We just saw how bandits are used for recommending the correct ad banner to the user. But the banner preference varies from user to user. User A likes banner type 1, but user B might like banner type 3. So we have to personalize ad banners according to user behavior. How can we do that? We introduce a new bandit type called contextual bandits.
In a normal MABs problem, we perform the action and receive a reward. But with contextual bandits, instead of just taking the actions alone, we take the environment state as well. The state holds the context. Here, the state specifies the user behaviors, so we will take actions (show ads) according to the state (user behavior) that will result in a maximum reward (ad clicks). Thus, contextual bandits are widely used for personalizing content according to the user's preference behavior. They are used to solve cold-start...
