Fraud detection ML system design на собеседовании Data Scientist

Постановка задачи

Real-time fraud detection — payments, account takeover.

Constraints:

• Score per transaction < 100ms.
• Imbalanced (0.1-1% fraud).
• Label delayed (chargeback after weeks).
• Adversaries adapt.

Real-time scoring

Transaction → Feature service → Model → Decision (allow / review / block).

Decision tiers:

• Allow. Low risk.
• Review. Manual queue.
• Block. High risk.
• Step-up auth. Send 2FA, Сlarify.

Threshold tuning depends на business cost (false positives expensive).

Features

Transaction. Amount, currency, country, merchant, time.

User history. Recent transactions, mean amount, country pattern.

Velocity. N transactions last 1h / 1day. Rapid changes — risk.

Device. Fingerprint, IP geolocation, mismatch с user country.

Network. Graph features (linked accounts, shared payment methods).

Historical risk. Score user based на past activity.

Feature store обязателен — pre-computed real-time.

Label delay

Fraud confirmed только когда chargeback (1-90 days).

Implications:

• Можно train только на «old enough» data.
• Recent transactions — labels missing → unable retrain immediately.

Mitigation:

• Weak labels. User reports / suspicious patterns как proxy.
• Sequence modeling. Predict pre-chargeback signals.

Class imbalance

99%+ legitimate.

Approaches:

• Sampling. Undersample majority, oversample minority (SMOTE).
• Class weights.
• Focal loss.
• Anomaly detection instead of supervised (если labels rare).

Adversarial nature

Fraudsters adapt. Models drift fast.

Defense:

• Frequent retrain (daily / hourly).
• Monitor model output drift.
• Manual review feedback loop.
• Adversarial training (simulate attacks).
• Layered defenses (rules + ML + manual).

Метрики

Offline:

• PR-AUC (better для imbalance).
• Precision @ K (capacity manual review).
• Recall @ K.

Online:

• $ saved (prevented fraud) vs $ lost (false blocks).
• Manual review queue size.
• Customer complaints / chargeback rate.

Связанные темы

• Class imbalance для DS
• Anomaly detection для DS
• Feature Store для DS
• Data drift для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на индустриальных fraud detection practices.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.