Comparison of machine learning methods for predicting genomic breeding values for growth traits in Braunvieh cattle

Abstract

Machine Learning (ML) algorithms have proven advantageous in addressing challenges associated with the quantity and complexity of information, discovering patterns, performing efficient analyses, and serving as a decision-making tool. The objective of this study was to compare four ML methods —artificial neural networks (NN), regression trees (RT), random forests (RF), and support vector machines (SVM)— for predicting genomic value in European Swiss cattle using phenotypic records of birth weight (BW), weaning weight (WW) and yearling weight (YW), as well as genomic information. The results indicate that the predictive ability of the models varies according to the features and the amount of information available. NN, RF, and SVM exhibited similar performances, while RT underperformed. The SVM methodology stood out as the tool with the greatest potential, achieving the highest values of Pearson correlation between corrected phenotypes and predicted genetic values for WW. Despite its higher computational cost, the NN performed reasonably well, especially for BW and YW. The selection of the final model depends on the specific requirements of the application, as well as on such practical factors as data availability, computational resources, and interpretability; however, in general, the NN and SVM emerged as solid choices in several categories.