A Deep Learning Ensemble With Data Resampling For Credit Card Fraud Detection (Adolescencia e Saude)

Summary

• The proposed deep learning ensemble combined with the SMOTE-ENN method achieved a sensitivity of 1.000 and a specificity of 0.997.p.1
• When evaluated on the original dataset, the extension Voting Classifier achieved performance scores of 1.000 across accuracy, precision, recall, and F1-score.p.6
• On the SMOTE-ENN dataset, both the Stacking Classifier and Voting Classifier recorded scores of 1.000 across all evaluation metrics.p.7
• Resampling training data only with ADASYN resulted in logistic regression obtaining a precision of 1.81% and an F1 score of 3.55%.p.3
• Deep reinforcement learning models performed weakly on imbalanced credit card datasets, achieving 34.8% accuracy.p.3

Key insights

1. The framework utilizes long short-term memory (LSTM) and gated recurrent unit (GRU) base learners combined with a multilayer perceptron (MLP) meta-learner to handle dynamic shopping patterns and class imbalance.p.1
2. A systematic review of 40 articles published from 2015 to 2021 revealed limited research on deep learning techniques for credit card fraud detection.p.2
3. Applying resampling techniques before the train/test split artificially inflates machine learning model outcomes to over 99% accuracy.p.3
4. Deep reinforcement learning models evaluated on imbalanced transaction data showed poor performance with 34.8% accuracy.p.3
5. The extension Voting Classifier consistently outperformed base models across both the original and SMOTE-ENN datasets.p.6
6. The implementation includes a web front-end built on the Flask framework integrated with SQLite for authentication and fraud testing.p.1