Data Wrangling and Preprocessing Flashcards

When concatenating training feature engineering, fitting an encoder on the full dataset before train/test split causes what problem?

Answer: Data leakage

Fitting any preprocessing on the combined data lets test information influence training, which is data leakage.

A target variable has 95% class A and 5% class B. Which preprocessing technique addresses this imbalance?

Answer: Resampling such as SMOTE or undersampling

Oversampling the minority class (e.g., SMOTE) or undersampling the majority class rebalances skewed class distributions.

Why should imputation of missing values inside a cross-validation loop be done within each fold rather than once beforehand?

Answer: To avoid leaking information across folds

Imputing before splitting lets statistics from validation folds leak into training, inflating performance estimates.

Two highly correlated numeric features (r=0.98) are kept in a linear model. What issue does this raise?

Answer: Multicollinearity

Highly correlated predictors cause multicollinearity, making coefficient estimates unstable and hard to interpret.

A free-text 'notes' column needs to become model features. Which preprocessing step is most appropriate first?

Answer: Text vectorization (e.g., TF-IDF or tokenization)

Raw text must be tokenized and vectorized (e.g., TF-IDF) into numeric features before most models can use it.

When using scikit-learn, bundling scaling, encoding, and the model into a single object that fits in order prevents leakage and simplifies deployment. What is it called?

Answer: A Pipeline

A Pipeline chains preprocessing steps and the estimator so transforms are fit only on training data within each fold.

You scale numeric columns but leave one-hot encoded binary columns unscaled. Why is this generally acceptable?

Answer: Binary 0/1 columns are already on a comparable bounded scale

One-hot columns are already bounded to 0 and 1, so they are on a scale comparable to standardized numeric features.