Credit Card Fraud Detection Using Metaheuristic Techniques

Credit Card Fraud Detection (CCFD) has become a critical challenge to financial security due to increasingly sophisticated fraudulent activities. This study investigates the effectiveness of Meta-Heuristic (MHT) optimization techniques in improving fraud detection (FD) through feature selection (FS) and model optimization. To address class imbalance, Random Under-Sampling (RUS) was applied. The selected feature subsets were evaluated using three machine learning (ML) classifiers—Decision Tree (DT), KNearest Neighbours (KNN), and XGBoost (Xgb-Tree)—across four benchmark datasets: European, Statlog (Australian Credit Approval), PaySim, and Credit Card Transactions Fraud Detection (CCTFD). Eleven binary MHT algorithms were implemented and compared. The comparative analysis shows that the Binary Particle Swarm Optimization (BPSO) algorithm consistently achieved the best balance between accuracy, robustness, and convergence stability across datasets and classifiers. In DT and KNN models, BPSO demonstrated superior performance with fewer features, while for XGBoost, metaheuristic-based feature selection achieved high accuracy (above 0.92). Other optimizers, including BSSA, BHGSO, and BAVO, also performed strongly, revealing synergistic effects and the potential of ensemble-based optimization. Overall, swarm-adaptive optimizers such as BSSA, BAVO, BASO, and BPSO proved most effective for binary feature selection when integrated with ensemble classifiers. The findings highlight the suitability of BSSA and BHHO as reference hybrid models for high-dimensional feature selection and classification, reinforcing the “No Free Lunch” theorem. This research supports the development of robust fraud detection frameworks by aligning the choice of optimizer and classifier with specific data environments.