Equitable port state control in maritime transportation: A data-driven optimization approach

Port state control (PSC) inspection, an important measure to prevent maritime accidents, prioritizes inspections for high-risk ships. The deficiency value of a ship is commonly used to quantify its potential risk, and the predicted deficiency value, derived from a prediction model trained through historical inspection records, is used to assist ship selection. Single prediction models are commonly used to estimate ships’ deficiency values. However, their performance can be affected by noise interference, which can decrease the stability and robustness of models. Ensemble learning is an effective method to enhance the prediction performance by combining results from multiple models. However, despite their advantages, current ensemble learning methods face challenges in effectively distinguishing ships with similar deficiency values, which can impact the effectiveness and fairness of the decision making. To address this issue, we propose a novel heterogeneous ensemble learning method that combines distributional estimation with lexicographic optimization to enhance ship selection performance. The model first constructs a comparison matrix between ships by distributional estimation. Subsequently, a lexicographic optimization model is applied to construct a discriminative weight vector based on the comparison matrix that assigns larger weights to higher-risk ships. Finally, we conduct a comprehensive evaluation of the effectiveness and fairness of the proposed method. We consider five base prediction models and select three widely used ensemble learning methods—averaging, stacking, and boosting—for further comparison. Experimental results from multiple test scenarios on real-world data from the port of Hong Kong demonstrate that the proposed method outperforms other methods in terms of both effectiveness and fairness.