Unbalanced data in multi label problem

1

hi, I have a multi-label problem, and they are very rare classes, and on the other hand, some very common classes.
I have a custom loss
I need o use all my data
how can I balance this with my loss?
tnx

2

There are some sample weighting techniques for such data.
See: https://arxiv.org/abs/1803.09050

This is in fact something I personally implemented a few days ago. Maybe my implementation can help you: https://github.com/TinfoilHat0/Learning-to-Reweight-Examples-for-Robust-Deep-Learning-with-PyTorch-Higher

3

tnx, but i dont see there reference to the multi label problem

4

Hi what do you mean? in the experiments, they have an example on mnist where frequency of samples is heavily skewed towards digit 9. isnt that like your scenario?

5

oh i see. yeah in that case each sample is associated with only a single label, different than your scenario.