Community topic usage in social networks

Abstract

When studying large social media data sets, it is useful to reduce the dimensionality of both the network (e.g. by finding communities) and user-generated data such as text (e.g. using topic models). Algorithms exist for both these tasks, however their combination has received little attention and proposed models to date are not scalable (e.g.: [4]). One approach to such combined modelling is to perform community and topic modelling independently and later combine the results. In the case of overlapping communities, this combination requires a method for attributing each users topic usage to the communities in which she participates. This paper presents a Bayesian model for attributing individual documents to communities which balances the users proportional community membership with community topic coherence. Community topic usage is modelled with a Dirichlet distribution with fixed concentration parameter, leading to a well defined conjugate prior. Thought the prior is computationally expensive, the already reduced dimensionality in both topics and communities make a tractable algorithm feasible, even for large data sets. The model is applied to a corpus of tweets and twitter follower relations collected on hash tags used by people with eating disorders [14].