Semi-supervised learning for multi-target regression in ship risk prediction

Limited port state control (PSC) inspection resources pose an urgent need to enhance PSC inspection efficiency. While existing PSC officers (PSCOs) assignment randomly assign ships with PSCOs, this study aims to predict ship deficiency numbers under various deficiency categories (types of non-compliance) using machine learning (ML) models. Moreover, not all foreign visiting ships to a port are inspected by PSC, which leads to a large amount of unlabeled data remaining unexplored. Using the port of Singapore as a case study, this paper utilizes both labeled and unlabeled data to predict ship deficiency numbers under the six deficiency categories of individual ships. A semi-supervised multi-target regression (SSMTR) framework is developed, which innovates in using prediction performance on the validation dataset to judge the reliability of unlabeled data. The SSMTR framework is extended to various ML regression methods, such as decision tree (DT), support vector regression (SVR), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP), resulting in DT-SSMTR, SVR-SSMTR, XGBoost-SSMTR, and MLP-SSMTR. Across four experiment groups with different numbers of labeled data samples, the mean squared error improves on average by 13.65% for DT-SSMTR, 1.62% for SVR-SSMTR, 4.39% for XGBoost-SSMTR, and 2.65% for MLP-SSMTR compared to models that only use labeled data.