A combinatorial bandit handles decisions where several actions are selected together and the payoff depends on the complete combination rather than one isolated action. section 5.6 section 5.7
A slate of recommendations
A homepage may fill several slots, a carousel may show ten product tiles, or a feed may rank items from top to bottom. Each complete slate determines the outcome, so the decision is about optimizing the set while avoiding five near-identical recommendations. section 5.6
Use the model when:
- Payoff depends on slates, bundles, or rankings rather than single actions. section 5.7
- Optimization must balance performance with diversity so selected items are not near-duplicates. section 5.6
- The action space is large enough that slate optimization must be performed efficiently. section 5.6
