"Databases of offenders" is different from "database of people who haven't yet done anything but statistically we think might". I could be reading it wrong, but it sounded to me like the parent comment was talking about predicting people who might post such videos _before_ they did. I don't think you'd need to theorize about using data science to ban people after they post some number of violating videos.
Social media companies can be passive, reactive, proactive, or predictive in their efforts to ban bad actors. How much effort are they putting in at each level? It seems more towards the least effort.
