BiUCB: A Contextual Bandit Algorithm for Cold-Start and Diversified Recommendation

In web-based scenarios, new users and new items frequently join the recommendation system over time without prior events. In addition, users always hold dynamic and diversified preferences. Therefore, cold-start and diversity are two serious challenges of the recommendation system. Recent works show that these problems can be effectively solved by contextual multi-armed bandit (CMAB) algorithms which consider the coldstart and diversified recommendation process as a bandit game. But existing methods only treat either items or users as arms, causing a lower accuracy on the other side. In this paper, we propose a novel bandit algorithm called binary upper confidence bound (BiUCB), which employs a binary UCB to consider both users and items to be arms of each other. BiUCB can deal with the item-user-cold-start problem where there is no information about users and items. Furthermore, BiUCB and k-ϵ-greedy can be combined as a switching algorithm which lead to significant improvement of the temporal diversity of entire recommendation. Extensive experiments on real world datasets demonstrate the precision of BiUCB and the diversity of switching algorithm.