Use of multivariate adaptive regression splines for animal husbandry

The aim of this study was to illustrate how to use MARS algorithm at a data set from Animal Science and to interpret the achieved results. Data of four sheep breeds (Balochi (48), Harnai (48), Beverigh (47) and Rakhashni (48)) reared in Pakistan were used in the prediction of body weight. Withers height, chest girth and body length as continuous predictors were used with the aim of introducing MARS data mining algorithm and developing the useful prediction equation for body weight. Breed was employed as a nominal predictor. With the package "earth" of R software, two MARS applications (26 and 38 term numbers) for the prediction of body weight were made by taking a basis for interaction degrees of 2, and v-ten-fold cross validation. For evaluating predictive accuracy of the MARS applications, goodness of fit criteria i.e. r (Pearson correlation coefficient between the actual and predicted values in body weight, GCV, RSS, GR2, R2, adjusted R2 and cross validation R2 were estimated. Goodness of fit criteria i.e. r (0.980 and 0.990), GCV(0.610 and 0.293), RSS(116 and 56), GR2 (0.960 and 0.980), R2 (0.960 and 0.980), adjusted R2 (0.954 and 0.975) and cross validation R2 (0.922 and 0.88) were estimated for both applications, respectively. Assignment of the suitable term number and interaction degrees producing the smallest GCV in R software is the most important points that are necessary to be taken into consideration for MARS algorithm. The achieved outcomes illustrated that MARS applications could be a notable option for proving breed standards and determining body measurements positively affecting the body weight of the sheep studied here.