Skip to content
mlmentorship

How would you do cold-start for a new user?

Cold-start is solved by combining minimal explicit signal, demographic and contextual fallbacks, and aggressive exploration in the first few sessions.

Published · 4 min read ·Role-specific ·Advanced

ML breadth · active recall

Practice before you read

8 minutes. Explain the mechanism, why it works, when it fails, and one alternative.

Back to my workbook How practice works

ML breadth · closed-book attempt

How would you do cold-start for a new user?

Explain the mechanism, why it works, when it fails, and one alternative.

08:00recommended time

Closing or reloading clears the scratchpad. Only score, weak rubric dimensions, attempt count, and retry date can be stored locally.

30-second answer map

Visual first · depth when needed

Trace how a recommender uses fallback signals before behavior exists, then gradually gives observed user behavior more influence without ending exploration.

Preparing the visual…

Asked in: recsys interviews. The classic open-ended sub-question.

The L4 candidate proposes “use popularity.” The L6 candidate stages a transition from generic to personalized signal across the user’s first several sessions.

What an L4 answer sounds like

“Show them popular items until they have some history, then switch to personalized.”

The right idea, no nuance. Misses how to make the transition smooth, what intermediate signals to use, and how to bootstrap quickly.

What an L5 answer sounds like

“Cold-start has three sources of signal, used in sequence:

  1. Onboarding signal: ask the user explicitly. Pick 5 favorite artists / genres / categories. Cheap to collect, surprisingly good signal. Many products skip this; they shouldn’t.

  2. Contextual fallbacks: time, location, device, referral source, demographic if available. A user signing up at 9am on weekday mobile has different intent than 11pm Friday on TV.

  3. Exploration in early sessions: explicitly diversify recommendations to gather signal across categories quickly. A bandit-style approach (epsilon-greedy or Thompson sampling on the candidate distribution) bootstraps faster than pure exploitation.

Transition: at session 1, almost all signal is from (1) and (2). After ~5-10 interactions, the personalization model starts to dominate. Hard threshold, gradual blend, or a learned mixing weight depending on signal volume.”

This is L5. Three signals, sequenced over user-lifetime.

Learning objective

How does a recommender hand control from fallback signals to learned user behavior?

1 · BEFORE BEHAVIOR EXISTS

Start with available evidence
Onboarding choices + request context + item content + safe population priors.

Compose the first slate
Anchor it with likely-relevant items, then reserve some positions for diverse exploration.

What is missing
There are no clicks, skips, saves, or watch-time observations for this user yet.

2 · EACH RESPONSE UPDATES THE MIX

Observe
Exploratory impressions produce clicks, skips, saves, and dwell-time evidence.

Shift influence gradually
Observed behavior gains influence; onboarding, context, and population fallbacks recede as confidence grows.

Keep learning
Do not turn exploration off: preferences, intent, and the item catalog can still change.

Read it this way: begin on the left with signals that exist at sign-up. The slate intentionally gathers feedback; carry that feedback to the right, where personalized behavior gains influence rather than switching on at one fixed interaction count. Exploration stays in the mix so the system can keep learning. Original schematic informed by Schein et al. on cold-start recommendation and Li et al. on contextual-bandit feedback.

What an L6 answer adds

“…practical things:

The first session matters disproportionately. Users decide whether to come back based on early experience. Cold-start that surfaces obvious popular hits is safe but boring; that loses long-term retention. The right balance: anchor on hits, sprinkle exploration.

Side information beats interaction signal at cold-start. Item content (genre, description, audio embedding for music, visual embedding for video) lets you recommend before any user interaction with that item. Two-tower models that use both content and collaborative signal handle this naturally.

Synthetic warm-up via similar users. A new user signing up with profile X can be matched to existing users with similar profiles; recommendations bootstrapped from those users’ aggregated history. Useful when explicit onboarding signal is weak.

Multi-armed bandits for shelf composition in early sessions. Try several recommendation strategies (popular, trending, content-based, demographic match) and use bandit logic to converge on what works for this user fast.

Cold-start as a metric, not just a feature. Track ‘time to good recommendations’ (sessions until a personalized model performs well for a new user) as a release-gating metric. New cold-start improvements should reduce this number.

The dual problem (cold items) is just as important. New items have no engagement signal; they need a content-based candidate source and an explicit boost in the first hours / days to gather signal.”

Tells that get you a strong-hire vote

  • You name three signals (onboarding, context, exploration) and sequence them.
  • You bring up content-based features for early personalization.
  • You mention bandits for shelf-level exploration.
  • You name cold-start as a metric to track over releases.
  • You connect to the dual cold-item problem.

Tells that get you down-leveled

  • “Just show popular items.”
  • No onboarding-signal collection.
  • No mention of exploration.
  • Only solving the user side; ignoring cold-items.

Common follow-up

“How would you measure how well your cold-start works?”

The L6 answer:

“Three metrics: (1) Day-N retention for new users (do they come back), (2) time to first ‘good’ recommendation (sessions until a personalized model meets quality threshold for that user), (3) per-cohort engagement on first session vs same cohort at session 10. A/B cold-start changes against the previous cold-start strategy on new users only; the population is small so power matters and tests need to run longer than usual.”

Related: Design YouTube’s recommender, Design Spotify’s homepage, System design case study: building personalized search ranking.

Found an error or stale claim? Report a correction or send a private note.