ISYE 6501: INTRODUCTION TO ANALYTICS MODELING
MIDTERM & FINAL PRACTICE EXAM 2026–2027
(90 Original Practice Questions with Correct Answers & Rationales)
Engineering & Analytics | Georgia Tech OMSCS-Style Practice
Key Domains: Analytics Modeling • Data Types • Classification (k-NN, Logistic Regression, SVM, Trees, Neural Nets) • Clustering •
PCA • Time Series • Regression • Optimization
IMPORTANT DISCLAIMER: This is an ORIGINAL educational practice resource for study purposes only. It is NOT an official Georgia Tech
exam, NOT an official OMSCS / ISYE 6501 assessment, and does NOT contain proprietary course exam items, quiz banks, or copyrighted
homework solutions. It is not affiliated with, endorsed by, or produced by Georgia Institute of Technology. Content is aligned with commonly
taught introductory analytics modeling themes (supervised/unsupervised learning, validation, classical ML methods, time series, and
optimization concepts). Course emphasis, software (R), and exact exam blueprints vary by term—always follow your current ISYE 6501
syllabus, lectures, textbooks/notes, and instructor guidance. This resource is for independent practice and concept reinforcement only.
Introduction
This 90-question practice set supports midterm and final review for Introduction to Analytics Modeling. Correct
answers appear in bold cyan with concise rationales explaining modeling reasoning and common misconceptions.
Focus on when to use each method, key assumptions, validation, and interpretation—not only definitions.
Suggested Study Use
• Attempt all 90 under timed conditions (~2–2.5 hours) before checking answers.
• Score by domain; rebuild weak areas with lecture notes and homework-style coding practice.
• For each method, recite: problem type → assumptions → key hyperparameters → validation metric → pitfalls.
• Always compare complex models to simple baselines and respect train/validation/test hygiene.
EXAM QUESTIONS (1–90)
Total: 90 original multiple-choice items. Correct answers in bold cyan with rationales.
PART A — Foundations, Data & Model Evaluation (Q1–12)
1. In analytics modeling, the primary goal of a predictive model is to:
A. Guarantee zero error on all future data
B. Prove causation with certainty in every observational dataset
C. Estimate an unknown response or class for new observations based on patterns learned from data
D. Replace domain expertise entirely
Correct Answer: C. Estimate an unknown response or class for new observations based on patterns learned
from data
Rationale: Predictive models learn relationships from historical data to score new cases. They do not automatically prove
causation or guarantee zero error.
2. Which statement best distinguishes supervised from unsupervised learning?
A. They are identical terms
B. Supervised learning never uses features X
C. Unsupervised learning always predicts a continuous y
D. Supervised learning uses labeled responses (y); unsupervised learning finds structure without labels
Correct Answer: D. Supervised learning uses labeled responses (y); unsupervised learning finds structure
without labels
Rationale: Classification/regression are supervised. Clustering and many dimensionality-reduction explorations are
unsupervised.
3. A continuous response variable (e.g., demand, temperature) is typically modeled with:
A. Only k-means clustering
B. Hard classification only with no probabilities
, C. Only association rules
D. Regression methods
Correct Answer: D. Regression methods
Rationale: Continuous y → regression; categorical y → classification. Method choice follows the response type and goal.
4. Which data type is categorical with no natural order?
A. Ratio continuous with true zero
B. Interval temperature only
C. Ordinal Likert ranks with meaningful order
D. Nominal (e.g., color, payment method)
Correct Answer: D. Nominal (e.g., color, payment method)
Rationale: Nominal = labels without order. Ordinal has order. Interval/ratio are numeric measurement scales.
5. Dummy (one-hot) variables are commonly used to:
A. Replace the need for a response variable
B. Encode categorical predictors for models that require numeric inputs
C. Always reduce dimensionality automatically
D. Compute PCA loadings only
Correct Answer: B. Encode categorical predictors for models that require numeric inputs
Rationale: Most linear models need numeric X. One-hot encoding creates binary indicators for categories (watch the dummy-
variable trap/collinearity).
6. Overfitting occurs when a model:
A. Fits noise in the training data and generalizes poorly to new data
B. Has too little flexibility by definition only
C. Performs equally well on train and test always
D. Never uses validation
Correct Answer: A. Fits noise in the training data and generalizes poorly to new data
Rationale: High train performance + poor test performance suggests overfitting. Regularization, simpler models, more data, and
validation help.
7. The main purpose of a holdout test set is to:
A. Provide an unbiased estimate of generalization performance on unseen data
B. Maximize training accuracy only
C. Tune hyperparameters repeatedly without leakage concerns
D. Replace cross-validation in all cases as identical
Correct Answer: A. Provide an unbiased estimate of generalization performance on unseen data
Rationale: Test data should be touched sparingly after model selection. Validation/CV is for tuning; test is for final assessment.
8. Cross-validation is primarily used to:
A. Increase training R² artificially without purpose
B. Eliminate the need for any preprocessing
C. Estimate out-of-sample performance and/or select models/hyperparameters more reliably than a single
split
D. Guarantee the global optimum of nonconvex models
Correct Answer: C. Estimate out-of-sample performance and/or select models/hyperparameters more
reliably than a single split
Rationale: k-fold CV rotates validation folds. Time series needs careful CV (e.g., rolling/forward chaining) to respect time order.
9. Bias–variance tradeoff suggests that:
A. Variance and bias always decrease together without tradeoff
B. More flexible models can reduce bias but may increase variance; simpler models often have higher bias
and lower variance
C. Only bias matters for prediction error
, D. Only variance matters for prediction error
Correct Answer: B. More flexible models can reduce bias but may increase variance; simpler models often
have higher bias and lower variance
Rationale: Expected prediction error decomposes into bias² + variance + irreducible noise. Model complexity balances bias and
variance.
10. Scaling/standardizing features is especially important for:
A. Tree-based splits that are completely invariant in all implementations always
B. Distance-based and regularized methods (e.g., k-NN, SVM with RBF, k-means, ridge/lasso) where feature
magnitude affects the algorithm
C. Reading CSV headers
D. Computing a contingency table only
Correct Answer: B. Distance-based and regularized methods (e.g., k-NN, SVM with RBF, k-means,
ridge/lasso) where feature magnitude affects the algorithm
Rationale: When algorithms use distances, penalties, or gradient steps sensitive to scale, standardize/normalize. Trees are more
scale-robust but scaling can still matter in pipelines.
11. A confusion matrix is used in classification to summarize:
A. Autocorrelation at lag 12 only
B. Counts of true positives, false positives, true negatives, and false negatives
C. Only residual sums of squares
D. Cluster centroids only
Correct Answer: B. Counts of true positives, false positives, true negatives, and false negatives
Rationale: From the confusion matrix we derive accuracy, precision, recall/sensitivity, specificity, F1, etc.
12. When classes are highly imbalanced, accuracy can be misleading because:
A. ROC curves cannot be drawn
B. Accuracy always equals recall
C. A trivial majority-class classifier can look good while missing the rare but important class
D. Imbalance makes precision undefined always
Correct Answer: C. A trivial majority-class classifier can look good while missing the rare but important
class
Rationale: Prefer precision/recall, F1, PR curves, balanced accuracy, or cost-sensitive metrics when class rarity matters (e.g.,
fraud, disease).
PART B — Classification: k-NN & Logistic Regression (Q13–25)
13. k-Nearest Neighbors classifies a new point by:
A. Maximizing a likelihood with no distance concept
B. Looking at the labels of the k closest training points (e.g., majority vote) in feature space
C. Fitting a global linear hyperplane only
D. Partitioning via recursive impurity splits only
Correct Answer: B. Looking at the labels of the k closest training points (e.g., majority vote) in feature space
Rationale: k-NN is instance-based/lazy: prediction uses local neighborhood. Choice of k, distance metric, and scaling are critical.
14. In k-NN, choosing a very small k (e.g., k=1) tends to:
A. Make the decision boundary perfectly linear
B. Remove the need for feature scaling
C. Increase variance and sensitivity to noise/outliers
D. Always maximize bias and underfit severely as the only effect
Correct Answer: C. Increase variance and sensitivity to noise/outliers
Rationale: Small k → flexible, jagged boundaries (high variance). Large k → smoother, more biased boundaries.
15. A practical reason to scale features before k-NN is that:
MIDTERM & FINAL PRACTICE EXAM 2026–2027
(90 Original Practice Questions with Correct Answers & Rationales)
Engineering & Analytics | Georgia Tech OMSCS-Style Practice
Key Domains: Analytics Modeling • Data Types • Classification (k-NN, Logistic Regression, SVM, Trees, Neural Nets) • Clustering •
PCA • Time Series • Regression • Optimization
IMPORTANT DISCLAIMER: This is an ORIGINAL educational practice resource for study purposes only. It is NOT an official Georgia Tech
exam, NOT an official OMSCS / ISYE 6501 assessment, and does NOT contain proprietary course exam items, quiz banks, or copyrighted
homework solutions. It is not affiliated with, endorsed by, or produced by Georgia Institute of Technology. Content is aligned with commonly
taught introductory analytics modeling themes (supervised/unsupervised learning, validation, classical ML methods, time series, and
optimization concepts). Course emphasis, software (R), and exact exam blueprints vary by term—always follow your current ISYE 6501
syllabus, lectures, textbooks/notes, and instructor guidance. This resource is for independent practice and concept reinforcement only.
Introduction
This 90-question practice set supports midterm and final review for Introduction to Analytics Modeling. Correct
answers appear in bold cyan with concise rationales explaining modeling reasoning and common misconceptions.
Focus on when to use each method, key assumptions, validation, and interpretation—not only definitions.
Suggested Study Use
• Attempt all 90 under timed conditions (~2–2.5 hours) before checking answers.
• Score by domain; rebuild weak areas with lecture notes and homework-style coding practice.
• For each method, recite: problem type → assumptions → key hyperparameters → validation metric → pitfalls.
• Always compare complex models to simple baselines and respect train/validation/test hygiene.
EXAM QUESTIONS (1–90)
Total: 90 original multiple-choice items. Correct answers in bold cyan with rationales.
PART A — Foundations, Data & Model Evaluation (Q1–12)
1. In analytics modeling, the primary goal of a predictive model is to:
A. Guarantee zero error on all future data
B. Prove causation with certainty in every observational dataset
C. Estimate an unknown response or class for new observations based on patterns learned from data
D. Replace domain expertise entirely
Correct Answer: C. Estimate an unknown response or class for new observations based on patterns learned
from data
Rationale: Predictive models learn relationships from historical data to score new cases. They do not automatically prove
causation or guarantee zero error.
2. Which statement best distinguishes supervised from unsupervised learning?
A. They are identical terms
B. Supervised learning never uses features X
C. Unsupervised learning always predicts a continuous y
D. Supervised learning uses labeled responses (y); unsupervised learning finds structure without labels
Correct Answer: D. Supervised learning uses labeled responses (y); unsupervised learning finds structure
without labels
Rationale: Classification/regression are supervised. Clustering and many dimensionality-reduction explorations are
unsupervised.
3. A continuous response variable (e.g., demand, temperature) is typically modeled with:
A. Only k-means clustering
B. Hard classification only with no probabilities
, C. Only association rules
D. Regression methods
Correct Answer: D. Regression methods
Rationale: Continuous y → regression; categorical y → classification. Method choice follows the response type and goal.
4. Which data type is categorical with no natural order?
A. Ratio continuous with true zero
B. Interval temperature only
C. Ordinal Likert ranks with meaningful order
D. Nominal (e.g., color, payment method)
Correct Answer: D. Nominal (e.g., color, payment method)
Rationale: Nominal = labels without order. Ordinal has order. Interval/ratio are numeric measurement scales.
5. Dummy (one-hot) variables are commonly used to:
A. Replace the need for a response variable
B. Encode categorical predictors for models that require numeric inputs
C. Always reduce dimensionality automatically
D. Compute PCA loadings only
Correct Answer: B. Encode categorical predictors for models that require numeric inputs
Rationale: Most linear models need numeric X. One-hot encoding creates binary indicators for categories (watch the dummy-
variable trap/collinearity).
6. Overfitting occurs when a model:
A. Fits noise in the training data and generalizes poorly to new data
B. Has too little flexibility by definition only
C. Performs equally well on train and test always
D. Never uses validation
Correct Answer: A. Fits noise in the training data and generalizes poorly to new data
Rationale: High train performance + poor test performance suggests overfitting. Regularization, simpler models, more data, and
validation help.
7. The main purpose of a holdout test set is to:
A. Provide an unbiased estimate of generalization performance on unseen data
B. Maximize training accuracy only
C. Tune hyperparameters repeatedly without leakage concerns
D. Replace cross-validation in all cases as identical
Correct Answer: A. Provide an unbiased estimate of generalization performance on unseen data
Rationale: Test data should be touched sparingly after model selection. Validation/CV is for tuning; test is for final assessment.
8. Cross-validation is primarily used to:
A. Increase training R² artificially without purpose
B. Eliminate the need for any preprocessing
C. Estimate out-of-sample performance and/or select models/hyperparameters more reliably than a single
split
D. Guarantee the global optimum of nonconvex models
Correct Answer: C. Estimate out-of-sample performance and/or select models/hyperparameters more
reliably than a single split
Rationale: k-fold CV rotates validation folds. Time series needs careful CV (e.g., rolling/forward chaining) to respect time order.
9. Bias–variance tradeoff suggests that:
A. Variance and bias always decrease together without tradeoff
B. More flexible models can reduce bias but may increase variance; simpler models often have higher bias
and lower variance
C. Only bias matters for prediction error
, D. Only variance matters for prediction error
Correct Answer: B. More flexible models can reduce bias but may increase variance; simpler models often
have higher bias and lower variance
Rationale: Expected prediction error decomposes into bias² + variance + irreducible noise. Model complexity balances bias and
variance.
10. Scaling/standardizing features is especially important for:
A. Tree-based splits that are completely invariant in all implementations always
B. Distance-based and regularized methods (e.g., k-NN, SVM with RBF, k-means, ridge/lasso) where feature
magnitude affects the algorithm
C. Reading CSV headers
D. Computing a contingency table only
Correct Answer: B. Distance-based and regularized methods (e.g., k-NN, SVM with RBF, k-means,
ridge/lasso) where feature magnitude affects the algorithm
Rationale: When algorithms use distances, penalties, or gradient steps sensitive to scale, standardize/normalize. Trees are more
scale-robust but scaling can still matter in pipelines.
11. A confusion matrix is used in classification to summarize:
A. Autocorrelation at lag 12 only
B. Counts of true positives, false positives, true negatives, and false negatives
C. Only residual sums of squares
D. Cluster centroids only
Correct Answer: B. Counts of true positives, false positives, true negatives, and false negatives
Rationale: From the confusion matrix we derive accuracy, precision, recall/sensitivity, specificity, F1, etc.
12. When classes are highly imbalanced, accuracy can be misleading because:
A. ROC curves cannot be drawn
B. Accuracy always equals recall
C. A trivial majority-class classifier can look good while missing the rare but important class
D. Imbalance makes precision undefined always
Correct Answer: C. A trivial majority-class classifier can look good while missing the rare but important
class
Rationale: Prefer precision/recall, F1, PR curves, balanced accuracy, or cost-sensitive metrics when class rarity matters (e.g.,
fraud, disease).
PART B — Classification: k-NN & Logistic Regression (Q13–25)
13. k-Nearest Neighbors classifies a new point by:
A. Maximizing a likelihood with no distance concept
B. Looking at the labels of the k closest training points (e.g., majority vote) in feature space
C. Fitting a global linear hyperplane only
D. Partitioning via recursive impurity splits only
Correct Answer: B. Looking at the labels of the k closest training points (e.g., majority vote) in feature space
Rationale: k-NN is instance-based/lazy: prediction uses local neighborhood. Choice of k, distance metric, and scaling are critical.
14. In k-NN, choosing a very small k (e.g., k=1) tends to:
A. Make the decision boundary perfectly linear
B. Remove the need for feature scaling
C. Increase variance and sensitivity to noise/outliers
D. Always maximize bias and underfit severely as the only effect
Correct Answer: C. Increase variance and sensitivity to noise/outliers
Rationale: Small k → flexible, jagged boundaries (high variance). Large k → smoother, more biased boundaries.
15. A practical reason to scale features before k-NN is that:
