Credit Card Fraud Detection Using Machine Learning Techniques: Dealing with Imbalanced Data using Over-Sampling and Under-Sampling Methods

Digital banking has eased financial transactions and significantly increased fraudulent activities. Credit card companies must identify fraudulent transactions to prevent customers from being charged for unauthorized purchases. Previously, several machine-learning methodologies and classifiers have been employed to identify fraudulent transactions. Nonetheless, due to the constantly evolving nature of fraud patterns, it is essential to examine new frauds and formulate a model based on these new patterns. Furthermore, authentic transactions are frequently higher than fraudulent ones, making it very difficult to distinguish between them. This issue must be addressed urgently to detect these fraudulent activities as they occur and notify the right individuals. This research proposes using machine learning techniques to identify and report fraudulent transactions automatically. We mainly address the class imbalance problem (fraud and non-fraud) by evaluating outcomes from several over-sampling and under-sampling methods. Several sampling techniques have been applied to the training set. However, SMOTE and ROS+RUS have been selected based on better precision and recall of the training data. Machine learning models can produce precise predictions and learn from a more balanced distribution. A performance comparison of Artificial Neural Network (ANN) and ensemble learning using test data is performed, and the comparison is made with baseline models, including Random Forest and XGBoost.