Predicting ship detention in port state control: An efficient machine learning approach with data balancing and explainable AI techniques

Port State Control (PSC) inspections play an important role in enhancing regulatory compliance and maintaining maritime safety. The increasing frequency of ship detentions indicates that substantial safety deficiencies continue to exist among arriving vessels, highlighting the need for more effective detention-risk identification strategies. However, detained ships account for less than 8% of PSC inspection records, leading to severe category bias that weakens the predictive power of machine learning models. To address this challenge, an integrated framework combining data balancing, machine learning, and explainable artificial intelligence techniques is developed using PSC inspection records from the Paris MoU covering the period 2015–2024. Four representative resampling methods are employed to mitigate class imbalance, while six machine learning algorithms are evaluated and optimized through Bayesian hyperparameter tuning. Subsequently, SHAP-based analysis is conducted to examine the influence of individual features on detention predictions and to improve model transparency. The results identify TomekLinks-XGBoost as the most effective modeling strategy, achieving an accuracy of 98.62% and an F1-score of 88.40%; the optimized framework improves accuracy by 1.09% and F1-score by 68.54%. The interpretability analysis further reveals that deficiency-related variables, particularly the total number of deficiencies, exert the strongest influence on detention outcomes. This study provides an effective predictive tool for identifying substandard ships and serves as a valuable reference for inspectors, port authorities, and shipping companies to enhance ship safety supervision and maritime safety.