A spatial LDA model for discovering regional communities

Abstract

Models and techniques for the extraction and analysis of communities from social network data have become a major area of research. Most of the prominent approaches exploit the link structure among users based on, e.g., information about followers or the exchange of messages among users. However, there are also other types of information that are useful for extracting communities from social network data, such as geographic information associated with postings and users. In this paper, we present a novel approach to discover so-called regional communities. Motivated by the fact that more and more postings to social networks include the geo-location of users, we claim that communities also form even if their users do not necessarily interact but are posting (similar) messages in both spatial and temporal proximity. To discover such regional communities we propose a generative probabilistic model based on spatial latent Dirichlet allocation (SLDA) that unveils not only regional communities but also topics associated with these communities. We demonstrate the effectiveness of our approach using Twitter data and compare the properties of communities detected that way with communities discovered by approaches using link graphs.