Investigation of social media representation bias in disasters: Towards a systematic framework

Due to its real-time and human-centered nature, social media posts have been widely applied to provide rapid situational awareness in disasters, particularly from a human-centered perspective. To generalize social media-derived insights on a population, a pre-requisite is that the employed social media posts are capable of revealing the information of disaster-affected population with-out bias. Such wide application and pre-requisite underscore the importance of investigating so-cial media bias for deriving reliable decision support insights in disaster management. However, a systematic framework that streamlines the investigation of social media representation bias is still missing. To address the research gap, we propose a framework comprising (1) the setting of an appropriate representation bias benchmark; (2) the modeling of the sampling uncertainty of social media-derived insights; and (3) the derivation and quantification of representation bias distribution across races/ethnicities. Public transit amid COVID-19 in the United States is studied for illustration purposes. Nation-level results show that the White group is over-represented, the Asian group is slightly over-represented, and the Hispanic and Black groups are under-represented throughout the studied period. The level of social media representation bias varies across the states of California, New York, Texas, and Florida, and it is inversely correlated with population ratios. Such findings are beneficial for decision-makers to use social media to derive reliable insights into disaster-affected population, thereby making informed operational deci-sions accordingly.