O(D)

O(N × D)

O(N)

O(K × D)

The Decision Tree will crash if class proportions differ between train and test

Stratified splitting always produces higher accuracy than random splitting

The test set may contain very few or zero positive examples, making performance metrics unreliable or undefined

No risk — random splitting is always valid regardless of class distribution

It depends on the number of samples

0

1

0.5

There is an error in the computation

Splitting on that attribute does not reduce entropy at all

The attribute has only one unique value

The attribute perfectly separates the classes

K=1 is the better model because training accuracy is more reliable than test accuracy

K=25 is underfitting because training accuracy dropped below 90%

Both models have the same bias-variance profile since both are KNN

K=1 is overfitting; K=25 has better generalisation but may still be suboptimal

0.0

1.0

0.5

2.0

Use a wrapper method: iteratively evaluate subsets of features using KNN's cross-validated accuracy as the scoring function, selecting the subset that maximises performance

Remove features one at a time based on their individual correlation with the target

Use PCA to reduce to 2 components regardless of explained variance, then run KNN

Keep all 50 features but increase K to compensate for noise

K=5, because it achieves the highest cross-validated accuracy

K=1, because it has the lowest bias

K=3, because odd numbers are always preferred

K=9, because larger K always generalises better