کنفرانس بین المللی فناوری اطلاعات و دانش

صفحه اصلی / شانزدهمین کنفرانس بین المللی فناوری اطلاعات و دانش

Improving Transition Cow Index Accuracy through CatBoost-Based Prediction of First Test-Day Milk Yield

نویسندگان :

Hoda Safaeipour¹ Sepehr Ebadi²

1- دانشگاه صنعتی اصفهان 2- دانشگاه صنعتی اصفهان

کلمات کلیدی :

Transition period،machine learning،Transition Cow Index (TCI)،dairy herd management،neural networks،milk yield prediction

چکیده :

Abstract— The transition period in dairy cows, encompassing three weeks pre- and post-calving, represents a critical physiological phase that significantly impacts subsequent milk production and overall herd health. Effective herd management during this period is indirectly assessed via the Transition Cow Index (TCI), which quantifies the deviation between predicted and actual first test-day milk yield. Traditionally, TCI prediction has relied on linear or heuristic statistical methods with limited accuracy and generalizability. In recent years, machine learning (ML) approaches have emerged as powerful alternatives, offering improved precision and robustness in complex agricultural decision-making contexts. This study developed and evaluated ML-based predictive models for first test-day milk yield in subsequent lactations, thereby enabling more reliable TCI computation. A comprehensive dataset from the Vahdat Cooperative Company, Isfahan Province, Iran, comprising 345,676 cow records across 99 herds collected from 2011 to 2022, was utilized. Various ML families—including regression-based models, tree-based ensembles, kernel methods, and neural networks—were comparatively tested, and the CatBoost Tuned model was identified as the best-performing approach. The proposed method demonstrated notable gains in predictive accuracy. Compared with the cooperative’s baseline model (R² ≈ 0.30), the CatBoost Tuned model improved the explained variance to 0.40 and reduced mean absolute error by nearly 10%, from above 7 kg to 6.4 kg per cow. Importantly, when aggregated at the herd level, errors were reduced to below 1.0 kg and R² exceeded 0.86, underscoring the practical utility of the ML-based framework for large-scale TCI benchmarking and herd management optimization.