Measuring Spatial Influence of Twier Users by Interactions

Œe three ways of interactions in TwiŠer–retweet, reply, and mention–comprise of a latent dynamic information ƒow network between users, which can be utilized to determine inƒuential users. Œis paper focuses on determining which TwiŠer users have great inƒuence on a query location Q in the sense that they are assumed to provide information that is of sucient interest to prompt people atQ to interact with them. Note that an inƒuential TwiŠer user who is of great inƒuence onQ may not be necessarily fromQ . Œerefore, we €rst de€ne generalized influential Twi„er users regardless of whether their location was known or not, meaning that such generalized inƒuencers onQ can be either from insideQ , or outside Q , or even unknown. A more interesting subset of generalized inƒuencers is the ones whose location is in Q , and termed as local influential Twi„er users. One potential application of €nding local inƒuencers (e.g., local news media) is to detect local events by tracking their tweets. Using a large amount of data collected from TwiŠer, we €rst build a large-scale directed interaction graph of TwiŠer users and present an analysis of the geographical characteristics of the edges in this interaction graph and make several interesting observations. Based on these €ndings, we propose two versions of PageRank that measure spatial inƒuence 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 e‚ectiveness of the proposed methods with regards to 3 di‚erent US cities “Boston, MA”, “Bristol, CT” and “SeaŠle, 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 inƒuential TwiŠer users as potential news seeders and showed that some types of inƒuential users have high credibility in outpuŠing local place-relevant tweets.