Prediction of breeding value of holstein offspring from parental ebu using a two-stage ensemble machine learning model

Abstract

Accelerating genetic progress in dairy cattle breeding requires methods for early prediction of animals’ breeding value. In practice, the mid-parent estimate is commonly used for this purpose, but it does not fully account for differences in parental reliability, inter-trait correlations, and parental compatibility. The aim of this study was to develop and validate a two-stage machine learning model for predicting breeding values of Holstein offspring from parental EBV. The analysis included data from 8,100 animals and 21 traits representing economic indices, milk production, functional, and conformation traits. At the first stage, an ensemble model combining XGBoost, random forest, Ridge regression, ElasticNet, and the classical mid-parent estimate was used; at the second stage, multi-trait residual correction was applied. To prevent information leakage caused by family structure, GroupKFold and sire-based data partitioning were implemented. The proposed approach improved prediction accuracy for 18 of 21 traits (85,7 %). The mean coefficient of determination (R2) increased from 0.6046 to 0.6083. The second stage showed the greatest effect for milk composition traits and the conformation trait “type”. The results confirm the potential of machine learning methods for breeding value evaluation and mate allocation in dairy cattle breeding.