Skip to content
Tutorials Center
My learning

Topics

Lesson 33 of 38 ยท Cold Start and Production

The Cold Start Problem

Solve the cold start problem: new users, new items and new systems. Popularity fallbacks, onboarding, content features, exploration, with runnable learning-curve examples.

Collaborative filtering learns from history. A user with no history, or an item nobody has touched yet, gives it nothing to work with. This is the cold start problem, and every real recommender has to face it every day: new accounts sign up, new products are listed, new videos are uploaded.

There are three versions of the problem:

  • New user: we know nothing about their taste.
  • New item: the catalog gained something no one has interacted with.
  • New system: a brand-new product with few users and few interactions overall (also called "bootstrapping").

Each has standard remedies, and they combine well with the hybrid designs from earlier.

Three kinds of cold start

New user No history yet
New item Nobody has used it
New system Little data anywhere

Fix them with other signals, then switch to behavior as it arrives.

1 / 2

New users

You have several signals to fall back on, roughly from cheapest to richest:

  1. Popularity (overall or in the user's country, device or language). It is hard to beat as a first list.
  2. Context: time of day, referrer, search query, the page being viewed. Even "what are people looking at right now" can personalize a little.
  3. Onboarding questions: ask the user to pick a few favorite genres, artists or topics. Three to five taps are more valuable than any default. Choose the questions that best split the user base, not only popular items.
  4. Demographics (where available and ethical to use): coarse and often weak, but better than nothing.
  5. First interactions: after only one or two clicks, an item-to-item recommendation ("people who liked this also liked") already works, because it needs the item, not the user history.
  6. Exploration: deliberately show a varied first list to learn the user's taste quickly (see the bandit section).

The learning curve below shows how quickly history pays off. For a group of test users, we reveal only m of their plays, rank items using item-item collaborative filtering built from other users, and measure how many of the remaining plays are found in the top 10. With m = 0 the only possible method is popularity.

Python How fast does a new user warm up?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Read down the column: most of the gain arrives with the first few interactions, which is why a good onboarding flow and fast model updates for new users matter more than polishing the long-run model.

New items

A new item has no interactions, so a collaborative model cannot score it and it would never be shown, which means it would never get interactions. That is a vicious circle. Solutions:

  • Content features: score the item through its attributes (genre, text, image embedding, price) against the user's profile, as in content-based filtering. Or train a model that maps item features to an embedding (this is what the item tower of a two-tower model does).
  • Exploration quota: reserve a few slots for new items and show them to a sample of users to collect first interactions. Boost fresh items temporarily, with a decaying bonus.
  • Creator or category priors: a new item by an established creator inherits some expected quality from their other items.
  • Editorial seeding: featuring a new item in a curated shelf gives it first impressions.

The next block models the "new item shelf". Twelve items have no training plays at all; only their genre labels (noisy) are known. We rank just those items for each user and measure the share of the user's real plays among the top 3, comparing a random order with a content-based profile match.

Python Recommending items with zero interactions

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The content model gives the new items a useful first ranking while they have no interactions of their own, and once enough interactions arrive you blend in (or switch to) the collaborative signal.

New systems

When the whole product is new, there is little behavior at all. Typical steps: start with rules and editorial picks, use item content and metadata, import data from a related source where you have rights to do so, run simple popularity and "similar items" features first, and instrument everything so that the logs you collect now (including impressions, not only clicks) train the real model later.

Evaluating cold start

Measure the cold segments separately. An overall NDCG hides that new users have a terrible experience, because they are a small share of the traffic. Slice your offline metrics by user activity (0, 1 to 5, more than 5 interactions) and by item age, and watch the share of recommendations given to items younger than a week.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which is a typical first recommendation list for a brand-new user with no information?

    Popularity is a strong, cheap fallback; context makes it somewhat personal.
  2. 2. Why does a purely collaborative model never show a brand-new item?

    This is the vicious circle of cold start: no exposure means no data, which means no exposure.
  3. 3. Which technique helps a new item get its first impressions?

    Reserving some exposure for new items collects the first interactions the model needs.
  4. 4. Why should cold-start segments be evaluated separately?

    New users are a small share of traffic, so a global metric can look fine while their experience is poor.

Summary

  • Cold start affects new users, new items and new systems.
  • New users: popularity, context, onboarding questions, item-to-item after the first click, exploration.
  • New items: content features, feature-based embeddings, exploration quotas and decaying boosts.
  • Most of the gain for a new user comes with the first few interactions, so adapt quickly.
  • Evaluate cold segments on their own.

Next: ANN Search and Serving