Comparing regression, naive Bayes, and random forest methods in the prediction of individual survival to second lactation in Holstein cattle

摘要

In this study, we compared multiple logistic regression,a linear method, to naive Bayes and random forest,2 nonlinear machine-learning methods. We usedall 3 methods to predict individual survival to secondlactation in dairy heifers. The data set used for predictioncontained 6,847 heifers born between January2012 and June 2013, and had known survival outcomes.Each animal had 50 genomic estimated breeding valuesavailable at birth and up to 65 phenotypic variablesthat accumulated over time. Survival was predicted at5 moments in life: at birth, at 18 mo, at first calving, at6 wk after first calving, and at 200 d after first calving.The data sets were randomly split into 70% trainingand 30% testing sets to evaluate model performancefor 20-fold validation. The methods were compared foraccuracy, sensitivity, specificity, area under the curve(AUC) value, contrasts between groups for the predictionoutcomes, and increase in surviving animals in apractical scenario. At birth and 18 mo, all methodshad overlapping performance; no method significantlyoutperformed the other. At first calving, 6 wk afterfirst calving, and 200 d after first calving, random forestand naive Bayes had overlapping performance, andboth machine-learning methods outperformed multiplelogistic regression. Overall, naive Bayes has the highestaverage AUC at all decision points up to 200 d afterfirst calving. Random forest had the highest AUC at200 d after first calving. All methods obtained similarincreases in survival in the practical scenario. Despitethis, the methods appeared to predict the survival ofindividual heifers differently. All methods improvedover time, but the changes in mean model outcomesfor surviving and non-surviving animals differed bymethod. Furthermore, the correlations of individualpredictions between methods ranged from r = 0.417 tor = 0.700; the lowest correlations were at first calvingfor all methods. In short, all 3 methods were able topredict survival at a population level, because all methodsimproved survival in a practical scenario. However,depending on the method used, predictions for individualanimals were quite different between methods.