Live Weight Prediction in Norduz Sheep Using Machine Learning Algorithms

Abstract

The objective of this study was to compare predictive performances of four machine learning (ML)\rmodels: Support Vector Machines with Radial Basis Function Kernel (SVMR), Classification and\rRegression Trees (CART), Random Forest (RF) and Model Average Neural Networks (MANN) to\rpredict live weight from morphological measurements of Norduz sheep (n=93). Seven\rmorphological measurements; chest girth (CG), chest width (CW), chest depth (CD), height at\rwithers (HW), body length (BL), heigth at rump (HR) and rump width (RW) were used to predict\rlive weigth (LW) of Norduz sheep. All morphological measurements were positively correlated to\rLW. Live weight had the highest correlation with CG and the lowest correlation with HR. Initially,\rhighly correlated predictors were removed from the data set. The remaining predictors were then\rsubjected to variable selection procedures using the Boruta algorithm. The results of Boruta\rconfirmed the importance of the four predictors HW, BL, CW, and CD. However, HR confirmed\rto be unimportant was excluded from the dataset. The ML models were trained on selected\rpredictors. The results showed that the prediction performance validated using the test dataset\rindicated that RF had the lowest values of Mean Absolute Error (MAE), Root Mean Squared Error\r(RMSE), and Mean Absolute Percent Error (MAPE). The permutation-based variable importance\rscores indicate that CW and CD were the most important variables in predicting LW. The actual\rLW had the highest significant positive correlations with the values predicted by SVMR and RF,\rand followed by ANN and CART models respectively. There were no differences between the\rmeans of actual and predicted LWs by machine learning models. The fact that the models\rgeneralized well on the testing data sets indicates that machine learning algorithms have valid\rpredictive patterns and are effective methods in LW weight of Norduz sheep. Considering runtime\rof the models, although the CART model had the lowest computational cost, it had the worst\rperformance. The MANN algorithm, on the other hand, required a longer runtime to process the\rsame dataset.