Comparison of data mining algorithms (Cart, Chaid and Mars) used to determine the effects of some characteristics on body weight in goats

The main aims of this study were to comparatively examine predictive performances of the CART, CHAID, Exhaustive CHAID and MARS data mining algorithms on the basis of a data set relevant to animal husbandry, and create R script files for estimating their goodness of fit criteria. To achieve this purpose, several morphological traits measured from 164 Teddy goats with the aim of predicting a continuous response variable (live body weight) were used as explanatory variables. To measure predictive performances of the examined algorithms, goodness of fit criteria i.e. Pearson correlation coefficient in the response variable between actual and predicted values (r), coefficient of determination (R2), adjusted coefficient of determination (Adj.R2), Akaike information criterion (AIC), the corrected Akaike information criterion (AICc), root-mean-square error (RMSE), mean error (ME), mean absolute deviation (MAD), standard deviation ratio (SDratio), global relative approximation error (RAE), coefficient of variation (CV%) and mean absolute percentage error, MAPE) were employed. To ensure the highest predictive accuracy of tree-based CART, CHAID and Exhaustive CHAID data mining algorithms, minimum animal numbers in parent and child nodes were set at 4:2. The highest predictive accuracy for the MARS algorithm was obtained with describing 4th interaction order and 45 terms. With the scope of predicting a quantitative trait (live body weight), the best performance was obtained with MARS algorithm, followed by the Exhaustive CHAID, CHAID and CART data mining algorithms. It was concluded in the light of the obtained results that use of MARS algorithm may be a better choice in both better describing the studied breed standards and determining significant predictors.