Common issue on One-vs-All approach
Balcanced classes do not mean having the same number of examples in all classes
- For example, a problem with a number of 9:1 for one of the classes. We probably want to do something to balance the classes, but we probably do not want the classes to have the same number (1:1), because in the real world the problem is probably something similar to 9:1. We can try to balance by 2:1.
Fixing by sampling
- Uses the notion of distance (cannot be applied to feature vectors with factor features)
- Ensemble methods (ensemble-learning)
- Data Augmentation