Evaluating the prediction performances of artificial neural network, nearest neighbor, and CART algorithms for body weight in Sujiang pigs using morphological measurements

Abstract

Abstract The objective of this study was to evaluate machine learning algorithms for predicting body weight in Sujiang pigs. Sujiang pigs originated from the Duroc and Jiangquhai blood lines to improve both the growth rate and lean percentage of native breeds. K nearest neighbor, decision tree (CART), and artificial neural network algorithms were used to predict body weight (BW) using morphological traits such as body length (BL), body height (BH), chest circumference (CC), hip width (HW), and backfat thickness (BFT). The age of the pigs (180±5) was also included as a nominal predictor. For this purpose, all morphological measurements taken from 365 Sujiang pigs in a previous study were used. In total, 219 individual measurements were obtained after data preprocessing. In the Sujiang pig dataset, BW had a significantly positive and high linear relationship with BH (0.66), BL (0.72), CW (0.81), HW (0.84), and CC (0.88) (p < 0.01). Overall, the ANN algorithm outperformed the KNN and DT algorithms in this pig dataset according to the goodness of fit criteria of R2 = 0.91 and RMSE = 3.1. Nevertheless, the KNN algorithm also demonstrated good predictions on the test dataset (R2 = 0.86 and RMSE = 3.57). In the ANN algorithm, several training algorithms were compared, such as the Levenberg‒Marquardt algorithm, Bayesian regularization, and scaled conjugate gradient. In addition, the number of neurons in the hidden layer was manipulated to 2, 3, or 4. All training algorithms yielded similar results. However, when the predictor variables were CC, HW, BL, and BH, the Levenberg–Marquardt network had a superior ability to predict body weight in Sujiang pigs, with R2 = 0.89 and RMSE = 3.05. Furthermore, when BL measurements were not included in the model, the model’s predictive ability decreased by approximately 6%. According to the results, the use of Levenberg‒Marquardt and Bayesian Regularization in the ANN algorithm could help to improve breeding strategies. The traits determined to be the best predictors of BW in Sujiang pigs via the ANN algorithm can be used as indirect selection criteria in the future. However, this study suggested that different age stages, breeds, and traits should be considered in the model to accurately predict BW.