I feel like these are just shitty models. A good recommendation model would get features like "is_mom" and learn that "is_mom" is a shitty predictor of relevan

I feel like these are just shitty models. A good recommendation model would get features like "is_mom" and learn that "is_mom" is a shitty predictor of relevance. Similarly with Amazon, products should have some sort of 'elasticity' score where it should learn that recommendations of inelastic products is a waste of screen real-estate. I mean, I doubt the model is giving a high % to most of those recommends - it's likely more a business/UX issue in that they've decided it's worth showing you low-probability recommends instead of a cleaner page (or something more useful). Youtube, on the hand, seems to be precision tuned to get you to watch easy to digest crap. You consume the crap voraciously but are generally left unfulfilled. This is a more difficult problem where you're rewarding naive views rather than a more difficult to discern 'intrinsic value' metric. As a 'long term' business play the model should probably weight more intellectually challenging content just like fast food restaurants should probably figure out how to sell healthier food because by pedaling crap you're only meeting consumer's immediate needs, not their long term ones.