Data analytics for improving shipping efficiency : models, methods, and applications

Abstract

This thesis aims to improve the effectiveness and efficiency of port state control (PSC) inspection, which is one of the most important international shipping policies, from the aspects of ship risk prediction and PSC inspector assignment and scheduling using data analytics and operations research models. In addition to a comprehensive summary and review of ship selection methods applied at ports over the world and proposed in existing literature, this thesis comprises three studies. In the first study, ship deficiency number, which is a ship risk indicator in the PSC inspection, is predicted using a state-of-the-art XGBoost model. The XGBoost model takes shipping domain knowledge regarding ship flag, recognized organization, and company performances into account to improve model accuracy and fairness. Based on the predictions, a PSC inspector scheduling model is proposed to help the ports optimally allocate inspection resources. According to the model structure, the concepts of inspection template and un-dominated inspection template are further proposed and incorporated in the optimization model to improve computation efficiency and model flexibility. In the second study, three two-step approaches that match the inspection resources with the ships' deficiency conditions are proposed, aimed at identifying the most deficiencies of them. The three approaches combine prediction models with optimization models, and the optimization models are equivalent in all the approaches while the prediction models differ from each other regarding their prediction targets or structure. Specifically, the first approach predicts the number of deficiencies in each deficiency category for each ship and then develops an integer optimization model that assigns the inspectors to the ships to be inspected. The second approach predicts the number of deficiencies each inspector can identify for each ship and then applies an integer optimization model to assign the inspectors to the ships to be inspected. The third approach is a semi-"smart predict then optimize" (semi-SPO) method. It also predicts the number of deficiencies each inspector can identify for each ship and uses the same integer optimization model as the second approach. However, instead of minimizing the mean squared error as in the second approach, it adopts a loss function motivated by the structure of the optimization problem in the second approach. The prediction results are then input to PSC officer (PSCO) assignment models such that the PSCOs' expertise and the ships' deficiency conditions can be matched, and the inspection efficiency can be improved. In the third study, a data-driven ship risk prediction framework using features the same as the current ship selection scheme is developed for high-risk ship identification and selection based on gradient boosting regression tree (GBRT). Like existing ship risk prediction models, the proposed framework is of black-box nature whose decision process and working mechanism are opaque. To improve model explainability, the explanation of the prediction of individual ships by the Shapley additive explanations (SHAP) method with the properties of local accuracy and consistency is provided. Furthermore, the local SHAP method is innovatively extended to a fully explainable near linear-form global surrogate model of the original black-box data-driven model by deriving feature coefficients and fitting curves of feature values and SHAP values. This demonstrates that the behaviour of black-box data-driven models can be as interpretable as white-box models while retaining their prediction accuracy.