Analyzing Pandemic-Induced Changes in Credit Card Fraud Using Auto encoder and Machine Learning Techniques

Abstract

Credit card fraud poses significant financial risks, necessitating advanced detection frameworks capable of adapting to evolving transaction behaviors. This study employs machine learning techniques to detect fraudulent activity using the publicly available PaySim synthetic dataset, which contains over 6.3 million transactions with features such as transaction type, amount, origin and destination balances, and fraud indicators. The dataset is preprocessed to handle missing values, encode categorical variables, and generate behavioral and temporal features, including transaction velocity, merchant and location diversity, device/IP consistency, and rolling statistics. Class imbalance is addressed using SMOTE to enhance minority class representation. Both supervised models (Logistic Regression, XGBoost, LightGBM) and unsupervised anomaly detection models (Autoencoder, Isolation Forest) are applied. A hybrid fraud alert system integrates model predictions with rule-based checks for high-value transactions, risky transaction types, abnormal velocity, and disproportionate amounttobalance ratios, enabling real-time actionable alerts. Among the models, XGBoost achieves the highest performance with an ROC-AUC of 0.9995, accuracy of 0.9969, precision of 0.9991, recall of 0.9969, and F1-score of 0.9978, outperforming Logistic Regression and LightGBM. The results demonstrate that ensemble boosting models effectively capture complex, non-linear fraud patterns. Overall, this study provides a robust framework for credit card fraud detection, combining behavioral analytics, anomaly detection, and adaptive machine learning, offering practical insights for financial institutions to monitor and mitigate fraudulent activities in dynamic transactional environments.