Credit Card Fraud Detection using Soft Voting Ensemble with Imbalance Treatment

Credit card fraudulent transactions are on the rise due to increased digitalization and growing reliance on online modes of payment. Credit card fraud detection may be classified as an anomaly detection problem, that calls for robust machine learning techniques such as ensemble models for accurate identification. This study utilizes a soft voting ensemble combining five powerful machine learning models comprising two bagging and three boosting ensemble classifiers-Random Forest, Extra Trees, XGBoost, LightGBM, and CatBoost, respectively. The five models’ class posterior probabilities are fused via averaging to determine the final prediction. Synthetic Minority Oversampling Technique (SMOTE) is applied to achieve a balanced representation of legitimate and fraudulent transactions. Performance metrics such as the Receiver Operating Characteristics, F1-score, and Mathews Correlation Coefficient (MCC) are employed to assess the performance of the soft ensemble model. The findings indicate that the soft voting classifier outperformed all individual models, highlighting its potential as a robust tool for mitigating fraud risks in financial transactions.