Recommender systems suggest items a user is likely to want: products, films, articles, courses.
Collaborative Filtering
Uses patterns of behaviour across many users. If people who liked A and B also liked C, recommend C to someone who liked A and B.
- User-based: find similar users and recommend what they liked.
- Item-based: find items similar to ones the user liked, based on who interacted with them.
- Matrix factorisation: learn a compact vector (embedding) for every user and item so that their dot product predicts interest.
Content-Based Filtering
Recommends items similar in their attributes to ones the user liked — genre, keywords, text embeddings. It works for new items with no interaction history.
Hybrid Systems
Most production systems combine both, often in two stages: candidate generation (quickly retrieving hundreds of plausible items) followed by ranking (a detailed model ordering them).
The Cold-Start Problem
New users and new items have no history. Content features, popularity, onboarding questions and exploration help.
Evaluation
- Offline: precision@k, recall@k, NDCG, measured on held-out interactions, split by time.
- Online: A/B tests of clicks, purchases or retention — the real test.
Beyond Accuracy
Consider diversity, novelty, fairness to item providers, and avoiding feedback loops where the system only recommends what it already promoted.
Try It
The MovieLens dataset on the glitchdata hub is the standard starting point for building a recommender.