KAN-Driven Fusion of Tree-based Ensembles for Intelligent Risk Assessment of Arctic Shipping

This study proposes a machine learning-based framework for assessing navigation risks in Arctic maritime environments. By leveraging real-world environmental and operational data, the model assists shipmasters in making informed decisions under extreme weather and sea-ice conditions. The predictive system comprises a dual-component architecture, integrating an ensemble of tree-based models, i.e. Random Forest, Extra Trees, XGBoost, LightGBM, and SnapBoost, with a model fusion module based on the Kolmogorov-Arnold Network (KAN). The system is validated through performance comparisons with MultiLayer Perceptron and standalone KAN models using evaluation metrics including RMSLE, MAE, MSE, and RMSE. Experimental results demonstrate that the proposed approach consistently outperforms conventional methods, achieving higher accuracy and robustness in both training and testing phases. This research contributes to safer Arctic shipping by combining machine intelligence with human expertise, offering a decisionsupport tool that adapts to complex and dynamic maritime environments.