Explainable AI and Voting Ensemble Model to Predict the Results of Seafood Product Importation Inspections

Background: As the volume of imported food flowing into South Korea rapidly increases due to the expansion of free trade agreements, improving inspection efficiency through artificial intelligence technology emerges as a critical task, particularly as time and cost expenditures for safety inspections conducted by the Korean Ministry of Food and Drug Safety concurrently increase rapidly. The lack of a generalizable machine learning model for predicting the safety of food for human consumption constitutes a significant challenge for policymakers and responsible authorities. Methods: This study developed an effective classification model for predicting non-conformance in customs inspection of imported seafood products. To address the severe class imbalance inherent in the inspection data, we applied class weight-based cost-sensitive learning and adopted an ensemble approach combining Decision Trees (DT), Random Forests (RF), Logistic Regression (LR), and Naive Bayes (NB) models. Results: Performance evaluation demonstrated that the soft voting ensemble technique achieved superior predictive performance in identifying non-conformance cases, with a recall of 75.57% and an Area Under the Curve (AUC) of 87.49%, significantly outperforming the hard voting method’s recall of 44.32% and AUC of 72.07%. Through SHapley Additive exPlanations (SHAP) analysis, we confirmed that various characteristics, including exporting country ratio, major product category, overseas manufacturer ratio, importer ratio, and seasonal variation, exerted substantial influence on the models’ decisions. Conclusion: Notably, the Naive Bayes model component provided a more comprehensive analysis for identifying non-conformance by considering multiple dimensions and potential seasonality. This research guide for predicting seafood product import inspection results contributes to enhancing inspection efficiency for securing the safety of imported aquatic products.The proposed methodology demonstrates potential applicability to other regulatory inspection domains confronting similar data imbalance challenges.