Three flavors of cold
Collaborative methods need history, so they break when history is absent. There are three cases:
- New user: no past interactions, so no neighbors and no learned vector.
- New item: never interacted with, so no collaborative signal exists.
- New system: a fresh product with almost no data anywhere.
Practical remedies
- Content fallback: lean on attributes. A new movie still has a genre and cast, so content based scoring fills the gap until interactions accumulate.
- Onboarding: ask the new user a few preference questions or have them pick favorites to seed a profile.
- Popularity priors: show trending or globally popular items as a safe default.
- Hybrid models: blend content features into a factorization model so a new item gets a vector from its attributes.
- Exploration: deliberately show some uncertain items, a bandit style approach, to gather signal fast.
The tradeoff
Cold start forces a balance between exploitation of what you know and exploration to learn about cold users and items quickly.
Key idea
Cold start hits new users, items, and systems; remedies are content fallback, onboarding, popularity priors, hybrids, and exploration.