Comparing Predictive Performances of Tree-Based Data Mining Algorithms and MARS Algorithm in the Prediction of Live Body Weight from Body Traits in Pakistan Goats

The main purpose of this investigation was to comparatively evaluate predictive performances of multivariate adaptive regression splines (MARS), chi-squared automatic interaction detector (CHAID), exhaustive CHAID and classification and regression trees (CART) data mining algorithms in predicting live body weight as a continuous response variable by means of morphological measurements i.e. live body weight (LBW), body length (BL), withers height (WH), rump height (RH), belly girth (BG) and chest girth (CG) as continuous predictors from 130 Pakistan goats. Also, sex factor was included as a possible nominal predictor in the current study. To measure predictive performances of the tested algorithms, model evaluation criteria such as the correlation coefficient between actual and predicted LBW values (r), Akaike's and corrected Akaike information criterion (AIC and AICc), root-mean-square error (RMSE), mean absolute deviation (MAD), standard deviation ratio (SDratio), and mean absolute percentage error (MAPE) were estimated. According to these criteria, MARS produced better predictive accuracy in explaining the variability in LBW compared with others. MARS produced the best fit for 3rd interaction order on the basis of the smallest generalized cross validation (GCV). In the MARS algorithm, BL and CG were the predictors that had the highest relative importance (100%) in the prediction of live body weight and these two predictors could be considered as indirect selection criteria for breeding schemes. It could be suggested that the CART, the CHAID, the Exhaustive CHAID and especially MARS algorithms in the prediction of live body weight were significant statistical tools in sophistically describing the studied breed standards for breeding purposes.