← Lessons

quiz vs the machine

Gold1410

Machine Learning

The Cold Start Strategies Deep

Recommending for new users and items that have little or no history.

4 min read · core · beat Gold to climb

The three cold starts

Collaborative methods need history, but cold start breaks that. There are three flavors: a new user with no clicks, a new item with no interactions, and a new system with almost no data at all.

Strategies for new items

  • Lean on content features like text, category, or image embeddings so a fresh item lands near similar known items.
  • Use a hybrid model that blends content with collaborative signal as interactions trickle in.

Strategies for new users

  • Ask a few onboarding questions or show a quick preference picker.
  • Fall back to popularity or demographic priors until signal accumulates.
  • Use contextual cues such as device, time, and location.

Bandits to learn fast

  • Treat early exposure as exploration, showing varied items to learn taste quickly.
  • A bandit balances exploring new options against exploiting what already works.

Trade off

  • Too much exploration annoys users, too little leaves cold items invisible forever.

Key idea

Cold start is solved by leaning on content features, popularity priors, and context for fresh entities, then using exploration to gather interactions that hand off to collaborative models.

Check yourself

Answer to earn rating on the learn ladder.

1. What is the best fallback for a brand new item with no interactions?

2. How do bandits help with cold start?