Geographical origin classification of milk using six-year stable isotope and multi-element fingerprints with SHAP-based model interpretation.

Accurately verifying the geographical origin of milk is increasingly crucial in global dairy markets. This study analyzed 1552 cow milk samples across six years collected from nine origins: Taiwan, Japan, New Zealand, Australia, United States, France, Poland, Denmark and Germany. In total, 24 elemental concentrations and two stable isotope ratios were measured to construct origin classification models using three machine learning algorithms, of which the random forest model achieved the best performance, including 98.36% accuracy for discriminating milk samples of Taiwan from all other origins. Moreover, SHAP-based variable importance analysis identified Tl, Ba, Rb, δ15N, Sr and δ13C as key markers for milk origin authentication. The six-year observation reveals the significant differences of elemental levels and isotopic profiles between particular years. The integration of multi-element and stable isotope profiling with explainable machine learning offers a practical framework for long-term traceability, regulatory monitoring and the protection of geographically labeled milk products.