DESIGN AND IMPLEMENTATION OF AN ADAPTIVE RISK SCORING AND FRAUD DETECTION SYSTEM USING A WEIGHTED RULEBASED MODEL

The exponential growth of financial technology and digital banking has led to a corresponding rise in fraudulent transactions, posing significant risks to individuals, businesses, and financial institutions. Traditional fraud detection methods, such as purely rule-based systems, provide transparency and interpretability but are often rigid and unable to adapt to the evolving strategies employed by fraudsters. Conversely, machine learning approaches such as decision trees, random forests, and gradient boosting classifiers are highly adaptive and accurate but suffer from limited interpretability and require substantial computational resources. This project addresses these limitations by designing and implementing a hybrid fraud detection system that integrates weighted rule-based reasoning with an advanced machine learning model (XGBoost). The hybrid approach ensures a balance between interpretability, adaptability, and predictive accuracy. The PaySim dataset, which simulates real-world mobile money transactions, was employed for training and evaluation. Preprocessing techniques such as feature extraction, handling class imbalance, and rule-based feature engineering were applied to prepare the dataset. Ten domaininformed rules were developed to capture suspicious transaction behaviors, such as abnormal balance differences and unusual transaction types, producing risk scores that were incorporated into the hybrid model. The XGBoost classifier was then trained on both the original features and the engineered rule-based scores. Hybrid probability fusion was implemented, where the final decision combined the rule-based component (weighted at 0.3) and the XGBoost probability (weighted at 0.7).