Two versions of the cold start problem

In recommender systems, cold start exists in two independent forms.

User cold start — a new visitor with no interaction history. Collaborative filtering has nothing
to work with: there is no one to compare them against. The system knows neither their interests, nor
their price range, nor their brand preferences.

Product cold start — a new item in the catalogue with no views, add-to-cart events or purchases.
The algorithms do not know who this product would appeal to. The result is that a new arrival is
invisible in recommendations and gets only the organic traffic that reaches it through the catalogue.

Strategies for solving it

For a new user

Strategy How it works Time to first signal
Session-based recommendations Built on views in the current visit Immediate — from the first click
Popularity-based fallback Bestsellers and trends No signal, so no personalization
Onboarding questions What are you interested in? After 2–3 answers
Geotargeting Local preferences by region On the first visit
Identity resolution Linking to offline history On sign-in

For a new product

  • Content-based algorithm — uses attributes (category, brand, price, specifications) to find
    similar products
  • Cluster boost — the item is assigned to the nearest cluster and inherits that cluster’s audience
  • Merchandising rules — manual promotion of new arrivals to the top of recommendations
  • Explorative strategies — epsilon-greedy shows new items in a small share of impressions to
    collect signals

The affinity profile as a solution

Instead of waiting for purchase history to accumulate, the system builds a profile of the shopper
from session behaviour: what they viewed, how long they spent on a page, what they added to the cart.

After as few as 5–10 viewed products, the algorithm has a read on the price range, the category of
interest and the brand preferences. That shortens the cold-start period substantially — from several
weeks to a single session.

Important: never leave cold start without a fallback strategy. Empty recommendation blocks, or
placeholders reading products coming soon, are worse than popular products used as filler.