MS-DS Master of Data science Exploratory Data Analysis 4 — Questions and Answers
Question 1: During EDA you discover that a numeric feature has a bimodal distribution. What does this most likely suggest?
- The variable should be dropped
- There may be two distinct subpopulations in the data (Correct answer)
- The data has been log-transformed
- The variable is normally distributed
Correct answer: There may be two distinct subpopulations in the data
Two peaks in a distribution often indicate the presence of two underlying groups or processes that were combined in the dataset.
Question 2: Which dimensionality reduction technique is commonly used in EDA to project high-dimensional data onto 2D for visual cluster discovery?
- Lasso regression
- t-SNE (Correct answer)
- Ridge regression
- ANOVA
Correct answer: t-SNE
t-SNE (t-distributed Stochastic Neighbor Embedding) is designed to preserve local neighborhood structure, making clusters visible in 2D projections.
Question 3: When examining a scatter plot of two variables, you observe a curved (non-linear) pattern rather than a straight line. Which correlation measure is MORE appropriate than Pearson's r?
- Phi coefficient
- Spearman's rank correlation (Correct answer)
- Point-biserial correlation
- Cramér's V
Correct answer: Spearman's rank correlation
Spearman's correlation measures monotonic association using ranks rather than raw values, handling non-linear but monotonic relationships better than Pearson's r.
Question 4: A feature has the values: [1, 2, 3, 4, 100]. The z-score of 100 is approximately 2.18. Should 100 be flagged as an outlier using the z-score method with threshold |z| > 3?
- Yes, because it is clearly separated from the rest
- No, because its z-score does not exceed 3 (Correct answer)
- Yes, because any value above the mean is an outlier
- No, because z-scores cannot detect outliers
Correct answer: No, because its z-score does not exceed 3
The z-score method only flags values with |z| > 3 as outliers; z ≈ 2.18 is below that threshold, so 100 would not be flagged by this rule alone.
Question 5: Which EDA technique is specifically designed to reveal how the marginal distribution of one variable changes across levels of a second categorical variable?
- Box plots grouped by category (Correct answer)
- Autocorrelation plot
- Scatter plot matrix
- Q-Q plot
Correct answer: Box plots grouped by category
Side-by-side box plots, one per category level, directly compare the center, spread, and skew of the numeric variable across groups.
Question 6: What is the primary risk of using only summary statistics (mean, std) without visualizing data during EDA?
- Summary statistics are always inaccurate
- Different datasets can share identical summary statistics while having very different distributions (Anscombe's quartet) (Correct answer)
- Visualizations are always misleading
- Summary statistics cannot be computed for skewed data
Correct answer: Different datasets can share identical summary statistics while having very different distributions (Anscombe's quartet)
Anscombe's quartet demonstrates that four datasets with nearly identical means, variances, and correlations can have completely different visual patterns and structures.
Question 7: In EDA, what does a 'long tail' in a distribution imply for downstream machine learning modeling?
- The data is normally distributed and ready for modeling
- Rare extreme values may disproportionately influence models sensitive to scale or magnitude (Correct answer)
- The feature should be dropped from the model
- The distribution has no outliers
Correct answer: Rare extreme values may disproportionately influence models sensitive to scale or magnitude
Long-tailed features can distort distance-based or linear models unless transformed, requiring EDA to flag them for preprocessing decisions.
MS-DS Master of Data science Exploratory Data Analysis 4
During EDA you discover that a numeric feature has a bimodal distribution.
What does this most likely suggest?