Solving the Cold Start Problem

New users have no interaction history. Strategies to bootstrap recommendations:

Onboarding preferences — Ask users to rate items or select interests during signup. Keep it quick (3-5 selections) to avoid abandonment. Spotify's artist/genre selection is a good example.

Demographic signals — Use age, location, device type, or referral source as weak signals. "Users in your area often like..."

Contextual recommendations — Time of day, day of week, device type can inform initial suggestions. Mobile users at lunchtime might want different content than desktop users in the evening.

Popularity fallback — Show trending or globally popular items. Not personalized, but better than nothing.

Exploration strategies — Intentionally show diverse items to quickly learn preferences. Use bandit algorithms to balance showing proven items vs. learning about the user.

New items have no interaction data. Strategies to surface them:

Content-based features — Recommend based on attributes (category, brand, tags). A new Nike shoe can be recommended to users who liked other Nike products.

Creator/brand similarity — New items from popular creators get initial boost. "New release from an artist you follow."

Editorial curation — Human editors surface quality new content. Expensive but high quality.

Exploration allocation — Reserve 5-10% of recommendations for new items. Accept lower short-term engagement for better long-term catalog coverage.

Bandits — Multi-armed bandit algorithms balance exploiting proven items with exploring new ones. Thompson Sampling and UCB are common choices.

Starting a new platform with zero data is the hardest case:

Import external data — Use public datasets or third-party data to bootstrap. Movie ratings from IMDB, product data from manufacturers.

Pre-trained embeddings — Use embeddings trained on similar domains. Transfer learning from public models.

Rule-based start — Begin with hand-crafted rules, transition to ML as data accumulates. "Show items in the user's selected category, sorted by recency."

A/B test baselines — Test different non-personalized strategies (popularity, recency, editorial picks) to learn what works while building data.

Leverage any signal — Search queries, page views, time spent — any behavior is signal. Don't wait for explicit feedback.

Design your product to generate feedback quickly:

Implicit signals — Track view duration, scroll depth, saves, shares. Much more abundant than explicit ratings.

Low-friction feedback — Quick rating prompts, "more like this" buttons, easy dismissal options.

Diverse exposure — Show variety to learn preferences faster. A user who only sees one category teaches you nothing about their other interests.

Session context — Use within-session behavior for immediate personalization, even for anonymous users.

The goal is reducing the time from "new user" to "personalized experience" from weeks to minutes.