Graph neural networks enabled accident causation prediction for maritime vessel traffic

Maritime vessel traffic accidents frequently result in significant casualties, property damage, and environmental pollution. The investigation of these accidents is crucial for determining responsibility and improving navigation rules. However, the complexity of the marine environment complicates the quick identification of accident factors by maritime authorities. In this study, the causes of maritime vessel traffic accidents were predicted. Firstly, 501 maritime vessel traffic accident investigation reports published on the official website of the China Maritime Safety Administration were collected and analyzed. Secondly, a maritime vessel traffic accident causation model was constructed, and information from the accident reports was extracted based on the model. Finally, textual information was converted into graph data, and a Mutual Information-based Deep Graph Convolutional Network (GCN) model (MIDG-GCN) was proposed to analyze accident characteristics and predict causation. A ranked list of accident causes was generated based on scores, demonstrating favorable comparisons with commonly used Graph Neural Networks (GNNs). The MeanRank of the model was found to be 9.5436, indicating effectiveness in ranking predicted causation. High accuracy in predicting the top-3 and top-10 causes was shown by Hits@3 and Hits@10 scores of 0.7505 and 0.9701, respectively. The proposed model provides valuable causation predictions to guide decision-making in accident investigations.