Milk Yield Prediction Using Animal and Production Features: Optimal Combinations and Machine Learning Approaches

Abstract

This study investigated the factors influencing milk yield in dairy cattle, the determination of optimal combinations, and the predictive performance of machine learning (ML) models. A total of 830 records collected in 2024 from a private dairy farm in Niğde, Turkey, were analyzed using breed, physiological status code, lactation number, age and days in milk as predictors. Statistical analyses confirmed that both breed and status code significantly affected milk yield (p < 0.05). Regression analysis indicated a positive effect of lactation number (β = 17.65, p < 0.001) and a negative effect of age (β = –0.80, p < 0.001). The highest yield (40.07 L/day) was recorded in Holstein Friesian cows with Cyclic status in the 3rd lactation. Six ML algorithms were compared using 5‑fold cross‑validation under two modelling scenarios designed to address multicollinearity between age and lactation number. Among the tested algorithms, Random Forest (R² ≈ 0.88, RMSE ≈ 5.2 L) and Gradient Boosting (R² ≈ 0.88, RMSE ≈ 5.2 L) provided markedly better predictive accuracy than linear regression models. SHAP (SHapley Additive exPlanations) analysis enhanced model interpretability and identified days in milk, physiological status and lactation number as the most influential predictors, followed by age and breed. These findings highlight the importance of integrating physiological and genetic factors in herd management and demonstrate the potential of ensemble ML techniques as reliable decision-support tools for optimizing milk yield in dairy farming.