A Constraint-Aware Soft Voting Ensemble Framework for Credit Card Fraud Detection under Operational Threshold Optimization

Credit card fraud detection is a challenging binary classification problem due to extreme class imbalance and the operational cost asymmetry between false negatives and false positives. Although high-performing classifiers can provide strong probability estimates, their practical effectiveness depends heavily on how these probabilities are converted into final fraud decisions. This study proposes a constraint-aware soft voting ensemble framework that explicitly separates model training from decision threshold selection. The framework combines the probabilistic outputs of XGBoost, LightGBM, and Decision Tree classifiers through soft voting and then determines the final decision threshold on a validation set under predefined minimum recall and precision constraints. Unlike conventional approaches that rely on fixed probability thresholds, the proposed method aligns the classification decision with operational fraud detection requirements. Experimental results on a highly imbalanced credit card fraud dataset show that the proposed ensemble achieves a precision of 0.9000, recall of 0.8514, F1-score of 0.8750, ROC-AUC of 0.9525, and PR-AUC of 0.8441. The confusion matrix further demonstrates that the framework maintains a low false-positive rate while detecting most fraudulent transactions. These findings indicate that constraint-aware threshold optimization can improve the practical decision quality of ensemble-based fraud detection systems without requiring retraining of the underlying classifiers.