Your recommendation system has a cold start problem for new items. How do you handle it — what signals can you use and what can't you use?
formulate your answer, then —
tldr
Cold start: new items have no interaction history, so collaborative filtering can't retrieve them. Immediate signals: content features (text, images, metadata), creator history. Strategies: content-based embedding fallback, warm start from similar item embeddings, exploration budget (epsilon-greedy exposure), creator-based initialization. Architecture fix: build item tower on content features, not just learned ID embeddings — new items representable from upload. Track coverage metrics separately since offline NDCG ignores new items.
follow-up
- How would you implement warm start for new item embeddings in a two-tower model without retraining from scratch?
- Your new-item recall is low in A/B tests. How do you debug whether the problem is retrieval (not fetched) vs ranking (fetched but ranked low)?
- What's the tradeoff between exploration (surfacing new items) and exploitation (showing proven content) in a recommendation system?