suggestion to implement undersample and oversample

Question

My dataset has the following class distribution

CLASS      FREQUENCY
2          22696
4          2541
1          2093
5          1298
3          1116
0          960
6          14

definitely I would want to generate a new sample, for that I will use python imblearn , I have three options:

1. oversample the minority classes
2. undersample the majority classes
3. choose a median class and apply both undersample the majority classes and oversamples the minority classes to equal the median class.

Later I will use the generated dataset to train three estimators RandomForest classifier, SVC and ensemble of both. I will choose the one with the best f-1 score. what would be the best option and which oversampling/undersampling algo should I use?

Answer 1

In general, undersampling is not very useful because you are throwing any information about the features.

SMOTE - Synthetic Minority Over-sampling Technique is useful for oversampling.

You might want to think about dropping Class #6. 14 is not a large enough sample for machine learning.