What a recommendation system actually predicts
A recommendation system ranks items using signals such as past interactions, item features or patterns across users. Its score predicts an objective, such as the likelihood of a click, rather than directly measuring personal value. What it recommends depends partly on what its designers choose to optimise.
Go deeper
Ranking follows a defined target
Recommendation systems estimate which items are suitable candidates for a user and then rank them. They may use item descriptions, interaction histories or similarities among patterns of behaviour.
The training target matters. Predicting a click, a completed video and an explicit positive rating are different tasks, even when they use some of the same information.
A high ranking is a prediction under a particular objective. It is not proof that an item is the best choice for the person's broader interests.
- Candidate selection finds possible items.
- Scoring estimates relevance or another target.
- Ranking and filtering determine what appears.
Recommendations also shape future data. An item shown prominently gets more opportunities for interaction than one never displayed, creating a feedback loop that system designers must account for.