DEEPALI GARG,UMA SHARMA
DOI: https://doi.org/Traditional detection techniques often struggle with highly imbalanced datasets and subtle patterns in fraudulent behavior and Credit card fraud depicts significant challenges to financial institutions because of its rarity, unpredictability, and rapid evolution. This study shows a deep learning-based approach utilizing a Multilayer Perceptron (MLP) which is trained on a strategically balanced dataset with a 1:4 fraud-to-non-fraud ratio. Extensive preprocessing and feature engineering steps were undertaken, which includes the generation of temporal, behavioral, geospatial, and probability-based features. To resolve class imbalance and focus on hard-to-classify fraudulent samples, focal loss with class weighting was employed during training. The MLP model, incorporating dropout regularization and ReLU activations, was inculcated and evaluated using precision, recall, F1-score, and PR-AUC metrics. With achieved results of precision of 97%, recall of 85%, F1-score of 90% and PR-AUC of 89%, it demonstrated significant improvements over traditional models on the original imbalanced test set. These findings emphasized the importance of engineered features and specialized training strategies in enhancing the detection of rare and costly fraud events.
