INSURANCE FRAUD DETECTION VIA MACHINE LEARNING AND EXPLAINABLE ARTIFICIAL INTELLIGENCE (XAI)

Date

2025-9-1

Author

Tezel, Cihan

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

3517
views

0
downloads

Cite This

This thesis studies insurance fraud detection with machine learning (ML) methods and explainable artificial intelligence (XAI) tools. The dataset contains insurance policies and their related claims, where fraud is marked as a binary variable. The preparation stage covers missing data handling, encoding of categorical fields, and several feature engineering steps. Since fraud cases are rare, the Synthetic Minority Oversampling Technique (SMOTE) is employed, and models are trained on both the original and the balanced data. The algorithms tested include Logistic Regression, Random Forest, Support Vector Machines (SVM), k-Nearest Neighbors (KNN), LightGBM, and XGBoost. For XGBoost, two variants are considered: one with manually designed features such as seasonality, polynomial terms, and interactions, and another with automated feature engineering tuned by cross-validated AUC. Model performance is evaluated with ROC-AUC, accuracy, recall, specificity, and balanced accuracy, with thresholds optimized via ROC analysis. In addition, DeLong tests are applied to assess whether the observed differences in AUC across models are statistically significant. For interpretability, SHAP (SHapley Additive exPlanations) is utilized. The results indicate that liability coverage and fault assignment are the strongest drivers of fraud detection. Timing and policyholder features, such as claim month and address changes, also contribute meaningfully, though at a lower level. The findings showed that boosting methods achieved the best overall performance. LightGBM and XGBoost outperformed the other models, with XGBoost slightly ahead on the SMOTE-balanced data (AUC ≈ 0.85, recall ≈ 90%). Both manual and automated feature engineering improved recall, thereby shifting the trade-off toward catching more fraud cases. These findings emphasized the importance of designing fraud detection models that combined strong predictive accuracy with transparent and interpretable decision-making.

Subject Keywords

Insurance Fraud Detection, Machine Learning, XGBoost, SHAP, Imbalanced Data, Explainable Artificial Intelligence (XAI)

URI

https://hdl.handle.net/11511/115659

Collections

Graduate School of Applied Mathematics, Thesis