This post has some suggestions for different sampling implementation when dealing with an imbalanced multi-label dataset.
