It's a direct reflection on what tiktok's algorithms think they know about you so far, modulo what they think they know about the contents of the videos they are showing. They have a good recommendation engine, sure, but it works on average over large populations through the limited funnel of video interactions, and their video understanding and inventory is similarly limited. This is even before considering that they clearly add some kind of extra exploratory weight to new content they don't have a lot of data about from you.
Their machine learning is basically fancy statistics on watch data, not a peering-into-your-soul Oracle, it's very possible it gets some people's preferences very wrong and is still profitable for larger population segments.