Level 1 — RecognitionData Preprocessing & Feature Engineering

Data Preprocessing & Feature Engineering

⏱ 20 minutes30 marksprintable — key stays hidden on paper

Time Limit: 20 minutes Total Marks: 30

Section A — Multiple Choice (1 mark each)

Choose the single best answer.

Q1. Which of the following is an example of ordinal data?

(a) Blood type (A, B, AB, O)
(b) T-shirt size (S, M, L, XL)
(c) Temperature in °C
(d) Customer ID number

Q2. Z-score normalization (standardization) transforms a feature so that it has:

(a) Minimum 0 and maximum 1
(b) Mean 0 and standard deviation 1
(c) Median 0 and range 1
(d) Sum equal to 1

Q3. Min-max scaling maps a value $x$ to which formula?

(a) $\dfrac{x - \mu}{\sigma}$
(b) $\dfrac{x - \min}{\max - \min}$
(c) $\dfrac{x}{\max}$
(d) $\log(x+1)$

Q4. One-hot encoding a categorical feature with 5 distinct categories produces how many new binary columns (no drop)?

(a) 1
(b) 4
(c) 5
(d) 25

Q5. SMOTE is primarily used to address:

(a) Missing values
(b) Outliers
(c) Class imbalance
(d) Multicollinearity

Q6. Which imputation strategy is most appropriate for a categorical feature with missing values?

(a) Mean imputation
(b) Mode (most frequent) imputation
(c) Log transformation
(d) Min-max scaling

Q7. Fitting a scaler on the entire dataset before splitting into train/test sets is an example of:

(a) Feature creation
(b) Data leakage
(c) Binning
(d) Undersampling

Q8. A log transformation is commonly applied to a feature that is:

(a) Normally distributed
(b) Right-skewed with positive values
(c) Binary
(d) Perfectly symmetric

Q9. Target encoding replaces a category with:

(a) A random integer
(b) The mean of the target variable for that category
(c) A one-hot vector
(d) The category's alphabetical rank

Q10. By the IQR method, an outlier is typically a value lying outside:

(a) $[Q1 - 1.5\,\text{IQR},\ Q3 + 1.5\,\text{IQR}]$
(b) $[\mu - \sigma,\ \mu + \sigma]$
(c) $[\min,\ \max]$
(d) $[Q1,\ Q3]$

Section B — Matching (1 mark each, 6 marks)

Q11. Match each technique in Column X to its correct purpose in Column Y.

Column X	Column Y
(i) Label encoding	(A) Reduce class imbalance by removing majority samples
(ii) Binning	(B) Assign integer codes to categories
(iii) Undersampling	(C) Convert continuous values into discrete intervals
(iv) Interaction term	(D) Product/combination of two features
(v) VIF	(E) Detect multicollinearity
(vi) Stratified split	(F) Preserve class proportions across sets

Section C — True / False with justification (2 marks each, 14 marks)

(1 mark correct T/F, 1 mark valid justification)

Q12. Standardization is bounded between 0 and 1. (T/F + justify)

Q13. Deleting all rows with any missing value can bias the dataset if data is not missing completely at random. (T/F + justify)

Q14. One-hot encoding is preferable to label encoding for a nominal feature fed into a linear model. (T/F + justify)

Q15. The test set should be used to decide which features to keep during preprocessing. (T/F + justify)

Q16. Two features with a correlation coefficient of $0.98$ indicate strong multicollinearity. (T/F + justify)

Q17. Min-max scaling is more robust to outliers than z-score standardization. (T/F + justify)

Q18. In EDA, examining distributions and missingness comes before modelling. (T/F + justify)

Answer keyMark scheme & solutions

Section A (10 marks)

Q1 — (b) T-shirt size. Why: Sizes have a meaningful order (S<M<L<XL) but no numeric magnitude → ordinal. Blood type is nominal; °C is numerical; ID is a nominal label. (1)

Q2 — (b) Mean 0, SD 1. Why: $z=(x-\mu)/\sigma$ centres at 0 and scales to unit variance. (1)

Q3 — (b) $(x-\min)/(\max-\min)$ . Why: This rescales the range to $[0,1]$ . (1)

Q4 — (c) 5. Why: One binary column per distinct category when none is dropped. (1)

Q5 — (c) Class imbalance. Why: SMOTE synthesises minority-class samples. (1)

Q6 — (b) Mode imputation. Why: Mean is undefined for categories; most-frequent value is the natural fill. (1)

Q7 — (b) Data leakage. Why: Test statistics leak into training via the scaler fit on all data. (1)

Q8 — (b) Right-skewed positive. Why: Log compresses large values, reducing right skew. (1)

Q9 — (b) Mean of target for that category. Why: That is the definition of (mean) target encoding. (1)

Q10 — (a) $[Q1-1.5\,\text{IQR},\ Q3+1.5\,\text{IQR}]$ . Why: Standard Tukey fence for outliers. (1)

Section B (6 marks)

Q11: (i)→B, (ii)→C, (iii)→A, (iv)→D, (v)→E, (vi)→F. (1 mark each)

Section C (14 marks)

Q12 — False. Justify: Standardization has no fixed bounds; values can exceed 1 or be negative. Min-max scaling is the bounded [0,1] method. (1+1)

Q13 — True. Justify: If missingness correlates with the target/features (MAR/MNAR), dropping rows removes a non-random subset → biased/unrepresentative sample. (1+1)

Q14 — True. Justify: Label encoding imposes a false ordinal ranking on nominal categories, which a linear model misinterprets as magnitude; one-hot avoids this. (1+1)

Q15 — False. Justify: Feature selection using the test set leaks information and inflates performance; selection must use only training (or validation) data. (1+1)

Q16 — True. Justify: $|r|=0.98$ is very high, meaning near-linear dependence between the features → multicollinearity. (1+1)

Q17 — False. Justify: Min-max uses min and max, which are extreme values directly set by outliers, so it is more sensitive; z-score (or robust scaling) handles outliers comparatively better/robust scaling best. (1+1)

Q18 — True. Justify: EDA (distributions, missingness, correlations) is performed before modelling to inform preprocessing choices. (1+1)

[
  {"claim":"Min-max scaling of x=7 with min=2,max=12 gives 0.5","code":"x,mn,mx=7,2,12; result=((x-mn)/(mx-mn))==Rational(1,2)"},
  {"claim":"Z-score of value equal to mean is 0","code":"mu,sig,x=10,3,10; result=((x-mu)/sig)==0"},
  {"claim":"One-hot of 5 categories yields 5 columns","code":"result=(5)==5"},
  {"claim":"IQR upper fence for Q1=10,Q3=20 is 35","code":"Q1,Q3=10,20; IQR=Q3-Q1; result=(Q3+Rational(3,2)*IQR)==35"}
]