Free preview

Cold Start

In one line: user cold start solves itself within a session, item cold start does not solve itself at all, and treating them as one problem gets the second one wrong.

Two problems, one name

A new user has no history. You know almost nothing about their taste, and every interaction they have teaches you something. The problem shrinks with every action — it is an information gap that closes on its own.

A new item has no interactions. Its collaborative embedding is untrained, so the strongest source cannot retrieve it, so it is not shown, so it gets no interactions. The problem does not shrink. It is self-sustaining, and it is the feedback loop's lower branch in its purest form.

Which is why item cold start is the business problem. A platform that cannot get new supply in front of anyone loses its suppliers, and then it has no new supply to fail at.

New users

Four mechanisms, roughly in order of what they cost.

Population defaults. Show what is broadly popular and safe. Weak, instant, and the right starting point — it is a floor, not a strategy.

Context. Device, locale, time of day, referrer, entry page. A user arriving from a search for a specific topic has told you a great deal before doing anything.

Onboarding. Ask directly. Pick a few topics, follow a few creators. Costs friction and completion rate, and it converts several sessions of learning into thirty seconds.

Fast in-session adaptation. The one that matters most and gets designed least. The first few actions in a session carry more information than anything else you have, so the system must react to them within the session — not at the next nightly retrain.

This has a real architectural consequence: session-scoped signals feed the ranker at request time, not through the training pipeline. The model does not need retraining; the features need to be fresh.

New items

Harder, and it needs a deliberate mechanism rather than a fallback.

Content-based retrieval from day one. Metadata, text, thumbnail, creator. An item's content embedding exists the moment it is created, so a content source can retrieve it before any interaction has happened. This is the main reason the candidate stage is a portfolio.

Creator and category priors. A new item from a creator with an established audience is not truly cold — it inherits a prior. Most new items are not from unknown creators.

A guaranteed exposure budget. The structural fix: reserve slots for items below an interaction threshold. Not a boost in the ranking — a quota, before ranking, because a boost still competes against items with real evidence and loses.

Graduate on evidence. Once an item has enough impressions to estimate its quality, it leaves the cold pool and competes normally. The threshold is a real parameter: too low and noise promotes bad items, too high and the budget is spent on items you already understand.

The cost, said honestly

Every cold-start slot is a slot not given to something with known-good expected value. That is a real, measurable engagement cost, paid now, for a benefit that arrives later and is harder to attribute.

This is why cold start gets underfunded: the cost lands on this quarter's metrics and the benefit lands on next year's supply. A candidate who names that tension — rather than presenting cold start as free — is describing a system somebody actually has to run.

The framing that makes it fundable is to treat the budget as a fixed percentage of impressions, agreed in advance, rather than something the ranking decides. A ranking optimising for engagement will always spend zero on cold start, correctly, by its own objective.

The two-sided version

On a marketplace, item cold start is supply retention. A creator whose first several posts reach nobody concludes the platform does not work for them and leaves. Their departure is invisible in every engagement metric — no user noticed a video that was never made.

Published work on centralised exploration frames this directly as accelerating audience building for new creators, and treats it as platform infrastructure rather than a per-recommender concern. Worth mentioning: it says the problem belongs at the platform level, not inside one ranker.

The metric that makes it visible: time from item creation to Nth impression, and the fraction of new items that never reach an audience at all. Neither appears on a standard recommender dashboard, and both are the thing the business cares about.

The one-line distinction

If an interviewer asks about cold start and you only discuss new users, you have answered the easy half. The sentence that shows you know the difference:

A new user is an information gap that closes on its own within a session. A new item is a structural gap that never closes without a deliberate exposure budget, because the thing that would generate its evidence is the thing its lack of evidence prevents.

Key takeaway

New users and new items are opposite problems. The user gap closes with every action, so the lever is in-session feature freshness rather than retraining. The item gap sustains itself, so it needs a guaranteed exposure quota before ranking — a boost loses to items with real evidence — plus a graduation rule and honest accounting for the engagement it costs today.

Next: exploration, and how to budget it.

Enjoying the preview?

Create a free account to unlock the rest of this course, the in-browser judge, and live AI mock interviews.

Sign up free to continue