A comparative evaluation of machine learning algorithms for predicting chicken egg weight from egg quality traits

This study compared the effectiveness of artificial neural network (ANN), K-Nearest Neighbors (KNN), Gradient Boost regression (GBoost), random forest regression (RFR), support vector regression (SVR), and classification and regression trees (CART) algorithms for developing predictive models to estimate the egg weight of Bovan brown chickens in Ethiopia based on egg quality traits. Data were collected from 600 consumption eggs at the Haramaya University poultry farm, with 300 eggs from each of the cage and deep litter production systems. The dataset included egg weight (EW) and ten egg quality metrics. Descriptive statistics revealed the strongest correlations with egg weight were albumen weight (AW) and egg length (EL), with correlation coefficients (r) of 0.86 and 0.64, respectively. Performance among the models varied, with GBoost achieving the highest predictive accuracy. It yielded the highest coefficient of determination (R² = 0.960), the lowest root means square error (1.112), Akaike information criterion (AIC = 1834.394), and Bayesian information criterion (BIC = 1847.544), as well as the highest correlation (0.98) between observed and predicted values. In contrast, the CART model demonstrated the weakest overall performance. A relative importance analysis identified AW as the most significant predictor across all models, accounting for 96.28 % in CART, 80.56 % in GBoost, 72.69 % in RFR, 30.6 % in ANN, 29.9 % in SVR, and 29.1 % in KNN. The results indicate that the GBoost algorithm is a reliable and superior method for predicting egg weight. This study suggests that, with the application of this model, egg producers can effectively anticipate egg weight based on key egg quality characteristics.