Decoding Maritime Encounters: Game‐Theoretic Feature Selection for Vessel Intent Prediction

ABSTRACT

Early intent recognition of nearby vessels is important for improving maritime navigation safety and for making timely decisions during vessel encounters. Since intent must be inferred from motion patterns and relative interactions, selecting the most informative features is critical for building accurate trajectory‐based prediction models. This work evaluates two cooperative game‐theoretic feature selection methods, the Shapley value and the nucleolus, for identifying influential motion and encounter features in maritime intent prediction. Using a dataset of expert‐defined vessel behaviours, the study compares these methods with standard feature selection approaches across multiple deep learning architectures and encounter distance regimes. The results show that Shapley‐based selection consistently ranks among the best‐performing feature selection methods across all dataset variants, achieving up to 98.16% accuracy and 98.16% 1‐score and providing statistically significant improvements over alternative methods in the majority of configurations. Overall, the study demonstrates an effective and interpretable framework for early vessel intent prediction on high‐fidelity simulated maritime encounter data; validation on real‐world data remains an important direction for future work.