PREDICTION OF VESSEL DETENTION CONSIDERING DATA IMBALANCE: METHODS AND COMPARISON

The safeguard of maritime transport is port state control (PSC) inspection that is responsible for inspecting foreign visiting ships to a port. As the port inspection resources are limited and expensive, while the vessels have their schedules associated with high delay cost, accurately identifying vessels with high risk for inspection is the key to guaranteeing efficient PSC inspection. Detention is the most severe outcome of PSC inspection, which is an intervention action taken by the port to prevent a vessel from proceeding to sea, and thus the probability of a ship to be detained can be regarded as its risk level. However, as the detention rate is very low in practice (which is usually no more than 5%), prediction of ship detention is not a trivial task, as the issue of imbalanced data set should be addressed either by data processing or by prediction algorithm design. This study aims to develop machine learning models to predict vessel detention in PSC based on the above two solutions. Especially, we adopt the combination of oversampling and undersampling to generate a balanced dataset, and classic classification models are developed for ship detention prediction. From the perspective of algorithm design, we develop isolation forest model where ship detention is viewed as anomaly behaviours to predict vessel detention. Numerical experiments using initial PSC inspection records at the Port of Hong Kong show that the second model is more effective in distinguishing detained ships from all foreign visiting ships to a port.