A k$-$nearest neighbour algorithm is used to develop a classifier. The data set contains noise. Which of the

following strategies will help to reduce sensitivity to noise? Simply write down the letter$($s$)$ of the correct

approach$($es$).$ Note that incorrect answers will be penalized.

$($a$)$ Reduce the value of $k$

$($b$)$ Use a larger value for $k$

$($c$)$ Find all Tomek links and remove both instances that form the Tomek link

$($d$)$ Nothing can be done to reduce sensitivity to noise

$($e$)$ Use SMOTE to oversample the minority class

Consider a regression problem, where the data set has the following characteristics:

There are $185$ instances

There is one categorical and five numeric descriptive features

The categorical feature has three possible values. One of these values occurs for $70\%$ of the instances,

and the other two values occur respectively for $13\%$ and $17\%$ instances

Four of the numeric features have values in the range $0,1,$ and the fifth numeric feature has values

in the range $100,100000$

For $1\%$ of the instances there are outliers for the target feature

One of the numeric descriptive features has a few outliers

One of the numeric descriptive features has $3\%$ missing values

A k$-$nearest neighbour algorithm is used and Euclidean distance is used as the similarity measure. Which

of the following statements are correct? Simply write down the letter$($s$)$ of the correct statement$($s$).$ Be

careful; marks will be subtracted for incorrect answers.

$($a$)$ The value for $k$ has to be large

$($b$)$ One$-$hot encoding has to be applied to the categorical feature

$($c$)$ The numerical features have to be scaled to the same range, e$.$g$.0,10$

$($d$)$ The outliers in the numeric descriptive feature have to be removed

$($e$)$ The missing values have to be imputed

$($f$)$ Under$-$sampling or over$-$sampling has to be applied to the categorical feature to balance the distri$-$

bution of possible values

$($g$)$ The predicted value is calculated using the average over the target values of the neighbors

What is the inductive bias of the $k-$nearest neighbour algorithm?

Is it necessary to normalize input features when a $k-$nearest neighbour algorithm is used? Motivate your

answer.

Explain how $k-$nearest neighbours can be used to impute missing values.

Can the $k-$nearest neighbour algorithm be applied to problems with categorical descriptive features?

Motivate your answer.

Discuss the consequences of different values for $k$ when $k-$nearest neigbours is applied to regression prob$-$

lerns.

AML$874$: Describe how $k-$nearest neighbours can be used to classify images.

AML$874$: Describe how $k-$nearest neighbours can be used to correct misspelt words using an example. How

would you decide how to modify the original word if $k$ is equal to three?