Think Build Implement Repeat
London, UK +44 7367 067226
WhatsApp FOLLOW f in X
  1. Home
  2. Blog
  3. The Cold Start Problem in Recommendations
AI & Machine Learning

The Cold Start Problem in Recommendations

New products, new customers and new sites all break collaborative filtering. What to show instead, and how to bootstrap towards personalisation.

Updated 2 min readBy SpiderHunts Technologies

Free estimateNo obligation

Get a free estimate

Tell us what you need. A senior engineer reads every enquiry.

Takes under a minute. We never share your details.

  • Free consultation
  • No commitment
  • NDA on request

Prefer to talk? Book a free 30-minute call →

Quick answer — TL;DR

Cold start has three separate forms - new item, new user, new system - and they need different answers. Content-based matching on attributes covers new items, popularity and context cover new users, and a staged plan moves you towards personalisation as data accumulates.

Three different problems with one name

Collaborative filtering works by finding patterns across many users and items. When either side is new, there is nothing to pattern-match against.

  • New item - a product just added, with no interactions. It cannot be recommended by behaviour, so it stays invisible and never accumulates the data it needs.
  • New user - a first-time visitor with no history. Common on most sites, and the majority of sessions for many businesses.
  • New system - you have just started collecting interactions at all.

Treating these as one problem produces a solution that handles none of them well.

Content-based matching for new items

A new product has no behaviour but plenty of attributes: category, price, brand, material, description text, images. Matching on those lets it be recommended from day one alongside similar established items.

This works best where your catalogue data is decent. If product attributes are inconsistent, half-populated or free-text, that is the thing to fix first - and it pays off across search, filtering and merchandising as well as recommendations.

Context beats nothing for new users

For an anonymous first-time visitor, you know more than you think: the page they landed on, the search term that brought them, the device, the time of day, and what they have clicked in this session.

Session-based recommendations built from within-visit behaviour often outperform personalised ones for new visitors, because immediate intent is a stronger signal than a thin profile.

SituationSensible default
No information at allPopular in this category, recently
Arrived on a product pageSimilar and complementary items
Arrived from a search termBest matches, then related
Two or three clicks inSession-based, updating live

A staged plan rather than a big build

Businesses regularly attempt a personalised recommender before they have the interaction data to support one, and conclude that recommendations do not work.

  1. Start with popularity by category and context. Unglamorous, immediately useful, and it generates the interaction data everything else needs.
  2. Add content-based similarity once attributes are clean. This covers new items permanently.
  3. Add session-based recommendations from sequences of views within a visit.
  4. Only then consider collaborative filtering, once there is enough repeat behaviour for it to learn from.

Each stage produces value and feeds the next, which is a better shape of project than a long build with one uncertain payoff.

Giving new items a chance to be seen

Even with content-based matching, popular items tend to crowd out new ones, and a recommender trained on its own output narrows the catalogue over time. Deliberately reserving a slot for newer or less-shown items counteracts this.

Keep the reserved share modest and measure its cost. The point is to gather information about items you would otherwise never learn about, and that has real value even when the immediate click rate is lower.

A recommender that only shows what already sells will slowly forget your catalogue.

FAQ

Frequently asked questions

The questions readers ask us after this guide.

Still have a question?

Ask us directly — a senior engineer will get back to you.

Ask about your project

How long until a new product has enough data?

It depends on traffic to it, not on elapsed time. Until it does, content-based matching should carry it.

Is popularity really a good recommendation?

For an unknown visitor it is a strong baseline and often hard to beat. It should be the benchmark any personalised system must exceed.

Do I need logins for recommendations?

No. Session-based recommendations work on anonymous behaviour within a visit, which covers most first-time traffic.

What if my catalogue changes constantly?

Then content-based and session-based approaches matter more, and collaborative filtering matters less, because items rarely accumulate long histories.

Keep reading

More on AI & Machine Learning

Start here

Want machine learning project details from us?

Tell us what you are trying to predict and roughly what data you hold. We will come back with an honest view on whether machine learning is the right tool, what the work would involve and a realistic cost range. If a spreadsheet would do the job, we will say so.

  1. You tell us what you needTwo minutes on the form, or a message on WhatsApp.
  2. A senior engineer reviews itAnd comes back with questions, a realistic range and an honest view on fit.
  3. Free 30-minute scoping callWe talk through scope, options and a realistic estimate — with no obligation.
Free estimateNo obligation

Talk to someone who builds this

Send a short brief and we will come back with an honest view and a realistic range.

Takes under a minute. We never share your details.

  • Free consultation
  • No commitment
  • NDA on request

Prefer to talk? Book a free 30-minute call →