Feature selection-driven machine learning for parallel species authentication and adulteration quantification in commercial dairy products via Raman fingerprints

The rising demand for premium non-cow dairy products has intensified the need for robust adulteration detection methods to ensure product authenticity and consumer safety. This study introduces a feature selection-driven machine learning (ML) framework integrating Raman spectroscopy for parallel species authentication and adulteration quantification in commercial dairy products. Raman fingerprints of milk and yogurt samples from cow, buffalo, yak, and goat were analyzed to identify matrix-agnostic biomarkers. Feature selection via Chi-square analysis enabled the identification of nine discriminative features, including four conserved biomarkers capable of effectively differentiating species origins across product matrices. Various ML classifiers were evaluated for their ability to classify dairy products, with Random Forest Classifier demonstrating superior accuracy in species identification (>96 % accuracy). The analysis also uncovered conserved features that serve as robust indicators of species origin and fermentation processes. Furthermore, regression models successfully quantified cow milk adulteration levels in buffalo/goat milk, with Gaussian-weighted k -nearest neighbor regression demonstrating superior performance (test set R 2 = 0.950/0.918, RMSE = 0.076/0.100). This work presents a scalable Raman-ML pipeline for dairy authentication, addressing critical gaps in cross-matrix generalizability and providing a rapid, non-destructive solution for industrial quality control.