ANALYTICS MODELING CERTIFICATION
ISYE 6501 Final Quiz A+
2026/2027
Official-Style Comprehensive Examination · Introduction to Analytics Modeling
A+ QUESTIONS VERIFIED 5 SECTIONS 100% RATIONALES
Original & Blueprint-Aligned Balanced Coverage Item-Specific Explanations
CATEGORIES
■ 1. Classification, Clustering & Validation
■ 2. Regression, Trees & Variable Selection
■ 3. Time Series, Change Detection & PCA
■ 4. Probability Models, Simulation & Optimization
■ 5. Data Preparation, Missing Data, Experimental Design & Model Integration
STUVIAACTUALEXAM
Advanced Application / Analysis Level · Passing Score 75 % · 1 Mark per Question
,SECTION 1: Classification, Clustering & Validation
Q1. A retail analytics team is building a model to flag high-risk credit applications as 'approve' or 'deny'. The data
set contains 12 continuous predictors that have very different scales (income in tens of thousands, age in years,
debt ratios between 0 and 1). The team plans to use a hard-margin support-vector machine. Before fitting the
model, which preprocessing step is essential for the geometric interpretation of the maximum-margin hyperplane
to remain valid?
A. Standardize each predictor to zero mean and unit variance
B. Apply principal-component analysis and retain only the first two components
C. Replace missing values with the column-wise mode rather than the mean
D. Convert every continuous predictor into a binary indicator via median split
Correct Answer: A
Rationale: Hard-margin SVM finds the maximum-margin separating hyperplane in Euclidean space. Predictors on different scales
distort distances, so the margin is dominated by large-scale features. Standardization restores a meaningful geometry. Median splits
discard information, PCA changes the feature space, and mode imputation is unrelated to the margin geometry.
Q2. An online streaming service wants to recommend content genres to new users. Historical data contain user
feature vectors and the genre each user ultimately preferred. The modeling team decides to use k-nearest
neighbors with Euclidean distance. After scaling the features, they observe that several users lie exactly on the
decision boundary between two genres for the chosen odd value of k. How should the final class label be
assigned in such a tie situation under the standard majority-vote rule?
A. Break the tie by choosing the class that appears more frequently in the entire training set
B. Randomly select one of the tied classes with equal probability
C. Increase k by one and recompute the neighborhood until the vote is unique
D. Assign the class of the single nearest neighbor regardless of the remaining k-1 neighbors
Correct Answer: B
Rationale: Standard k-NN majority voting treats a perfect tie as unresolved; the conventional practical resolution is a uniform random
choice among the tied labels. Using only the single nearest neighbor effectively changes the algorithm, global frequency introduces
bias, and automatically increasing k alters the hyper-parameter chosen by cross-validation.
Q3. A hospital is validating a binary classifier that predicts 30-day readmission risk. The data set is moderately
imbalanced (18 % positive cases). The analytics team randomly splits the data into training (60 %), validation (20
%), and test (20 %) sets, then tunes the decision threshold on the validation set to maximize F1 score. After
locking the threshold, they evaluate the final model on the test set and obtain an F1 of 0.71. Which statement
correctly describes the statistical role of the reported F1?
A. It is optimistically biased because the threshold was chosen to maximize F1 on a related data partition
B. It is pessimistically biased because the training set was smaller than the full sample
C. It cannot be interpreted because F1 is inappropriate for imbalanced data
D. It is an unbiased estimate of future performance because the test set was never used for model selection
Correct Answer: D
Rationale: Once the model (including the threshold) is completely locked, the held-out test set provides an unbiased estimate of
generalization performance. The validation set was used for threshold selection, so any metric computed on validation would be
optimistically biased; the test metric is not. F1 remains a valid summary for imbalanced problems when the positive class is the focus.
STUVIAACTUALEXAM | Page 2 of 19
, Q4. A marketing analytics group is segmenting 50 000 customers into mutually exclusive groups for targeted
campaigns. They apply k-means with Euclidean distance after standardizing all numeric features. After trying
several values of k, the within-cluster sum of squares (WCSS) curve shows a clear elbow at k = 6, yet the
average silhouette width peaks at k = 4. The business stakeholders insist that the clusters must be roughly equal
in size. Which diagnostic should receive the highest priority when selecting the final k?
A. The elbow in WCSS, because it is the classical criterion for k-means
B. A plot of cluster sizes across candidate k values, because equal-size clusters are a hard business constraint
C. The gap statistic, because it is theoretically preferred for spherical clusters
D. The silhouette peak, because it directly measures cohesion versus separation
Correct Answer: B
Rationale: When stakeholders impose a hard operational constraint (roughly equal cluster sizes), that constraint dominates purely
statistical criteria. Both the elbow and silhouette may point to different k, but only the size distribution can confirm whether a
candidate k satisfies the business requirement. The gap statistic is useful but secondary to the stated constraint.
Q5. An insurance firm trains a soft-margin SVM classifier on a large data set of policy applications. After
cross-validation the optimal cost parameter C is found to be very large. What does this large value of C imply
about the resulting decision boundary relative to a hard-margin SVM?
A. The soft-margin boundary is nearly identical to the hard-margin boundary because misclassifications are heavily
penalized
B. The soft-margin boundary will ignore many support vectors and become smoother
C. The soft-margin boundary will be less sensitive to outliers than the hard-margin solution
D. The soft-margin boundary is guaranteed to achieve zero training error regardless of separability
Correct Answer: A
Rationale: In soft-margin SVM the parameter C controls the trade-off between margin size and classification error. As C → ∞ the
penalty on slack variables becomes infinite, recovering the hard-margin formulation (provided the data are linearly separable). A
large C therefore produces a boundary nearly identical to the hard-margin solution.
Q6. A data-science team is comparing five candidate classification algorithms on a fixed training set. They
perform 10-fold cross-validation for each algorithm and record the mean validation AUC. The algorithm with the
highest mean AUC is then retrained on the entire training set and evaluated once on a held-out test set. Why is
this procedure preferable to simply reporting the cross-validation AUC of the winning algorithm as the final
performance estimate?
A. The test-set evaluation after final retraining provides an unbiased estimate that accounts for the model-selection
step
B. Ten-fold cross-validation is known to be unstable for AUC, so an independent test set is mandatory
C. Cross-validation AUCs are always optimistically biased when model selection is performed on the same folds
D. Retraining on the full training set guarantees better generalization than any cross-validation estimate
Correct Answer: A
Rationale: When the same cross-validation folds are used both to select the best algorithm and to estimate its performance, the
reported CV metric is optimistically biased. Evaluating the ultimately chosen model on a completely held-out test set after retraining
yields an unbiased estimate of future performance.
STUVIAACTUALEXAM | Page 3 of 19
ISYE 6501 Final Quiz A+
2026/2027
Official-Style Comprehensive Examination · Introduction to Analytics Modeling
A+ QUESTIONS VERIFIED 5 SECTIONS 100% RATIONALES
Original & Blueprint-Aligned Balanced Coverage Item-Specific Explanations
CATEGORIES
■ 1. Classification, Clustering & Validation
■ 2. Regression, Trees & Variable Selection
■ 3. Time Series, Change Detection & PCA
■ 4. Probability Models, Simulation & Optimization
■ 5. Data Preparation, Missing Data, Experimental Design & Model Integration
STUVIAACTUALEXAM
Advanced Application / Analysis Level · Passing Score 75 % · 1 Mark per Question
,SECTION 1: Classification, Clustering & Validation
Q1. A retail analytics team is building a model to flag high-risk credit applications as 'approve' or 'deny'. The data
set contains 12 continuous predictors that have very different scales (income in tens of thousands, age in years,
debt ratios between 0 and 1). The team plans to use a hard-margin support-vector machine. Before fitting the
model, which preprocessing step is essential for the geometric interpretation of the maximum-margin hyperplane
to remain valid?
A. Standardize each predictor to zero mean and unit variance
B. Apply principal-component analysis and retain only the first two components
C. Replace missing values with the column-wise mode rather than the mean
D. Convert every continuous predictor into a binary indicator via median split
Correct Answer: A
Rationale: Hard-margin SVM finds the maximum-margin separating hyperplane in Euclidean space. Predictors on different scales
distort distances, so the margin is dominated by large-scale features. Standardization restores a meaningful geometry. Median splits
discard information, PCA changes the feature space, and mode imputation is unrelated to the margin geometry.
Q2. An online streaming service wants to recommend content genres to new users. Historical data contain user
feature vectors and the genre each user ultimately preferred. The modeling team decides to use k-nearest
neighbors with Euclidean distance. After scaling the features, they observe that several users lie exactly on the
decision boundary between two genres for the chosen odd value of k. How should the final class label be
assigned in such a tie situation under the standard majority-vote rule?
A. Break the tie by choosing the class that appears more frequently in the entire training set
B. Randomly select one of the tied classes with equal probability
C. Increase k by one and recompute the neighborhood until the vote is unique
D. Assign the class of the single nearest neighbor regardless of the remaining k-1 neighbors
Correct Answer: B
Rationale: Standard k-NN majority voting treats a perfect tie as unresolved; the conventional practical resolution is a uniform random
choice among the tied labels. Using only the single nearest neighbor effectively changes the algorithm, global frequency introduces
bias, and automatically increasing k alters the hyper-parameter chosen by cross-validation.
Q3. A hospital is validating a binary classifier that predicts 30-day readmission risk. The data set is moderately
imbalanced (18 % positive cases). The analytics team randomly splits the data into training (60 %), validation (20
%), and test (20 %) sets, then tunes the decision threshold on the validation set to maximize F1 score. After
locking the threshold, they evaluate the final model on the test set and obtain an F1 of 0.71. Which statement
correctly describes the statistical role of the reported F1?
A. It is optimistically biased because the threshold was chosen to maximize F1 on a related data partition
B. It is pessimistically biased because the training set was smaller than the full sample
C. It cannot be interpreted because F1 is inappropriate for imbalanced data
D. It is an unbiased estimate of future performance because the test set was never used for model selection
Correct Answer: D
Rationale: Once the model (including the threshold) is completely locked, the held-out test set provides an unbiased estimate of
generalization performance. The validation set was used for threshold selection, so any metric computed on validation would be
optimistically biased; the test metric is not. F1 remains a valid summary for imbalanced problems when the positive class is the focus.
STUVIAACTUALEXAM | Page 2 of 19
, Q4. A marketing analytics group is segmenting 50 000 customers into mutually exclusive groups for targeted
campaigns. They apply k-means with Euclidean distance after standardizing all numeric features. After trying
several values of k, the within-cluster sum of squares (WCSS) curve shows a clear elbow at k = 6, yet the
average silhouette width peaks at k = 4. The business stakeholders insist that the clusters must be roughly equal
in size. Which diagnostic should receive the highest priority when selecting the final k?
A. The elbow in WCSS, because it is the classical criterion for k-means
B. A plot of cluster sizes across candidate k values, because equal-size clusters are a hard business constraint
C. The gap statistic, because it is theoretically preferred for spherical clusters
D. The silhouette peak, because it directly measures cohesion versus separation
Correct Answer: B
Rationale: When stakeholders impose a hard operational constraint (roughly equal cluster sizes), that constraint dominates purely
statistical criteria. Both the elbow and silhouette may point to different k, but only the size distribution can confirm whether a
candidate k satisfies the business requirement. The gap statistic is useful but secondary to the stated constraint.
Q5. An insurance firm trains a soft-margin SVM classifier on a large data set of policy applications. After
cross-validation the optimal cost parameter C is found to be very large. What does this large value of C imply
about the resulting decision boundary relative to a hard-margin SVM?
A. The soft-margin boundary is nearly identical to the hard-margin boundary because misclassifications are heavily
penalized
B. The soft-margin boundary will ignore many support vectors and become smoother
C. The soft-margin boundary will be less sensitive to outliers than the hard-margin solution
D. The soft-margin boundary is guaranteed to achieve zero training error regardless of separability
Correct Answer: A
Rationale: In soft-margin SVM the parameter C controls the trade-off between margin size and classification error. As C → ∞ the
penalty on slack variables becomes infinite, recovering the hard-margin formulation (provided the data are linearly separable). A
large C therefore produces a boundary nearly identical to the hard-margin solution.
Q6. A data-science team is comparing five candidate classification algorithms on a fixed training set. They
perform 10-fold cross-validation for each algorithm and record the mean validation AUC. The algorithm with the
highest mean AUC is then retrained on the entire training set and evaluated once on a held-out test set. Why is
this procedure preferable to simply reporting the cross-validation AUC of the winning algorithm as the final
performance estimate?
A. The test-set evaluation after final retraining provides an unbiased estimate that accounts for the model-selection
step
B. Ten-fold cross-validation is known to be unstable for AUC, so an independent test set is mandatory
C. Cross-validation AUCs are always optimistically biased when model selection is performed on the same folds
D. Retraining on the full training set guarantees better generalization than any cross-validation estimate
Correct Answer: A
Rationale: When the same cross-validation folds are used both to select the best algorithm and to estimate its performance, the
reported CV metric is optimistically biased. Evaluating the ultimately chosen model on a completely held-out test set after retraining
yields an unbiased estimate of future performance.
STUVIAACTUALEXAM | Page 3 of 19
