I have an imbalanced class dataset (0.3%-99.7%), I have used SMOTE (https://arxiv.org/pdf/1106.1813.pdf) to create a proportion of 30%-70%.
But in the SMOTE dataset my product-segment mapping seems to have changed. Earlier there was 1:1 mapping but then it changed to many:many.
- Will synthetic samples not retain original data properties?
- Is there a way to main original data mapping while using SMOTE?
- If the mapping gets changed, will classifier give correct predictions?
- As my minority class proportion is very small, what would be a ideal proportion criteria ?
Any help is appreciated. Thanks!