Practice English vocabulary for ML feature engineering: normalization, one-hot encoding, feature interactions, feature selection, and feature importance.
0 / 5 completed
1 / 5
What is the difference between 'min-max normalization' and 'z-score normalization'?
Feature normalization prevents features with large scales from dominating distance-based models. Min-max: (x - min) / (max - min). Z-score: (x - mean) / std. Z-score is more robust to outliers since extreme values don't compress the entire distribution into a tiny range.
2 / 5
What is 'one-hot encoding for categorical variables'?
Assigning integers to categories (Red=1, Blue=2, Green=3) implies ordering (Green > Blue > Red) that doesn't exist. One-hot encoding creates separate binary features for each category, letting the model treat them as independent. High-cardinality features (1000+ categories) may use embedding instead.
3 / 5
What is a 'feature interaction' and what signal does it capture?
Some predictive signal only exists in feature combinations. Tree-based models (XGBoost, random forests) automatically learn interactions, but linear models require them to be engineered explicitly. Creating interaction features: multiplication, ratio, or concatenation of two features.
4 / 5
What is 'feature selection' and why does it reduce dimensionality?
High-dimensional feature spaces cause the 'curse of dimensionality' — models see sparse data and overfit. Feature selection methods (filter: correlation/chi-squared; wrapper: recursive feature elimination; embedded: LASSO regularization) remove noise features, improving generalization.
5 / 5
What does 'the feature importance score ranks predictors' mean?
Feature importance (available in Random Forest, XGBoost, etc.) helps practitioners understand the model, identify redundant features for removal, and explain predictions to stakeholders. Common measures: impurity-based importance, permutation importance, and SHAP values.