KNLR: A heterogeneous ensemble learner for predicting Foie gras weight grade in mule ducks (Anas platyrhynchos × Cairina moschata)

Ante-mortem prediction of foie gras weight grade remains an unsolved challenge in commercial duck production. We developed KNLR, a novel heterogeneous ensemble learner that accurately predicts foie gras weight classification in mule ducks using pre-overfeeding morphometric measurements and post-overfeeding live weight, enabling producers to optimize feeding strategies and improve grading consistency. KNLR integrates Heterogeneous Ensemble Feature Selection (HEFS) with Weighted Area Under Curve Evaluation (WAUCE) to enhance predictive robustness. Comparative evaluation with four base learners (LightGBM, Naïve Bayes, Random Forest, and K-Nearest Neighbors) indicated that KNLR achieved the best overall performance across multiple machine-learning and statistical metrics. Using six features, KNLR achieved the highest precision (0.6425 ± 0.0869), significantly outperforming all base learners. Feature importance analysis indicated that overfeeding liver weight, breast depth, and body slope length were the most important predictors of foie gras grade. The proposed heterogeneous ensemble model may allow early identification of mule ducks with high-quality livers, providing a basis for precision feeding strategies aimed at optimizing feed efficiency and foie gras quality. By supporting grade-specific feeding management during the overfeeding period, KNLR offers a data-driven approach for breeding enterprises to potentially reduce production costs and improve economic returns through more accurate liver grade prediction.