What are common approaches in order to deal with unbalanced multi-label multi-class classification problems in deep learning? Furthermore there is correlation between the labels. I tried two approaches.

  • Using weights for the loss function. I used a sperate weight for each class and label. The weight was indirectly proportional to the frequency of the target class given a label.

  • Training $q$ classifiers for $q$ labels. There I sampled the target label with an equally distributed class distributions. I used weights as weighting between different labels. The target label got a weight higher compared to other labels.

In the beginning I thought that these two approaches should result in similar results. But Approach $2$ performed much better. It would be nice to see either it was just a random result in my experiment or there is more systematic in this kind of problems.

Is there any literature for this problem or does anybody have some experience?

  • $\begingroup$ I don't have any experience with this problem, but you can try the following article (and look for who cites it to find additional papers): link.springer.com/content/pdf/… $\endgroup$
    – Mark.F
    Dec 22, 2018 at 10:58


Your Answer

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