Artificial Intelligence for Detecting Irregularities in Banking Operations (International Journal of Industrial Engineering)

Summary

• The dataset used for fraud recognition research contains 50,000 samples, consisting of 47,573 normal transactions, 2,423 fraudulent transactions, and four unlabeled samples.p.1
• A character-level convolutional neural network designed to detect phishing websites from Turkish domain names achieves an accuracy of 90%.p.3

Key insights

1. Combining behavioral transaction characteristics with class imbalance techniques establishes a framework for evaluating machine learning classifiers in banking fraud detection.p.1
2. Deep learning autoencoder models reconstruct valid transactions to detect abnormalities without relying on tagged fraudulent training data.p.2
3. Reframing fraud detection to minimize total business costs balances fraud mitigation benefits against human review expenses and customer friction.p.2
4. Variational autoencoders learn underlying data distributions to effectively extract features and identify anomalies in large-scale transactions.p.3
5. Unsupervised machine learning on Coin-Join updated Bitcoin user graphs enables identification of illicit network activity and money laundering patterns.p.4
6. Testing classifiers on unaltered imbalanced data reveals high false positive rates, indicating the necessity of operating threshold calibration.p.7