New-User Cold Start
Every personalised feed has to answer the same awkward question in the first second of a new account's life: what do we show someone we know nothing about? Collaborative filtering has no answer, because a user's embedding is a summary of their past interactions and this one has none.
What actually breaks
New-user cold start is the absent-signal problem on the user side. There is no row in the interaction matrix to factorize, no neighbourhood to average over, and no history for a sequential model to extend. What the system falls back on is the popularity prior — the catalogue's global bestsellers — which is safe, uninformative, and identical for everyone.
The usual repairs
- Onboarding elicitation: ask for a handful of genre or creator picks, buying a few explicit preferences at the cost of friction.
- Side features: device, locale, referral source and signup context feed a user tower that produces a usable embedding before any interaction exists.
- Session-based models: condition on the clicks of the current visit alone, so personalisation starts within minutes — what GRU4Rec and SASRec are built for.
- Exploration: treat the first slates as information-gathering, where the bandit framing earns its keep.
Why it is the expensive one
New users are the cohort most likely to leave, and they judge the product on the slate handed to them before the model knows anything. A recommender tuned on an offline split of established users optimises for the people least at risk, which is why cold-start performance is reported separately, on interaction-count buckets rather than in aggregate.
New-user cold start is the absent-signal problem on the user side: a new account has no interaction history, so there is no row to factorize and no neighbourhood to average, leaving only catalogue-wide popularity to serve. It is addressed with onboarding preference questions, signup-context features in a user tower, or session-based models that personalise from the current visit's clicks. It matters because new users decide whether to stay on exactly those first slates.
Walk into a bar you have never visited and ask the bartender for something you will like. He cannot know. The best he can do is pour whatever is popular and watch your face.
New-user cold start is that moment for recommendation software. It works out what you enjoy from what you have watched, played or bought before — and on day one there is none of that. So it falls back on what everyone else likes, which is rarely what you came for. That is why apps ask you to pick a few favourite genres when you sign up, and why they change so quickly once you start tapping things.