Computer Vision Techniques -Based Financial Fraud Detection Using Ensemble Learning and Explainable Artificial Intelligence

Abstract

Financial fraud is a growing concern for the global economy, with hundreds of billions of dollars lost every year, and the traditional rule-based fraud detection systems are no longer effective because they are unable to cope with the increasing complexity of fraud schemes. In this paper, we propose FraudShield-XAI, an ensemble learning framework to produce high fraud detection performance and transparent/interpretable decision making, which uses a stacking-based framework consisting of XGBoost, random forests, and an adaptive neural network meta-learner. We evaluate and train the framework using two widely used datasets, the IEEE-CIS Fraud Detection dataset and the PaySim simulated mobile money transactions dataset, and in order to address the class imbalance issue that is typical in fraud detection problems, we apply the Synthetic Minority Oversampling Technique with Edited Nearest Neighbours (SMOTE-ENN). Our experimental results show that FraudShield-XAI outperforms traditional single-model based approaches with an AUC-ROC of 98.72%, an MCC of 0.923, and an F1-score of 97.84%, and we use SHAP and LIME to explain the predictions of FraudShield-XAI, providing the most influential features for each prediction. Identified key factors include transaction velocity, merchant category, and geographical deviations, which can offer actionable insights to compliance teams and fraud analysts, therefore, FraudShield-XAI bridges the performance vs. interpretability gap, which is essential for regulatory approval and real-world deployment in fintech settings.