Title: Utilizing Social Media to Optimize Disaster Response
Speaker: Christie Nelson, Rutgers University
Date: Tuesday, October 14, 2014 12:00 - 1:00pm
Location: DIMACS Center, CoRE Bldg, Room 431, Rutgers University, Busch Campus, Piscataway, NJ
From 1980 to 2011, the total loss from weather-related catastrophes was $1,060 billion in 2011 USD and 30,000 people lost their lives in North America. The number of weather-related catastrophes is only increasing, with an increase factor of nearly 5 in North America over the past 30 years. In real-time scenarios an accurate picture of the situation is needed quickly. Often during large-scale disasters, cell towers become overloaded, and the only way of communication is through text messages. It becomes important to gather information from text messages sent to emergency numbers in order to respond quickly and efficiently with life-saving efforts. In addition, responders are unable to manually handle the large volume of incoming texts. Real-time information from streaming data is needed, and responders would benefit from text classification of incoming messages. To add to this difficult problem, these data sources tend to be microtext, which makes the problem of modeling the data more challenging.
The goal of this research was to develop a methodology to summarize text messages sent during an emergency for use by responders, including analysis of locations to identify geospatially potentially new areas of population in need of emergency assistance. The real-time disaster needs were then input into a mixed integer programming resource allocation model for distribution of resources for disaster aid. Prior research included resource allocation and text modeling, but the combination of the two was a novel application not only in this arena, but more broadly across domains. The model found the emerging real-time needs by geolocation. Two methods were evaluated for determining these emergency needs: a supervised method modeled the data with a variation of Naive Bayes, Higher-Order Naive Bayes (HONB), and an unsupervised approach modeled the data with a variation of Latent Dirichlet Allocation, Higher-Order Latent Dirichlet Allocation (HO-LDA). It was found that HONB performed better on domain relevant data than Naive Bayes, and HO-LDA performed better than LDA. Also, the use of Higher-Order Learning in conjunction with clustering geolocations to determine emerging population centers during an emergency centralized response, which reduced the unmet humanitarian aid need.
DIMACS/CCICADA Interdisciplinary Series, Complete Calendar