AI Recommendation System — AI Glossary

Reference details

Related terms

• AI Personalization
• Embedding Model
• Dense Retrieval

How it works

1. Interactions are logged as a sparse user-item matrix — views, clicks, purchases, dwell time — and that matrix, not any model, is the system's primary asset.
2. Collaborative filtering factorises it into user and item embeddings, so that similarity in the learned space predicts interaction. It needs interaction history and therefore says nothing about a new item or a new user.
3. Content-based scoring embeds the item's own attributes instead, which covers exactly the gap collaborative filtering leaves: a new item has features even with no interactions.
4. Retrieval and ranking are separate stages for cost reasons. A cheap retriever reduces millions of candidates to hundreds, and an expensive ranker scores only those, because scoring the full catalogue per request is not affordable.
5. Language models enter at three points: describing new items so content-based retrieval has something to match, generating explanations for a ranked result, and supporting conversational refinement of a request.
6. The feedback loop is the structural hazard. The system trains on interactions it caused, so items it stops showing stop accumulating evidence, and the catalogue narrows unless exploration is deliberately funded.

Example: Why the two-stage split exists, in one calculation

A catalogue of 1,000,000 items and a ranker costing 1ms per item. Scoring everything for one request takes 1,000,000 x 1ms = 1,000 seconds — nearly 17 minutes, for a page that must render in under a second. Retrieve first with an approximate-nearest-neighbour lookup costing about 5ms total to return 500 candidates, then rank those: 5 + 500 x 1 = 505ms, which fits. The split bought a factor of 1,000,000 / 500 = 2,000 in ranker cost, and what it bought it with is a hard ceiling — anything the retriever misses cannot be ranked, so the retriever's recall at 500 caps the whole system no matter how good the ranker becomes. That is the same structure as a RAG pipeline, and it implies the same priority: a ranker improvement is bounded by retrieval recall, so measuring recall at the candidate-set size is the first diagnostic, not the last.

Commonly confused with

A recommender is not a search engine, and the difference is where the intent comes from. Search starts with a query that states what the user wants, so relevance is measured against that statement. Recommendation has no query and must infer intent from behaviour, which makes the target ambiguous — immediate engagement, long-term retention and catalogue discovery are different objectives that frequently conflict. This is why a recommender optimised purely on click-through can measurably succeed while making the product worse, and why the metric choice is the design decision rather than the model.

When to use it

Reach for it when:

• Catalogues too large to browse, where the alternative to recommendation is the user not finding the item at all.
• Rich interaction history, which is what collaborative filtering requires and what makes it strong.
• Alongside content-based scoring for cold-start coverage, since the two cover each other's gaps rather than competing.
• With language models for new-item description and explanation, which is where they add signal collaborative filtering cannot.

Reach for something else when:

• Small catalogues, where a well-designed list or filter beats inferred intent and is fully predictable.
• When the user has stated intent. That is search, and inferring around a clear query degrades it.
• Optimised on engagement alone without a retention or diversity counterweight, which is how a feedback loop narrows a catalogue.
• Where a wrong recommendation carries real cost — clinical, financial, legal — without a human decision in the path.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Recommendation System — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path