iav ML Study Group Pop Quiz 2

K-Nearest Neighbours, Random Forests and Gradient Boosting are all algorithms which...

• ...can only be used for Classification if there are fewer features than samples

• ...can only be used for Regression if there are fewer features than samples

• ...can and will be used for both Regression and Classification

• ...are only academic theories, and are not practical for real ML solutions

ROC stands for

• Receiver Operating Curve

• the mythical Roc, a giant bird from the legend of Sinbad the sailor

• Regressive Orthogonal Curve

• Random Omissions Cumulative

For a dataset with IMBALANCED classes, we would generally choose as our error metric:

• Precision-Recall Curve, because we would expect a trade-off between precision and recall

• ROC Curve, because it is accepted industry practice

• Parabolic Curve, because the data is frequently quadratic

• Straight-Line Entropy, because we want the most accurate measurement of error

How do we determine the correct value of K for KNN?

• Trick Question: KNN provides this value.

• As a hyperparameter, K is extrinsic to the dataset and must be arrived at via an iterative series of experiments.

• As a parameter, K is intrinsic to the dataset and can simply be read out of it.

• K can be derived synthetically from other statistical measures like sigma.

We can describe upsampling and downsampling for imbalanced datasets as follows (pick all correct options):

• Downsampling adds importance to the MINOR class, sending recall up and precision down

• Downsampling adds importance to the MAJOR class, sending recall down and precision up

• Upsampling will mitigate excessive weight on the MINOR class. Recall will still be higher than precision, but the gap will lessen

• Upsampling will mitigate excessive weight on the MAJOR class. Precision will still be higher than recall, but the gap will lessen

These are the four steps in blagging, possibly out-of-order:

1) BALANCE each sample by downsampling

2) MAJORITY vote

3) BOOTSTRAP samples from the population

4) LEARN a decision tree

To answer the question, put all 4 steps in order with no spaces in the box e.g. "6958"

The broadest and most flexible ensemble method, which involves the mixing of an arbitrary number of models without any limits on their types is called:

• Piling

• Stacking

• Chaining

• Combining

The terminal points of a decision tree are called:

• Leaves

• Nodes

• Branches

• Roots

A decision tree will continue splitting until

• The nodes are balanced around the middle of the tree

• All leaf nodes are PURE, containing only one class

• A maximum depth, set as a hyperparameter, is achieved

• A certain performance metric, set as a hyperparameter, is achieved

The formula for the entropy equation which has be minimized to determine the best split for each node in a decision tree is

• 
• 
• 
•