Sense and Focus: Towards Effective Location Inference and Event Detection on Twitter

Twitter users post observations about their immediate environment as a part of the 500 million tweets posted everyday. As such, Twitter can become the source for invaluable information about objects, locations, and events, which can be analyzed and monitored in real time, not only to understand what is happening in the world, but also an event's exact location. However, Twitter data is noisy as sensory values, and information such as the location of a tweet may not be available, e.g., only 0.9i¾?% of tweets have GPS data. Due to the lack of accurate and fine-grained location information, existing Twitter event monitoring systems focus on city-level or coarser location identification, which cannot provide details for local events. In this paper, we propose SNAF Sense and Focus, an event monitoring system for Twitter data that emphasizes local events. We increase the availability of the location information significantly by finding locations in tweet messages and users' past tweets. We apply data cleaning techniques in our system, and with extensive experiments, we show that our method can improve the accuracy of location inference by 5i¾?% to 20i¾?% across different error ranges. We also show that our prototype implementation of SNAF can identify critical local events in real time, in many cases earlier than news reports.