Advancing Milk Quality Prediction: Dynamic Stacking Ensemble with Quality-Aware Adaptive Sampling

Abstract

Predicting milk quality is essential for the dairy industry, as it directly impacts production efficiency, safety standards, and consumer satisfaction. This study presents a novel hybrid machine learning framework that combines a Modified Stacking Ensemble (MSE) with Quality-Aware Adaptive Sampling (QA-ADASYN) to enhance predictive performance and address class imbalance. The MSE integrates diverse base learners through a dynamically weighted meta-model informed by feature importance. At the same time, QA-ADASYN generates synthetic samples near decision boundaries of underrepresented classes, such as high-grade milk. To enhance model transparency and support explainable AI, we apply SHAP (SHapley Additive Explanations) to interpret feature contributions. The SHAP analysis confirms that features like pH and temperature are dominant predictors of milk quality. Furthermore, we conduct 5-fold cross-validation, yielding an average accuracy of 99.6% ± 0.05%, confirming the model's robustness and generalizability. Our framework achieves 100% classification accuracy on the validation set and 98.9% accuracy on an external noisy dataset, demonstrating both precision and resilience. This interpretable and scalable solution sets a new benchmark for intelligent dairy quality control, with strong potential for real-time deployment and sensor integration.