Learning curves на собеседовании Data Scientist

Что такое learning curve

Plot model performance vs training set size.

X: training samples (1k, 5k, 10k, ..., full).
Y: train score, validation score.

Иллюстрирует, как модель learns с more data.

High bias signature

Score
 ^
1.0|
   |
0.6|---train------val---  (both low, plateau)
   |
0.0|________________ N samples

Both curves low and converged. More data не помогает.

Action: more model capacity (deeper, more features), не more data.

High variance signature

Score
 ^
1.0|----train-----------
   |    big gap
0.7|         val ↗ (rising с more data)
   |
0.0|________________ N samples

Big train-val gap. Adding data closes gap.

Action: more data, regularization, dropout, simpler model.

Sample size decisions

Learning curve helps decide.

If val plateauing. No benefit от more data — focus на model improvements.

If val rising. More data would help. Justifies labeling investment.

Project planning:

• Need 5% more accuracy?
• Curve says: gain 1% per 10k samples.
• Need ~50k more labeled samples → budget annotation.

Validation curves

Different — plot performance vs hyperparameter.

X: regularization strength.
Y: train, val score.

Underfitting region. Both low, large reg.

Sweet spot. Val peaks.

Overfitting region. Train high, val drops.

Pick HP near sweet spot.

Связанные темы

• Bias-variance trade-off для DS
• Cross-validation для DS
• Hyperparameter tuning для DS
• Active learning для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на классических ML diagnostics.

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