Improving maritime accident severity prediction accuracy: A holistic machine learning framework with data balancing and explainability techniques

Abstract

• A holistic machine learning framework is developed to conduct maritime accident severity prediction. • The framework integrates eight well-established models, employing rigorous hyperparameter tuning and cross-validation. • Six advanced data balancing techniques are employed to mitigate class imbalance. • A novel dual interpretability approach combining SHAP and LIME provides both global and local insights. • The proposed framework offering a robust basis for risk-informed maritime safety management. Accurately predicting the severity of maritime accidents is crucial for enhancing safety management and minimizing operational risks. Traditional prediction models, however, often suffer from the challenges resulted from unbalanced datasets and the complexity of multidimensional factors. This study aims to develop an integrated prediction framework incorporating six data balancing techniques to effectively address category imbalance and enhance model predictive robustness. Additionally, eight well-established machine learning models are utilized, with their performance optimized through hyperparameter tuning and cross-validation. To interpret the model results, SHapley Additive exPlanations (SHAP) are applied for global feature contribution analysis, while Local Interpretable Model-agnostic Explanations (LIME) provide local interpretations, enabling an in-depth understanding of feature-specific impacts on predictions. The results indicate that the combination of RandomOverSampler and CatBoost achieves optimal performance across all metrics, with an accuracy of 86.45%, precision of 84.38%, recall of 89.70%, F1-score of 86.81%, and ROC AUC of 93.69%. The analysis identifies accident type, ship type, engine power and gross tonnage as the key features influencing accident severity prediction. Furthermore, the integrated explanatory framework combining SHAP and LIME elucidates both the individual contributions and the collective impact of these features, along with the direction and magnitude of their influence on individual predictions, ensuring model transparency and interpretability. This study advances the prediction of maritime accident severity and provides a robust scientific basis for decision-making in maritime safety, enabling policymakers and industrial stakeholders to make accurate and reliable risk-informed decisions. The source code is publicly available at : https://github.com/AdvMarTech/BalancedMaritimeAccidentXAI .

Topics & Concepts