Mining Social Media for Improved Fairness: A Case Study of Hurricane Harvey

Keywords: disaster resilience, social media, hurricane Harvey, vulnerability, Twitter, fairness
Abstract Type: Paper Abstract

Abstract

Communities in the U.S. and the world have been suffering from natural disasters more and more frequently. It is important and effective to use an emergency informatics system to better coordinate search and rescue operations. The system needs in-time, various and big data to make accurate predictions. With the rapid development of social media, researchers start to mine social media data to meet the data requirements. However, social media data has a limited representativeness for a whole community. Thus, it is important to answer questions: (1) For communities that are overserved/underserved in terms of sending rescue request through social media during a disaster, what geographical, socioeconomic, and digital access characteristics do they have? (2) And with social media representation bias, what suggestions can we provide when we try to build fairness-aware AI prototype rescue informatics system? This study collected Twitter data during 2017 Hurricane Harvey and 35 geographical, socioeconomic, digital access indicator data. We developed a fairness measurement index, conducted spatial patterns analysis, descriptive statistics, and regression analysis. We found that block groups that consist of higher portions of people who are young, house-owners, vehicle-owners, health insurance covered, having access to internet infrastructures, with high education, have a better tendency of being overserved in terms of sending rescue request through social media. To build a fairness-aware AI rescue prediction model, researchers can increase the fairness and accuracy of rescue prediction model by adding weighted variables such as housing type, age, and digital access infrastructure level.