Tuning and Application of Random Forest Algorithm in Genomic Evaluation

One of the most important issues in genomic selection is using a decent method for estimating marker effects and genomic evaluation. Recently, machine learning algorithms which are members of non-parametric and non-linear methods have been extended to genomic evaluation. One of these methods is Random Forest (RF) on which this research was focused. Important parameters in RF algorithm are the number of SNPs selected randomly at each tree node (mtry), the number of trees to grow` (ntree) and the minimum size of terminal nodes of trees (node size) which need to be pre-defined before analyses and for them the model should be tuned. A genome comprised of five chromosomes, one Morgan each, on which 10000 bi-allelic SNP were arrayed was simulated and the efficiency of different combinations of mtry, ntree and node size was tested and the best combination was selected based on comparison of accuracy of predicted genomic value as well as OOB error estimates. For the simulated data in the current study the least OOB error as well as the maximum prediction accuracy was related to a model with 6000 mtry, 1000 ntree and 5 node size. Other combinations did not increase the accuracy of prediction while led to an increase in time of analyses for those which used more trees. Since the accuracy of prediction is a function of mtry, ntree and node size, in genomic evaluation, different combinations of these parameters should be used and the combination which caused the maximum prediction accuracy should be used for genomic evaluation.