The encyclopedia · Engineering & Operations · Technical decision · 2023
Kuaishou trained its recommender to keep you coming back, not just to get a click.
Kuaishou's RLUR recommender optimized the time between sessions instead of next-item clicks, a long-horizon target that raised user retention and DAU.
Kuaishou
the move
Recommendation models traditionally maximize the chance of a click or a watch on the next item. But a platform's real goal is that the user opens the app again, which depends on many interactions and arrives late.
Kuaishou's RLUR reframes the problem as an infinite-horizon Markov decision process where the reward is time to the next session. Shortening the gap between sessions is the same as lifting DAU, but that signal is biased and delayed, so the model is designed to learn from it despite never observing a user who returns later.
The result brought the model's objective in line with retention itself, and Kuaishou deployed it fully in its short-video app, reporting consistent gains in retention and DAU.
why it works
- Retention is the outcome the platform actually wants, not a proxy
- Modeling the whole journey lets the policy plan for the long run
- The correction mechanism handles the delayed, biased return signal
- Aligning objective and metric prevents optimizing the wrong thing
what transfers
A metric you can measure per action is not always the metric you want; when success is a deferred, aggregate outcome, change the objective and the technique, not the tune.
what came after
The paper became a widely-cited example of using reinforcement learning to optimize long-term retention rather than immediate clicks, and it pushed recommender research toward session-level and retention-aware objectives across the industry.
references
spotted an error? The archive wants to know.