The cold start problem is the difficulty a recommender system faces when it must make suggestions for new users, new items or an entirely new platform that have little or no interaction history. In recommender systems built on collaborative filtering, predictions come from patterns in past behavior, so an entity with no clicks, ratings or purchases gives the model nothing to learn from.
Every production recommender meets this problem continuously, because new people sign up and new products, articles or videos arrive every day. How well a system handles cold start often decides whether new users stay and whether new items ever get discovered.
Why collaborative filtering fails without history
The cold start problem arises because collaborative filtering represents users and items only through their interactions. Matrix factorization, for example, learns an embedding for each user and each item from the observed entries of the user-item matrix. A user or item with no entries receives no meaningful embedding, so its predicted scores are essentially random or default to the global average.
Cold start is closely related to, but different from, the data sparsity problem. Sparsity describes a matrix in which most cells are empty even for established users; cold start describes specific rows or columns that are almost completely empty. A system can have moderate overall sparsity and still suffer severe cold start for its newest entities.
Practitioners usually distinguish three forms, summarized below.
| Form | What is missing | Typical situation |
|---|---|---|
| User cold start | Interaction history of a new user | First visit after sign-up |
| Item cold start | Interactions with a newly added item | New product, song or article |
| System cold start | History for almost everyone | Launch of a new platform or market |
Strategies for new users
Systems address user cold start by substituting other signals for missing behavior until enough interactions accumulate. The simplest fallback is popularity-based ranking, which is safe but generic and reinforces popularity bias. Better approaches use whatever is known at sign-up or within the first session.
Onboarding questionnaires ask users to pick favorite genres or rate a few items, giving an explicit starting point. Contextual signals such as device, locale, time of day and referral source support context-aware recommendation from the first request. Session-based models react to clicks within the current visit and need no long-term profile at all.
| Strategy | Signal used | Trade-off |
|---|---|---|
| Popularity fallback | Global or regional top items | Safe but impersonal |
| Onboarding preferences | Explicit choices at sign-up | Adds friction to registration |
| Context features | Device, location, time, referrer | Weak signal on its own |
| Session-based recommendation | Clicks in the current visit | Needs a few interactions first |
| Demographic attributes | Age band or region if provided | Privacy and fairness concerns |
Strategies for new items
Item cold start is solved mainly through content features that exist before any interaction. Content-based filtering compares item attributes such as text, category, price, audio or images with what a user already likes, so a new item can be recommended from the moment it is added to the catalog.
Modern systems often train a two-tower model or a hybrid recommender in which the item tower takes content features as input. A new item then receives a usable item embedding immediately, placed near similar existing items. Pretrained text or image encoders, and increasingly large language models, generate these feature vectors without manual tagging.
Content alone cannot reveal whether an item is actually good, so many platforms add deliberate exploration: they show new items to a small, controlled share of traffic and learn from the response. Multi-armed bandit algorithms such as Thompson sampling or upper confidence bound methods balance this exploration against showing items already known to perform well.
Comparing cold start techniques
No single technique solves every form of cold start, which is why production systems usually combine several layers. The following table compares the main families by the form they address and their main limitation.
| Technique | Form addressed | Main limitation |
|---|---|---|
| Content-based filtering | Item, partly user | Tends toward overly similar suggestions |
| Hybrid models | User and item | More complex to train and serve |
| Bandit exploration | Mainly item | Costs some short-term engagement |
| Cross-domain transfer | User | Requires data from a related service |
| Meta-learning | User and item | Research-heavy, harder to operate |
| LLM-based recommendation | User and item | Higher latency and serving cost |
Cross-domain recommendation deserves particular mention: a user who is new to a video service but has history on a sister music service can start with a transferred profile. Meta-learning methods train a model to adapt quickly from very few interactions, treating each new user as a small learning task.
Measuring and avoiding common mistakes
Cold start performance must be measured separately from overall accuracy, because aggregate metrics are dominated by established users and popular items. Teams typically report metrics such as recall or NDCG for users with fewer than a set number of interactions, often five or ten, and track how quickly new items receive their first engagements.
| Mistake | Consequence | Correction |
|---|---|---|
| Evaluating only on random splits | Cold start weakness stays hidden | Hold out new users and items explicitly |
| Relying on popularity forever | New items never surface | Reserve traffic for exploration |
| Long mandatory onboarding | Users abandon registration | Ask for few, optional choices |
| Ignoring catalog coverage | Feedback loop narrows the catalog | Monitor coverage and item age |
A subtle risk is the recommendation feedback loop: items that start with no exposure gather no interactions, so they keep receiving no exposure. Without exploration or content-based boosts, cold start becomes permanent for part of the catalog, which harms both diversity and the creators or sellers behind those items.
In plain words
Think of a bookseller who recommends books by remembering what each regular customer bought before. When a stranger walks in, the bookseller has nothing to go on. When a new book arrives, nobody has bought it yet, so the bookseller does not know whom it would suit.
That is the cold start problem. A good bookseller asks the stranger a couple of questions, looks at the new book’s cover and topic, and occasionally suggests it to see how people react. Recommendation software does the same: it uses descriptions, context and small experiments until it has enough real behavior to learn from.
A real-world example
An online grocery service adds about 400 new products every week and gains several thousand new customers each month. Its collaborative filtering model works well for regular shoppers, but new products rarely appear in recommendations, and first-time customers see the same generic best-seller list.
The team introduces a hybrid two-tower model whose item tower uses product name, category, brand, price band and an image embedding. New customers are asked to select three dietary preferences at sign-up, and the first basket immediately updates their profile through a session-based component. The team also reserves 5 percent of recommendation slots for bandit-driven exploration of items added in the last 14 days.
Over the following quarter, recall for customers with fewer than five orders rises noticeably, and the share of new products that receive at least one recommended purchase in their first two weeks more than doubles, without lowering engagement among established shoppers.