A reward signal is the observed outcome of a contextual-bandit action. In a recommendation system, displaying a post is the action, the user is the context, and viewing time is the reward: more viewing time indicates a better immediate outcome. section 5.1.1
The reward is measured after the system acts rather than chosen directly. A prediction model estimates viewing time from user features and a candidate post, while production logs record context, action, and observed reward for later model fitting. section 5.1.1
Viewing time as a measurable signal
A simulator treats measurement as a black box: it combines context and action weights, adds random variation, and returns a positive viewing-time value. A logged sample stores the context, selected action, and resulting reward.
Use reward signals with these constraints:
- Evaluate a greedy recommender with mean viewing time per post over each model-fitting period. section 5.1.3
- Require a short-term reward caused by one action, such as an ad click or a play click. section 5.1.3
- Do not treat a multistep return as a single-action reward when later outcomes depend on earlier actions. section 5.1.3
- Validate contextual-bandit changes with an A/B test, because exploration and feedback effects can alter measured performance. section 5.5
