Pairwise Regression with Upper Confidence Bound for Contextual Bandit with Multiple Actions

Abstract

The contextual bandit problem is typically used to model online applications such as article recommendation. However, the problem cannot fully meet certain needs of these applications, such as performing multiple actions at the same time. We defined a new Contextual Bandit Problem with Multiple Actions (CBMA), which is an extension of the traditional contextual bandit problem and fits the online applications better. We adapt some existing contextual bandit algorithms for our CBMA problem, and developed the new Pair wise Regression with Upper Confidence Bound (PairUCB) algorithm which addresses the new properties of the new CBMA problem. Experimental results demonstrate that PairUCB significantly outperforms other approaches.