Fraud Detection System

Design an ML system to detect fraudulent transactions in real-time.

Requirements:

• Real-time scoring (<100ms latency)
• High precision (minimize false positives)
• Handle class imbalance (fraud is rare)
• Adaptive to new fraud patterns

Key Components:

1. Feature Store (real-time + batch features)
2. Real-time Feature Pipeline
3. Model Serving Infrastructure
4. Human Review Queue
5. Active Learning Pipeline

Discussion Points:

• Feature engineering (velocity, device fingerprint, behavioral)
• Model choice (XGBoost, neural networks, ensembles)
• Class imbalance handling (SMOTE, focal loss, oversampling)
• Online learning vs batch retraining
• Model calibration for different risk thresholds
• Feedback loops for continuous improvement