How do I set up recommendations for new users with no history?

For new users with no interaction history, the problem is called the cold start problem. The usual approach is to combine a few strategies so you can make useful recommendations immediately and improve them as soon as the user starts interacting.

Common ways to handle it

1. Use popularity-based recommendations

Start with items that are:

• Most viewed
• Most purchased
• Highest rated
• Trending recently

This is the simplest fallback and works well as a default baseline.

2. Ask for onboarding preferences

When a user signs up, ask them to choose:

• Categories/topics they like
• Brands they prefer
• Goals or use case
• Budget range
• A few “liked” examples

Then recommend items similar to those choices.

3. Use contextual signals

Even without history, you may know:

• Location
• Device type
• Time of day
• Referral source
• Language
• Signup channel

These can help tailor recommendations, though usually only lightly.

4. Recommend based on content features

If your items have metadata, use content-based filtering:

• Tags
• Descriptions
• Categories
• Embeddings from text/images

For example, if the user picks “running shoes,” recommend similar products from that category.

5. Use collaborative filtering with side information

Classic collaborative filtering needs history, so for new users you can:

• Fall back to item popularity
• Use user attributes if available
• Use hybrid models that combine collaborative + content features

6. Explore and learn quickly

Once the user starts interacting:

• Track clicks, saves, purchases, skips
• Update recommendations in near real time
• Mix exploitation with exploration to learn preferences faster

A common tactic is to show a diverse first page to gather signals quickly.

---

Practical setup recommendation

A good production approach is:

1. Default to popular/trending items
2. Collect onboarding preferences
3. Use content-based matching to personalize
4. Switch to behavior-based recommendations after a few actions
5. Continuously retrain or update user profiles

---

Example policy

You might define rules like:

• 0 interactions → trending + onboarding-based + diverse categories
• 1–5 interactions → content-based personalization
• 5+ interactions → hybrid recommender using behavior history

---

Important tips

• Don’t show only the most popular items; add diversity so users can discover their interests.
• Measure success with metrics like click-through rate, conversion rate, and time to first meaningful action.
• Test onboarding carefully; too many questions can hurt signups.
• Keep recommendations fresh so new users don’t see stale popular items.

If you want, I can also help you design:

• a simple rule-based cold-start system
• a machine learning approach
• or a sample implementation in Python / SQL / Spark