Integrating Ensemble Learning and Model-Agnostic Explain ability for Reliable Credit Card Fraud Prediction

Abstract

Because of highly imbalanced datasets and rapidly changing fraud patterns, there is still a lack of effective methods for credit card fraud detection in the context of modern financial systems. We propose a novel + explain ability→ machine-learning framework that integrates both Random Forest and XGBoost classifiers with explain ability techniques to enhance not just predictive accuracy, but critical interpretability for audience⇄ the consumers of insights→. This study employs the publicly available credit card fraud dataset (284,807 transactions), with only 0.17% of them being fraudulent samples. Following pre-processing and standardization, models were trained and evaluated on the basis of 70–30 train–test split. Random Forest classifier scored an overall accuracy of 99.99% whilef0.93, recall f0.82, and f1-score f0.87 among the fraud class. From the confusion matrix, we see that there are 111 true fraud detected but only 25 missed cases, and the ROC–AUC of 0.986 indicates, we have great discriminatory power. Finally, with the application of XGBoost, we were able to reach a precision of 0.94, a recall of 0.82, and an F1-score of 0.87 for the minority class, thus also confirming very good generalization and robustness between the metrics. For the sake of interpretability, SHAP and LIME were performed to catch the main contributing features. SHAP bee swarm and bar plots indicated that V14, V17, V12, and V10 were the most important variables for predicting fraud, while LIME gave case-level explanations of the feature-values pointing a transaction to being fraud or non-fraud. XGBoost achieved a robust performance in low-prevalence conditions represented by the Precision–Recall curve, where it retains a high precision over a large range of recall. In summary, these results indicate that ensemble classifiers and model-agnostic explain ability tools make a natural combination that leads to a powerful and transparent fraud detection approach. This provides the framework for the financial institutions to make fast yet reliable and interpretable decisions on a continuous basis. The possible extensions for the visual representation will cover the deep learning architectures, anomaly detection hybrids, and cost-sensitive learning to maximize the performance and minimize the false negative rate.