Social media modeling of human behavior in natural emergencies

During natural emergencies (e.g., hurricanes, tornadoes, storms), individuals can choose to avoid or leave areas of risk. Yet, often people choose to stay or travel to danger areas. Some may underestimate the danger; others may want to protect their property or families. Widespread social media use by these individuals can help us understand their motives and quantify their likelihood to engage in risky travel or decisions to stay. Social media data in such situations is not unlike sensor data; by tracking where individuals go and what they tweet about we can discover both temporal and spatial trends in human emotion and behavior during weather events. In this paper, we describe our extensible, distributed, real-time data collection and analysis pipeline that combines public streaming data from the National Weather Service and Twitter for subsequent exploration and analysis, including risk behavior modeling. Our pipeline leverages the open-source Apache Storm framework and the ELK (Elasticsearch, Logstash, Kibana) stack to process, filter, augment and index this streaming data for subsequent efficient retrieval. This work, which can be expanded to other social media (Facebook, Flickr, Instagram) is pathbreaking in several respects; first, it represents a novel integration of weather and social media data; second, our pipeline can be easily adapted to other analyzes by adding or removing processing components; and finally, this work represents the first (to our knowledge) quantification of human risk behavior using social media data in the form of average vectors and individual risk behavior indicators.