Applicability of machine learning methods for classifying lightweight pigs in commercial conditions

Area of Science:

• Animal Science
• Agricultural Economics
• Data Science

Background:

• Varying pig growth rates challenge all-in-all-out systems.
• Accurate identification of slow-growing pigs is crucial for industry efficiency.

Purpose of the Study:

• To evaluate statistical methods for classifying pigs at risk of growth retardation.
• To assess the performance of machine learning algorithms in different production stages.

Main Methods:

• Analysis of a large dataset (26,749 pigs) across three production stages.
• Comparison of ordinary least squares, decision tree, random forest, and generalized boosted regression models.
• Evaluation of classification performance using the area under the curve (AUC).

Main Results:

• Random forest and generalized boosted regression showed superior performance (AUC 0.772-0.861).
• Parametric linear models had acceptable performance (AUC 0.752-0.818).
• Decision trees were ineffective (AUC 0.608-0.726); birth weight and prior weight were key predictors.

Conclusions:

• Machine learning algorithms can accurately identify pigs at risk of growth retardation.
• These methods offer potential for improved decision-making and efficiency in pig production.
• Predictive factors for growth retardation vary depending on the production stage.