

• XGBoost, an ensemble learning algorithm, achieved the highest accuracy (92%) and AUC (0.94) in classifying feeding patterns.
• Yak milk fat content (27.8%) and lactose (23.1%) were identified as the most significant discriminators.
• Biologically relevant interactions between fat, lactose, and freezing point were highlighted as important factors.

Conclusions:

• An interpretable, practical, and low-cost framework using ordinary dairy analyzers was developed for yak milk authentication.
• The study provides a methodological basis for establishing standardized grazing (GZ) certification systems.
• This approach enhances the feasibility of authenticating yak milk quality based on feeding practices.