ML System Design

Design Cold-Start Personalization for a Marketplace

Personalize recommendations when users, items, sellers, or regions have little historical data.

AdvancedCold startExplorationContent-based retrievalBanditsMarketplace fairness

Prompt

Design cold-start personalization for a marketplace where new users, new items, and new sellers need useful exposure before collaborative signals exist.

Evaluation lens

New-user activationNew-item exposureConversionDiversityLong-term retention

Split cold start into three problems

New users, new items, and new sellers need different strategies. A generic "use popularity" answer is weak because it can lock the marketplace into incumbents.

Candidate generation

Use content-based retrieval from item text, images, category, price, geography, seller metadata, and onboarding signals. Add popularity priors, but cap them so they do not dominate every request.

For new users, ask lightweight onboarding preferences or infer from acquisition source, locale, and first-session behavior.

Exploration strategy

Use an exploration budget for new supply and uncertain user preferences. Contextual bandits can help, but start with controlled randomization if the team lacks reliable logging.

Metrics

  • new-user activation and retention
  • new-item first meaningful exposure
  • seller liquidity
  • conversion by cohort
  • diversity and concentration metrics
  • regret or opportunity cost from exploration

Failure modes

  • Popularity trap: popular items get all exposure and new supply never learns.
  • Low-quality exploration: users see irrelevant items and churn.
  • Delayed labels: conversion can take time, so early proxies must be validated.
  • Fairness tension: exposure fairness can conflict with immediate relevance.

What the architect signal looks like

End with a ramp plan: start with interpretable content-based priors, add exploration, measure new-cohort health, then introduce learned bandits once logging is trustworthy.