How do you implement recommendation systems using machine learning?

Recommendation systems suggest items to users based on preferences and behavior. Approaches: (1) Collaborative filtering: user-based (find similar users, recommend what they liked) or item-based (find similar items to what the user liked). Matrix factorization (SVD, ALS) decomposes the user-item interaction matrix into latent factors. (2) Content-based: represent items by features (genre, keywords, embeddings), build user profiles from interacted items, and recommend items similar to the user profile. (3) Deep learning: neural collaborative filtering (NCF) learns nonlinear user-item interactions, Two-Tower models (separate user and item encoders for efficient retrieval), and sequential models (GRU4Rec, SASRec) capture temporal patterns in user behavior. (4) Hybrid: combine collaborative and content-based signals for robustness. Production architecture: candidate generation (retrieve 100-1000 candidates using fast approximate methods), ranking (score candidates with a detailed model considering features like recency, diversity, and user context), and re-ranking (apply business rules: diversity, freshness, fairness, deduplication). Cold start: for new users, use popularity-based recommendations and content-based filtering until sufficient interaction data accumulates. For new items, use content features and metadata. Evaluation: offline (precision@k, recall@k, NDCG, coverage, diversity), online (A/B testing: click-through rate, conversion rate, engagement time). Training: handle implicit feedback (clicks, views, purchases) using techniques like BPR or weighted matrix factorization, implement negative sampling (users didn't interact doesn't mean they wouldn't like it), and update models regularly (daily or weekly retraining with incremental updates). Scale: use approximate nearest neighbor search (FAISS, ScaNN) for candidate generation from millions of items.