Comparative evaluation of optimized Bayesian versus improved machine learning methods for genomic selection of milk production traits

Abstract

This study aimed to enhance the genomic selection (GS) for milk production traits in Holstein cows within a digital breeding framework. We evaluated three optimized Bayesian (BB-opt, BL-D, BRR-G) and three improved machine learning (GB-Ens, RF-S, RKHS) methods using multiple cross-validation strategies (3FCV, 5FCV, 10FCV, 20FCV) on three traits: milk fat percentage, milk yield, and somatic cell score. Results under the optimal 20FCV strategy showed that BB-opt achieved the highest accuracy for milk fat percentage (0.885±0.004), BL-D for milk yield (0.808±0.007), and RKHS for somatic cell score (0.778±0.006). Model unbiasedness followed the same order. Machine learning methods (GB-Ens, RKHS) were significantly faster (145-175 min) and had higher digital fitness scores (85±2) than Bayesian methods (292-715 min, 58-62 points). The study confirms the trait-specific advantages of BB-opt, BL-D, and RKHS, and highlights GB-Ens and RKHS as suitable for real-time digital breeding platforms due to their efficiency and compatibility.