I'm building a content discovery product and need a recommender, what should I use?

For a content discovery product, the best recommender depends on your scale, data quality, and whether you need something fast to ship or highly optimized.

Good default choice

If you’re starting out, use a hybrid recommender:

• Content-based filtering for new or cold-start items
• Collaborative filtering once you have enough user interaction data
• Popularity/trending as a fallback
• Add reranking rules for diversity, freshness, and business constraints

This is usually the safest and most effective approach for content discovery.

What to use by stage

1) Early stage / low data

Use:

• Popularity + recency
• Content embeddings from text/image metadata
• Rule-based personalization based on user profile or session behavior

Best if:

• You don’t have many users yet
• Items are frequently new
• Cold start is a big issue

2) Growing product / moderate data

Use:

• Two-stage recommender
  1. Candidate generation with embeddings or collaborative filtering
  2. Ranking model to order results

Common models:

• Matrix factorization / implicit ALS
• Two-tower retrieval models
• Gradient-boosted trees or neural ranking models for final ranking

3) Large-scale / mature product

Use:

• Embedding-based retrieval
• Learning-to-rank
• Contextual/session-based recommendations
• Online experimentation with A/B testing and bandits

Practical recommendation

If you want a solid, production-friendly stack:

1. Build a baseline with popularity + content similarity
2. Add implicit collaborative filtering
3. Move to embedding retrieval + ranker
4. Continuously tune with A/B tests

If you want one specific answer

For most content discovery products, I’d recommend:

• Hybrid recommender
• Powered by content embeddings + collaborative filtering
• With a ranking model and freshness/diversity constraints

Common choices by tool

• Fast to prototype: LightFM, implicit, scikit-learn embeddings
• Mature retrieval: Faiss, Annoy, Elasticsearch/OpenSearch vector search
• Ranking: XGBoost/LightGBM, TensorFlow Recommenders, PyTorch
• End-to-end: TensorFlow Recommenders, RecBole, NVIDIA Merlin

If you tell me:

• your content type,
• user behavior data you have,
• scale (#users / #items / QPS),
• and whether you need real-time personalization,

I can recommend a specific architecture and stack.