Odd K reduces computational cost

To avoid tied votes between the two classes

Even values of K cause the algorithm to crash

Odd values always produce higher accuracy

The Decision Tree will be robust because it considers global structure via information gain

The Decision Tree will perform worse because it cannot handle continuous features with noise

KNN with K=1 will be robust to noise because it looks at local structure

Both will perform poorly because both will memorise the noisy labels

Diagonal lines at arbitrary angles through the feature space

Axis-parallel (rectilinear) rectangles, because each split thresholds on a single feature

Smooth curved boundaries similar to SVMs with RBF kernels

Circular boundaries centred on cluster centroids

0.14

0.398

0.602

0.86

Train multiple Decision Trees on different bootstrap samples of the data and average their predictions (bagging)

Train one very deep Decision Tree and increase min_samples_leaf to 1

Increase the number of features used at each split to include all available features

Train one Decision Tree and use it to generate synthetic training data for a second tree

Red — because point A is the closest

Blue — because 2 out of 3 nearest neighbours are Blue

Red — because the two nearest red points outweigh the single blue

Cannot be determined without normalisation

Normalise all features to speed up Euclidean distance calculations

Switch to Manhattan distance because it's faster than Euclidean

Use an approximate nearest neighbour (ANN) structure like KD-trees or locality-sensitive hashing to avoid brute-force distance computation

Set K=1 to minimise computation during prediction

KNN — because showing similar past patients is more intuitive

KNN — because it has no parameters to explain

Neither — both models are equally interpretable

Decision Tree — its if-then rules can be directly explained as a flowchart of questions and answers

0.48

0.36

0.97

0.50

Assign arbitrary numeric codes (A=1, B=2, AB=3, O=4) and use Euclidean distance

Remove the categorical feature and use only the numeric ones

One-hot encode Blood Type and use standard Euclidean distance on everything

Use a composite distance: normalised Euclidean for numeric features and Hamming distance for categorical features, combined with appropriate weighting