Predicting food safety problems among catering service operators using longitudinal regulatory data.

Full-coverage on-site inspection of catering service operators is difficult to sustain when regulated entities are numerous and risk is unevenly distributed. Using administrative regulatory data from Shanghai, China (2020-2024), we constructed a longitudinal dataset comprising 46,614 catering service operators and 150,467 inspection intervals. Machine-learning models were developed to predict whether at least one problem would be detected at the subsequent inspection, and model predictions were interpreted using SHAP. We further examined cross-period persistence and transition among categories. The selected primary model, LightGBM, achieved an AUC of 0.866, average precision of 0.806, and F1 score of 0.761 on the test set, with the highest recall among the individual models (0.805). SHAP analysis indicated that current-inspection findings and cumulative inspection history were the dominant predictors, with complaint and whistleblower records providing complementary information. Cross-period transition analysis revealed that recurrence patterns differed by problem category: environmental hygiene and raw material control problems recurred frequently but with low category specificity, whereas qualification and labeling compliance problems were rarer but showed stronger self-persistence. These findings demonstrate that routine regulatory records can support finer stratification of operators according to their probability of inspection-detected problems and more targeted follow-up by distinguishing operators with general recurrent problems from those with persistent category-specific deficiencies.