ML для детекции фрода

ML — золотой стандарт для fraud detection в iGaming. Rule-based системы дополняются моделями с высокой accuracy.

Задачи

Payment fraud: stolen cards, chargeback prediction. Account takeover detection. Bonus abuse identification. Bot detection. Multi-accounting linkage. Money laundering pattern recognition. Каждая задача имеет свою target-переменную и feature set.

Модели

Gradient boosting: XGBoost, LightGBM, CatBoost — производительный дефолт. Neural networks: для sequences (LSTM транзакций) и графов (GNN для network fraud). Anomaly detection: Isolation Forest, autoencoders для unknown patterns. Ensemble: комбинация нескольких моделей.

Features

Temporal: time since registration, activity patterns. Behavioral: click patterns, game preferences. Network: связи с другими users через IP/device. Financial: deposit/withdrawal patterns, velocity. External: IP intelligence, device reputation. 200–2000 features в production модели.

Deployment

Real-time scoring: <100ms в hot path. Feature store: consistent features online/offline. Shadow mode: model выдаёт scores без action, validation. Champion-challenger testing. Human-in-the-loop для high-value decisions. Continuous monitoring drift и performance.