Most of the bang for their buck was when they used a plain old sorting of those ratings. That already hid a lot of the sludge. The ML stuff is just a minor optimization compared to that.

Everything is just my opinion, of course, they're probably following a bunch of metrics I don't have access to and which could prove that the ML effort was worth it.