A UCB-Like Strategy of Collaborative Filtering


Atsuyoshi Nakamura ;
Proceedings of the Sixth Asian Conference on Machine Learning, PMLR 39:315-329, 2015.


We consider a direct mail problem in which a system repeats the following process every day during some period: select a set of user-item pairs (u,i), send a recommendation mail of item i to user u for each selected pair (u,i), and receive a response from each user. We assume that each response can be obtained before the next process and through the response, the system can know the user’s evaluation of the recommended item directly or indirectly. Each pair (u,i) can be selected at most once during the period. If the total number of selections is very small compared to the number of entries in the whole user-item matrix, what selection strategy should be used to maximize the total sum of users’ evaluations during the period? We consider a UCB-like strategy for this problem, and show two methods using the strategy. The effectiveness of our methods are demonstrated by experiments using synthetic and real datasets.

Related Material