MulSetRank: Multiple set ranking for personalized recommendation from implicit feedback

Learning user preferences from implicit feedback through collaborative ranking is an approach that has received increased attention from researchers in recent years. The existing approaches focus on modeling the ranking differences between observed and unobserved items for independent individuals, ignoring their collaborative effects on the social recommendation. However, homophily theory suggests that, for a specific user, potential preference items can be mined from its friends or users with similar interests. Motivated by this observation, this paper presents a novel setwise ranking model that considers users’ preference rankings among multiple sets of items. Unlike existing models, which only consider the preference differences between observed and unobserved items, our model assumes that some potential preference items are in-between. We propose a collaborative method that exploits users’ behavioral similarities to mine users’ potential preference items. Our approach allows us to capture users’ collaborative signals at a finer granularity. We also develop a sampling method to efficiently compute the setwise preference probability. Finally, we conduct extensive experiments to evaluate the effectiveness and efficiency of the proposed model based on several benchmark datasets. The experimental results demonstrate that our approach outperforms the state-of-the-art methods. •We propose a novel setwise recommendation framework.•We develop a new scheme to mine users’ potential preference items.•We propose a sampling scheme to approximate the top one probability.•We conduct extensive experiments to validate the effectiveness.