Glossary
Recommendation systems
In short
A recommendation system is software that suggests products, films, music or articles that a user is likely to like. It learns from the behavior of that user and of similar users.
AI basics · recommender system
MMOOZZGG.com editorsChecked: 1 min read
For a child
The librarian knows you love books about dinosaurs. When a new book about ancient lizards arrives, she puts it aside for you. She also notices that kids who read about dinosaurs often like books about space. So she suggests something new to you. A computer helper in a toy store or a cartoon app does the same, only for a huge number of children at once.
In plain words
When a website shows a “you may also like” block, a recommendation system stands behind it. It works in two main ways. The first finds people with similar tastes and offers what they liked and you have not seen yet. This is collaborative filtering. The second looks at the properties of an item, such as the genre of a film or the brand of sneakers, and picks something similar to what you already chose. Usually the two are mixed. The system counts clicks, purchases and viewing time and updates all the time, so every feed is different.
For business
For a store or a service, recommendations bring revenue without new customers: the average order value, the number of items per order and repeat purchases all grow. Measure the effect with an A/B test that compares a group with recommendations and a group without them. The price depends on the catalog size and the sales history. Ready-made cloud modules connect quickly; your own model makes sense when the catalog is large and accuracy matters. The risks are the cold start for new items and buyers, a bias toward popular items, and pushy suggestions that annoy people. Purchase history is personal data, so check its storage and processing against 152-FZ, Russia’s personal data law. AI for an online store shows where such suggestions pay off.
For a specialist
Recommender systems rank items for a user using implicit feedback (clicks and views) and explicit feedback (ratings). The classic approaches are content-based filtering, collaborative filtering by neighbors or by matrix factorization, and hybrid schemes. Matrix factorization approximates the user-by-item table with a product of matrices of hidden factors. Industrial systems work as a cascade: P. Covington and co-authors (RecSys 2016) described for YouTube a deep model that generates candidates, selecting a small set of videos from the whole catalog, and a separate ranking model. Modern versions use several modalities at once, for example the text description and the photo of a product. They also use sequential models and transformers over the action history, graph neural networks, and embeddings with vector search. Quality is measured offline with ranking metrics (precision@k, NDCG) and online with A/B tests, and the diversity of results is tracked separately.
Example
An online cinema notices that viewers who finish a detective series often turn on Scandinavian thrillers. The system starts offering them after the final episode, and viewers leave the service less often between seasons.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map