The Cold Start Problem in Recommender Systems and How to Solve It

Solving cold start for new users and new items: popularity and content-based fallbacks, onboarding signals, metadata embeddings, exploration and bandits.

Intermediate 3 min read · Updated 29 October 2026

The cold start problem is what happens when a recommender has no history to learn from: a new user who has never clicked anything, a new item nobody has seen, or a new system with no logs at all. Collaborative signals ("people like you watched this") don't exist yet, so the model falls back to guesses and new items never get the exposure they need to collect data. Every recommendation design question expects a cold start plan.

Three kinds of cold start

Kind	What's missing	Main levers
New user	Interaction history	Context, onboarding choices, popular-in-segment, fast session learning
New item	Engagement statistics and an id embedding	Content features, metadata embeddings, exploration boosts
New system or market	All logs	Heuristics and editorial curation first; log everything; transfer from similar markets

New users

A new user's journey from cold to warm. The system leans on context and popularity at first, then shifts weight to personal signals as the session and history fill in.

Context features: location, language, device, referral source and time all carry signal on the first request.
Popularity within a segment: "trending among new users in your country" beats global popularity.
Onboarding questions: ask for a few interests or let the user pick creators; keep it short, every step loses users.
Session-based models: recommend from the clicks in the current session (sequence models or item-to-item similarity), so personalisation starts within minutes.
Imported signals: with consent, contacts or a linked account can seed a profile.

New items

A new video or product has no clicks, so an id-based model gives it a near-random score, and it may never be shown. Fixes combine content understanding with deliberate exposure:

Content-based embeddings: embed the title, description, image, audio or creator, so the item lands near similar items in the embedding space from the moment it's uploaded. Two-tower item towers that use content features (not just the id) handle this naturally; see embeddings.
Creator and category priors: start the item's predicted CTR at its creator's or category's average instead of zero.
Exploration budget: guarantee new items a slice of impressions (for example 5% of slots) and let a bandit decide which earn more.
Freshness boost: a time-decaying bonus in re-ranking for items under a certain age.
Fast feedback: stream early engagement into item features within minutes, so a good new item is recognised quickly.

A new item's first day. Content features give it a starting score, the exploration budget earns it early impressions, and streamed engagement lets the main ranker take over within hours.

A brand-new system

Launch with heuristics: popular, new, editor's picks, simple content similarity.
Log impressions, positions and outcomes from day one; that log is your future training set.
Train a first simple model after a few weeks; keep the heuristic as a fallback.
Transfer from related data if available: another market, another surface, or pretrained content encoders.

Measuring cold start

Report offline and online metrics separately for new users (first 7 days) and new items (first 24 to 72 hours).
Track new-item reach: the share of new items that receive at least N impressions in their first day.
Track new-user retention: day-1 and day-7 return rates by recommendation variant.

Where it’s used

Short-video apps

Every upload gets a small, targeted exposure batch; strong early engagement earns larger batches.

E-commerce

New products inherit priors from brand and category and are matched by image and text embeddings.

Music streaming

Audio embeddings place new tracks near similar songs before anyone has played them.

In an interview

Treat new users and new items separately, with a plan for each.
Use content features in the item tower so new items have an embedding on arrival.
Reserve an exploration budget and explain how it's allocated.
Measure cold start with slice metrics for new users and new items.

Frequently asked questions

What is the cold start problem?

When a recommender has no interaction history for a new user, a new item or a new system, so collaborative signals do not exist and new items may never get exposure.

How do you solve cold start for new items?

Use content-based embeddings from text, images and metadata so the item has a representation on arrival, start from creator or category priors, reserve exploration impressions for new items, and stream early engagement into features quickly.