A supervised machine learning approach for the decision-making process on data-based culling in dairy farms

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Dairy farm managers make culling decisions based on various factors.
• Predicting these decisions can optimize herd management and resource allocation.

Purpose of the Study:

• To develop a supervised machine learning (ML) approach for predicting dairy cow culling.
• To identify ML algorithms capable of learning farm-specific culling criteria.

Main Methods:

• Utilized farm data (milk yield, lactation number, etc.) from dairy cows in Mexico.
• Evaluated six ML algorithms (logistic regression, Gaussian naïve Bayes, k-NN, SVM, random forest, MLP) using binary classification.
• Employed hyperparameter optimization with tenfold stratified cross-validation to handle class imbalance.

Main Results:

• All classifiers achieved accuracy greater than 0.90.
• Random Forest, k-NN, SVM, and MLP demonstrated superior prediction performance (MCC > 0.91).
• Gaussian Naïve Bayes and Logistic Regression showed lower performance (MCC = 0.50 and 0.72, respectively).

Conclusions:

• Supervised ML algorithms can effectively learn and predict dairy cow culling criteria.
• Random Forest is the top-performing algorithm, with k-NN, SVM, and MLP as viable alternatives for on-farm use.
• Further validation across diverse farms and scenarios is recommended to enhance model reliability.