Domain adaptation на собеседовании Data Scientist

Зачем domain adaptation

Train на source domain (много labeled). Apply на target domain (different distribution).

Examples:

• Sentiment model на product reviews → applied на tweets.
• Object detection на synthetic data → real images.
• Sales prediction на USA → adapt to Russia.

Без adaptation — accuracy drops significantly.

Setting types

Supervised DA. Target имеет labels (мало). Standard fine-tuning.

Unsupervised DA. Target — только unlabeled. Сложнее.

Semi-supervised DA. Target часть labeled, часть unlabeled.

Domain generalization. Train на multiple sources, evaluate на unseen domain.

Fine-tuning strategies

Full fine-tune. Update all weights с small learning rate.

Linear probe. Freeze backbone, train classifier head. Good если target small.

LoRA. Add adapter, update only adapters.

Layer-wise. Freeze early layers (general features), tune later (domain-specific).

DANN

Domain-Adversarial Neural Network. Add domain classifier с gradient reversal.

Feature extractor → label classifier (helps task)
                  → domain classifier (gradient reversed, makes features domain-invariant)

Forces features that don't reveal source / target domain — generalize.

MMD

Maximum Mean Discrepancy. Measure distance между distributions в RKHS.

Loss penalize feature distributions source vs target — push close.

L = task_loss + λ · MMD(features_source, features_target)

Used в transfer learning, generative models.

Связанные темы

• Self-supervised learning для DS
• Few-shot learning для DS
• Active learning для DS
• Data drift для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на работах Ganin 2015 (DANN), Long 2015 (MMD).

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