There are two types of familys for recsys

User-item and item-item, recommendations are often a mix of both, yt in particular has a skew to use user-item based on your history ( maybe it's because they found out that people like to view the same thing or more from creator they already know) but when the mix is skewed liked that you tend to miss on de serendipity of the recsys

It all depends of the target metric that their models where trained, if you look at their papers they said that they want to maximize view time