Operating recommenders in production.
WHAT THE SERVING PATTERN USUALLY IS
Candidate generation, then ranking.
WHY TWO STAGES
Scoring every item for every user is infeasible; a cheap stage narrows to hundreds, and an expensive stage orders them.
WHAT CANDIDATE GENERATION USES
Similarity retrieval Popularity Recent behaviour Business rules
WHAT RANKING USES
A model scoring each candidate for the specific user and context.
WHAT TO PRECOMPUTE
User and item representations, refreshed on a schedule.
WHAT TO COMPUTE AT REQUEST TIME
Only what depends on the current context.
WHAT TO HANDLE
New users with no history New items with no interactions Items becoming unavailable
WHY THAT LAST POINT
Recommending something out of stock is worse than recommending nothing.
WHAT TO APPLY AFTER RANKING
Filtering for availability and eligibility Diversity constraints Business rules
WHY AFTER
They are constraints, not learned preferences.
WHAT TO MONITOR
Coverage of the catalogue Click and conversion rates Whether the same items dominate
WHAT CATALOGUE CONCENTRATION INDICATES
A feedback loop narrowing what anyone sees.
