AI glossary · Beginner · Classic ML, recsys & search
Recommender system
A recommender system is software that picks and shows products, videos, articles and other items a person is likely to enjoy, based on the user's past behavior and the characteristics of the content.
About this term
- Level
- Beginner · Core ideas for first-time readers
- Category
- Basics
- Added
- 2026-10-08
How it works
There are two main approaches. Collaborative filtering recommends "what people with tastes similar to mine liked", and it works using only behavior records such as clicks, purchases and views. Content-based recommendation recommends "things with features similar to what I liked", using content information such as genre, descriptions and images. Today, deep learning approaches that convert both users and content into embeddings (numeric vectors) and then find ones close to each other are widely used, and the two approaches are often combined.
Two basic recommendation approaches
Collaborative filteringFollow people with similar tastes
- Works with behavior records alone
- Good at surprising discoveries
- Weak with new items and new users
Content-basedSimilar to what you liked
- Uses features like genre and descriptions
- Can recommend new items too
- Tends to repeat similar items
공통 Real services often combine the two approaches
How it runs in large services
In services with a very large number of products or videos, it is impossible to score every item for each person. So a fast method first narrows the pool down to a large set of candidates, a more sophisticated model then ranks only those candidates, and finally rules such as deduplication and diversity are applied to decide the list shown on screen. Vector DBs and ANN search are often used to narrow down candidates. This staged structure also resembles the RAG flow, where search results are re-sorted by a reranker.
Recommendation stages that narrow down candidates
Why it matters
In services with too many options for people to browse themselves, recommendations help users quickly find what they want. The home screens of many services, including shopping, video, music, news and advertising, are filled with recommendation results. So recommendation quality directly affects user satisfaction and service revenue. Whether a change to recommendations actually made things better is usually checked with A/B testing.
Things to keep in mind
Newly registered users and newly listed products have no behavior records, which makes recommendation difficult; this is called the 'cold start' problem. Recommending only similar things over and over can trap users in narrow tastes, creating a 'filter bubble'. Optimizing only to increase clicks also risks over-recommending sensational content. Because personal behavior records are used, privacy protection and transparency in explaining why something was recommended must also be considered.
Example
The home screen of a video service is a typical output of a recommender system. Candidates are chosen based on videos the user recently watched to the end, videos they skipped, and the choices of other people with similar viewing histories, and the screen is filled in order of how likely the user is to watch them. The "customers who viewed this item also viewed" section of an online shopping mall is built on the same principle. Recently, approaches that use LLMs to explain the reason for a recommendation in sentences, or to ask about tastes in conversation and then recommend, are also being tried.