I am building classification model for bio (scRNA) data. Datasets in this field, for example, dataset A has 1, 2 classes, dataset B has 2, 3 classes kind of that. So I integrated datasets for training dataset, which is imbalanced. (class A has 52% proportion(which is the highest), some other classes have under 1% of proportion.)

So I am trying to do some augmentation right now,but the worst part is that I don't know what distribution the data you need to test or predict has. (like dataset A, B above)

I was taught that the distribution of the training dataset doesn't matter, but the distribution of the validation and test data should be the same.

Is there any method or paper for this kind of problem?

  • $\begingroup$ Welcome to Data Science! Try and provide some more context in your question. Describe the input data, how many sets, and classes? What are you trying to do? Why are you combining the datasets? $\endgroup$
    – fswings
    Oct 26, 2023 at 20:09
  • $\begingroup$ The question as written implies that you had no choice in the distribution of the test and validation set. You alone have to choose what goes in each one. So perhaps reframe the question to make it clearer what your after. $\endgroup$
    – fswings
    Oct 26, 2023 at 20:11


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.