Classification, Trees, Neighbours and Clusters

Section A

Circle the letter of the correct answer. One mark each.

1. 1What is a decision boundary?1
   1. AThe edge between regions a classifier assigns to different classes
   2. BThe point where training stops improving
   3. CThe split between training and test data
   4. DThe largest value a feature is allowed to take
2. 2In kNN with k = 5, a new point's nearest neighbours are 3 cats and 2 dogs. What is predicted?1
   1. A60% cat
   2. BA tie, so no prediction
   3. CDog
   4. DCat
3. 3What is the risk of using k = 1 in k-nearest neighbours?1
   1. AFeatures no longer need to be scaled
   2. BEvery point is given the most common class
   3. COne mislabelled example creates wrong answers around it
   4. DThe model becomes far too slow to train
4. 4What does the sigmoid function do in logistic regression?1
   1. ASplits the data into a training set and a test set
   2. BSquashes a score into a probability between 0 and 1
   3. CFinds the k nearest training examples to a point
   4. DRemoves features that the model does not need
5. 5A group contains 3 apples and 1 orange. What is its Gini impurity?1
   1. A0.25
   2. B0.375
   3. C0.5
   4. D0.75
6. 6How does a random forest reduce overfitting?1
   1. AIt grows a single tree much deeper than usual
   2. BIt removes every feature except the most useful one
   3. CMany trees trained on random samples vote, so their errors cancel
   4. DIt trains one tree on the test set as well
7. 7A model finds 40 of the 50 real fraud cases and flags 10 genuine payments by mistake. What is its recall?1
   1. A20%
   2. B40%
   3. C50%
   4. D80%
8. 8For a cancer screening test, which mistake is usually worse, and so which measure is favoured?1
   1. AA false negative, so recall is favoured
   2. BA true negative, so the F1 score is favoured
   3. CA true positive, so accuracy is favoured
   4. DA false positive, so precision is favoured
9. 9In k-means, what happens in the update step?1
   1. AEach centroid moves to the mean of the points assigned to it
   2. BThe number of clusters k increases by one
   3. CEach point moves to the centroid it is nearest to
   4. DThe points with the largest distances are deleted
10. 10Why is k-means usually run several times with different starting centroids?1
    1. AIt needs several runs before it can find any clusters
    2. BEach run adds one more cluster to the result
    3. CDifferent starts can settle on different, worse clusterings
    4. DRunning it again makes the clusters more equal in size