JoWUA

Abstract

Financial fraud detection is still a major problem due to having highly imbalanced transaction data, changing fraud patterns, and the need for explainable AI in financial decision-making systems. In this study, a hybrid framework for fraud detection is proposed based on the IEEE-CIS Fraud Detection dataset, which is a combination of Isolation Forest anomaly detection, boosting-based classification models, and SHAP explainability. In the first step, Isolation Forest is used to detect unusual behaviors in transactions and to create anomaly scores in the absence of class labels. These anomaly scores are then combined with features of the original transactions to form an enhanced feature representation for supervised learning. The most popular boosting algorithms to predict fraud are compared: XGBoost, LightGBM, and CatBoost. The results of experiments indicate that the proposed hybrid approach yields significantly better performance in detecting fraud than the stand-alone boosting approaches. The best-performing of the evaluated approaches was Hybrid IF-CatBoost with accuracy scores of 98.73%, precision scores of 95.61%, recall scores of 93.42%, F1-scores of 94.50%, ROC-AUC of 0.992, and PR-AUC of 0.966. The results of statistical comparison showed that the use of anomaly scores in the boosting models resulted in the improvement of the average F1-score by 4.49%. In addition, SHAP gave both general explanations and local explanations, which indicated that the most important fraud indicators were anomaly scores, the number of transactions, device information, and card-related attributes. The proposed framework strikes a beneficial balance between the predictive power, robustness, and interpretability of the model and can be used in practical financial fraud monitoring systems.