Learning to Predict Pellet Quality: A Machine Learning and Feature Engineering Approach.

Pellet quality is a major concern for feed mills, as low-quality pelleted feeds impact both mill and downstream animal production efficiency and economics. The pellet quality is influenced by multiple factors, which makes it difficult to control in a commercial setting. Therefore, computational predictions of pellet quality are desirable and have been explored by this group. This study extended previous work by exploring the application of feature engineering approaches to enhance the predictive performance of machine learning models and to gain insights into the importance of individual and groups of factors relevant to pellet mills. Feature engineering techniques were implemented to generate an additional 74 features, such as the pairwise interactions between Nutrient Composition features. A Recursive Feature Elimination approach was also implemented for 10 different ML algorithms to select optimal feature subsets. Non-overfitting models were selected, and model comparison was conducted for seven well-fitted base models and two ensemble learning models that combined the capabilities of the others. The SVR model outperformed all other models. While the SVR model had the second lowest MAE (1.814 ± 0.084) and the second highest CCC (0.594 ± 0.036) on the training set, it ultimately used fewer features (42 features) and had the highest CCC (0.608) and R (0.650), the lowest MAE (1.868), RMSPE (2.359) and MSPE (5.573) using a bootstrap resampling strategy on the testing set. Among all nine models, the top 10 features based on average absolute SHAP values largely aligned with feature importance rankings. Several features consistently emerged as key predictors across multiple models, including Dehydrated Bakery Meal (%), Ambient Temperature (°C), Crude Protein Content (%) × Starch Content (%) and Conditioning Temperature (°C) × ADF (%). While the SVR model did not significantly surpass previous studies in predictive capacity, it provided valuable insights into key predictive factors and their interactions for prediction of PDI. In addition, this study introduced a useful pipeline for handling datasets with large numbers of features for prediction, and highlighted opportunities for feature engineering.