The model that trained on its own opinions
The recommender is retrained on interaction logs. Users interact with what they are shown. What they are shown is what the model chose.
So each retrain learns from a world the previous version built. Items the model never surfaced have no positive signal, and the absence of signal is read as evidence against them. The metric improves because the model is getting better at predicting its own behaviour.
Outside every boundary: Interaction logs (clicks and skips; FAILED: only what was shown), Offline metric (improving), Business metric (flat; FAILED: unmoved), Retrain (on those logs), Recommender (picks the slate), What users see, an empty slot for the a fraction chosen another way, an empty slot for the recorded with each impression Connections: Recommender calls What users see (step 1) What users see publishes to Interaction logs (step 2) Interaction logs publishes to Retrain (step 3) Retrain publishes to Recommender (step 4) a fraction chosen another way controls What users see (step 5) recorded with each impression controls Interaction logs (step 6) Interaction logs controls Offline metric — on the same bias