AI Cold Start Recommendation System

Recommendations for Brand-New Users

Design the part of a streaming service's recommender that serves people who just signed up and have no history yet, using sign-up context, a short onboarding step, careful exploration and the events of the first session.

Overview

A regular recommender learns from history: people who played what you played also liked this. A user who signed up a minute ago has no history, so that method has nothing to compare. This is the cold start problem. It matters because the first few sessions decide whether a new user stays, and a service pays to acquire every one of them.

The fix is not one model but a ladder of weaker signals, used in order as they appear: the context of the sign-up (country, language, device, the ad or page they came from), a short onboarding step where they pick a few titles, and then the plays and skips of the first session. Netflix's help center describes exactly this first step: new accounts and profiles are asked to choose a few titles they like to jump-start recommendations, the step is optional, and people who skip it start from a diverse and popular set of titles.

Questions to ask first

What kind of catalog? — A streaming service: 10 million tracks or titles, grouped into about 500 taste clusters
How many new users? — 1 million sign-ups a day, peak 3 times the average
What do we know at sign-up? — Country, language, device, sign-up channel, time of day; no age or gender
Is there an onboarding step? — Yes, optional: pick at least 3 items from a grid
How fast must picks react? — The next request after a play or skip reflects it
When is a user no longer new? — There is no switch; the regular model's share grows with history
How is success measured? — First-session plays and 7-day and 30-day return, in A/B tests

What you will be able to do

Explain why collaborative filtering fails for new users and list the signals that exist instead.
Design an onboarding step that learns a lot from a few taps, and a fallback for people who skip it.
Build a cohort prior with smoothing, and explore with Thompson sampling instead of always showing the most popular items.
Update a user's taste inside the first session and blend it into the regular model without a visible jump.
Measure the whole thing with experiments on new-user cohorts.

Key concepts

Cold start
Onboarding picks
Cohort prior
Bayesian smoothing
Explore and exploit
Thompson sampling
Contextual bandit
Session vector
Blending by history size
Diverse fallback

Functional requirements

Return a ranked list of items for a user with zero history, within the first request after sign-up
Offer an optional onboarding grid that covers many tastes, and record the picks
Build a starting profile from sign-up context: country, language, device, channel
Explore: reserve some slots to learn which kinds of items this user likes
Update the ranking after each play, skip or save in the same session
Blend smoothly into the regular recommender as history grows
Fall back to a diverse popular list when any signal or service is missing
Log each shown list with the strategy and the probabilities used, for training and experiments

Non-functional requirements

p99 latency of 200 ms for a recommendation request (a target we chose)
A play or skip visible to the next request in the same session
Always an answer: every failure degrades to the diverse popular list, never an empty screen
Each user stays in the same experiment group on every visit
No sensitive attributes (age, gender) used unless the user gave them for this purpose

Interview questions

Why does collaborative filtering fail for new users, and what do you use instead?
How do you design the onboarding step?
Why use a bandit instead of showing the most popular items for the cohort?
How do you update recommendations during the first session without retraining?
When does a user stop being cold, and how do you hand over?
How do you handle new items, not new users?
How do you evaluate a cold-start change?
What if the bandit stats or session store are unavailable?