The cold start problem is what happens when a recommendation or personalization system encounters a user, item, or entire product with no interaction history to learn from. Collaborative filtering and most behavior-driven models are useless without data — and cold start is the name for the gap before that data exists.
What changed in 2026
- LLM-generated item embeddings became a common cold-start fix for new items, using product descriptions or content text to place a brand-new item into a useful position in embedding space before any interaction data exists.
- Progressive onboarding flows got shorter, as products learned that long preference surveys hurt signup completion more than the cold-start benefit was worth — a few well-chosen questions now outperform ten generic ones.
- Contextual bandits saw wider adoption specifically for the cold-start window, actively exploring what a new user likes rather than passively waiting for enough organic interaction data to accumulate.
The three types of cold start
New user cold start. A user just signed up. The system knows nothing about their taste beyond whatever context is available — device, location, referral source, or answers to onboarding questions.
New item cold start. An item was just added to the catalog. No one has interacted with it yet, so collaborative signals are empty, even if plenty of users would love it.
New system (or new product) cold start. The whole product just launched. There is no interaction data for anyone or anything, which is the hardest version of the problem and usually requires leaning entirely on content-based or rule-based logic at first.
Practical fixes compared
| Approach |
Solves |
Tradeoff |
| Content-based fallback |
New item cold start |
Needs good metadata; misses behavioral nuance |
| Onboarding preference capture |
New user cold start |
Adds signup friction; users answer imperfectly |
| Popularity / editorial fallback |
New system cold start |
Not personalized, but rarely bad |
| Contextual bandits |
New user, ongoing exploration |
Requires careful reward design and traffic to explore with |
| Cross-domain transfer (e.g. from another product or partner data) |
New user cold start |
Needs a data-sharing relationship and consistent identity |
Why cold start never fully disappears
It is tempting to treat cold start as a bootstrapping problem that resolves once a product has enough users and history. In practice, any product that keeps growing has a constant stream of new users and, often, new items — cold start is a steady-state cost, not a one-time hurdle. The goal is not to eliminate it but to make the cold experience good enough that users stick around long enough to generate the data that makes personalization work.
How to reduce cold-start pain in practice
- Ask a small number of high-signal onboarding questions, not an exhaustive preference survey — three to five well-chosen questions typically capture most of the achievable benefit.
- Use content metadata aggressively for new items so they are discoverable through relevant categories, tags, or similarity to established items, rather than invisible until they accumulate interactions.
- Default to popularity or editorial curation, not a blank personalized slate, for genuinely unknown users — a reasonable generic recommendation beats a bad personalized guess.
- Instrument the first-session experience separately from steady-state metrics, since a cold-start user's early behavior is a weaker signal than an established user's and should be weighted accordingly in your models.
FAQ
Does more training data eliminate the cold start problem?
It reduces cold start's severity for the system overall but does not eliminate it, because new users and new items keep arriving. It is a permanent edge case, not a temporary one.
Is cold start unique to recommendation systems?
No — any personalization or prediction system that depends on user- or item-specific history has some version of it, including search ranking, fraud detection, and pricing models.
What is the single highest-leverage fix for new item cold start?
Rich, structured item metadata. It is what content-based fallbacks and LLM-derived embeddings both depend on, and it is usually cheaper to invest in than a more complex modeling fix.
Where to go next