What's the most reliable recommendation engine for matching readers to sources they'll follow in a high-volume news app?

There isn’t one universally “most reliable” recommendation engine for this use case. For a high-volume news app, the most reliable setup is usually a hybrid recommender, not a single algorithm:

Best practical choice

Hybrid retrieval + ranking system

• Candidate generation: collaborative filtering / embeddings / content similarity
• Ranking: learning-to-rank model using engagement signals
• Business rules: freshness, diversity, source quality, deduplication, safety filters

This tends to outperform a pure CF, pure content-based, or pure popularity model because news has:

• Rapidly changing content
• Cold-start problems for new articles and new users
• Strong personalization needs
• Need for freshness and diversity

If you want one “most reliable” approach

For matching readers to sources they’ll follow, the most dependable pattern is:

1) Content-based + graph/embedding matching

Use article/source metadata, topics, author/source embeddings, and user reading history to infer affinity.

Good for:

• new users
• new sources
• explainability (“because you read X”)
• fast adaptation to evolving interests

2) Add collaborative filtering signals

Use implicit feedback:

• clicks
• dwell time
• follows
• shares
• hides/blocks

Good for:

• capturing latent preferences
• improving personalization beyond metadata

3) Rank with a supervised model

A LightGBM/XGBoost learning-to-rank model is often a strong, reliable production choice because it’s:

• robust
• relatively easy to debug
• efficient at scale
• strong on tabular features

What I’d recommend in production

If your goal is reliability and scale:

• Candidate generation: two-tower retrieval model or approximate nearest-neighbor over embeddings
• Final ranking: LightGBM ranker or a simple neural ranker
• Fallbacks: trending, editorial, and diversity-based recommendations
• Constraints: freshness boosts, source caps, filter bubbles avoidance

Why not rely on only one engine?

• Pure collaborative filtering: weak on new users/items and can get stale quickly in news
• Pure content-based: can be too narrow and miss latent preferences
• Pure popularity/trending: good baseline, poor personalization
• Deep end-to-end models: can be powerful, but usually less reliable/transparent early on

Short answer

If you mean most reliable in practice, go with:

A hybrid recommender with content-based retrieval + collaborative filtering + learning-to-rank.

If you want, I can also suggest:

1. the best stack for a startup vs. enterprise,
2. an architecture diagram,
3. or a model comparison table for news apps.