Feature engineering quiz

Question 1

What is feature engineering in machine learning, and why is it important?

Accepted Answer

Creating, transforming, and selecting input variables to improve model performance

Answer

Choosing the learning rate of a neural network

Answer

Splitting data into train and test sets

Answer

Tuning the number of epochs during training

Question 2

Why is feature scaling (e.g., standardization) often required before training models like k-NN or logistic regression?

Accepted Answer

So features with large ranges do not dominate distance or weight calculations

Answer

So the dataset takes less memory to store

Answer

Because models cannot read decimal numbers

Answer

To convert categorical data into numbers

Question 3

What is the difference between standardization (z-score scaling) and min-max normalization?

Accepted Answer

Standardization gives mean 0 and std 1; min-max maps values to a fixed range like [0,1]

Answer

They are identical operations with different names

Answer

Standardization always maps to [0,1]; min-max gives mean 0

Answer

Min-max removes outliers; standardization adds them

Question 4

What is one-hot encoding and when would you use it?

Accepted Answer

Turning a nominal category into binary indicator columns, one per category

Answer

Assigning increasing integers to ordered categories

Answer

Replacing categories with their mean target value

Answer

Scaling numeric features to the range [0,1]

Question 5

Why can label encoding (assigning integers 0,1,2,... to categories) be problematic for nominal data?

Accepted Answer

It introduces a false sense of order and magnitude among unordered categories

Answer

It always produces too many columns to store

Answer

It cannot handle string categories at all

Answer

It deletes rare categories automatically

Question 6

When is ordinal encoding appropriate instead of one-hot encoding?

Accepted Answer

When the categories have a genuine inherent order (e.g., Low < Medium < High)

Answer

When there are too many categories to one-hot encode

Answer

When the feature is numeric and continuous

Answer

When the target variable is categorical

Question 7

What is feature selection, and how does it differ from feature extraction?

Accepted Answer

Selection keeps a subset of original features; extraction builds new combined features

Answer

Selection creates new features; extraction deletes features

Answer

They are the same process

Answer

Selection only applies to images; extraction only to text

Question 8

How does Principal Component Analysis (PCA) help in feature engineering?

Accepted Answer

It creates fewer uncorrelated components that capture most of the data's variance

Answer

It removes all categorical columns from the dataset

Answer

It increases the number of features to improve accuracy

Answer

It fills in missing values using the mean

Question 9

What is a common strategy for handling missing values in a numeric feature, and what is a risk of it?

Accepted Answer

Imputing with mean/median, which can reduce variance and distort relationships

Answer

Deleting the target column to avoid the issue

Answer

Converting all numbers to text strings

Answer

Multiplying missing values by zero

Question 10

What is data leakage in the context of feature engineering, and why is it dangerous?

Accepted Answer

Using information not legitimately available at prediction time, inflating performance

Answer

Losing data due to a corrupted file

Answer

Having too many features in the dataset

Answer

Forgetting to scale numeric features

Question 11

Why should feature scaling parameters (like mean and standard deviation) be computed only on the training set?

Accepted Answer

To avoid leaking test-set information into the training process

Answer

Because the test set has no numeric features

Answer

To make the test set larger

Answer

Because scalers only work on small datasets

Question 12

What is binning (discretization) of a numeric feature, and give one reason to use it?

Accepted Answer

Grouping a continuous variable into intervals to capture non-linearity or reduce noise

Answer

Removing all numeric columns from the data

Answer

Encoding text into numbers

Answer

Increasing the precision of decimal values

Question 13

Why might applying a log transformation to a skewed feature improve a model?

Accepted Answer

It reduces right-skew and the impact of large outliers, making data more symmetric

Answer

It converts categorical data into numbers

Answer

It always increases the number of features

Answer

It removes missing values automatically

Question 14

What is an interaction feature, and why might it help a model?

Accepted Answer

A combined feature (e.g., product of two features) capturing their joint effect

Answer

A feature that interacts with the user via a UI

Answer

A duplicate copy of an existing feature

Answer

A feature that is always set to zero

Question 15

How can a datetime column be turned into useful features?

Accepted Answer

Extracting parts like month, day-of-week, hour, and weekend/holiday flags

Answer

Converting the date into a single random number

Answer

Deleting the column because dates are unusable

Answer

Storing the date only as text

Question 16

Why are sine and cosine transformations used to encode cyclical features such as the hour of day?

Accepted Answer

To represent the wrap-around so values like hour 23 and 0 are treated as close

Answer

To make the feature larger in scale

Answer

To convert hours into categories

Answer

To remove seasonal patterns from the data

Question 17

What is the 'curse of dimensionality' and how does it relate to feature engineering?

Accepted Answer

As features increase, data becomes sparse and models need much more data

Answer

Models run faster as dimensions increase

Answer

Adding features always improves accuracy

Answer

It refers to too few rows in the dataset

Question 18

What problem does high cardinality (very many categories) cause for one-hot encoding, and name one alternative?

Accepted Answer

It creates too many sparse columns; alternatives include target encoding or grouping

Answer

It always loses category names; the only fix is deletion

Answer

It converts numbers to text

Answer

It has no downsides for any dataset size

Question 19

What is target (mean) encoding for a categorical feature, and what is its main risk?

Accepted Answer

Replacing categories with the mean target value; main risk is leakage/overfitting

Answer

Replacing the target with category names; risk is none

Answer

Scaling the target variable; risk is slow training

Answer

Deleting categories with low frequency; risk is memory

Question 20

How can outliers affect feature engineering, and what is one robust way to reduce their impact?

Accepted Answer

They distort statistics and models; clipping or robust scaling reduces their impact

Answer

They always improve accuracy and should be kept untouched

Answer

They only affect categorical features

Answer

They are removed automatically by all models

Question 21

What are polynomial features and when can they be useful?

Accepted Answer

Powers and products of features that let linear models fit non-linear patterns

Answer

Features that contain only polynomials of text data

Answer

A method to delete correlated features

Answer

A way to scale features to [0,1]

Question 22

Why can highly correlated (redundant) features be a problem, especially for linear models?

Accepted Answer

They cause multicollinearity, making coefficients unstable and hard to interpret

Answer

They always increase accuracy substantially

Answer

They reduce the number of rows needed

Answer

They convert categorical data automatically

Question 23

What is the difference between filter, wrapper, and embedded methods of feature selection?

Accepted Answer

Filter uses statistics, wrapper tests model subsets, embedded selects during training

Answer

All three are identical statistical tests

Answer

Filter trains models, wrapper uses statistics, embedded deletes rows

Answer

They differ only in programming language

Question 24

How does L1 (Lasso) regularization act as a feature selection method?

Accepted Answer

It shrinks some coefficients exactly to zero, dropping those features

Answer

It doubles every coefficient to emphasize features

Answer

It only works on categorical features

Answer

It increases all coefficients equally

Question 25

What is the bag-of-words representation in text feature engineering?

Accepted Answer

Representing text as word-count vectors over a vocabulary, ignoring order

Answer

Storing text as a single integer hash

Answer

Encoding only the first word of each sentence

Answer

Translating text into another language

Question 26

What does TF-IDF do, and why is it often better than raw word counts?

Accepted Answer

It weights terms by frequency and rarity, emphasizing distinctive words

Answer

It counts only the longest words in a document

Answer

It removes all numbers from text

Answer

It translates words into vectors of random numbers

Question 27

What is a feature, and what is a target (label) in a supervised dataset?

Accepted Answer

A feature is an input variable; the target is the value being predicted

Answer

A feature is the output; the target is the input

Answer

Both are inputs with no difference

Answer

A feature is the model and the target is the data

Question 28

Why is domain knowledge valuable in feature engineering?

Accepted Answer

It lets you craft meaningful features that capture real-world relationships

Answer

It is irrelevant because models learn everything automatically

Answer

It only matters for choosing the learning rate

Answer

It is used only to label the test set

Question 29

What is the difference between fitting and transforming a scaler, as in scikit-learn's fit and transform?

Accepted Answer

Fit learns the parameters; transform applies them to the data

Answer

Fit applies the change; transform learns the parameters

Answer

They are identical and interchangeable

Answer

Fit deletes data; transform restores it

Question 30

What is a count/frequency encoding for categorical variables?

Accepted Answer

Replacing each category with its count or frequency in the dataset

Answer

Replacing each category with the target mean

Answer

Creating one binary column per category

Answer

Sorting categories alphabetically

Test yourself