A Hybrid Approach to Fraud Detection: Combining Machine Learning and Symbolic Rules for Enhanced Performance and Explainability

This paper introduces an extensive framework for detecting vehicle claim fraud by combining a supervised machine learning model with symbolic rules. The project examines the “Vehicle Claim Fraud Detection” dataset available on Kaggle, which consists of insurance claim data. The methodology employs a varied approach, involving data exploration and preprocessing, a comparative analysis of supervised learning models (Logistic Regression, Random Forest, and XGBoost), and the utilization of explainable AI (XAI) techniques (SHAP and LIME) to identify the best-performing model. This research revolves around an innovative hybrid approach that involves creating a new feature called symbolic_flag, which is based on recognized fraud patterns and then used to retrain the XGBoost model. This combination harnesses both the predictive capabilities of a data-driven model and the critical insights derived from domain knowledge. The findings indicate that this hybrid model outperforms others and offers a level of explainability that a purely black-box model does not, rendering it a more reliable and trustworthy option for high-stakes financial applications.