Bandit feedback records the reward for the action a policy selected while omitting outcomes for actions it did not sample. In a recommender, the logged data therefore contains context, chosen action, and reward—not complete labels for every possible recommendation. section 5.2.2
This creates three practical consequences:
- Treat the data as policy-collected rather than randomly sampled, so model estimates can be biased toward previous decisions. section 5.2.2
- Use exploration to select actions the current model would not choose and collect their previously missing counterfactual rewards. section 5.2.2
- Refit on the expanding logs so broader action and context coverage can reduce bias and improve future rewards. section 5.2.2
Recommender exploration
An epsilon-greedy recommender randomly selects an action with probability epsilon and otherwise follows its prediction-based choice. The random selections expose posts that a greedy policy would omit, adding counterfactual observations to later training logs. section 5.2.2
