Mind the gaps: Mapping and mitigating exclusionary data bias in crisis informatics

Abstract

Increasingly large numbers of people are living in areas susceptible to catastrophic disasters because of urban sprawl and worsening extreme weather patterns from climate change. The extent of the impact on humans and society can be mitigated through real-time information on human location, activity, and responsiveness. Social media has been proposed as one source of this information. However, although many potential uses for social media have been identified, a number of both ethical and computational biases have been identified as well. An important area for research is identifying these biases, the effects they have on disadvantaged populations, and how to mitigate that bias in the growing body of work designed to utilize social media in crisis situations. Within this work, I examine the prevalence and significance of decreases in social media activity from a normal state to crisis conditions, the influence of geographic scale on the statistical reliability of social media data and the correlation between social media and infrastructural damage, and the effect of social vulnerability factors on the presence or absence of social media data during a disaster. The results of this dissertation inform the reliable extent of social media data and its sensitivity to external factors (i.e., infrastructure damage and the presence of vulnerable populations) and analytical factors (i.e., spatiotemporal scale, aggregation, and bursting behavior). By pinpointing disparities in the representational capacities of social media data and proposing alternative methods of analysis, I hope to improve the usability and equity of social media data for crisis response.