Recommender systems often show users only the same content, creating filter bubbles. Trying to fix this by boosting under-explored items usually fails because the system cannot predict how users will react to unfamiliar items, and it often hurts engagement.
Researchers built a system called URED that recycles items the ranking model scored highly but never showed due to competition from other items. These items are both relevant and predictable, unlike cold items. The system collects them over a 30-day window and reintroduces them in a simple lookup, without retraining the model.
In live tests, the system increased the variety of categories users saw by 3.47% and overall content richness by 8.47%, while engagement was preserved and completed video views actually increased. This shows that a careful data pipeline can improve discovery without sacrificing the metrics that matter.