Rooting out deception : the application of tree-based learners for motor insurance fraud detection

Abstract

This dissertation investigates motor insurance fraud detection in the Spanish market by implementing and comparing tree-based methods, renowned for their performance and interpretability. The study begins with basic Decision Trees and progresses to tree-based ensemble methods, including Random Forests, Gradient Boosting machines, and Newton-based boosting techniques such as LightGBM, XGBoost, and CatBoost. A significant challenge in motor insurance fraud detection is addressing the class imbalance. To address this issue, the dissertation evaluates cost-sensitive learning approaches and resampling techniques to optimize model performance. The analysis concludes that a cost-sensitive LightGBM model is the most effective method for this scenario, achieving a balanced accuracy of 81% and successfully identifying 83% of fraudulent cases. The findings of this study provide valuable insights into the effectiveness of tree-based methods in detecting motor insurance fraud and highlight the potential of LightGBM in efficiently identifying fraudulent cases. By presenting a rigorous comparison of different techniques and addressing the class imbalance issue, this research contributes to the ongoing development of interpretable solutions for combating insurance fraud.