Enhancing food collection inspection efficiency using a Bayesian network model

Ensuring food safety requires efficient and targeted inspection strategies. To this end, a Bayesian network model that prioritizes food products with a high likelihood of noncompliance was developed, and its predictive performance was subsequently evaluated. To construct the model, variables related to noncompliance in food collection inspections were selected. A total of 2523 records from the Integrated Food Safety Information Network (IFSIN) in Korea were used to train a data-driven Tree-Augmented Naive Bayes (TAN) Bayesian network model, with compliance set as the root node. The selected variables contributing to noncompliance status included manufacturer-related characteristics, such as past noncompliance history, number of employees, annual sales, and annual exports, as well as product-related factors, such as vulnerability to noncompliance by food type, annual product sales, and HACCP certification status. When the TAN Bayesian network model was applied to 1081 test samples excluded from training, the decision threshold was adjusted to enhance predictive performance and increase the likelihood of selecting noncompliant products for inspection. At a threshold of 0.021, the recall reached 0.7667, and the likelihood of actual noncompliance among inspected products was 9.83 %—approximately 3.5 times higher than the baseline noncompliance rate of 2.75 %. The number of items to inspect can be determined based on this threshold, which may be adjusted according to available resources such as budget and manpower. The results indicate that inspecting only 21.7 % of all items can identify 76.67 % of noncompliant products, demonstrating a substantial improvement in inspection efficiency.