DUET: Data-Driven Approach Based on Latent Dirichlet Allocation Topic Modeling

Abstract

Social networking platforms have been widely employed to detect and track physical events in population-dense urban areas. They can be effective tools to understand what happens and when and where it happens, either retrospectively or in real time. Correspondingly, a variety of approaches have been proposed for detecting either targeted or general events. However, neither type of event detection technique has been developed to detect urban emergencies that happen in specific geographic locations and with unpredictable characteristics. Therefore, we propose a spatial and data-driven detecting urban emergencies technique (DUET) for natural hazards, manmade disasters, and other emergencies. The method addresses both geographic and semantic dimensions of events using a geotopic detection module and evaluates their crisis levels on the basis of the intensity of negative sentiment through a ranking module. DUET was designed specifically for georeferenced tweets from a Twitter streaming application programming interface (API). To validate the technique, we conducted multiple experiments with geotagged tweets in different urban environments over a period of four to six consecutive hours. DUET successfully identified emergencies of different types among all the candidate geotopics. Our future work focuses on enabling online-mode detection with high scalability with large volumes of streaming data and providing interactive visualization through a GIS system. DUET can identify emergencies of general types and provide timely emergency reports both to first responders and to the public. The technique contributes to building an efficient and open disaster information system through a crowdsourcing effort and adding agility to urban resilience regarding crisis detection, situation awareness, and information diffusion.