Harnessing the power of machine learning: Can Twitter data be useful in guiding resource allocation decisions during a natural disaster?
Principal Researchers: Manish Shirgaokar
Urban and Regional Planning Assistant Professor Manish Shirgaokar published an article in March 2019 in Transportation Research about the potential use of social media during natural disasters to pinpoint the most devastated areas. The full article, "Harnessing the power of machine learning: Can Twitter data be useful in guiding resource allocation decisions during a natural disaster?" is available online here.
Location-based social networks like Twitter can relay not just what is happening but where. To enable efficient resource allocation during natural disasters, managers need to know what citizens are experiencing and where. During Hurricane Irma in 2017 in Florida over six million people were asked to evacuate. The hurricane caused widespread damage. We utilized geospatial and machine learning techniques to categorize geolocated tweets, which were both about Hurricane Irma and located within Florida. We employed sentiment analysis to classify tweets about damage and/or transportation into negative, neutral, or positive groups. We used the American Community Survey (ACS) 2017 geography to provide socio-economic context, and relied on a multinomial logit specification to examine which features of the tweet, tweeter, or location were likely to be associated with negative or positive sentiments. We found that longer tweets were more likely to have useful sentiment-based content. We also discovered that popular tweeters were likely to tweet positive sentiments, and that popular tweets were less likely to have useful information about the disaster. The ACS backdrop suggested that the likelihood of tweets with negative sentiments was higher in census block groups with younger families. The probability of tweets with useful positive or negative sentiments dropped in locations with higher property values or median rents. This paper offers a methodology for extracting information embedded within social media data, and suggests ways to develop a probability-based understanding of needs to guide disaster managers.