Brand-Centric Recommendation with Inter-brand Similarities

With the increasing popularity of mobile web, social networking services (SNSs) are an integral part of our everyday lives, since they are used for communicating with friends, for gaining information on other people or some items of interest, and even for business profits. Social network information is incorporated into recommender systems to improve their performance, but most of existing work is focused on usercentric cases in which items or venues are recommended for users. On the other hand, brands are also important social objects. For instance, in Foursquare, which is a location-based online SNS, brands provide information on venues and share the tips with their followers. Thus, it is important to recommend venues for brands so that brands select interesting venues for their followers, leading to brand-centric recommendation where the targets for recommendation are brands (not users). In contrast to user-centric recommendation, brands have few social links to other brands, so trust between brands is difficult to use. In this paper we present a method for brand-centric recommendation where inter-brand similarities are implicitly determined by decomposing a brand-follower matrix. This social information on inter-brand similarity is incorporated into probabilistic matrix factorization to reveal brand latent factors as well as venue latent factors. Experiments on the dataset collected from Foursquare (by web crawling) demonstrate that our method improves the recommendation performance over existing matrix factorizations.