Inferring Shared Tastes with Network Topic Models

It is widely assumed that people tend to gather in groups of shared interests, where such interests drive friendships and vice versa. Thanks to online social networking platforms, information about a users’s friends as well the items he is interested in are available, but represent only an incomplete picture. We study probabilistic network topic models that distill common shared interests of friends from this data. So far, a popular choice is based on the mixed-membership stochastic blockmodel which draws inference on the absence of edges. In this paper we give theoretic and empirical evidence that absence of friendship does not refer to difference in taste. Rather, we present the shared taste model which is agnostic towards absent edges. The model’s success over blockmodels is demonstrated on data sets from LibraryThing, Zune Social, and CiteSeer. The shared taste model has many practical applications such as improving content subscription, visualizing the structure of contacts, or connecting people who otherwise would not have someone to share their interests with.