Does there exist a method can generally oversample training data?
As I know, when we want to oversample a training dataset, we need to have some basic understanding about features being used. Take processing images as an example, if we want to oversample an image, we can add noises on it, rotate it, or resize it. We can't oversample an image without knowing it is an image and apply oversample technics on it.
Now, I have a training dataset that is not 'raw'. This dataset is derived from applying some feature extraction technics to a 'raw' data. But I can't access on 'raw' data.
In this condition, is this possible to oversample this training dataset just by observing its attribute (value/distribution)?


Yes, there are numerous approaches for oversampling, as well as undersampling. There are simple approaches, such as randomly re-sampling (or dropping) cases, as well as more sophisticated methods like ROSE (Random Over Sampling Examples) and SMOTE (Synthetic Minority Over Sampling Technique), but none are without disadvantages. Just as with images, oversampling comes with risks that must be weighed against the potential advantage of equating class balance. Sampling methods are often not necessary, or can be dealt with, unless your class imbalance is quite severe. After attempting some training on your unaltered dataset, do you have reason to believe that class imbalance is preventing a stable solution?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.