Handle the two sides separately. For new items, use content features (text, category, embeddings derived from the item itself) so they can be scored before any interactions exist, and give them exploration budget. For new users, fall back to popularity or segment level recommendations, then use early session signals and move to personalized collaborative signals once enough interactions accumulate.
Why interviewers ask this
Cold start is the first thing that breaks pure collaborative filtering in production, so interviewers use it to see whether you have shipped a recommender rather than trained one on a static dataset. They want the user side and item side treated separately, a hybrid content plus collaborative answer, and some acknowledgement of exploration, since a system that never shows new items never learns about them.
How to structure your answer
- Split the problem into the new item side and the new user side.
- Use content features to score items with no interaction history.
- Describe the fallback ladder for a brand new user.
- Add explicit exploration so new items can accumulate signal.
Example answer
They are two different problems, so I treat them separately. New items are the easier half: I use a two tower setup where the item tower takes content features, so title text, category, price band, an embedding of the description or the thumbnail. That means a freshly uploaded item has a usable vector from minute one, before anyone has clicked it. New users get a ladder. Nothing at all means popularity, ideally popularity within whatever context I do know, like country and device. As soon as they interact a couple of times I can build a session based representation from those items, which in my experience is worth more than anything from a signup questionnaire. Then once they cross an interaction count they move onto the fully personalized path. The piece people forget is exploration. A pure exploit ranker never surfaces a new item, so it never gets impressions, so it never gets signal, and the catalog quietly ossifies. I reserve a slot in the feed for a candidate chosen with an upper confidence bound style rule, and I track catalog coverage as a real metric alongside click rate.
Walking into this interview soon? GhostPilot listens to your live call, spots the question the moment it is asked, and puts a structured answer on your screen in real time. Try it on your next mock, or grab a $29 Session Pass, no subscription, for the real thing.
See how it worksFollow-up questions to expect
- How would you evaluate a recommender offline before it goes live?
- What is popularity bias and how does it compound over time?
- How would you decide how much exploration traffic to spend?
Related machine learning engineer questions
Your interviewer will ask their own version of this. Paste your actual job description into the free Question Predictor and get the 20 questions that role is most likely to ask, with what each one is really probing.
Predict my questions