Assessing the predictive performance of the Bagging algorithm for genomic selection

The aim of the present study was to compare the predictive performance of the Bagging algorithm with other decision tree-based methods, including regression tree (RT), random forest (RF) and Boosting in genomic selection. A genome including ten chromosomes for 1,000 individuals on which 10,000 single nucleotide polymorphisms (SNP) were evenly distributed was simulated. QTL effects were assigned to 10% of the polymorphic SNPs, with effects sampled from a gamma distribution. Predictive performance measures including accuracy of prediction, reliability and bias were used to compare the methods. Computing time and memory requirements of the studied methods were also measured. In all methods studied, the accuracy of genomic evaluation increased following increase in the heritability level from 0.10 to 0.50. While RT was the most efficient user of time and memory, it was not recommended for genomic selection due to its poor predictive performance. The obtained results showed that the predictive performance of Bagging was equal to RF and higher than RT and Boosting. However, it required significantly higher computational time and memory requirements. Considering the overall performance, Bagging was recommended for genomic selection, especially when due to the size and structure of the genomic data, the use of RF is limited.

Assessing the predictive performance of the Bagging algorithm for genomic selection Keywords:

gamma distribution , genomic selection , heritability , regression tree , SNP