ENSEMBLE AND HYBRID LEARNING FOR DIGITAL TWIN MILK CLASSIFICATION

Abstract

Сhanges in milk composition and quality affect consumer health, production efficiency, and supply chain safety, which is why fast and reliable assessment is essential. Lately, there’s increasing interest in digital models for food systems. However, the number of studies comparing ensemble, hybrid, and classical machine learning models to directly classify milk quality is still limited. The aim of this study was to evaluate and compare the performance of classical, ensemble, and hybrid machine learning models for milk quality classification within a Digital Twin framework. The Kaggle Milk Quality dataset (1,059 samples, seven physicochemical attributes) was used. Six conventional algorithms, three ensemble methods (Random Forest, XGBoost, LightGBM), and three hybrid models (RF+KNN, XGBoost+RF, and Stacking) were assessed using stratified train–test splitting and five-fold cross-validation. The results showed that ensemble and hybrid methods consistently outperformed traditional machine learning methods. The Stacking ensemble achieved the highest results, 100.0%, Accuracy, Precision, Recall, F1-score, and ROC-AUC. Among the high-performing models, RF+KNN showed the lowest prediction latency (1.50 ms for 1,000 samples). The findings indicate that ensemble and hybrid machine learning models are well suited for Digital Twin–based real-time milk quality monitoring and can support intelligent decision-making in smart dairy production.