Which baselines must a model beat, and why are they so hard to beat?

The mandatory set: most popular in a window, recently viewed by this user, similar to the last item (item-to-item), and random as a floor. Popularity is hard to beat because it exploits genuine demand statistics and is self-reinforcing: the system shows popular items, which then collect even more clicks. Recency is hard to beat because in most products returning to something already seen is the single most common behaviour. If a sophisticated model cannot beat these on an honest temporal split, the problem is usually the framing or a leak, not the model.