Recommendation System Cold Start: A Fix List for 4 Scenarios

Summary: Cold start is not a single problem but four scenarios: new users with no history, new items with no behavior, a new system with no data, and a new scenario in an old system with no samples. Each has its own main fix: new users rely on signup signals and migration, new items rely on content vectors, a new system relies on rule-based fallbacks, and a new scenario relies on real-time features. Sort the scenario correctly and the fix list lines up.

Every model in a recommendation system is built on historical behavior data, and cold start is exactly the moment when there is no data: a new user has just registered and the system does not know what they like; a new item has just gone live and there are no clicks or purchases to learn from. Handled badly, the cost runs in both directions: users see irrelevant recommendations and never look a second time, while new items sit in inventory and miss their best exposure window. Many articles treat cold start as one vague problem, but in practice you must first split it by scenario: user cold start, item cold start, system cold start, and scenario cold start. The four differ in cause, in the signals available, and in the fix. This article gives the fix list and the acceptance criteria for each of the four, so you can hold them against your own system and pick directly.

Key Takeaways

Split cold start by scenario before talking about fixes: the signals available to users, items, systems, and scenarios are completely different, and mixing them together leads to a mismatched plan.
The core of new-user cold start is trading the smallest asking cost for the strongest preference signal: keep the signup prompt to three to five tags at most.
The main channel for new-item cold start is the content vector: similarity recall over semantic vectors built from text and images does not depend on historical behavior, so an item can be distributed the moment it goes live.
For a new system, start with rules and popularity as a fallback while making sure event tracking follows the spec from day one: data quality is the most valuable asset during cold start.
Scenario cold start, a newly launched feed for example, can reuse the profiles and vectors of older scenarios and accumulates samples for the new scenario quickly through real-time features.
The effect of a cold-start fix must be measured on its own: next-day retention for new users, first-week exposure for new items, and ramp-up speed of click-through rate for new scenarios each have their own acceptance criterion.

1. Four Types of Cold Start: Sort Them Out First

The boundaries between the four types have to be drawn clearly, because the same trick produces completely different results in different scenarios.

User cold start: the system has data, the user is new. The challenge is that nobody knows the new user's preferences. The signals available are registration information, device features, the acquisition channel, and initial selection behavior.

Item cold start: users have profiles, the item is new. The challenge is that the new item has no behavior data, so collaborative-filtering recall naturally leaves it out. The signal available is the content itself: title, description, images, category, and price band.

System cold start: the whole recommendation system starts from zero, a new business or a first build. The challenge is that users and items both lack data, so the fix has to be global: hold the line with operational rules and popularity ranking first, while making sure data collection follows the spec.

Scenario cold start: the system has data, but a new recommendation scenario has been opened, such as a new homepage feed or a new cart recommendation slot. The challenge is that the sample distribution of the new scenario differs from the old one. The assets available are site-wide profiles and item vectors.

Once the four are separated, most of the "cold start has no fix" dead ends can be traced back to a specific scenario and broken one at a time.

2. User Cold Start: Spend the Asking Cost Where It Counts

When a new user opens the product for the first time, the system knows nothing about them. The workable fixes, in order of priority:

Targeted guidance during signup. Let the user pick three to five categories or roles they care about during the registration flow. The point is to control the asking cost: faced with more than five options, most users pick at random and signal quality drops instead. Choosing the questions matters too: the options must be far apart, category level rather than item level, so the answer can feed recall filtering directly.

Device and channel signals. The acquisition channel itself carries intent: a user arriving from a category campaign page should get that category as the starting point of the first recommendation. Audience attributes targeted by ad delivery can also serve as a prior, as long as compliance holds.

Popularity fallback plus fast convergence. Use best sellers inside the category as the fallback for the first screen, guaranteeing something that is not relevant but probably not annoying. Every click afterwards corrects the profile, and the first ten behaviors should carry far more weight than usual: the model needs a better memory during cold start.

Cross-device account migration. When an existing user signs in on a new device, the profile migrates along and the cold-start problem dissolves by itself. This is also a hidden benefit of pushing registration and sign-in: the higher the sign-in rate, the smaller the share of cold-start users.

3. Item Cold Start: Let the Content Speak for Itself

A new item has no behavior, but the item carries its own content. The main line is content vector recall: map the title, description, and images into a semantic vector and match it against the user's historical interest vector by similarity, so the item enters the candidate pool the instant it goes live and bypasses the loop that demands behavior data. On the engineering side, the vector search capability of Elasticsearch turns unstructured content such as text into vectors and recalls them by similarity, while FAISS, open sourced by Meta, provides an approximate nearest neighbor retrieval library for large-scale vectors. Both are common foundations for this channel.

Three supporting lines:

Operational cold-start exposure pool. Reserve a fixed initial exposure quota for new items, a fresh-pick slot for example, and trade real exposure for the first batch of behavior data. The quota has to be controlled: a cold-start quota is essentially trading mainline conversion rate for exploration, so a cap on the total volume is required.

Riding along with similar items. Place the new item next to a similar older item that already has behavior data, the similar-products slot on a detail page for example, inheriting part of the traffic and part of the users.

Grouped rollout during a test period. Test the click-through rate of a new item in a small traffic pool first and scale it up step by step once it passes. This is both a traffic allocation mechanism and a filter that catches items whose data cannot be trusted, such as image fraud or clickbait titles.

4. System Cold Start: Rules as a Fallback, Data First

When you build a recommendation system from zero, models are off the table for now. The right order is rules first, data in parallel, and models taking over later.

Phase one holds the experience with deterministic rules: best-seller lists, top-rated lists, category navigation. These rules need no personalization data and immediately deliver recommendations that beat random, stabilizing the baseline experience.

Phase two is data first: instrument events to spec from day one, covering event naming, required attributes, and globally unique user and item IDs. The most expensive thing in cold start is not the algorithm but data quality; messy event tracking is repaid with interest as rework in the model phase. Unifying system-level behavior logs is also the foundation for later collaboration between search and recommendation. The architectural principle that both share the same behavior event stream is covered in detail in an earlier collaboration feature on this site.

Phase three is the model handover: once data has accumulated to a trainable scale, run the simplest collaborative filtering through the full pipeline first, then upgrade the model step by step. The acceptance criterion for the handover is explicit: the model version has to stay at or above the rule baseline on core metrics in an A/B test before it counts as taking over.

5. Scenario Cold Start: Reuse Assets and Real-Time Features

When you launch a new recommendation slot, you do not have to start from zero. Two main lines:

Reuse site-wide assets. User profiles, item vectors, and category relations are shared across scenarios. Recall for the new scenario reuses these assets directly, ranking starts with simplified rules, and a model arrives once there is enough data. When an e-commerce site extends from a guess-what-you-like slot on the product detail page to a cart recommendation slot, the vectors and profiles are all reused and only the candidate set boundary needs redefining, for example excluding already purchased items in the cart scenario.

Real-time features to speed up the ramp. The sample distribution of a new scenario differs from the old one, so a model trained offline goes through an adjustment period. Adding in-session real-time behavior, this session's search term or the category just browsed, as features lets recommendations track the user's immediate intent inside the new scenario and shortens the ramp.

6. Acceptance Criteria and Common Pitfalls

Each of the four scenarios has its own acceptance criterion. User cold start looks at next-day retention and first-week behavior depth of new users. Item cold start looks at first-week exposure and sell-through rate of new items. System cold start looks at how far the model beats the baseline when it takes over. Scenario cold start looks at the number of days the click-through rate of the new scenario needs to ramp. Cold-start optimization without an acceptance criterion easily turns into something that merely looks like a lot was done.

Four common pitfalls:

Treating exploration as pure loss and squeezing it: the exploration quota is cut tighter and tighter, short-term conversion looks good, but new items never break out and the ecosystem dries up.
Asking too much in the signup flow: past five options, users pick at random and the signal is polluted, so fewer is better than more.
Failing to guard content vector quality: vectors built from clickbait titles or images that do not match the text drag similarity recall off course. Vector quality and product detail quality are the same thing.
Withdrawing the rules the moment the model ships: rule fallbacks such as popularity and operational pinning should stay for the long term as a safety net for the long tail and for anomalies, rather than being fully replaced by the model.

FAQ

Q: For new-user cold start, should I use tag selection or default recommendations?
A: Use both together. Three to five high-discrimination category questions buy you a prior at a controllable cost with a strong signal, and right after the selection the user moves into category popularity plus fast personalization. Skipping the questions entirely and converging on behavior alone suits scenarios with low signup willingness, and the price is that the first few screens depend entirely on the popularity fallback, with personalization only showing after ten or more behaviors.

Q: A new item has no behavior data at all, so how does it get into the recommendation candidate pool?
A: The main channel is content vector recall: map the title, description, and images of the item into a semantic vector and match it against the user interest vector by similarity, so the item can be distributed as soon as it goes live without any historical behavior. Support it with an operational cold-start exposure pool and with riding along next to similar items, trading a small amount of deterministic traffic for the first batch of real behavior, then transitioning to regular recall naturally.

Q: Will the popularity fallback during cold start make recommendations look identical for everyone?
A: Yes, which is why you have to control how long and how wide the fallback runs. The popularity fallback should only cover the first few screens plus the first few sessions, and every real click should raise the weight of personalization. Watch two numbers: how fast the share of personalized recall for new users rises with the number of behaviors, and the click-through rate of the fallback slot. If a week later new users are still being shown the same batch of popular items, the convergence mechanism is not working.

Q: Building a recommendation system from zero, should I start with a model or with rules?
A: Rules first, with data in parallel. Rules such as best sellers, top rated, and category navigation need no personalization data to deliver a passing experience, and at the same time instrument events to spec from day one, covering event naming, unique IDs, and complete attributes. Once data has accumulated to a trainable scale, ship the simplest model first and let it take over only when it beats the rule baseline in an A/B test. The reverse, starting with a model, means doing the most expensive construction on the thinnest data foundation.

Q: How much exposure should the cold-start quota, the fresh-pick slot, take?
A: Cap it at the opportunity cost of exploration you can afford, and the common practice is a low single-digit percentage of recommendation exposure for new items, adjusted dynamically with fluctuations in the overall conversion rate. The point is to have both a total-volume control and an exit mechanism: items whose test-period click-through rate stays below the threshold leave the cold-start pool and hand the quota to the next candidate.

Q: How do I know whether cold-start optimization is actually working?
A: Look at scenario-specific metrics rather than global ones. For new users, next-day retention and average behaviors per user in the first week. For new items, first-week exposure and sell-through rate. For new scenarios, the number of days click-through rate needs to ramp to a stable value. The global conversion rate is affected by too many factors, so cold-start improvements get diluted out of sight and you have to build a separate observation metric for each scenario.

About Tengence

Tengence Search serves the recommendation and retrieval scenarios described in this article, providing content vector recall and behavior event collection, so new items are distributed the moment they go live and new scenarios reuse site-wide profiles quickly. Tengence (TENGENCE) is the enterprise growth engine for the AI era, focused on public-traffic acquisition and private-domain conversion. Its core products are Tengence GEO, a dual-engine GEO plus SEO product, and Tengence Search.

Data Sources

Elasticsearch, "What is vector search?": vector search maps unstructured content such as text into dense vectors and recalls them by similarity, supporting the content-vector cold-start recall channel for new items.
FAISS (Meta AI Research) official repository: a clustering and approximate nearest neighbor retrieval library for large-scale dense vectors, a common engineering foundation for the vector recall channel.
Netflix TechBlog, "System Architectures for Personalization and Recommendation": the division into offline, nearline, and online computing layers, which supports the layered cold-start design of rule fallback and model handover.