Predicting types of human-related maritime accidents with explanations using selective ensemble learning and SHAP method

Abstract

Maritime accidents frequently lead to severe property damage and casualties, and an accurate and reliable risk prediction model is necessary to help maritime stakeholders assess the current risk situation. Therefore, the present study proposes a hybrid methodology to develop an explainable prediction model for maritime accident types. Based on the advantages of selective ensemble learning method, this study pioneers to introduce a two-stage model selection method, aiming to enhance the predictive accuracy and stability of the model. Then, SHAP (Shapley Additive Explanations) method is integrated to identify effective mapping associations of seafarers’ unsafe acts and their risk factors with the prediction results. The results demonstrate that the model developed achieves good prediction performance with an accuracy of 87.50 % and an F1-score of 84.98 %, which benefits stakeholders in assessing the type of maritime accident in advance, so as to make proactive intervention measures.

Keywords

• Maritime accidents
• Seafarers' unsafe acts
• Risk prediction
• Selective ensemble learning