How many instances should be synthesized for each class when using over-sampling techniques?

As for an imbalanced multi-class dataset, how many instances should be synthesized for each class if we use over-sampling techniques such as SMOTE?

For example, there is 4 class including ‘A’, ‘B’, ‘C’ and ‘D’. The number of samples for each class is as follows:

• A: 5
• B: 50
• C: 100
• D: 500

Are the following methods meaningful?

1. Make all class contain 500 instances.
2. Set a threshold and over-sample the class number of whose instances is smaller than it. For example, set threshold = 100, and make A and B to 100.