Constrained by limited information and recommendation opportunities: An exploration and exploitation problem for recommender systems

In this paper, we investigate a resource allocation problem in recommendation systems where unknown new items keep coming to the system at different time. The task is to recommend (allocate) each user a limited number of new items. The objective is to maximize the overall positive response rate. This problem is non-trivial because, on one hand, we need to allocate these news items to users who are helpful for learning new item feature profiles using the limited recommendation opportunities (resources) in order to improve prediction accuracy; on the other hand, allocate these items to users who would most likely purchase them on the basis of the new item information gathered so far in order to maximize positive response rate. In this paper, we propose a two-stage batch solution to approximately optimize the objective, using group buying as a working example. During the first stage, we estimate the user purchase decisions towards new items by allocating some resources for exploration. During the second stage, we optimally allocate the remaining resources for exploitation according to the prediction and the operational constraints using the binary integer programming technique. Our experiments indicate that the proposed approach significantly improves the positive response rate.