Ensemble Learning Based on Bagging and Hybrid Sampling for Food Safety Risk Prediction.

Area of Science:

• Food Science
• Machine Learning
• Data Science

Background:

• Food safety sampling faces challenges with rare high-risk samples, hindering accurate prediction in complex supply chains.
• Conventional machine learning models struggle with severe class imbalance, limiting effectiveness in food safety risk assessment.

Purpose of the Study:

• To develop and evaluate a unified Bagging-Stacking framework to overcome class imbalance for improved food safety risk prediction.
• To enhance the detection of minority high-risk samples within food safety inspection data.

Main Methods:

• Proposed a unified Bagging-Stacking framework integrating stacking ensembles, bagging, and SMOTE-Tomek resampling.
• Employed a stacking ensemble with five tree-based base learners and Logistic Regression meta-learner.
• Incorporated probability-based threshold optimization and SHAP analysis for interpretability.

Main Results:

• The framework significantly improved high-risk recall and precision, achieving the highest F1 score among compared models.
• Demonstrated robust and consistent generalization performance across varying test set proportions.
• SHAP analysis identified key risk predictors including storage conditions, production month, and shelf life.

Conclusions:

• The proposed Bagging-Stacking framework offers accurate, robust, and interpretable support for food safety risk prediction.
• Provides practical value for proactive risk prevention and efficient allocation of regulatory resources.
• Addresses limitations of conventional models in handling imbalanced food safety data.