Catalog

Before a recommender can choose anything, something has to define what there is to choose from. The catalogue is that set: every item the system is permitted to put in front of a user.

A scoring universe, not an inventory list
The word carries a narrow sense here. The catalogue is the universe a model's scores are defined over, and two of its properties shape every design decision that follows. Its size decides whether exhaustive scoring is even possible — nobody scores fifty million videos per request. Its composition changes continuously, with items arriving, selling out and being withdrawn, so some share of it always has no interaction history at all.

What follows from it
The catalogue is the origin of the field's characteristic problems:

• Scale forces the retrieve-then-rank split — candidate generation narrows millions of items to hundreds before a ranking model sees them.
• Churn makes new-item cold start a permanent condition rather than a launch-day inconvenience.
• Skewed interaction counts create the long tail, and coverage metrics exist to ask what fraction of the catalogue a system ever recommends at all.

The eligible catalogue
In production the number that matters is rarely the headline count. Availability, region, licensing and business rules filter the catalogue per request, so the set a ranker actually chooses from is smaller and differs between users. That gap is worth remembering when an offline experiment on a fixed MovieLens item set looks healthier than the live system it was supposed to predict.