Can a new item be recommended before anyone interacts with it?

Yes. Content, editorial rules or other side information can support an initial recommendation, though uncertainty remains.

Is cold start only a launch problem?

No. Every new user or item can create it, and sparse activity can keep it relevant.

Does asking users questions eliminate the problem?

It provides useful evidence if the questions are relevant and answered meaningfully, but the resulting profile is still incomplete and can change.

AI Glossary / Retrieval and knowledge systems / Intermediate

Cold start (recommender systems)

Definition

Cold start in recommender systems is the difficulty of making useful recommendations when there is little or no relevant interaction history. It can affect a new user, a new item or an entirely new service. The missing evidence limits methods that depend on patterns of previous ratings, clicks, purchases or other interactions.

What it means

A recommender often learns that users with similar histories prefer related items. A new user has no established history, and a new item has not yet accumulated responses. The system therefore lacks some of the signals it normally uses.

Cold start is different from a computer taking time to boot or a model loading into memory. It is an information problem. It can persist for infrequent users and rarely exposed items, even when the recommendation service itself has operated for years.

How it works

First identify which evidence is missing. New-user approaches may ask for a few preferences, use current-session signals or provide a carefully chosen general starting set. New-item approaches can use descriptions, categories or other content features. Hybrid systems combine available signals rather than relying entirely on interaction history.

The system may also explore by showing eligible new items and learning from responses. Exposure must be recorded because an item cannot receive feedback if it is never shown. No method creates reliable personal preferences from nothing: its starting recommendations rely on assumptions, shared patterns or additional information supplied by the user or item.

Three practical examples

1. A new reader selects several topics during onboarding, giving a news recommender initial preferences before a browsing history exists.

2. A newly added training course has no enrollment interactions, but its subject and difficulty metadata allow it to enter content-based candidate selection.

3. A small catalog launches with a curated starting list and gradually incorporates observed feedback. The team evaluates new-user experience separately from established-user performance.

Think of it this way

Think of helping a new library visitor choose a book. You cannot use their borrowing history yet, but you can ask what they enjoy and inspect the books' descriptions. Those clues support an initial suggestion without guaranteeing a match. Over time, actual choices provide additional evidence that can revise the initial impression.

Worked example

Consider a simplified content-based starting rule for a user who explicitly chooses two interests, gardening and photography. Represent those interests as the set {gardening, photography}. Three new items have tags: A = {gardening, outdoors}, B = {photography}, and C = {cooking, outdoors}.

Use Jaccard similarity, intersection size divided by union size, only for this toy example. A shares one tag with the user and has a union of three tags, scoring 1/3. B shares one tag with a union of two, scoring 1/2. C shares none and scores zero. The rule ranks B above A above C despite none of the items having interaction history.

This is a transparent initial heuristic, not a complete recommender. It ignores item quality, repetition, changing intent and whether the tags capture the content accurately. A later click is evidence of interaction, not proof of satisfaction. The example demonstrates how side information can support a start while leaving substantial uncertainty.

Illustrative case study

Fictional case: Cedar Courses introduces new lessons every week. Its interaction-only candidate generator rarely includes them because they have no enrollment history. The team adds content-based candidates using verified lesson metadata and separately tracks how new lessons are exposed.

Evaluation distinguishes first-time visitors, established learners and newly published lessons. Reviewers check whether a gain in clicks corresponds to useful learning choices rather than misleading titles. This invented case illustrates a measurement plan; it reports no actual learner experiment, financial result or guarantee that content features solve every cold-start problem.

Common misconceptions and mistakes

Cold start is not solved simply by assigning a default embedding or copying a popular item's score. Those choices provide an initial representation but may reveal little about the new user or item.

Popularity can be a useful baseline, yet relying on it exclusively can prevent new items from gaining exposure. Content-based methods still need meaningful content and some way to connect it with user interests. Another mistake is treating every missing interaction as dislike. The user may never have seen the item, making absence of feedback different from negative feedback.

Business use

The issue affects onboarding, catalog growth and the experience of occasional users. Plan a starting experience deliberately instead of leaving it as an accidental fallback from a model trained on established users. Keep the information requested from users relevant and make its purpose understandable.

Measure recommendation quality by cold-start scenario rather than reporting only an aggregate. Use evaluation splits that reflect items or users genuinely unseen during training. Track exposure, feedback availability and how quickly the system improves with appropriate evidence. A strong result on familiar users does not establish good performance for newcomers.

Check your understanding

How can item B rank first in the worked example without any interaction history?

Answer: The illustrative rule compares its content tags with interests the user supplied. B has the highest set similarity under that rule. This supports an initial ranking but does not prove that the user will prefer B after seeing it.

Related terms

Recommender System
Collaborative filtering
Exploration
Popularity Bias
content-based recommendation
side information
onboarding
item exposure
hybrid recommendation

Learn next

Separate new-user, new-item and new-system scenarios. List which signals are available in each, then design a transparent baseline and suitable evaluation split. Study how exposure creates feedback and distinguish observed interactions from the deeper outcomes the recommendations are intended to support.

Sources

Google, Recommendation Systems, collaborative-filtering summary and cold-start limitations: https://developers.google.com/machine-learning/recommendation/collaborative/summary

Google Research, Expediting exploration by attribute-to-feature mapping for cold-start recommendations, author research record: https://research.google/pubs/expediting-exploration-by-attribute-to-feature-mapping-for-cold-start-recommendations/

The tag-similarity calculation and Cedar Courses case are original hypothetical illustrations.

Previous: Cold languages Next: Collaborative filtering