Ensemble Methods - Other Modeling Aspects

Question 1

Binary Target

Answer 1

When there is a disproportionate number of observations in one of the classes, a classification model will likely struggle to make proper predictions for the minority class. Thus, in light of sensitivity and specificity, such models tend to have a high value for one of the metrics and a low value for the other one.

Question 2

Oversampling

Answer 2

Sampling technique for duplicating observations in the minority class to reduce the imbalance.

Only apply to training set

possible to overfit the minority class since duplicating observations

Question 3

Undersampling

Answer 3

Sampling technique for detecting observations in the majority class to reduce the imbalance.

Only apply to training set

possible to underfit the majority class since observations are removed -> can lose useful info with a smaller dataset

Question 4

Oversampling and Undersampling for cv

Answer 4

k-fold cv -> divide the training ste into k groups.

perform over/under on all k training sets before fitting the model