I am working on a multilabel classification in which I am having 206 labels. When I saw the percentage of the number of 1's in each label they are way less than 0.1% for each label. The maximum percentage of ones in labels is 0.034%.
Below is the distribution of percentage of one's in each labels If I simply build a multilabel classification single model. The score it gives may be high but it got biased towards zeros very much so, it doesn't give probability of a label to be one very high. And if I want to build for each label different model, I can treat it as a bunch of imbalanced data and apply smote algorithm to each model, But I have a doubt whether can smote produce a good amount of data to balance because we know how imbalance my data is. Now, doubt is can I gave a try to autoencoders, which I heard good at fraud detection when the data is having a percentage of one's less than 1% or such. Will it perform better in my case? because if it can work well, then I will study autoencoders.