Recommender systems

In short: Recommender systems suggest what users might like: films, products, videos, posts. They learn from behaviour (clicks, purchases, viewing time) and from properties of the content.

In more detail: The goal is usually to maximise attention or sales. The system sorts thousands of candidates in milliseconds.

In detail

• Collaborative filtering: “users like you liked …”. A user-item table is decomposed into small vectors (matrix factorisation), users and items live in the same space.
• Content-based: items with similar properties (genre, text embedding).
• Two-stage system: first fetch candidates (embedding search, see Vector databases and similarity search), then sort with a precise model (ranking).
• Cold start: new users or items have no history.
• Reinforcing the same: filter bubbles and feedback loops, see Bias, fairness and data quality.
• Measurement: click rate, dwell time, A/B tests. High metrics do not automatically mean benefit for people.
• Law: the Digital Services Act requires transparency and options without profiling on large platforms.

See also: k-nearest neighbours, k-means and clustering, Embedding (vector representation)