5.2 Interpretation

Question 1

Consider the skin cancer screening test which is a binary classification problem. We record two binary features from 1000 patients, and we build some models with the following approaches. Which method gives the model with the largest size?

Answer 1

The model of Nearest Neighbour Classifiers is just the data set itself! For this problem with only two binary features and 1000 instance this model is the largest.

For 0-R the model is simply the label of the most frequent class (O(1)). The model in Naive Bayes is a set of prior probabilities and conditional probabilities (O(2+2*2)). In SVM model is the maximum-margin hyperplane, defined by w and b (O(2+1)).

Question 2

Which of the following are hyperparameters of the associated approach (select all hyperparameters)?

Answer 2

Refer to the lecture note. Note that the

1. k for Nearest Neighbour,
2. the kernel in SVM
3. the stopping criterion for decision tree

act as hyperparameters, which are tuneable “knobs”.

Question 3

Which statements are incorrect? (select all that are incorrect)

Answer 3

The model on 0-R is simply the label of the most frequent class which does not deliver any information about the features.

The weight of a zero feature does not play any role in the classification by SVM.

Question 4

Which statements are incorrect? (select all that are incorrect)

Answer 4

PCA is an unsupervised approach.

Question 5

What is a nearest prototype classifier?

Answer 5

Calculate the centroid of each class, and classify each test instance according to the class of the centroid it is nearest to