A Scalable MLOps Framework for Intelligent Fraud Detection and Identification in Financial Transactions (International Journal of Advanced Research in Science Communication and Technology)

Summary

• The study develops a scalable MLOps framework evaluated on 100,000 transactions sampled from the Credit Card Transactions Fraud Detection Dataset.p.1
• Applying SMOTE balancing increased the training set fraudulent transactions from 478 to 7,952, creating a balanced training dataset of 87,474 transactions.p.7
• The Hybrid Stacking Model achieved the highest accuracy of 99.49%, precision of 55.23%, and F1-score of 65.07%.p.1
• LightGBM and HistGradientBoosting models demonstrated the highest recall rate for fraud detection at 82.50%.p.1
• Demographic analysis showed that male consumers accounted for 56.19% of the reported fraud rate, while female consumers accounted for 43.81%.p.6
• Operational monitoring across 20,000 test predictions recorded an average prediction latency of 0.000014 seconds.p.12

Key insights

1. LightGBM achieved the highest ROC-AUC score of 0.982 among all evaluated fraud detection models.p.11
2. Fraud occurrence varies by day of the week, peaking on Friday at a rate of 0.79% and reaching its lowest on Sunday at 0.44%.p.6
3. The Kolmogorov-Smirnov test detected data drift in key features including amt, category, hour, cc_num, and age.p.13
4. The Hybrid Stacking Model achieved the lowest classification error rate of 0.0051 across all tested models.p.10
5. Transaction amount (amt) displayed the strongest linear correlation with the fraud target variable among numerical features, with a Pearson correlation of 0.26.p.7
6. Random Forest feature importance ranking was used to select the top 20 predictive features from an initial 24 features.p.7
7. Operational model health monitoring identified 212 fraud predictions out of 20,000 test records, representing a fraud prediction rate of 1.06%.p.12