.

i would not do that : imagine you have images and augment the minority class with some kind of blurring filter ; unless you are sure that similar blurring "naturally" occurs in your majority class, your new training set will be close to useless (blurred image => minority class, whatever the image content)

the best way to be sure that "similar blurring occurs in the majority class" is to not rely on this happening "naturally" (unless you have very strong prior knowledge about that), but to apply the blurring filter to the majority class too

.

Its a really common problem in the natural data set. If you want to have balance results. You can under-sample or oversample the train data. Under sample is you take the lowest class and reduce the other class to that number. Oversample is the opposite of taking the highest and create a copy of the other class. Sometimes noise or other thing is added to the signal to give model little bit different input of the same data. In the case of image rotation, blurring changing the contrast could be done. Do not do it for testing because that will change the natural distribution of the data. Another way around it is by using the F1 score or weighted F1 score for the deciding factor of the model.

It really depends on the nature of your data and the problem you are precisely working on. However, in general, if your training samples are majorly containing samples from only 1 class, then the model would not really generalise, but would memorise. There is a danger of the model to introduce extreme bias towards the class with majority samples. What I would do in this scenario is to oversample the minority data points, holding out a specific unchanged portion for testing. Thus, the new training samples would have a balance in the 2 classes, and at the same time, you wouldn't risk making predictions on oversampled data, but on the data which you already separated for testing purposes before the oversampling. This may not be the best way, but is definitely one way we could try approaching the situation. Hope this helps.

Hi Friend, it depends on the nature of the problem you are solving and the percentage distribution of minority and majority class. For instance, I am currently undergoing a similar research with imbalance data set with distribution ratio of majority and minority class 99:1 respectively. In this case, I only applied SMOTE to the minority class because training with the raw data may introduce bias into the result.

Hi, the reason of imbalanced learning is about not only sample size but also learning ability. Please read the following paper:

AdaBalGAN: An Improved Generative Adversarial Network With Imbalanced Learning for Wafer Defective Pattern Recognition

You do not need to do this unless your model (incorrectly) learns to always predict the majority class. However, if the resulting model is biased in this way, then there are several techniques to deal with the problem. Data augmentation is one of several techniques on the data side, which include undersampling the majority class, over-sampling the minority class, etc.. Alternatively, you could compensate for the problem in the cost function (cost-sensitive learning), where you would penalize the wrong prediction of the minority class more than the majority class. Another way to deal with the class-imbalance issue is to design the classification algorithm in such a way that the class imbalance problem does not affect it's learning. So the model would be immune to class imbalance by design.

How fair would a model be if you over-sampled the minority class? First of all, this depends on the model. Let's use an example. If your data is univariate and normally distributed and you model it using a univariate Gaussian by learning it's parameters from the data. So long as you learned the correct mean and variance and corrected for the actual class-prior at prediction time, your model would perform optimally. Another issue with over-sampling has to do with generalization. If you don't have enough samples of the minority class to represent the underlying distribution, no matter what you do on the data side, your model will not learn the correct parameters, unless you get very lucky and artificially change the training data in such a way that the resulting data represents the actual distribution well.

if you don't have to use it try to make under sampling on the other classes

I don't really know the probable answers but you can study Generative Adversarial Network(GAN) that generates samples for image data.