Measuring Spatial Influence of Twitter Users by Interactions

The three ways of interactions in Twitter-retweet, reply, and mention-comprise of a latent dynamic information flow network between users, which can be utilized to determine influential users. This paper focuses on determining which Twitter users have great influence on a query location Q in the sense that they are assumed to provide information that is of sufficient interest to prompt people at Q to interact with them. Note that an influential Twitter user who is of great influence on Q may not be necessarily from Q. Therefore, we first define generalized influential Twitter users regardless of whether their location was known or not, meaning that such generalized influencers on Q can be either from inside Q, or outside Q, or even unknown. A more interesting subset of generalized influencers is the ones whose location is in Q, and termed as local influential Twitter users. One potential application of finding local influencers (e.g., local news media) is to detect local events by tracking their tweets. Using a large amount of data collected from Twitter, we first build a large-scale directed interaction graph of Twitter users and present an analysis of the geographical characteristics of the edges in this interaction graph and make several interesting observations. Based on these findings, we propose two versions of PageRank that measure spatial influence on the interaction graph: Edge-Local PageRank (ELPR), and Source-Vertex-Locality PageRank (SVLPR), which takes into account the spatial locality of edges and the spatial locality of source vertices in edges, respectively. In addition, a Geographical PageRank (GPR) is also proposed trying to incorporate both of these two factors together. In the experimental evaluation, we examine the effectiveness of the proposed methods with regards to 3 different US cities "Boston, MA", "Bristol, CT" and "Seattle, WA", and the results show that our algorithms outperform their baseline approaches including the topological network metrics and the original PageRank. In addition, we also explored the possibility of using local influential Twitter users as potential news seeders and showed that some types of influential users have high credibility in outputting local place-relevant tweets.