Let's say I have a binary classification dataset skewed towards negative samples. Let's say it's 1000 positives and 100000 negatives. Let's say it's image data.

I'm training a classifier and I'm using data augmentation on positives to attenuate the imbalance issue.

At each training epoch I sample my positives from a pool of random transforms, some with random parameters itself (e.g. horizontal flip, vertical flip, random rotation, random zoom, random crop).

I also sample an equivalent number of negatives from the whole negative pool. So at each epoch I have the same number of positive and negative examples.

I can tell from validation the thing is somewhat working, but how do I estimate how much imbalance am I effectively removing?

Some source says I apply N transforms to positives so I am increasing the positive set size N times. But each transform is random itself and it returns a different (albeit correlated) sample each time, so the effective set is bigger... but how much?

Maybe I could compare variances of the whole negative and transformed positive pool?


Your Answer

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

Browse other questions tagged or ask your own question.