! S8: Under- & Oversampling

Question 1

Sampling

Answer 1

= statistical process where subset of data taken taken (to estimate characteristics of the whole population)

Question 2

Undersampling

Answer 2

• taking a subset of the over-represented category in the data

Question 3

Ways of handling unbalanced data sets

Answer 3

• collecting more data
• modifying class weights
• oversampling
• under sampling

Question 4

Oversampling

Answer 4

• uses synthetic data generation to increase the number of under-represented category in data

Question 5

SMOTE - Steps

Answer 5

1. Find kNNs for each sample in minority class
2. Select samples randomly from kNN
3. Find new samples = original sample + distance_difference *number_btw0&1
4. Add new samples to minority class

Question 6

Oversampling techniques

Answer 6

• SMOTE
• ADASYNE

Question 7

ADASYNE

Answer 7

• extension of SMOTE
• Adaptive Synthetic Sampling
• adaptively generating minority data samples according to their density using KNN
• adaptively updates the distribution
• no assumptions made for the underlying distribution of the data
• dynamic adjustment of weight

Question 8

SMOTE - Application

Answer 8

• takes entire dataset but just increases minority class examples
• SMOTE percentage parameter = 100 -> doubles minority class examples

Question 9

SMOTE vs ADASYN

Answer 9

• SMOTE: generated arbitrary nr of new minority eg. -> shift classifier learning bias toward minority class
• ADASYN: generated new minority eg via linear interpolation btw existing minority class -> shift classifier decision boundary to be more focused on those difficult to learn -> improves learning performance
• ADASYN: density distribution = criterion to automatically decide nr of snythetic samples for each minority sample
• SMOTE: generates the same nr of synthetic samples for each sample