How Do You Build Balanced Sampling for a Long-Tail Video Dataset?

I am working on a video dataset with a long-tail distribution across topics, languages, creators, and source platforms. The current concern is that random sampling may overrepresent popular categories and creators while leaving rare but important groups underrepresented. I am considering a sampling pipeline with: Multi-label taxonomy fields Per-category sample counts Language-aware sampling Creator or channel caps Platform-level balancing Weighted sampling during training A separate coverage...