05 Chapter

Recommender-System Algorithms

Match users to items they're most likely to engage with.

Recommender systems match users to items across large catalogs, powering feeds, search, and personalization. The algorithms below progress from classic collaborative filtering through matrix factorization to the two-tower and ranking models used at production scale.

  • Use matrix factorization or collaborative filtering for classic recommenders.
  • Use two-tower neural networks plus ranking models for large-scale production systems.
#AlgorithmBest forCommon fields
1Collaborative Filtering User-item recommendations
  • E-commerce
  • streaming
  • news feeds
2Matrix Factorization / SVD / ALS Sparse rating or interaction data
  • Movies
  • products
  • music
  • marketplaces
3Content-Based Filtering Recommending similar items
  • Retail
  • jobs
  • real estate
  • articles
4Hybrid Recommenders Combining user, item, and context signals
  • Most production recommender systems
5Learning-to-Rank Ranking search/recommendation results
  • Search engines
  • ads
  • marketplaces
6Deep Recommenders / Two-Tower Models Large-scale retrieval and ranking
  • YouTube-style feeds
  • ads
  • social platforms
7Bandit Algorithms Online recommendation optimization
  • Ads
  • personalization
  • experimentation