An Interpretable Ensemble Framework for Credit Card Fraud Detection on a Highly Imbalanced Dataset

The rapid growth of digital payment systems has made credit card fraud a serious global concern, causing substantial financial losses. Detecting such fraudulent activities is challenging due to the severe class imbalance in transaction datasets, where fraudulent cases are extremely rare. Many existing models suffer from overfitting or produce misleading outcomes under these conditions. To solve this problem, this study introduces a two-stage hybrid framework that focuses on both accuracy and clear interpretability. First, an unsupervised LSTM Autoencoder is trained only on normal transactions to learn regular spending behavior. Transactions with high reconstruction errors are marked as suspicious, and these errors are turned into a new feature called the Anomaly Score. Next, the SMOTE is used to balance the dataset, and a LightGBM classifier is trained on this improved data. The proposed system performs strongly on unseen test data, achieving an accuracy of 99.94% and an F1-score of 82.65%. Further analysis shows that the Anomaly Score plays a key role in the model’s decision process. Overall, this work offers a dependable, interpretable, and effective method for detecting credit card fraud in highly imbalanced datasets.