New-Item Cold Start
A recommender learns an item's embedding from the people who interacted with it. A film uploaded ten minutes ago has no such people, so there is nothing to learn from — and the model will score it near the bottom of every list, which is what keeps it there.
Why the loop closes on itself
This is new-item cold start, and the damaging part is the feedback loop. Collaborative filtering can only rank what it has evidence about; an item with no interactions carries a default or near-random embedding, never enters a top-N list, and therefore never collects the interactions that would fix the embedding. Catalogues where new supply arrives constantly — marketplace listings, news, user-uploaded video, podcast episodes — hold most of their inventory in this state at any moment.
How systems break the loop
- Content features: embed the item from its metadata, text or thumbnail so a tower can place it near similar items on day one — a main reason production retrieval is a two-tower model rather than pure ID factorization.
- Deliberate exploration: reserve a slice of impressions for under-observed items, borrowing the bandit framing to buy information instead of clicks.
- Fallback placement: editorial or popularity-based slots until an item clears an interaction threshold.
What it costs when ignored
Offline evaluation hides the problem completely: held-out test items are by construction items that already have interactions, so a model can look excellent while burying everything new. The visible damage is on the supply side, where sellers and creators whose items get no distribution leave — which is why fresh-item coverage is tracked as its own metric alongside accuracy.
New-item cold start is the inability to rank an item that has no interaction history, because collaborative filtering derives an item's representation from the users who engaged with it. Unranked items collect no feedback, so the gap sustains itself. Systems break the loop with content-based features that embed an item from its metadata, or with reserved exploration traffic. Without that, fresh inventory stays invisible however accurate the model looks offline.
A bookshop assistant who has worked there for years can tell you which regulars will love a particular novel, because she has watched people buy it. Hand her a book published this morning and she has nothing to go on.
That is new-item cold start. Recommendation software works out what an item is like from the people who clicked, watched or bought it, so a brand-new item starts out invisible. Worse, being invisible means nobody clicks it, so it stays invisible. Systems get around this by reading the item itself — its title, description, cover picture — or by deliberately showing new things to a few people to see what happens.