Fraud Detection & Anomaly Detection

Fraud Detection Challenges

Imbalanced Data:

• Frauds are <1% of transactions
• Solutions:
  • SMOTE (synthetic minority oversampling)
  • Class weights in loss function
  • Ensemble methods (Random Forest robust to imbalance)
  • Anomaly detection (treat fraud as outlier)

Features:

• Transaction: Amount, time, location, merchant
• Behavioral: Average spend, frequency, velocity
• Network: Device fingerprint, IP address
• Historical: Past fraud flags, dispute rate

Interview Q: "Credit card fraud detection - what model?" A:

1. Start: Logistic Regression or Random Forest (interpretable for compliance)
2. Handle imbalance: Class weights, SMOTE, or anomaly detection
3. Features: Transaction velocity, location change, amount deviation
4. Real-time: Low latency (<100ms), use cached user profiles
5. Evaluation: Precision-Recall (not accuracy), AUC-PR
6. Monitoring: Concept drift (fraudsters adapt)

Anomaly Detection

Techniques:

• Statistical: Z-score, IQR
• Isolation Forest: Isolates outliers faster
• Autoencoders: Reconstruction error for normal vs anomaly
• One-class SVM: Learn boundary of normal data

Interview Q: "Detect unusual login activity" A:

• Features: Login time, location, device, failed attempts
• Baseline: User's historical patterns
• Model: Isolation Forest or autoencoder
• Alert: Threshold on anomaly score + contextual rules

:::