Cold Start Problem

Collaborative filtering learns from history. The cold start problem is what happens when there is none — a user who signed up a minute ago, or an item uploaded this morning, offers nothing for the model to learn from.

Three versions of one gap
A new user has no interactions, so there is no neighbourhood to find and no learned factor vector to score with; the system can fall back only on what is popular. A new item is the harder case, because the system causes its own blindness: an item that is never shown is never clicked, and an item that is never clicked is never shown. A new system has an empty interaction matrix and cannot use collaborative signal at all. Sparsity is the milder, chronic form of the same condition.

How it gets handled
Every remedy substitutes some other signal for missing history:

Side information — genre, brand, text, images — gives a new item a representation before it has interactions, which is why content-based and hybrid designs persist alongside collaborative filtering.
A two-tower retrieval model that embeds items from attributes rather than ids can place a brand-new item in the index immediately.
Onboarding asks a new user for a few favourites, with demographic or popularity priors covering the gap until real signal accumulates.
Exploration inverts the problem: a bandit treats an unknown item as an uncertain arm and spends impressions deliberately to learn its value.

Why it never ends
Audiences and catalogues churn continuously, so cold users and cold items are a standing share of traffic. A system that quietly declines to recommend anything unproven still looks healthy on offline metrics computed over well-covered items, while its coverage shrinks and new hits never surface.

The cold start problem is a recommender's inability to serve users or items with no interaction history: a new user has no neighbourhood and no learned factors, and a new item is never shown, so it never earns the clicks that would get it shown. Content features, onboarding questions, popularity fallbacks and deliberate exploration are the standard remedies. Because catalogues and audiences churn constantly, it is a permanent share of traffic, not a launch-day issue.

Recommendation software learns from what people have already done. Cold start is the awkward moment when there is nothing to learn from.

Picture your first evening on a new streaming service. It knows nothing about you, so it cannot find people with similar taste to borrow suggestions from. The same thing happens in reverse to a film uploaded this morning: nobody has watched it, so the service does not know who to show it to — and until it shows it to someone, it never finds out. The usual fixes are to ask you a few questions up front, lean on whatever is popular, or judge a new item by its description and cast rather than its audience.