What should I use for recommendations with collaborative filtering and embeddings?

If you want recommendations using both collaborative filtering and embeddings, a strong practical choice is:

Best default approach

Hybrid recommender = matrix factorization / collaborative filtering + item/user embeddings

Common options

1. Two-tower model (best modern choice for scale)

   • One tower embeds the user
   • One tower embeds the item
   • Trained so matching user/item pairs are close in vector space
   • Works well with implicit feedback and retrieval at scale

2. Matrix factorization with embeddings

   • Classic collaborative filtering method
   • Learns latent user and item vectors
   • Simple, strong baseline, easy to interpret

3. Hybrid neural recommender

   • Combine:
     • collaborative signals
     • item text/image embeddings
     • user features
   • Good when you have side information and cold-start problems

If you have different needs

• Small/medium dataset, want simplicity: use implicit matrix factorization (ALS/BPR)
• Large-scale retrieval: use a two-tower model
• Need content + CF + cold start: use a hybrid model with embeddings from metadata/content
• Need ranking after retrieval: add a learning-to-rank model on top

Popular libraries/tools

• Implicit (ALS, BPR) for classic CF
• LightFM for hybrid recommender systems
• TensorFlow Recommenders for two-tower and retrieval/ranking
• PyTorch + custom model if you want flexibility
• FAISS / ScaNN / Annoy for nearest-neighbor search over embeddings

Practical recommendation

If you’re starting fresh:

• Use LightFM for a quick hybrid baseline
• Move to TensorFlow Recommenders or a PyTorch two-tower model if you need scale
• Use FAISS for fast lookup of embedding similarities

If you want, I can recommend the best specific method for your case if you tell me:

• size of data
• explicit ratings vs clicks/purchases
• whether you have item/user features
• whether you need real-time retrieval or just offline recommendations