Abstract
Traditional feeding strategy during lactation can result in nutrient deficiencies and negatively impact long-term productivity, compromising both the sustainability and profitability of the swine industry. Precision feeding, supported by decision-making systems built on advanced predictive models, offers a promising solution to address these challenges. This study aimed to develop prediction models for weaning performance, focusing on key indicators such as weaned litter weight (WLW), weaned litter size (WLS), dry matter in milk (DMm), and nitrogen in milk (Nm). The models integrate farm management practices and feed nutrient composition, providing a data-driven framework for optimizing performance. A total of 10,089 observations were collected from 17 trial pig farms across eight provinces in China. Eleven statistical and machine learning (ML) regression algorithms were employed, incorporating stratified sampling and the recursive feature elimination method for feature selection. The findings demonstrated that the ensemble learning models, specifically random forest and gradient boosting decision tree regression, delivered the best overall performance, with a coefficient of determination (R (2)) ranging from 0.40 to 0.80 and a mean absolute error (MAE) between 0.11 and 4.36. The shapley additive explanations (SHAP) heatmap used for feature importance analysis revealed that, although the key predictors of weaning performance varied across models, this study newly identified lactation duration, birth litter weight, parity, and backfat thickness on the 7th day of lactation (L.d7BF) as consistently important features across different models. The discrepancies between correlation analysis and feature importance suggest the presence of non-linear relationships, feature interactions, and multicollinearity within the dataset. This study presents a novel framework that provides valuable insights into the factors influencing weaning performance under diverse management practices and feed nutrient conditions. The optimized prediction model can be employed to guide real-time sensor-based precision feeding systems, thereby enhancing efficiency and sustainability in swine production.