Recsys ML system design на собеседовании Data Scientist

Постановка задачи

E-commerce / streaming recsys.

Constraints:

• 100M users, 10M items.
• Real-time recommendations.
• Continuous catalog growth.

Hybrid approaches

Collaborative filtering. «Users like you also liked». Strong на established items / users.

Content-based. Item embeddings (BERT, image). Strong для cold start.

Combined. Multi-stage:

Stage 1: Candidate generation
  - Two-tower CF (top-1000).
  - Content-based для new items (top-100).
  - Trending / popular для cold start (top-50).
  Merged → ~1500 candidates.
Stage 2: Ranking model on combined features.
Stage 3: Re-ranking — diversity, business rules.

Cold start

New user.

• Onboarding survey.
• Demographic-based default.
• Trending recommendations.
• Popular items.

New item.

• Content embedding (text / image).
• Two-tower model with item features (не только ID).
• Bandit exploration — show к users randomly, learn rapidly.

New domain / category. Transfer learning от similar.

Online learning

В addition to batch training.

• Real-time feature updates (last action seconds ago).
• Streaming model updates (Vowpal Wabbit, FTRL).
• Bandit exploration / exploitation.

Real-time updates

User clicks → event → feature store updated.

User: viewed_item_recently=item_xyz, last_active=now.
Next request: ranker uses fresh features.

Time-decay engagement signals — recent matters more.

Метрики

Offline:

• NDCG, Recall@K, Hit@K.
• Coverage (catalog diversity).

Online:

• CTR.
• Conversion rate.
• Revenue per user.
• Retention long-term.
• Diversity / serendipity.

Beware — clickbait optimization. Long-term retention важно.

Связанные темы

• Collaborative filtering для DS
• Two-tower DSSM для DS
• Feed ranking system design для DS
• Multi-armed bandit для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на индустриальных recsys practices (Spotify, Netflix, YouTube).

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