1.
Explain how sample size influences generalization in supervised
learning.
2. Why does improper feature scaling affect distance-based algorithms
more than tree-based algorithms?
3. Explain how hypothesis restriction reduces search complexity.
4. Describe a real situation where probability estimation improves
classification decisions.
5. Why does normalization not affect decision trees significantly?
6. Explain how model complexity relates to training error trends.
7. Describe a scenario where multi-label classification is necessary.
8. Why is independence assumption sometimes beneficial even if
unrealistic?
9. Explain why KNN does not require explicit training.
10.Describe a case where logistic regression may fail but SVM may
succeed.
11.Explain why decision trees are prone to overfitting.
12.How does adding features sometimes degrade model performance?
13.Why is version space reduced when contradictory examples are
added?
14.Explain how margin maximization improves robustness.
15.Describe the intuition behind least squares minimization.
16.Why does high dimensionality affect distance measures?
17.Explain how bias-variance tradeoff appears graphically.
18.Why does increasing K in KNN increase bias?
19.Describe a practical example of semi-supervised learning.
20.Explain why outlier removal may sometimes harm performance.
4 Marks Q
21.Given a small dataset, explain how you would choose between KNN
and logistic regression.
22.Explain how entropy changes as class distribution becomes balanced.
23.Derive the gradient descent update rule conceptually for minimizing
squared error.
24.Discuss how PAC learning connects probability and generalization.
25.Explain why SVM optimization is a convex problem.
26.Compare hinge loss and log loss in terms of decision boundary
behavior.
27.Explain how multi-layer networks overcome linear separability
limitation.
28.Discuss how cross-validation reduces variance in model evaluation.
29.Construct a small example to demonstrate information gain
calculation.
30.Explain why high correlation between features affects Naive Bayes
performance.
31.Discuss trade-offs between interpretability and accuracy among
regression, trees, and SVM.
32.Explain why kernel trick avoids explicit high-dimensional mapping.
33.Analyze computational complexity of KNN during prediction.
34.Discuss the effect of regularization parameter on decision boundary
shape.
35.Explain why logistic regression outputs probabilities while perceptron
does not.
36.Discuss mathematical meaning of error bounds in learning theory.
37.Explain difference between empirical risk and expected risk.
38.Describe a workflow for selecting hyperparameters systematically.
39.Explain how imbalance in dataset affects evaluation metrics.
40.Compare generative and discriminative classifiers with reasoning.
5 marks Q
1. A dataset is linearly separable but contains noise. Analyze which
algorithm is most suitable and justify mathematically.
2. Given a dataset with overlapping classes, explain how soft margin
SVM handles misclassification.
3. Design a machine learning solution pipeline for spam detection and
justify algorithm choice.
4. Explain mathematically why entropy is maximum when class
probabilities are equal.
5. Given a dataset, manually compute one iteration of gradient descent.
6. Construct a small dataset where Naive Bayes gives incorrect
classification due to feature dependence.
7. Compare logistic regression and SVM from optimization perspective.
8. Analyze how increasing model depth in decision trees affects
variance.
9. Explain kernel trick using geometric interpretation.
10.Derive logistic regression cost function from likelihood principle.
11.Given a small numeric dataset, perform Z-score normalization and
interpret results.
12.Discuss theoretical limitations of PAC learning assumptions.
13.Explain how multi-label classification modifies loss functions.
14.Analyze trade-offs between hard margin and soft margin SVM.
15.Compare instance-based and model-based learning approaches.
16.Given a dataset, show step-by-step how KNN classifies a test point for
two different K values and compare results.
17.Explain how regularization changes weight magnitude
mathematically.
18.Discuss situations where high accuracy does not imply good model
performance.
19.Design a strategy to handle high-dimensional sparse data.
20.Evaluate three supervised learning algorithms for a credit scoring
application and justify selection criteria.