Learning to predict pellet quality: a machine learning and feature engineering approach.

Area of Science:

• Agricultural Science
• Machine Learning
• Data Science

Background:

• Pellet quality is crucial for feed mills, impacting animal production efficiency and economics.
• Controlling pellet quality is challenging due to multiple influencing factors.
• Computational prediction models for pellet quality are needed to aid commercial settings.

Purpose of the Study:

• To enhance machine learning model predictive performance for pellet quality using feature engineering.
• To identify key individual and interacting factors influencing pellet quality.
• To develop a robust pipeline for handling large feature datasets in predictive modeling.

Main Methods:

• Implemented feature engineering to create 74 new features, including pairwise interactions.
• Utilized Recursive Feature Elimination (RFE) with 10 machine learning algorithms for optimal feature selection.
• Compared seven base models and two ensemble models, selecting non-overfitting models.
• Employed bootstrap resampling for robust evaluation on the testing set.

Main Results:

• The Support Vector Regression (SVR) model achieved the best performance on the testing set, with the highest Concordance Correlation Coefficient (CCC) of 0.608 and lowest Mean Absolute Error (MAE) of 1.868.
• Key predictive features identified include Dehydrated Bakery Meal (%), Ambient Temperature (°C), Crude Protein Content (%) × Starch Content (%), and Conditioning Temperature (°C) × ADF (%).
• SHAP values confirmed the importance of these features and their interactions across multiple models.

Conclusions:

• Feature engineering significantly enhances machine learning model performance for predicting pellet quality.
• The SVR model, utilizing a carefully selected feature subset, provides valuable insights into critical factors affecting pellet quality.
• The study presents an effective pipeline for predictive modeling with high-dimensional datasets in the feed industry.