What's the best recommendation engine for personalizing article feeds in a consumer reading app with cold-start users?

There isn’t one universal “best” recommendation engine for this use case, but for a consumer reading app with cold-start users, the strongest practical choice is usually a hybrid recommender:

Best overall approach

Hybrid content-based + collaborative filtering + exploration

• Content-based handles cold-start well by using article metadata and embeddings.
• Collaborative filtering improves personalization once you have enough user behavior.
• Exploration/bandits help you learn fast from new users without overcommitting too early.

What I’d recommend in practice

1) For cold start: content + embeddings

Use article features such as:

• title, headline, summary
• topic/category
• author/publication
• entities/keywords
• text embeddings from a transformer model

Then match users based on:

• onboarding-selected topics
• clicks/reads/saves
• inferred preferences from early interactions
• session context

This is typically the most robust starting point because new users can still get relevant results immediately.

2) For mature personalization: two-stage ranking

A strong architecture is:

• Candidate generation: retrieve a few hundred relevant articles
  • content similarity
  • trending/popular
  • collaborative neighbors
• Ranking model: reorder candidates using user/article interaction features

Common ranking models:

• gradient-boosted trees
• deep ranking models
• two-tower retrieval models

3) For better cold-start learning: contextual bandits

If your app wants to adapt quickly to sparse signals, use a contextual bandit layer to balance:

• exploitation: show likely relevant articles
• exploration: test new topics and diversify

This is especially useful when you have short sessions and limited feedback.

If you want a simple answer

If you’re choosing one “best” practical setup:

Use a hybrid recommender built around article embeddings + collaborative signals, with a contextual bandit for exploration.

Good options by stage

• Early-stage app / low data: content-based recommender with embeddings
• Moderate data: hybrid system
• Large-scale app: two-tower retrieval + learning-to-rank + bandit exploration

Things that matter a lot for reading apps

• freshness/recency
• diversity across topics
• avoiding filter bubbles
• explanation quality (“Because you read X”)
• session intent vs long-term profile

My direct recommendation

For most consumer reading apps, I’d start with:

1. Article text embeddings
2. User interest profile from onboarding + recent activity
3. Hybrid ranking with popularity and freshness features
4. Contextual bandit for cold-start exploration

If you want, I can also suggest:

• a concrete architecture,
• model choices by traffic scale,
• or a build-vs-buy comparison of recommender platforms.