Unit 3: SUPERVISED LEARNING: CLASSIFICATION - Practice Quiz

1 What is the main idea behind the -nearest neighbors method?

2 In -nearest neighbors, what does represent?

3 Naïve Bayes is based mainly on which type of probability?

4 What simplifying assumption does Naïve Bayes make about features?

5 What is the basic structure of a decision tree?

6 In a decision tree, what does a leaf usually represent?

7 A decision rule commonly has which form?

8 What does a support vector machine try to find for classification?

9 In an SVM, what is the margin?

10 What does classification accuracy measure?

11 If a classifier makes 90 correct predictions out of 100, what is its accuracy?

12 What does logarithmic loss evaluate?

13 For logarithmic loss, which result is generally better?

14 What does the area under the ROC curve measure?

15 What does an AUC value of usually indicate?

16 Precision answers which question?

17 Recall is also commonly called which of the following?

18 The F1 score combines which two measures?

19 What information is shown in a confusion matrix?

20 What is a true positive?

21 A 3-nearest neighbors classifier finds the three closest training instances with labels A, A, and B. Assuming uniform voting, how will the new instance be classified?

22 A dataset has two features: annual income measured in dollars and age measured in years. Before applying k-nearest neighbors with Euclidean distance, what is the most appropriate preprocessing step?

23 For a message containing words and , a Naïve Bayes classifier uses , , and , where means spam. What quantity is proportional to the posterior probability of spam?

24 In a text classifier, a word appears in a test document but never appeared in the training documents of one class. Which technique prevents that class probability from becoming zero?

25 A decision-tree node contains 40 positive and 40 negative examples. Split X produces two pure child nodes, while Split Y produces two child nodes with the same 50-50 class ratio as the parent. Which split has the larger information gain?

26 A fully grown decision tree achieves 100% training accuracy but performs poorly on validation data. Which modification is most likely to improve generalization?

27 A root-to-leaf path in a decision tree tests Age > 30, then Income > 50000, and predicts Approve. Which rule correctly represents this path?

28 Two linear classifiers separate the training data perfectly. Classifier M has a wider margin between the classes than classifier N. Which classifier would a standard hard-margin support vector machine prefer?

29 A dataset cannot be separated effectively by a straight line in its original feature space. Which SVM approach can create a nonlinear decision boundary without explicitly constructing all transformed features?

30 A classifier produces , , , and . What is its accuracy?

31 In a dataset where 98% of transactions are legitimate, a model predicts every transaction as legitimate and obtains 98% accuracy. What is the main limitation of using accuracy here?

32 For a binary example whose true class is positive, a model predicts a positive-class probability of . Approximately what is the logarithmic loss for this example?

33 Two models correctly classify the same positive example at a threshold of . Model A predicts , while Model B predicts . Which statement about their log losses is correct?

34 A classifier has an ROC AUC of . Which interpretation is most appropriate?

35 Two positive examples receive scores and , while two negative examples receive scores and . Assuming no ties, what is the ROC AUC?

36 A spam filter marks 40 messages as spam. Of these, 30 are actually spam. What is the filter's precision?

37 A medical dataset contains 30 patients with a disease. A classifier correctly identifies 24 of them and misses 6. What is its recall for the disease class?

38 A classifier has precision and recall . What is its F1 score?

39 In a disease-detection confusion matrix, an infected patient is classified as healthy. How is this prediction recorded?

40 The decision threshold of a binary classifier is lowered from to . Assuming all predicted probabilities remain fixed, which change is generally expected in the confusion matrix?

41 A one-dimensional training set contains , , , and . Using 1-nearest-neighbor classification with Euclidean distance, what is the leave-one-out cross-validation accuracy?

42 Independent, irrelevant binary features are appended to data used by a Euclidean -NN classifier. Each added feature is equally likely to match or differ between two observations. What is the most likely asymptotic effect as many such features are added?

43 A binary feature is duplicated exactly, and both copies are supplied to a Naïve Bayes classifier. Assuming no smoothing issues, what is the principal consequence?

44 For classes and , suppose , , , , and . Under the Naïve Bayes assumption, what is ?

45 A node contains positive and negative examples. A split produces children with class counts and , where each pair is (positive, negative). Using base-2 entropy, what is the information gain?

46 A dataset contains all four combinations of two binary attributes , with target , and each combination occurs equally often. What happens if a greedy tree algorithm stops whenever no candidate split has positive information gain?

47 Positive examples are and , while negative examples are and . For a hard-margin linear SVM in canonical form , which parameters define the maximum-margin separator?

48 Every training observation is duplicated exactly times in the soft-margin SVM objective , while is left unchanged. Relative to the original dataset, this is equivalent to using which penalty parameter?

49 A classifier has sensitivity and specificity . It is deployed where the positive-class prevalence is . Assuming these rates remain stable, what is its expected accuracy?

50 For three labels , the true label sets are , , and . Predictions are , , and . What is the multilabel subset accuracy?

51 For binary observations with labels , a model assigns positive-class probabilities . Using natural logarithms, what is the mean logarithmic loss?

52 Among observations with identical features, the true conditional positive rate is . Which constant predicted probability uniquely minimizes the expected binary logarithmic loss for these observations?

53 Three positive examples receive scores , and three negative examples receive scores . Treating a tied positive-negative pair as half a correct ranking, what is the ROC AUC?

54 A strictly increasing nonlinear transformation is applied to every classifier score. No scores were tied originally. Which statement is necessarily true?

55 A classifier has true-positive rate and false-positive rate in a population with positive prevalence . What precision should be expected?

56 At one threshold, a classifier has and precision . After increasing the threshold, true positives decrease by and false positives decrease by . What is the new precision?

57 For a fixed set of continuously scored observations, the positive decision threshold is lowered without changing any scores. Which statement about positive-class recall is necessarily correct?

58 A classifier has precision and F1 score . What must its recall be?

59 A three-class confusion matrix has rows as actual classes and columns as predicted classes: What is the macro-averaged F1 score?

60 A test set has observations and positive prevalence . A classifier has recall and precision . Which confusion matrix counts are correct?