Regression metrics на собеседовании Data Scientist

MAE

Mean Absolute Error.

MAE = (1/N) Σ |y_i - ŷ_i|

Robust к outliers. Same units как target.

MSE

Mean Squared Error.

MSE = (1/N) Σ (y_i - ŷ_i)²

Penalizes large errors more. Differentiable — used training. Squared units (problematic interpretation).

RMSE

Root MSE.

RMSE = √MSE

Same units as target. Sensitive к outliers.

Most popular для general regression.

MAPE / sMAPE

MAPE. Mean Absolute Percentage Error.

MAPE = (100/N) Σ |y_i - ŷ_i| / |y_i|

Percent units. Issues:

• Undefined when y=0.
• Asymmetric (under-predict bounded к 100%, over unlimited).

sMAPE. Symmetric MAPE — uses average of actual + predicted в denominator. Fixes asymmetry.

R²

Coefficient determination.

R² = 1 - (SS_res / SS_total)

Proportion variance explained. 1.0 = perfect, 0 = mean predictor, < 0 = worse than mean.

Useful для comparing models на same dataset.

Choosing

Outliers important. RMSE / MSE — sensitive.

Outliers should be ignored. MAE / median.

Percentage interpretation needed. MAPE / sMAPE.

Variance explained. R².

Quantile predictions. Quantile loss / pinball.

В практике — RMSE + R² often reported. MAPE для business context.

Связанные темы

• Linear vs logistic regression для DS
• Loss функции для DS
• Bias-variance trade-off для DS
• Forecasting system design для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на классических ML metrics.

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