Prediction of ship risk by a monotonic decision tree

Ship inspections as part of the port state control (PSC) process can ensure that major international conventions and regulations are complied with by foreign visiting ships. Due to the scarcity of inspection resources and concerns over prolonged inspection time, accurate identification of ships with higher risk is necessary for PSC. While previous studies have developed data-driven models to predict vessel's risk profile, domain knowledge is not adequately integrated into existing models. The gap can challenge the model's performance, as well as the trustworthiness, which can subsequently affect industry adoption. To bridge the knowledge gap, this study develops a monotonic regression decision tree model to predict ships’ risk profiles. The monotonicity is realized by first constructing a normal regression decision tree. Then, the outputs of the tree are revised by an optimization model whose objective is to minimize the prediction error with monotonicity constraints to guarantee that the outputs follow domain knowledge while retaining the tree structure. Real inspection records at the Port of Hong Kong are used to validate model performance in terms of monotonicity and accuracy. In addition to the enhanced interpretability and trustworthiness from monotonicity, improvement on accuracy performance is also observed on the proposed model. Moreover, the proposed model is applicable to a wide range of regression problems, such as shipping emission prediction, where monotonicity constraints shall be applied.