Machine Learning Based Fraud Detection System (INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND MANAGEMENT)

Summary

• The proposed machine learning fraud detection pipeline achieves an ROC-AUC of 0.984 and a PR-AUC of 0.886 across 284,807 European cardholder logs.p.1
• The experimental model captures 90.82% of illegitimate transfers while suppressing false customer flags to 0.061%.p.1
• The European Card Benchmark evaluation corpus consists of 284,807 transaction records with 492 fraudulent occurrences, reflecting a 0.172% positive rarity.p.2
• The IEEE-CIS E-Commerce Repository dataset encompasses 590,540 commercial transactions characterized across 394 behavioral and device dimensions.p.2
• Under simulated stress trials at 10,000 TPS, the compiled C++ LightGBM runtime completes full transaction scoring with a P99 response of 24.8 ms.p.3
• When transaction load is increased to 20,000 TPS, tail latency stabilizes at 43.8 ms.p.3

Key insights

1. Legacy static conditional filters struggle under operational friction, triggering false alarms while failing to intercept shifting illicit stratagems.p.1
2. Deep neural topologies like LSTM and Graph Neural Networks often surpass 150 ms latencies, breaching the sub-50 ms clearing gateway service agreements.p.1
3. Supervised tree outputs are merged with unsupervised anomaly indices using an arbitration weight beta of 0.85.p.2
4. Both evaluation benchmarks maintained a strict chronological 70:30 train-test partitioning to prevent look-ahead contamination.p.2
5. TreeSHAP integration provides exact, additive feature attribution scores to comply with Reserve Bank of India directives and GDPR Article 22 requirements.p.3
6. Continuous Population Stability Index monitoring triggers automated model retraining whenever drift metrics exceed 0.25 over rolling 24-hour intervals.p.3
7. The hybrid framework preserved 88.4% detection recall under simulated adversarial evasion attacks.p.3