Design a Recommendation System for a New Product (the Cold-Start Problem)

Asked by: Hugging Face

"How would you design a recommendation system for a new product?" is a Hugging Face ML-system-design question, and the word "new" is the whole interview. Any candidate can recite "collaborative filtering." But collaborative filtering learns from user–item interactions — and a new product has none. No clicks, no purchases, no co-occurrence. The question is testing whether you understand both the standard multi-stage recommender and how to bootstrap it from a cold start.

The crux (spend ~60% of your time here). Two things: (1) the standard recommender is a funnel — cheap retrieval narrows millions of items to hundreds, then an expensive ranker orders them — and you must explain why it's staged. (2) At launch there's no interaction data, so you can't learn collaborative signal; you must lean on content/metadata features, popularity priors, and exploration (deliberately showing uncertain items to gather data), then transition to learned models as the data flywheel spins up. Naming cold-start as the crux and designing the bootstrap-to-steady-state path is what distinguishes a strong answer.